asian development review...of the philippines); and martin alexander cruz (university of the...

Asian Development Review

ADB COVID-19 Policy Database: A GuideJesus Felipe and Scott Fullwiler

Evaluating Strategies to Reduce Arsenic Poisoning in South Asia:A View from the Social SciencesMatthew Krupoff , Ahmed Mushfi q Mobarak, and Alexander van Geen

Does the All-China Federation of Industry and Commerce Align Private Firmswith the Goals of the People’s Republic of China’s Belt and Road Initiative?Jeff rey B. Nugent and Jiaxuan Lu

The Borrowing Puzzle: Why Do Filipino Domestic Workers in Hong Kong, China Borrow Rather than Dissave?Wooyoung Lim and Sujata Visaria

Impacts of an Information and Communication Technology-Assisted Program on Attitudes and English Communication Abilities: An Experiment in a Japanese High SchoolYuki Higuchi, Miyuki Sasaki, and Makiko Nakamuro

Analyzing the Sources of Misallocation in Indian Manufacturing: A Gross-Output ApproachSujana Kabiraj

Trade Volatility in the Association of Southeast Asian Nations Plus Three:Impacts and DeterminantsThi Nguyet Anh Nguyen, Thi Hong Hanh Pham, and Thomas Vallée

Could Weather Fluctuations Affect Local Economic Growth? Evidence from Counties in the People’s Republic of ChinaChengzheng Li, Jiajia Cong, and Haiying Gu

ADB Distinguished Speaker’s ProgramHuman Capital as Engine of Growth: The Role of Knowledge Transfersin Promoting Balanced Growth within and across CountriesIsaac Ehrlich and Yun Pei

Volume 37 2020 Number 2

ADEV3702-Cover.indd 1ADEV3702-Cover.indd 1 8/25/20 9:38 AM8/25/20 9:38 AM

EDITOR

YASUYUKI SAWADA, Asian Development Bank

MANAGING EDITOR

JESUS FELIPE, Asian Development Bank

EDITORIAL TEAM

GEMMA ESTHER B. ESTRADA, Asian Development BankMARIA SUSAN M. TORRES, Asian Development BankMARY ANN M. MAGADIA, Asian Development Bank

EDITORIAL BOARD

XIN MENG, Australian National UniversityAHMED MUSHFIQ MOBARAK, Yale UniversityNANCY QIAN, Northwestern UniversityKRISLERT SAMPHANTHARAK, University of

California, San DiegoKUNAL SEN, UNU-WIDER and The University

of ManchesterAYA SUZUKI, The University of TokyoMAISY WONG, University of Pennsylvania

KYM ANDERSON, University of AdelaidePREMA-CHANDRA ATHUKORALA,

Australian National UniversityKLAUS DESMET, Southern Methodist UniversityJESUS FELIPE, Asian Development BankNEIL FOSTER-MCGREGOR, UNU-MERITSHIN-ICHI FUKUDA, The University of TokyoSUNG JIN KANG, Korea UniversityHONGBIN LI, Stanford University

The Asian Development Review is a professional journal for disseminating the results of economic and development research relevant to Asia. The journal seeks high-quality papers done in an empirically rigorous way. Articles are intended for readership among economists and social scientists in government, private sector, academia, and international organizations.

The views expressed in this publication are those of the authors and do not necessarily refl ect the views and policies of the Asian Development Bank (ADB), the Asian Development Bank Institute (ADBI), the ADB Board of Governors, or the governments they represent.

ADB and ADBI do not guarantee the accuracy of the data included in this publication and accept no responsibility for any consequence of their use.

By making any designation of or reference to a particular territory or geographic area, or by using the term “country” in this document, ADB and ADBI do not intend to make any judgments as to the legal or other status of any territory or area.

Please direct all editorial correspondence to the Managing Editor, Asian Development Review, Economic Research and Regional Cooperation Department, Asian Development Bank, 6 ADB Avenue, Mandaluyong City, 1550 Metro Manila, Philippines. E-mail: [email protected].

HONORARY BOARD

Chair: MASATSUGU ASAKAWA, Asian Development Bank

MONTEK SINGH AHLUWALIA, Former Deputy Chairman of the Planning Commission, India

PETER DRYSDALE, Australian National UniversityJUSTIN LIN, Peking University MARI ELKA PANGESTU, World Bank

HAN SEUNG-SOO, UN Secretary-General’s Special Envoy for Disaster Risk Reduction and Water

LAWRENCE SUMMERS, Harvard University, John F. Kennedy School of Government

Notes: In this publication, “$” refers to United States dollars, unless otherwise stated.ADB recognizes “Korea” as the Republic of Korea.

For more information, please visit the website of the publication at www.adb.org/publications/series/asian-development-review.


Asian Development ReviewVolume 37 • 2020 • Number 2

September 2020

Volume 37 2020 Number 2

ADB COVID-19 Policy Database: A Guide 1Jesus Felipe and Scott Fullwiler

Evaluating Strategies to Reduce Arsenic Poisoning in South Asia:A View from the Social Sciences 21

Matthew Krupoff , Ahmed Mushfi q Mobarak, and Alexander van Geen

Does the All-China Federation of Industry and Commerce Align Private Firms with the Goals of the People’s Republic of China’s Belt and Road Initiative? 45

Jeff rey B. Nugent and Jiaxuan Lu

The Borrowing Puzzle: Why Do Filipino Domestic Workers in Hong Kong, China Borrow Rather than Dissave? 77

Wooyoung Lim and Sujata Visaria

Impacts of an Information and Communication Technology-Assisted Program on Attitudes and English Communication Abilities: An Experiment in a Japanese High School 100

Yuki Higuchi, Miyuki Sasaki, and Makiko Nakamuro

Analyzing the Sources of Misallocation in Indian Manufacturing: A Gross-Output Approach 134

Sujana Kabiraj

Trade Volatility in the Association of Southeast Asian Nations Plus Three: Impacts and Determinants 167

Thi Nguyet Anh Nguyen, Thi Hong Hanh Pham, and Thomas Vallée

Could Weather Fluctuations Affect Local Economic Growth? Evidence from Counties in the People’s Republic of China 201

Chengzheng Li, Jiajia Cong, and Haiying Gu

ADB Distinguished Speaker’s ProgramHuman Capital as Engine of Growth: The Role of Knowledge Transfersin Promoting Balanced Growth within and across Countries 225

Isaac Ehrlich and Yun Pei

ADB COVID-19 Policy Database: A GuideJesus Felipe and Scott Fullwiler∗

The ADB COVID-19 Policy Database displays the measures taken andmonetary amounts announced or estimated by the 68 members of the AsianDevelopment Bank, two institutions, and nine other economies (i.e., a totalof 79 entries) until May 2020, to fight the coronavirus disease (COVID-19)pandemic. Measures are classified according to (i) the path a given measuretakes to affect the financial system, spending, production, and so forth, i.e.,provide liquidity, encourage credit creation by the financial sector, or directlyfund households; and (ii) the effects on the financial statements of households,businesses, government, i.e., whether the measures create more debt or moreincome. This gives a total of nine categories. When the information is available,we report the amounts that governments have announced (intentions) they willallocate to each measure (in many cases, no amount is provided because themeasure does not entail spending, e.g., interest rate reductions). These are amix of actual amounts and estimates, today and in the future (without specifyingwhen). The database will be updated, revised, and expanded as information isreleased. It is available at https://covid19policy.adb.org/.

Keywords: Asian Development Bank, COVID-19, databaseJEL codes: A10, C82

I. Introduction1

This paper explains the key concepts that underlie the taxonomy used toconstruct the Asian Development Bank’s (ADB) ADB COVID-19 Policy Database(1 June 2020 version). The database is available at https://covid19policy.adb.org/.It was built to keep track of the measures that the 68 members of ADB haveimplemented to fight the coronavirus disease (COVID-19) pandemic that has

∗Jesus Felipe (corresponding author): Asian Development Bank (ADB). E-mail: [email protected]; Scott Fullwiler:University of Missouri–Kansas City. E-mail: [email protected]. This is a project of the Economic Research andRegional Cooperation Department (ERCD) of ADB. We are deeply grateful to Gemma Estrada (ADB); Maria SusanTorres (ADB); Mary Ann Magadia (ADB); and Donna Faye Bajaro (ADB) for their outstanding research assistancefor countless hours and dedication. Other members of the team that contributed to the database are Maria Hanna Jaber(ADB); Remrick Patagan (ADB); Al-Habbyel Yusoph (University of the Philippines); Simon Alec Askin (Universityof the Philippines); and Martin Alexander Cruz (University of the Philippines).

1This policy database provides information on the key economic measures that governments are taking tocombat the COVID-19 pandemic. The policy database might not fully reflect all the measures implemented by theeconomies considered. Errors and omissions will be corrected in successive versions. The policy database includespublicly available information and its intent is solely to inform the public, and it does not make any judgment. Pleasesend questions about the database to [email protected]. The usual ADB disclaimer applies.

Asian Development Review, vol. 37, no. 2, pp. 1–20https://doi.org/10.1162/adev_a_00147

© 2020 Asian Development Bank andAsian Development Bank Institute.

Published under a Creative CommonsAttribution 3.0 International (CC BY 3.0) license.

2 Asian Development Review

affected the world since early 2020. In this article, we provide details of themethodology we use to classify these measures. The version of the database thispaper refers to incorporates the available information until early May 2020. It coversthe 68 members of ADB plus the European Central Bank (ECB), the EuropeanUnion (EU), Argentina, Brazil, Mexico, the Russian Federation, South Africa,Nigeria, the Arab Republic of Egypt, Saudi Arabia, and the Islamic Republic ofIran, that is, a total of 79 entries.

Apart from national sources, the main sources of information used are listedin the references at the end of the paper. The rest of the document is organized asfollows. Section II provides the rationale behind the exercise and the frameworkused. Section III discusses the categorization that we use to classify the economicmeasures and to construct the database. Section IV shows the worksheet of eacheconomy and discusses the macroeconomic impact of the different measures.Section V provides some final notes and clarifications on the liquidity measuresthat economies are undertaking.

II. Rationale and Framework

Not all forms of macroeconomic stabilization or stimulus are created equal.Economists have long recognized, for instance, that different types of governmentspending or tax cuts will have different macroeconomic impacts (or multipliereffects). Rarely, however, do they ground their analysis in operations and financialstatement effects that are fundamental in every transaction. One regularly readsin textbooks or the financial press about central banks “pumping money into theeconomy” through open market purchases or lending through standing facilities,as if these were identical to a direct government transfer to households. Thesetransactions, however, are entirely different in terms of operations and financialstatement effects. An individual quite obviously is not indifferent to such choicesas (i) an end-of-year salary bonus, (ii) an equal amount conversion of part of herretirement portfolio from bonds to a money market fund, or (iii) a line of creditof equal value. By the same token, neither should the economy be indifferentto macroeconomic policies that often have far greater variety in their operationaldetails and financial statement effects.

For the purpose of understanding different policy actions in response toCOVID-19, the approach here is to categorize these actions according to theirdifferences in operational details and/or financial statement effects. Operationaldetails in this context define the path a given measure takes to affect the financialsystem, spending, production, and so forth. For the COVID-19 policy responses,these fall into the following categories:

• Provide liquidity to financial and non-financial businesses and/or state/local/regional governments

ADB COVID-19 Policy Database 3

• Encourage credit creation by the financial sector

• Directly fund households, businesses, and/or state/local/regional governments

Financial statement effects of a given measure answer one of the followingquestions:

• Who, if anyone, bears the financial risk of the measure and what kind?

• Does the measure create more debt or more income (e.g., net worth or equity,other things being equal) for the recipients?

These financial statement effects enable an expansion or elaboration ofthe operational detail categories shown in Table 1. The left column repeats thethree bulleted categories for operational details. The respective potential financialstatement outcomes of a given measure are to the right of the correspondingoperational detail categories. In order to provide liquidity, for instance, governmentsor central banks can (i) lend (expanding the borrowers’ liabilities in order to obtaincentral bank liabilities) via existing or expanded standing facilities; (ii) purchasefinancial assets (exchanging the sellers’ financial assets for central bank liabilities);or (iii) relax regulations (such as lowering required minimum liquidity ratios),expand the range of acceptable collateral for secured loans from the central bank,and so on, which do not directly alter private sector financial statements in thesense that there are no accompanying transactions (though they may encourageor enable financial institutions’ subsequent actions and thereby lead to changes intheir financial statements indirectly, of course). The effects on the financial positionsderived from credit creation and direct funding are likewise discussed in detail.

This taxonomy allows us to classify the measures taken by most economies,often not presented using identical criteria. Also, it is wider and richer than aclassification into fiscal and monetary policies. Finally, one would like to go asdeep as possible, but the information provided by many economies does not allowit.

III. A Categorization of COVID-19 Macroeconomic Measures

In this section, we elaborate on the measures that this database considers.The framework in Table 1 shows five measures in the taxonomy of COVID-19macroeconomic measures. We note that Measures 01–04 mostly correspond tomonetary policy, while Measure 05 corresponds to fiscal policy. Three additionalmeasures are effectively double counting from an accounting perspective but arenonetheless important measures. We label them Measures 06–08. These three


Table 1. Categorization of Measures according to Operational Details and FinancialPosition Effects

Operational Details Financial Positions

Provide liquidity Measure 01Loans from the central bank or government to the private sector and

state/regional/local sectorGovernment or central bank purchases of short-term assets from

the private sectorRegulatory or other changes that do not directly alter private sector

financial statements

Encourage credit creation by Measure 02the financial sector Increases in liabilities of the private sector and state/regional/local

sector to the government or central bank through loans to thefinancial sector (to enable further lending to the financial andnon-financial sectors) or secondary market purchases ofsecurities issued by the financial sector, businesses, orstate/regional/local governments

Interest rate changes, loan guarantees, forbearances, and regulatorychanges that do not directly affect financial positions toencourage private credit creation

Directly fund Measure 03Increases in recipients’ liabilities through direct loans from the

government or central bankMeasure 04

Increases in ownership claims of the government or central bankthrough equity investments in the business and/or financialsectors

Measure 05Increases in income or reductions in costs or obligations through

government transfer payments, loan cancellation, tax cuts,forbearances, and so forth

Source: Authors’ elaboration.

measures are sources or funds, while Measures 01–05 are uses of funds. Measure09 is the mirror image of Measure 08 (see explanation below). We add Measure 010to take into account those actions for which the current information is unclear aboutthe particular measure they should be added to.

In what follows, we elaborate on the specific actions that each of the measuresrefers to and provide examples:

Measure 01: Support the normal functioning of the money markets andshort-term finance

• 01A. Lending to the private sector or state/local/regional governments, andasset purchases to provide liquidity• Additional standing facilities or increased provision for normal lending to

money markets


• Short-term loans to nonfinancial businesses to refinance maturing obligationsor otherwise finance short-term operations

• Direct short-term loans (1 year or less to maturity) to state/local/regionalgovernments or purchases of their short-term securities (1 year or less tomaturity)

• Direct purchases of short-term financial assets in secondary markets• Repurchase agreements

• 01B. Non-lending actions and regulatory adjustments: collateralrequirements, payments system policies, liquidity regulations, reserverequirements, etc.

• 01C. Foreign exchange operations or domestic lending in foreign currency• Loans in foreign currency or foreign exchange swaps from a central bank or

government to the domestic private sector or into domestic currency markets

Examples of Measure 01:° People’s Republic of China: Expansion of re-lending and re-discounting

facilities by CNY1.8 trillion to support manufacturers of medical suppliesand daily necessities; micro-, small-, and medium-sized firms; and theagriculture sector at low interest rates

° EU: The European Investment Bank (EIB) dedicated liquidity linesto banks to ensure additional working capital support for small andmedium-sized enterprises (SMEs) and mid-caps of EUR10 billion.

° Denmark: The Danmarks Nationalbank announced the launch of an“extraordinary lending facility” which will make full-allotment, 1-week,collateralized loans available to banks at –0.5% interest rate.

Measure 02: Encourage private credit creation

• 02A. Secondary market purchases of securities (greater than 1 year tomaturity), and loans to financial sector• Purchases of mortgage-backed securities• Purchases of corporate bonds, collateralized loan obligations (CLOs), or bond

exchange-traded funds (ETFs)• Purchases of new financial sector loans to the non-financial sector in full or

less than full

• 02B. Interest rate reductions and other regulatory adjustments: capitalrequirements, credit and lending standards, oversight, etc.• Interest rate reductions


° Announced reductions in policy rates° Attempts to indirectly reduce interest rates via purchases of securities (or

bond/fixed-income ETFs) in secondary markets• Reduced capital requirements° Temporary or permanent reductions in risk-weighted capital requirements,

supplementary leverage ratio requirements, countercyclical capital bufferrequirements, etc.

° Temporary or permanent omitting or reduced weighting of certain financialassets in calculating required capital

° Temporary or permanent sheltering of losses for lenders from equityimpairment

° Regulatory forbearance• Oversight° Reductions in macroprudential margins of safety (such as loan-to-value

ratios, debt-service ratios, etc.)° Relaxations in microprudential oversight (such as bank examinations)

• 02C. Loan guarantees

Examples of Measure 02:° Australia: The government has allocated up to AUD15 billion to invest in

residential mortgage-backed securities and asset-backed securities.° Cambodia: The National Bank of Cambodia has delayed additional increases

in the capital conservation buffer, cut the interest rate in its LiquidityProviding Collateralized Operations to decrease banks’ funding costs indomestic currency, and lowered the interest rate on negotiable certificates ofdeposit to encourage banks to disburse loans.

° Spain: EUR100 billion government loan guarantees for firms and self-employed; EUR2 billion public guarantees for exporters

° Thailand: Soft loans by the Bank of Thailand to financial institutionsamounting to THB500 billion to be on-lent at 2% interest to SMEs withoutstanding loans.

Measure 03: Long-term direct lending to businesses, households, and state/local/regional governments, and forbearance

• 03A. Long-term direct lending to businesses, households, and state/local/regional governments• Direct loans to the non-financial sector (more than 1 year)• Primary market purchases of private debt securities with maturities greater

than 1 year (corporate bonds, mortgages or mortgage-backed securities,bonds issued by state/local/regional governments, etc.)


• 03B. Forbearance

Example of Measure 03:° Malaysia: MYR1 billion fund allocated by Bank Negara Malaysia (central

bank) for SMEs involved in food production; limited to MYR5 million perSME at 3.75% per annum, for a maximum tenure of 8 years

Measure 04: Equity claims on the private sector

• Purchases of equities and/or equity ETFs• Direct investments in non-financial corporations• Direct investments in banks and other financial institutions

Example of Measure 04:° Germany: (i) EUR100 billion under the newly created economic stabilization

fund to directly acquire equity of larger affected companies and strengthentheir capital position; (ii) EUR2 billion to expand venture capital financingto start-ups, new technology companies, and small businesses during thecoronavirus crisis; and (iii) EUR10 billion fund by the state of Bavaria tobuy stakes in struggling companies

Measure 05: Government support to income/revenue

• 05A. Health• Healthcare-related additions to non-national government income

(households, businesses, state/local/regional government)

• 05B. Non-health• Non-healthcare-related additions to non-national government income

(households, businesses, state/local/regional government)• General examples:° Direct purchases (infrastructure, goods, services, etc.)° Direct financial assistance for payroll, non-payroll expenses, reductions in

revenues/income° Direct transfer payments° Direct income support for the unemployed, poor, etc.° Direct job creation° Tax cuts, credits, exemptions, delayed payments, etc.° Reduction in payment commitments (utilities, rent, etc.) with government

assistance/subsidies to private payees° Moratoria on debt collections and late payment collections with

government assistance/subsidies to private payees or creditors° Assistance for private production and supply chains (including “wartime

powers” command over industry)


Examples of Measure 05:° India: The finance minister announced on 26 March a stimulus package

valued at approximately 0.8% of GDP, which included in-kind and cashtransfers to lower-income households, insurance coverage for workers inthe health-care sector, and wage support to low-wage workers.

° Republic of Korea: The National Assembly approved the budget ofKRW2.1 trillion for disease control, i.e., epidemic prevention andtreatment, and support for medical institutions and quarantined people.

The next three measures are also consistent with Table 1 but effectivelydouble count from an accounting perspective:

Measure 06: Redirecting or reallocating previously budgeted spending

Measure 07: Central bank financing government operations

• 07A. Direct lending or government reserve drawdown• Central bank direct loans to government• Central bank purchases of government securities in the primary market• Other, e.g., deficit without bond issues or central bank direct lending (e.g.,

Singapore spending from reserves)

• 07B. Secondary market purchases of bonds• Central bank purchases of government treasury bonds in the secondary market

Examples of Measure 07:° Japan: Purchases of Japan government bonds for “yield curve control”° United States (US): Federal Reserve purchases US Treasury securities in

secondary markets (quantitative easing)° ECB: Purchases of national government bonds° Philippines: To further support the Filipino people during the COVID-19

pandemic, the Monetary Board authorized the Bangko Sentral ng Pilipinas(BSP) to purchase government securities from the Bureau of Treasury(BTr) under a repurchase agreement in the amount of PHP300 billion witha maximum repayment period of 6 months.

Measure 08: International assistance (borrower/recipient)

• 08A. Swaps and clearing arrangements (borrower)• Central bank currency swaps and repurchase agreement facility for official

foreign accounts° Currency swap lines to other central banks (loan collateralized by the

borrowing economy’s currency) or official international organizations(e.g., the International Monetary Fund [IMF])


° Non-currency swap repurchase agreement facilities for official foreignaccounts

• 08B. International loans and grants (recipient)

• International aid and grants/donations° Grants or loans that are related to the COVID-19 pandemic in support of

developing member countries (DMCs).° Loans or aid from the World Bank, United Nations (UN), World Health

Organization (WHO), etc.° IMF special drawing rights (SDRs) granted° Donations/aid to specific nations

• 08B1. Asian Development Bank

• 08B2. Others

Measure 06 is double counting because it is previously budgeted spending(already allocated/budgeted) that is redirected or reallocated and has beenpreviously accounted for in government budget position projections and therefore,in theory, should not affect subsequent projections to the budget position. Measure07 is double counting because it is just the funding for Measure 05. Central bankpurchases of government securities or direct loans to the government double countgovernment deficits (except to the degree that the purchases or loans become greaterthan COVID-19-related deficits). Finally, international assistance (Measure 08) isdouble counting because it is receiving funds, not spending, lending, or investingthem.

Measure 09: International assistance (lender/donor)Measure 09 is the mirror image of Measure 08, from the point of view of the

donor economy. It is not double counting from this economy’s perspective.

• 09A. Swaps and clearing arrangements (lender)

• 09B. International loans and grants (lender/donor)

Measure 010: No breakdownThis category captures actions or announced measures that do not yet clearly

fit into one or more of the other measures.

• Amounts from measures that cannot be clearly allocated according to theirpurposes (e.g., amounts that are intended to cover several measures).

Example of Measure 010:° EU: EUR37 billion unallocated funds of cohesion policy funding 2014–2020

will be eligible for Coronavirus-crisis-related expenditure within the CoronaResponse Investment Initiative. Member states can use them to support publicinvestment for hospitals, SMEs, labor markets, and stressed regions.


Figure 1. COVID-19 Measures and Their Funding

COVID-19 = coronavirus disease.Note: The width of the arrows is intended to give an idea of the approximate relative size of the funding. The thickerthe larger.Source: Authors.

As noted earlier, there is a use and funding relationship between Measures01–05 (and Measure 010, discussed below) and Measures 06–08, respectively(which is the accounting corollary of the just explained “double counting” for thesemeasures). From the point of view of the uses, Measures 01–04 are mostly fundedby the central bank (self-financed) and also partly by the government. Measure 05is funded by the government’s bond sales to the non-government sector (which maybe purchased in the secondary market by the central bank in Measure 07B), centralbank loans or primary market purchases of government bonds (Measure 07A),drawdown of existing reserves (Measure 07A), and also partly by internationalassistance (Measure 08B). From the point of view of the funding sources, Measure06 is also a source of government spending, lending, or investing, but is mutuallyexclusive from Measures 01–05 in this taxonomy since “where” the spending hasbeen reallocated to is already in Measure 06. As noted, in Measure 07, the centralbank directly or indirectly funds the government, which then appears in the latter’sactions across Measures 01–05. Measure 08A directly goes to the central bank,providing funding for activities in Measure 01C. Finally, as noted, Measure 08Bis a source of funds for the government and likely ends up in Measure 05. Theserelationships are summarized in Figure 1.

What is the appropriate combination of measures that capture a nation’s totalCOVID-19 policy response?

For an individual country, Measures 01–05 and 09, together, capture thefinancial positions the central bank and government have taken relative to the privatesector and state/local governments across their cumulative policy responses—lending to the private sector and state/local governments, contingent liabilities,


equity investments, foreign exchange intervention, lending in domestic markets inforeign currencies, lending domestic currencies to other central banks, and directtransfer of income.

For aggregation across countries, however, Measures 01–05 are theappropriate ones to sum. Measure 09 must be dropped in this case because it doublecounts both Measure 08 and, more importantly, Measure 01C, in the context of acompilation across countries.

We emphasize again, however, that this is not to suggest that any of theseindividual measures are qualitatively the same. If they were, there would be noreason to have categories in the first place. Summing across countries for Measures01–05 and summing Measures 01–05 and 09 together for an individual countrygives the total financial positions that have been assumed vis-à-vis the private sectorand state/local government sectors; a larger sum tells us a response was likely tohave been larger, but does not necessarily tell us that the response was better.

Finally, we again note that there are significant differences in what countriesreport, the quality of reporting across countries, whether reporting is accompaniedwith numerical estimates, and that significant portions of Measures 01 and 02 do notalways lend themselves to such reporting of numerical estimates (such as relaxationof liquidity or capital requirements). So, while it is useful to calculate a nation’stotal policy response or to compare responses across countries, it is also importantto understand what is included and not included in the calculation.

Given the rationale above, each worksheet in the database provides the sumof the amounts in Measures 01–05 (in US dollars), plus the amount in Measure010 because this refers conceptually to Measures 01–05 and the amount in Measure09 for the lenders. We refer to this as the Total Package provided. We stress thatMeasures 01–05 (and 010) include aspects as diverse as central bank or governmentpurchases of assets (Measure 01), the expected impact of lower interest rates interms of credit creation (Measure 02), and actual government spending (Measure05). The reason for adding them up is that, as we show in the next section, thesemeasures are consistent with either stimulus (i.e., results in multiplier effects greaterthan 0) or prevention of further macroeconomic decline (i.e., similar to automaticstabilizers but discretionary in this case). They are all “response measures.” Further,we of course recognize that any monetary sum cannot fully represent the measurestaken, given that authorities are adjusting interest rates, liquidity regulations, andcapital regulations. It is also important to recognize that measures (and amounts)announced are changing very often. Finally, some economies do not provide figuresfor measures that could be represented in monetary terms and have instead issuedpolicy statements without monetary amounts.2

2For these reasons, we ask users of the Policy Database to take this into account and exercise caution whenmaking comparisons.


For those measures that involve actual spending from the government, suchas lending and direct income transfers, the amounts reflected are either actual orestimates, depending on the availability of data. The following four are estimates:(i) effects of non-lending actions; (ii) forbearances; (iii) loan guarantees; and (iv)tax deferrals.

We also show the ratio of the total package to 2019 gross domestic product(latest available), the ratio of the economy’s package to the total package of theregional grouping the economy belongs to (according to ADB’s classification),as well as the ratio of the package to population (in US dollars). These ratiosvary significantly across economies as a consequence of the monetary amounts inMeasures 01–05. In some cases (typically, but not only, developing economies),the ratios are very small because economies are dedicating small amounts to thosemeasures, and instead they are passing laws that do not involve payments, e.g., lowerreserve requirements for banks, or ask banks to restructure loans or not to distributedividends.3

Recall that for the European countries that use the euro, we separate themeasures taken by the governments from those taken by the ECB and those of theEU institutions such as the EIB. Hence, some of these countries will not record anyliquidity measures under Measure 01. ECB measures, as well as those recorded inthe European institutions, cannot be apportioned by country.

Finally, we stress that the type of information that some economies provide(qualitative and, at times, not clear) requires some judgment in order to assign someactions into particular measures. This means that some actions may be reclassifiedin future versions as more information becomes available.

IV. A Classification of Macroeconomic Impacts

Table 1 and the measures in section II allow a classification ofmacroeconomic measures and their macroeconomic impacts. The latter are shownin Table 2.

First, every measure’s operational details for the main macroeconomicMeasures 01–05, are consistent with either stimulus (i.e., results in multiplier effectsgreater than 0) or prevention of further macroeconomic decline (i.e., similar toautomatic stabilizers but discretionary in this case). Whether there is stimulus orprevention for Measures 06 and 07 depends on the context. Measures 08 and 09 areprevention.

Second, the measures we consider could have one or several of the followingeffects: (i) change/support asset prices; (ii) private debt creation; (iii) delay paymentobligations; (iv) government/central bank claims on private sector; (v) contingent

3See previous footnote.


Table 2. Sample Worksheet for the ADB COVID-19 Policy Database and MacroeconomicImpacts of Each Measure

Measures Classification of Macro Impacts

Measure 01A Main purpose:Lending to the private sector or Preventionstate/local/regional governments, and asset

purchases to provide liquidity:Effects:Change/support asset pricesGovernment/central bank claims on private sector(i) Loans by the central bank or government

(standing facilities, loans to enablerefinance)

Measure 01ALending to the private sector or

state/local/regional governments, and assetpurchases to provide liquidity:

Main Purpose:PreventionEffects:Change/support asset prices

(ii) Asset purchases by the central bank orgovernment

Measure 01B Main Purpose:Non-lending actions and regulatory

adjustments: collateral requirements,payments system policies, liquidityregulations, reserve requirements, etc.

PreventionEffects:Change/support asset pricesPrivate debt creation

Measure 01C Main Purpose:Foreign exchange operations or domestic

lending in foreign currencyPreventionEffects:Change/support asset prices

Measure 02A Main Purpose:Secondary market purchases of securities

(greater than 1 year to maturity), and loansto financial sector

StimulusEffects:Change/support asset pricesPrivate debt creationGovernment/central bank claims on private sector

Measure 02B Main Purpose:Interest rate reductions and other regulatory

adjustments: capital requirements, credit andlending standards, oversight, etc.:

(i) Interest rate reductions

StimulusEffects:Change/support asset pricesPrivate debt creation

Measure 02B Main Purpose:Interest rate reductions and other regulatory

adjustments: capital requirements, credit andlending standards, oversight, etc.:

(ii) Capital requirements, lending oversight

StimulusEffects:Private debt creation

Measure 02C Main Purpose:Loan guarantees Stimulus

Effects:Private debt creationContingent liabilities of government/central bank

Continued.


Table 2. Continued.

Measures Classification of Macro Impacts

Measure 03A Main Purpose:Long-term direct lending to businesses,

households, and state/local/regionalgovernments

StimulusEffects:Private debt creationGovernment/central bank claims on private sector

Measure 03B Main Purpose:Forbearance Prevention

Effects:Delay payment obligations

Measure 04 Main Purpose:Equity claims on the private sector Prevention(i) Direct equity or preferred equity

investmentsEffects:Change/support asset pricesGovernment/central bank claims on private sector

Measure 04 Main Purpose:Equity claims on the private sector (ii)

Secondary market, ETFs, etc.StimulusEffects:Change/support asset pricesGovernment/central bank claims on private sector

Measure 05 Main Purpose:Government support to income/revenue Stimulus

Effects:Direct increase in private sector net financial assets

Measure 06 Stimulus vs. prevention depends on contextRedirecting or reallocating previously

budgeted spendingEffects:Double counting

Measure 07 Stimulus vs. prevention depends on contextCentral bank financing government operations Effects:

Change/support asset pricesDouble counting

Measure 08 Main Purpose:International assistance (borrower/recipient) Prevention

Effects:Change/support asset pricesDouble counting

Measure 09 Main Purpose:International assistance (lender/donor) Prevention

Effects:None

ADB = Asian Development Bank, COVID-19 = coronavirus disease, ETFs = exchange-traded funds.Source: Authors.

liabilities of government/central bank; (vi) direct increase in private sector netfinancial assets; and (vii) double counting. Every measure in Measures 01–05involves some combination of asset price changes/support and/or financial positioneffects for the private sector, while Measures 06, 07, and 08 are double counting,


Table 3. Summary of (Typical) Differences among Measures 01–04

Actions to SupportNormal

Functioning ofMoney Markets

(i.e., liquidityprovision)

EncouragingPrivate Credit

Creation

Lending toNonfinancial

Sector(unrelated to

liquidityneeds)

Equity Claimson the Private

Sector01 02 03 04

Maturities <1 year �1 year �1 year N/AMarkets Any short-term credit

marketSecondary debt

markets orloan purchases

Primary debtmarkets, directloans

Equities (primaryand/orsecondary,ETFs, etc.)

Borrowers Financial institutions,nonfinancialbusinesses,state/regional/localgovernments, centralbanks and officialaccounts (currencyswaps and similararrangements)

Financialinstitutions(who then lendto the privatesector)

Nonfinancialbusinesses,state/regional/localgovernments,households

N/A

Lenders Central bank andgovernment

Central bank andgovernment



ETFs = exchange-traded funds, N/A = not applicable.Source: Authors.

and asset price changes or support is possible for Measures 07 and 08. The effectof Measure 09 depends on the recipient economy. These effects are all consistentacross economies implementing the same measure.

Finally, Table 3 shows the typical differences in maturities, markets, lenders,and borrowers across Measures 01–04. These characteristics are “typical” or“usual,” but not necessarily universal or present in every circumstance.

V. Additional Clarifications

We end this paper with a collection of additional clarifications on somecategories and subcategories.

Liquidity

As in the impact classification above for Measure 01, “liquidity” provisionin a financial crisis like the current one, or the global financial crisis (GFC),operationally is not stimulus. These are operations for preventing a still worse crisis,or, so to speak, putting a floor underneath it. Under normal circumstances, moneymarkets enable payment settlement, short-term funding of financial institutions’operations, and funding of working capital for non-financial businesses mostly


on private sector balance sheets. A central bank’s role is usually to backstop thelargest banks and/or government bond dealers through overdrafts, direct loans (oftensecured), repurchase agreements, and outright asset purchases, sales, or issuance.Banks and dealers in turn, in hierarchical fashion, provide the same to otherfinancial institutions and corporations. While the central bank is certainly active inthe money markets in various ways, it is itself normally involved in a small subsetof liquidity provision with a subset of financial institutions.

In a financial crisis, the money markets cease normal functioning. In theGFC, the failure of Lehman Brothers in September 2008 was obviously central.Lehman was a market maker in several money markets, and was also holdingbillions of US dollars for its clients, who could not get to their accounts tomeet their commitments. Even before Lehman’s failure, though, the significantdisruption to commercial paper and Eurodollar markets led the Federal Reserve(Fed) to create several new standing facilities to lend to banks and dealers. Later,the Fed’s currency swap operations with other central banks, which were lines ofcredit, enabled those central banks to support their domestic US dollar markets.The Fed’s subsequent commercial paper standing facility directly funded short-term operations of issuers, while its Term Asset-Backed Securities Loan Facility(TALF) operations to purchase asset-backed securities (ABS) in secondary marketsencouraged mortgage lending by ensuring a market buyer for the securitized loans.

Central banks’ liquidity operations during the GFC provided the funding thatis, under normal circumstances, provided by private financial markets. As financialinstitutions could not carry out their normal roles in these markets, and/or as thequantity of trades in some markets overwhelmed what would be “normal,” centralbanks simply took these markets onto their balance sheets temporarily, partiallyor in full. These operations are nothing like “flooding markets with liquidity.” Ifanything, the “amount of liquidity” available in these markets was less than innormal times.

The “supply shock” brought on by COVID-19 has disrupted money marketsagain at an even faster rate and a larger scope than the GFC. Almost overnight,every business told to cease operations and every worker told to stay home becamea substantially greater financial risk. The same goes for the businesses regularlyin the supply chain of the closed businesses, the places those workers shop, thelenders they owe debt service to, the landlords they owe rent to, the utilities they paymonthly, the governments they pay taxes to, and so on. And most of these entitiestypically finance their short-term operations in money markets or through short-term credit arrangements.

It should be obvious that operations by central banks and governments toprovide this short-term finance when normal liquidity provision is severely impairedis preventive, not stimulative. The mix of quickly devised and scaled standingfacilities has not been comparable to the speed, depth, and scope of liquidityprovision that occurs daily during normal times.


Measures 01A, 02A, and 03A

It is admittedly difficult to apply precise categorizations of different measuresto “liquidity” versus “encouraging private credit creation” versus “direct lending.”The Fed’s TALF operations noted above were clearly providing liquidity to thesecondary markets for ABS; such ensuring of short-term refinance in moneymarkets to lenders obviously encourages private credit creation. In other words,there is no precise, clear line that can demarcate encouraging credit creation fromproviding liquidity. As shown in Table 3, our approach here in Measures 01A, 02A,and 03A is to separate them by maturity (1-year or less is 01A; greater than 1year is 02A or 03A) and by primary or secondary market for 02A (secondary) and03A (primary). So, the Fed’s earlier TALF facility would be in Measure 02A sincethe ABS purchased were in the secondary market and had more than 1 year tomaturity remaining. Consistent with this, Measure 02A is qualitatively the better fitfor TALF purchases of ABS: although they provided liquidity to the ABS secondarymarket, the purpose of the facility was to encourage credit creation more than toenable businesses and financial institutions to settle payment commitments, rollover maturing short-term liabilities, or finance or refinance working capital.

Measures 01B and 02B

Here again there is no clear dividing line between policy and regulatorymeasures that enable liquidity versus providing incentive for credit creation. Areduction in, say, the supplementary leverage ratio (02B) can enable liquidityprovision by opening up balance sheet space for large banks to lend into moneymarkets, while a reduction in liquidity coverage ratio requirements (01B) can reducebanks’ regulatory costs to providing longer-term credit. The approach here is simplyto separate policy measures that tend to require lenders to enter money markets tomeet the requirements (01B) from policy measures that tend to directly enable orrestrict credit creation either by encouraging borrowers (reduced interest rates, forinstance) or offering lenders greater balance sheet space, less oversight, and so forth(02B).

Loan Guarantees

A significant difference from other taxonomies is our treatment of loanguarantees, especially from the government sector. Loan guarantees are contingentliabilities—effectively insurance policies or put options—not government spendingper se (unless the government or central bank normally sell the guarantees to lendersbut are now making them available at a lower cost or at no cost). Only in the eventof a loan default will government spending increase. Consequently, whereas it hasbeen common for others such as the Organisation for Economic Co-operation and


Development (OECD) to categorize loan guarantees as part of government deficitincreases devoted to the COVID-19 response, we initially place loan guaranteesin Measure 02C; thereafter, for those government-guaranteed loans that fail, wewill subtract the appropriate value from Measure 02C and add it to Measure 05(our measure that most nearly resembles the typical understanding of a governmentdeficit).

Our approach is similar for government loan programs that call for thecancellation of loans that meet certain criteria (such as loans to businesses thatqualify for cancellation if the borrower uses the funds to pay employees and soforth)—a government transfer of income to the private sector occurs only when theloan is cancelled, and so it is at that point (and not before) that the measure’s relevantmonetary value is debited from its current measure and credited to Measure 05.

Consequently, we fully expect that our Measure 05 is not equal to theprojected fiscal packages of countries—in our view, there is an important differencebetween measures that have raised the private sector’s net income directly and thosethat do not, especially if the measure instead raises the private sector’s liabilities(even temporarily).

Forbearances

Forbearances are frequently reported incompletely for precise categorizationfrom an accounting perspective. If the government offers a delay in paymentrequired of its borrowers or taxpayers, it is clear the measure belongs in our Measure05. But if legislation requires banks to allow their borrowers to delay payments, forinstance, more information is required for precise categorization. If the governmentis subsidizing the banks for lost or delayed interest, that spending also belongs inour Measure 05. On the other hand, if the government is not subsidizing banks forlost or delayed interest, then the forbearance adds a financial burden for lenderswhile reducing financial burdens for borrowers. This is similar for delays in rent,utility payments, and so forth. Unfortunately, incomplete reporting is the normfor forbearances; where it is clear that the government or central bank is bearinga financial burden, we record this in Measure 05. For any other portion of aforbearance, or if there is not enough information to make such a determinationin the first place, we record this in Measure 03C.

Measure 07B and “Liquidity”

When a central bank purchases government bonds in the secondary market(Measure 07B) in normal times, some refer to this as an increase in “liquidity.”Similarly, they may want to refer to these operations in the same way in thecurrent pandemic scenario. The central bank’s purchase has two sides to it—thepurchase of the government security is added to the central bank’s assets,


while the purchase is paid for by adding reserves (that is, central bank settlementbalances) to the purchaser’s bank’s account at the central bank (and the purchaser’sbank subsequently credits the purchaser’s account). The added reserves are theincrease in central bank’s liabilities corresponding to the increase in its assets.Our approach, implicit in Measures 01–04, is to categorize loans and assetpurchases according to what claims on the private sector have been acquired bythe government or central bank—that is, according to what assets the governmentor central bank has acquired. The one measure that is based on what the privatesector acquires from the government or central bank is Measure 05, because it isa direct increase in the private sector’s income and (by definition) a decrease inthe income of the government or central bank. Measure 07B does not increase theprivate sector’s income but is rather an exchange of a government liability for acentral bank liability. Therefore, our approach here to use Measure 07B for centralbanks’ secondary market purchases of government bonds is consistent with ourcategorization of central bank actions in Measures 01–04 that likewise account forthe assets the central bank acquires rather than the liabilities it has created to makethe acquisitions.

References

Policy trackers:European Bank for Reconstruction and Development (EBRD). Coronavirus Policy Response.

https://www.ebrd.com/what-we-do/coronavirus/coronavirus-policy-response.EY. COVID-19: How Are Governments Responding to the Call for Stimulus? https://www

.ey.com/en_gl/tax/how-covid-19-is-causing-governments-to-adopt-economic-stimulus–.Institute of International Finance (IIF). Global Responses—Developed Markets. https://www

.iif.com/Portals/0/Files/Databases/COVID-19_responses.pdf.International Monetary Fund (IMF). Policy Responses to COVID-19. https://www.imf.org

/en/Topics/imf-and-covid19/Policy-Responses-to-COVID-19.KPMG. Government Response—Global Landscape. https://home.kpmg/xx/en/home/insights

/2020/04/government-response-global-landscape.html.Official Monetary and Financial Institutions Forum (OMFIF). Central Bank Policy Tracker.

https://www.omfif.org/policy-tracker/?utm_source=omfifupdate.Organisation for Economic Co-operation and Development (OECD). https://www.oecd.org

/coronavirus/en/.Yale School of Management. Program on Financial Stability: COVID-19 Crisis. https://som

.yale.edu/faculty-research-centers/centers-initiatives/program-on-financial-stability/covid-19-crisis.

Central sources for international assistance received:African Development Bank Group (AFDB). https://www.afdb.org/en/countries.Asian Development Bank (ADB). COVID-19 (Coronavirus): ADB’s Response. https://www

.adb.org/what-we-do/covid19-coronavirusADB. Internal document.Government of the United States, Human Resources & Services Administration. https://bphc

.hrsa.gov/emergency-response/coronavirus-cares-FY2020-awards


Inter-American Development Bank (IDB). The IDB Group in Response to COVID-19(coronavirus). https://www.iadb.org/es/coronavirus.

International Monetary Fund (IMF). Emergency Financing and Debt Relief. https://www.imf.org/en/Topics/imf-and-covid19/COVID-Lending-Tracker.

Islamic Development Bank (ISDB). Covid-19 Funding Overview. https://www.isdb.org/covid-19-overview.

United Nations Development Programme (UNDP). UN COVID-19 Response and Recovery Fund.http://mptf.undp.org/factsheet/fund/COV00.

USAID. News and Information—Facts and Information. https://www.usaid.gov/news-information/coronavirus/fact-sheets/may-29-2020-update-united-states-continues-lead-global-response-covid-19

World Bank (WB). Operational Response to COVID-19 (coronavirus)—Projects List. https://www.worldbank.org/en/about/what-we-do/brief/world-bank-group-operational-response-covid-19-coronavirus-projects-list.

For trade measures:World Trade Organization (WTO). COVID-19 and World Trade. https://www.wto.org/english

/tratop_e/covid19_e/covid19_e.htm.

For US dollar swaps:Federal Reserve Bank of New York. Central Bank Liquidity Swap Operations. https://apps

.newyorkfed.org/markets/autorates/fxswap.

Evaluating Strategies to Reduce ArsenicPoisoning in South Asia: A View

from the Social SciencesM K, A M M,

A G∗

The World Health Organization has labeled the problem of arseniccontamination of groundwater in South Asia as “the largest mass poisoning inhuman history.” Various technical solutions to the problem fall into one of twobroad categories: (i) cleaning contaminated water before human consumptionand (ii) encouraging people to switch to less contaminated water sources.In this paper, we review research on the behavioral, social, political, andeconomic factors that determine the field-level effectiveness of the suite oftechnical solutions and the complexities that arise when scaling such solutionsto reach large numbers of people. We highlight the conceptual links betweenarsenic-mitigation policy interventions and other development projects inBangladesh and elsewhere, as analyzed by development economists, that canshed light on the key social and behavioral mechanisms at play. We concludeby identifying the most promising policy interventions to counter the arseniccrisis in Bangladesh. We support a national well-testing program combined withinterventions that address the key market failures (affordability, coordinationfailures, and elite and political capture of public funds) that currently preventmore deep-well construction in Bangladesh.

Keywords: arsenic, health behavior, water qualityJEL codes: I12, O15, Q53

I. Introduction

Much of the world’s disease burden is due to environmental threats (Pruss-

Ustun and Corvalan 2006). People often respond to environmental health risks

by adopting technologies that reduce the risk (Pattanayak and Pfaff 2009). For

example, people can invest in preventive health products such as bed nets to reduce

∗Matthew Krupoff: Yale University, Yale Research Initiative on Innovation and Scale (Y-RISE); Ahmed

Mushfiq Mobarak (corresponding author): Yale University, Y-RISE; Deakin University; Center for Economic Policy

Research; and National Bureau of Economics Research. E-mail: [email protected]; Alexander van Geen:

Columbia University. Alexander van Geen’s arsenic-related research has been supported by NIEHS grant P42

ES010349, NSF grant ICER1414131, and several grants from the Earth Institute at Columbia University. We would

like to thank the managing editor and the anonymous referee for helpful comments and suggestions. The Asian

Development Bank recognizes “Orissa” as Odisha. The usual ADB disclaimer applies.




22 A D R

their risk of malaria or chlorine tablets to reduce the risk of acute gastrointestinal

diseases like diarrhea. Arsenic contamination of drinking water is one such

important challenge, and this paper describes the scope of that problem, technical

solutions that can reduce contamination, and the design of policies to encourage

widespread adoption of a solution that could effectively address this public health

threat.

An estimated 45 million Bangladeshis consumed drinking water with arsenic

concentration levels exceeding what is deemed dangerous to the human body

according to a report published in 2009 (Bangladesh Bureau of Statistics and

UNICEF 2011). The World Health Organization (WHO) referred to chronic

exposure to arsenic from drinking well water in Bangladesh as “the largest mass

poisoning of a population in history” (Smith, Lingas, and Rahman 2000). As

a response, the government and various nongovernment organizations (NGOs)

have implemented strategies to mitigate exposure to arsenic. Some of the initial

attempts at arsenic mitigation focused on the technological aspects of arsenic

removal. These efforts can only be successful to the extent that the technology

is widely implemented by policy makers and/or adopted and used by households

drinking contaminated water. Complexities in implementation, the political calculus

of policy makers, coordination failures in the community, or simply household

aversion to behavior change can undermine the promise of technically effective

solutions.

Certain fields within social science, such as development economics and

behavioral economics, have developed insights that can help us understand the

sources of aversion to behavior change and the challenges of implementing

technically effective solutions. For example, economic analysis can shed light on

the reasons for low demand for point-of-use filters despite their apparent large

benefits. Mechanism design can be used to overcome collective action failures. And

randomized controlled trials and other techniques can be used to rigorously evaluate

the effects of policy interventions and advise policy makers on the strategies that

work best.

This paper analyzes the behavioral, economic, and institutional challenges of

implementing arsenic mitigation interventions and identifies solutions that appear

most promising according to the evidence base. The interventions we review fall

under two broad classes of strategies: (i) either remove arsenic from contaminated

water before it enters the human body or (ii) encourage consumers to switch to a

different water source with a lower arsenic concentration. The paper also discusses

the complexities of scaling up arsenic-mitigation interventions to address the needs

of tens of millions of people.

This paper is organized as follows. Section II provides a background on

the arsenic poisoning crisis in South Asia and other parts of the world. Section

III discusses the two thematic behavioral strategies to reduce arsenic exposure,

potential solutions that fall under these strategies, challenges in implementation,

E S R A P S A 23

and interventions that overcome those challenges backed by empirical evidence.

Section IV discusses the complexities of scaling interventions that address the issue.

Section V concludes with policy recommendations.

II. Background

Arsenic contamination is not unique to Bangladesh, but it is the most affected

country in the world by far. Arsenic is naturally released into groundwater by

Himalayan sediments. As a result, the groundwater in many countries in South

and Southeast Asia (including India, Myanmar, Nepal, Pakistan, Cambodia, the

Lao People’s Democratic Republic, and Viet Nam) is contaminated to some degree

(Ravenscroft, Brammer, and Richards 2009). Bangladesh is especially affected,

with an estimated 45 million Bangladeshis consuming drinking water with arsenic

concentration exceeding the WHO guideline of 10 micrograms per liter (Smedley

and Kinniburgh 2002; Fendorf, Michael, and van Geen 2010).

There was a massive shift toward groundwater in Bangladesh in the 1970s

and 1980s due to public health concerns about bacterial contamination of surface

water sources. Excess infant mortality from diarrheal diseases, cholera, and other

waterborne illnesses led governments, international donors, and NGOs to undertake

massive programs promoting shallow tube-well installation across the country to

reach aquifers free of pathogens.

The presence of arsenic in groundwater was first noted in the early 1980s

in the geologically similar neighboring Indian state of West Bengal, when visible

manifestations of the disease were identified and attributed to water from shallow

tube wells (Chakraborty and Saha 1987). It was not until the late 1990s when the

scale of the problem was fully understood, prompting massive public health action

by the Government of Bangladesh and multinational organizations like the World

Bank to test tube wells across the country (Dhar et al. 1997). By 2005, 1.4 million

shallow wells with groundwater with an arsenic concentration above Bangladesh’s

drinking water standard of 50 micrograms per liter were painted red; another 3.5

million wells that were below the contamination threshold were painted green. Most

tube wells have been replaced since then and very few were ever retested after the

national testing campaign ended in 2005 (Ahmed et al. 2006, van Geen et al. 2016).

Some early efforts to mitigate the arsenic crisis focused on switching from

groundwater to surface water from hand-dug wells, rainwater storage devices,

and (filtered) pond and river water (Ahmad, Khan, and Haque 2018). Whereas

switching to surface water sources can reduce arsenic consumption, it can also

have the unintended consequence of increasing the risk of disease through fecal

contamination (Lokuge et al. 2004, Howard et al. 2006, Johnston et al. 2014).

The health impacts of chronic arsenic exposure are severe (Vahter et al.

2010). It is estimated that 6% of total mortality in Bangladesh is due to chronic

24 A D R

exposure to arsenic (Flanagan et al. 2012). The main cause of the excess mortality

is cardiovascular disease and not the types of cancer that researchers have linked

to arsenic elsewhere (Smith, Lingas, and Rahman 2000; Chen et al. 2011). Chronic

exposure has also been linked to increases in stillbirths, infant mortality, and motor

and intellectual impairment of children (Wasserman et al. 2004, Parvez et al. 2011,

Quansah et al. 2015).

Arsenic exposure negatively affects productivity. Pitt, Rosenzweig, and

Hassan (2020) estimate that reducing Bangladeshi arsenic retention to United

States levels would, on average, increase household income by 9% per male

worker. Flanagan et al. (2012) estimate that arsenic-related mortality is expected

to cost Bangladesh $12.5 billion from lost productivity over the next 20 years. The

authors base this estimate on the productivity loss associated with deaths from the

types of cancer known to be related to arsenic poisoning. However, this may be

an underestimation because this economic loss does not account for health-care

expenditures and other costs to society.

III. Strategies to Reduce Arsenic Consumption

Solutions that reduce arsenic in the water supply involve different categories

of interventions as well as coordination between policy makers, implementers,

communities, and end users. There are two broad strategies to address arsenic

poisoning. The first is to clean the contaminated water before it enters the body

by means of technological solutions like filtration systems. The second is to have

people switch to clean sources of water by means of well testing and building low-

arsenic deep tube wells. Both strategies require households to change their behavior.

Barriers to a household’s willingness to invest in preventive health products,

coordination failures, and political economy factors are all challenges that must

be addressed through careful policy design.

Removing arsenic from water or inducing people to switch to cleaner sources

may require households to invest resources into buying water filters or installing

deeper wells. Research by development economists in a variety of settings has

found puzzlingly low rates of preventive health investments among poor households

despite the long-run benefits (Kremer and Miguel 2007; Ashraf, Berry, and Shapiro

2010; Meredith et al. 2013). Factors such as liquidity constraints, information

failures, peer effects, and intra-household conflicts over health are found to be

responsible for the low demand (Brown, Mobarak, and Zelenska 2014). These

barriers to technology adoption will be discussed in more detail in the following

sections as they relate to specific arsenic mitigation approaches. We will highlight

successful policy interventions that have managed to overcome such barriers in

other settings.

E S R A P S A 25

Some solutions require community members to make collective decisions on

the locations of clean water sources and to coordinate community contributions

in cash, labor, and maintenance (Cocciolo, Habib, and Tompsett 2019). Failure

to coordinate between group members, such as free riding, can hurt the long-run

sustainability of community-based programs.

Public goods are sometimes delivered in a decentralized way where

investments are delegated to local governments. For example, decisions about

deep tube wells in Bangladesh—wells over 150 meters that are low in arsenic—

are delegated to and financed by local governments. Decentralization of service

delivery is thought to be efficient because local governments may have more

accurate local information to better target services (World Bank 2003). However,

taking a decentralized approach in rural communities with poverty, socioeconomic

inequality, and a lack of political awareness can lead to distortions in targeting

toward elites (Bardhan and Mookherjee 2000).

A. Cleaning Up before It Enters the Body

Filtering methods to clean contaminated water was promoted by the National

Arsenic Mitigation Policy in response to the discovery of arsenic in well water.

Pond sand filters and small community slow sand filters were promoted because

they could purify readily available surface water from ponds and rivers. However,

support for sand filtration diminished because of the susceptibility to fecal

contamination (Howard et al. 2006). Early efforts to remove the arsenic from

groundwater using large arsenic removal plants were ineffective in reducing arsenic

poisoning due to technical problems and poor maintenance (Hossain et al. 2005).

Some household-level filtration devices may be effective, but demand for such

products is low. Community filtration systems that serve large numbers of people

are promising, provided that maintenance efforts are properly coordinated. This

section will go over these options, their challenges, and recommendations.

1. Point-of-Use Treatment

Point-of-use arsenic purification filters—such as SONO water filters, three-

pitcher filters, and READ-F filters—have been shown to effectively reduce arsenic

levels (Hussam and Munir 2007, Sutherland et al. 2002). However, field tests have

found disappointing results on their adoption and usage (Johnston, Hanchett, and

Khan 2010). One example is Sanchez et al. (2016), who provided households

with READ-F filters—an easy-to-use device that filters arsenic from shallow well

water—and encouraged their use over the 6-month duration of the intervention.

Initially, participants showed a reduction in urinary arsenic levels, which is an

objective indicator of intake and exposure. However, the benefits eroded over time

26 A D R

and arsenic in urine returned to preintervention levels by the end of the study period.

After 1 year, 95% of the filters had been abandoned.

More research is needed to ascertain how much households are willing to pay

for point-of-use filters for arsenic removal and how to encourage their use. Research

on other water purification products have shown that demand has been low among

poor households (Ahuja, Kremer, and Zwane 2010). In Ghana, for example, Berry,

Fischer, and Guiteras (2019) measured the demand for Kosim water filters, which

are effective at removing more than 99% of E. coli in trials. Their assessment found

that households are willing to pay only 10%–15% of the cost of manufacturing and

delivery. Similarly, Ahuja et al. (2010) found low willingness to pay for point-of-

use chlorine treatment in Kenya when households were given coupons to redeem at

local stores.

Liquidity constraints are cited as a key reason why demand for health

products in developing countries is low despite their high benefits. People in poor

rural areas may not have the liquidity necessary to pay large lump-sum costs

for preventive health products. For example, SONO filters which remove arsenic

through chemical reactions with iron, cost about $40 (Hussam and Munir 2007).

High prices and the low willingness to pay suggest that price subsidies may

be a sensible policy to increase the adoption of point-of-use filters. However—

in addition to concerns about the fiscal capacity to provide subsidies—there

are concerns that lower prices may affect how people value the product and

subsequently use them. The psychological bias called the sunk cost fallacy posits

that higher prices cause people to value a product more than if they got it free.

Screening effects are when higher prices screen buyers who place a relatively high

valuation on a product and thus would likely use it more than someone who is less

willing to pay (Thaler 1980, Bagwell and Riordan 1991).

The Read-F filters used in Sanchez et al. (2016) were provided for free

and the low usage they observe may lend support to concerns about sunk cost

fallacy and screening effects. However, without observing adoption decisions under

experimentally varied prices, this remains inconclusive. Field experiments that

explicitly test for sunk cost fallacy and screening effects suggest that these concerns

are unfounded (Ashraf, Berry, and Shapiro 2010).

Information failures may cause people to underestimate the true

benefits of certain decisions from school choice or adopting new agricultural

technology. Households may thus underinvest in preventive health decisions

because they lack information about health risks (Somanathan 2010). There is some

evidence that providing information about water quality increases adoption of water

filters. In India, Jalan and Somanathan (2008) found that 45% of those surveyed

did not equate contaminated water with diarrhea. The researchers tested the water

and informed a randomly selected group of households about the contamination

status and the various purification methods that are available. Households with

contaminated water increased efforts to purify water before consumption once

E S R A P S A 27

they were informed. Information programs have also been designed for arsenic

mitigation in Bangladesh and have proven to be highly effective in reducing

consumption through inducing households to switch to cleaner wells (Madajewicz

et al. 2007).

2. Community Filtration Systems

Centralized community-based water treatment systems are an alternative to

household point-of-use filters. These can supply arsenic-free water to around 100–

200 families (German et al. 2019, Sarkar et al. 2010). Current units can produce up

to 1 million liters of clean water before needing replacement (Sarkar et al. 2010).

Community filtration systems have certain advantages over household filters. For

example, arsenic levels are easier to monitor with centralized filtration systems

because the tests only need to be administered at one community unit, instead

of household filter units that are spread out. Centralized systems also make it

easier to coordinate proper waste disposal compared to household filters (Johnston,

Hanchett, and Khan 2010). However, the high cost and regular maintenance needs

lead to concerns about long-run sustainability. This has led to concerns about the

capacity of governments and NGOs to successfully deliver services.

Certain institutional arrangements in which community members organize

funds and provide maintenance may address sustainability issues with rural water

infrastructure. In such arrangements, village water committees collect small fees

from villagers that contribute to the cost of maintenance. Maintenance itself is

conducted by caretakers who are appointed by the committee. In some models,

committees have little explicit public authority for revenue collection, but such

cases do not show promising results. For example, Miguel and Gugerty (2005)

report that 50% of borehole wells in Kenya that were maintained using a

community-based maintenance model on a voluntary basis were inoperable by

2000. In rural Tanzania, free riding and a lack of coordinated maintenance

decisions decreased the functionality rate of NGO-installed clean water pumps

and consequently lowered rates of child survival and school attendance (O’Keeffe-

O’Donovan 2019).

Clean water is a public good and maintaining it has positive externalities

for other people in the community. If there are coordination failures and

free riding, then it becomes difficult to maintain quality under community-

based arrangements. Many community-level interventions experience coordination

difficulties. One example is community toilets in India, where a study showed

that one in six toilet seats was entirely nonusable (J-PAL 2012). Communal

arrangements must be structured to ensure that incentives are correctly aligned, and

the community can monitor its members (Duflo, Galiani, and Mobarak 2012).

Some evidence suggests that private contracting maintenance systems are an

efficient way of maintaining water sources (Kremer et al. 2011). For point-of-source

28 A D R

chlorine dispensers in Kenya, Ahuja et al. (2010) found that paying contractors

to maintain the system increased the level of maintenance significantly. Water

collection fees can discourage free riding, leading to an increase in functionality

(O’Keeffe-O’Donovan 2019). Current community-based filtration systems that

charge user fees as low as $0.15–$0.30 a month and compensate unit caretakers have

been found to be financially sustainable and lead to local job growth (German et al.

2019). Complementing a system with delivery services can also increase demand

and revenue generation (Johnston, Hanchett, and Khan 2010; Sarkar et al. 2010;

German et al. 2019).

B. Switch to Groundwater That Is Already Low in Arsenic

The second strategy to mitigate arsenic poisoning is to encourage people to

switch from a high-arsenic water source to a clean water source. Individuals choose

their water source to maximize their welfare subject to the constraints they face and

their information set. Consuming arsenic-contaminated water may be indicative of

information failures or a lack of alternative clean water sources. For example, since

arsenic levels in groundwater vary greatly over small distances, informing people

of the status of their wells can induce them to switch to neighboring clean wells.

Fortunately, concentrations of arsenic usually do not change over time, although

some aquifers and wells need to be monitored more frequently than others (Fendorf,

Michael, and van Geen 2010). Increasing a household’s access to clean water by

installing new low-arsenic deep tube wells is also a strategy worth considering.

1. Information and Testing

People may drink from contaminated wells if they lack information about

the arsenic concentration in their shallow well relative to other nearby wells. The

distribution of arsenic in groundwater varies greatly, even over small distances

and most owners live within walking distance of an uncontaminated well. Testing

the groundwater concentration is therefore essential to provide the necessary

information for people to switch (van Geen et al. 2002).

Arsenic tests are attractive because of the low cost to administer them. In

previous interventions, the cost of a simple test was as low as $2.30, with the cost of

supplies only amounting to $0.30 per test. Because of the large health consequences

of chronic exposure to arsenic, simply providing information through arsenic tests

can therefore be a highly cost-effective intervention as long as people respond to the

new information. Evaluations show that providing test data to households, in some

cases along with various forms of reinforcement, has induced between one-quarter

and one-half of exposed households to stop using contaminated wells (Madajewicz

et al. 2007, Bennear et al. 2013, Balasubramanya et al. 2014, Pfaff et al. 2017).

E S R A P S A 29

One issue with arsenic tests is how they should be provided and who should

provide them. Public provision has not met the needs for testing. Recent estimates

show that despite the national well-testing campaign between 2000 and 2005, at

least one-half of the currently used tube wells in Bangladesh have never been tested

for arsenic (van Geen et al. 2014, Jamil et al. 2019). National testing campaigns

have not been repeated and most wells have by now been replaced and therefore

were never tested.

Private testing may be a useful complement to public provision. The prospect

of a private market for arsenic testing can induce local entrepreneurs to identify

untested wells and market their services (Barnwal et al. 2017). Despite the low

cost, poor households may not be able to afford arsenic test kits. An evaluation

in the neighboring Indian state of Bihar shows that while demand for test kits is

substantial, it is also highly price sensitive: the take-up level falls from 69% to 22%

when cost increases from $0.16 to $0.80. This steeply downward-sloping demand

curve is reminiscent of the elastic demand for other effective preventive health-

care products such as insecticide-treated bed nets and deworming pills (Kremer and

Miguel 2007, Cohen and Dupas 2010). Subsidizing testing kits may be efficient

policy if encouraging initial usage helps neighbors learn about the value of testing

and increases the demand for future testing. Barnwal et al. (2017) find that demand

for test kits rose from 27% to 45% within 2 years of the initial subsidy campaign

without any change in the nominal sales price.

Households will switch away from contaminated to cleaner wells after testing

only if they know about the health consequences of arsenic in the first place.

Interventions that combine tests with education about arsenic poisoning have been

shown to increase switching (George et al. 2013, Chen et al. 2007, Pfaff et al. 2017).

For example, Khan et al. (2015) found higher switching rates among children after

an arsenic education curriculum designed to raise awareness of arsenic poisoning

was administered in elementary schools in Araihazar, Bangladesh.

Tests are commonly provided by representatives from organizations outside

of the village who leave once tests are administered, leaving little opportunity

to reinforce that information. Training community members to deliver arsenic

education concurrently with testing may be a more cost-effective way to monitor

arsenic levels and reinforce information about health consequences. Such types

of community health worker programs are a widely used intervention to improve

the quality of health-care services—from health education to family planning and

distribution of preventive care products—around the world. However, in one study,

engaging community members did not decrease arsenic exposure any more than

outside testers (George et al. 2012). Poor monitoring and a lack of incentives—

common problems with other community health worker programs—may have been

a reason why there was no difference. Providing monetary incentives to health

workers, or better monitoring, may help improve performance and lead to better

outcomes (Björkman et al. 2017, BenYishay and Mobarak 2019).

30 A D R

2. Well-Sharing Arrangements

Many shallow wells are privately owned, and arsenic concentration levels

vary between wells; therefore, exposure varies from household to household.

Sharing arrangements between owners of clean shallow wells and owners of dirty

wells can increase the proportion of the population consuming clean water. Such

arrangements are possible in areas where houses are geographically close to one

another and people interact on a regular basis, which is often the case in small

village economies (Barnwal et al. 2017). However, households may not be willing

to share with people outside their social network and low-income households may

be less able to barter for access to a neighbor’s clean well than households that

are better off (Madajewicz et al. 2007). Social constraints may also be important

determinants of water source usage (Mosler, Blochliger, and Inauen 2010; Inauen

et al. 2013). Households with unsafe wells have also been found to purposefully

conceal the results of the test, suggesting that social stigma could partially be to

blame (although this could also be explained by concerns that the reveal would

lower property value) (Barnwal et al. 2017).

These results suggest that we need to design mechanisms that are

cognizant of such social constraints. For example, combining testing with a group

commitment component where groups of households make a public commitment to

their group before seeing test results—that if their well is tested and found clean,

then they would promise to share water with those who have unclean wells—can

address free riding and aversions to water sharing. If households are risk averse, then

such a “risk-sharing contract” with ex ante commitments can improve joint welfare

for the group of households and help to develop positive social norms about water

sharing. Tarozzi et al. (2020) test this theory through a randomized controlled trial

in Sonargaon, Bangladesh in which groups of buyers were offered tests and asked

to sign an informal agreement about sharing water from their clean wells with those

who had negative well-testing results. This form of soft commitment showed higher

switching rates to clean water from dirty water among those who received a negative

result compared to the treatment group where well tests were done at the individual

level.

Public commitments have been shown to be effective in changing behavior,

having been tested for other public health goals such as latrine adoption. For

example, community-led total sanitation programs are an intervention aimed at

changing social norms about open defecation by having communities pledge

to become open-defecation-free. Bakhtiar, Guiteras, and Mobarak (2019) show

that combining a form of a community-led total sanitation program, in which

community members make public pledges in front of their neighbors, was effective

in increasing the adoption of latrines when compared to private pledges and group-

level financial incentives. In the context of arsenic, Inauen et al. (2014) show that

public commitments enhance the effects of information on well switching.

E S R A P S A 31

3. Installing Deeper Wells

In addition to well testing and well sharing, another promising approach is

to install wells that reach deeper aquifers where arsenic concentrations are lower.

In some areas, such aquifers are accessible at a depth of less than 90 meters and

therefore reachable by local drillers using manual methods (Gelman et al. 2004).

In one study area, many households switched to private intermediate-depth wells

in response to early well testing in 2003 (Jamil et al. 2019). However, the cost

of well construction increases linearly with well depth, and installation costs at

this depth reach $200 per well. Estimates from blanket well tests in Araihazar,

Bangladesh suggest that digging these expensive wells may still have positive net

benefits since around 60,800 inhabitants experienced reduced exposure through this

form of mitigation at an average cost of $28 per person.

Interventions to alleviate liquidity constraints may be necessary to help

households afford the installation costs. In particular, providing microcredit to

finance large purchases can enable households to invest and reap the long-run

benefits. Credit provision can increase investments in preventive health. In India,

providing access to microconsumer loans for insecticide-treated bed nets led to a

large increase in uptake (Tarozzi et al. 2014). In Morocco, providing households

access to credit to purchase a home water connection from a local water utility

company led to 69% of households buying a connection, compared to just 10% in a

control group (Devoto et al. 2012). In Cambodia, microloans significantly increased

the willingness to pay for household latrines by 45 percentage points compared to a

control group without the option to finance (BenYishay et al. 2017).

If there are learning externalities, then subsidies can induce others to

subsequently adopt. Understanding the social dynamics of demand is useful

to target subsidies efficiently. For example, targeting well subsidies to certain

groups, such as highly influential people in a social network or people of lower

socioeconomic status, can lead to greater subsequent adoption if neighbors learn

more about the benefits and costs of the new technology, or by changing social

norms. Social learning appears to have been important for nontraditional cookstove

adoption in Bangladesh, where households made inferences about the new stoves

based on information from people in their social network (Miller and Mobarak

2015). It was also relevant for hygienic latrines in Bangladesh (Guiteras, Levinsohn,

and Mobarak 2019).

Deep tube wells that are deeper than 150 meters are more consistently low in

arsenic but beyond the financial reach of most households. A deep tube well, when

properly located, can meet the needs of several hundred villagers for years while

requiring little maintenance (van Geen et al. 2003). Over 200,000 deep wells were

installed as of 2007 by both NGOs and the Government of Bangladesh (Department

of Public Health Engineering and Japan International Cooperation Agency 2009).

Despite their engineering promise, the installation costs, inclusive of labor and

32 A D R

materials, can reach up to $850, which is beyond what most rural households can

afford (Ravenscroft et al. 2014). Private deep wells therefore do not seem to be a

feasible solution absent financial support or encouraging community members to

pool their resources to jointly invest and share the well water. Khan et al. (2014)

find that households are willing to pay on average 5% of their disposable annual

household income for a communal deep-well fund. Variation in willingness to

pay across households implies that one needs to solve a complicated problem to

determine how much each member should be asked to contribute.

There are nontrivial challenges to successfully coordinating investments

across households. Cocciolo, Habib, and Tompsett (2019) found that in a

community-based program where members collectively made funding, location,

and maintenance decisions for deep tube wells, larger groups led to fewer

households participating in community meetings and less time spent deliberating

over source location. In addition, they found that fewer households contributed to

the cost of installation. As a result, larger groups saw smaller increases in the use of

deep wells compared to smaller groups. More empirical evidence is needed about

the drivers of collective action failure and on how community networks change as

interventions scale.

The choice of where to place deep wells creates important complexities.

Over half of the deep wells that have been installed by governments and NGOs

were sited in areas where the prevalence of contaminated shallow wells is modest

(Department of Public Health Engineering and Japan International Cooperation

Agency 2009, van Geen et al. 2016). Households in heavily affected areas live

too far from installed deep wells, beyond the 100–150-meter walking distance that

previous studies have found to be the maximum that members of rural Bangladeshi

households are willing to walk to fetch water (van Geen et al. 2003, Opar et al.

2007). From a blanket survey of all wells across Araihazar, van Geen et al. (2016)

find that less than one-third of arsenic-contaminated shallow wells are located

within walking distance (100 meters) of at least one of the 915 deep or intermediate-

depth wells in the study area. If deep wells had been more evenly distributed, the

percentage of shallow wells covered could have increased to 74%. Even when the

engineering and financing constraints are addressed, there still appears to be some

issue with the spatial distribution of deep-well placement (Figure 1).

One possible explanation for this inefficient deep-well placement is elite

capture of this valuable public resource. Local government officials in Bangladesh

have large discretionary authority over the siting of deep wells. In Araihazar, a

subdistrict where much arsenic research has been conducted, the central government

allocated funds to local government officials to install 50–100 deep wells each year

over a decade. The location of a well is determined on the basis of input from the

bureaucrat in charge of the subdistrict (Upazila Nirbahi officer), the senior local

government official (Upazila Parishad chairman) who is directly elected, and the

12 Union Parishad chairmen who are also elected (van Geen et al. 2016). This

E S R A P S A 33

Figure 1. Clustering of Deep Wells

Notes: Panel (a) shows the overlapping circles with a radius of 100 meters around installed deep wells and shows that

29% of contaminated well (red) were within 100 meters of a deep well. Panel (b) shows the optimal placement of a

subset of the wells from a regular-spaced grid.

Source: van Geen, Alexander, Kazi Matin Ahmed, E.B. Ahmed, and Imtiaz Choudhury. 2016. “Inequitable Allocation

of Deep Community Wells for Reducing Arsenic Exposure in Bangladesh.” Journal of Water Sanitation and Hygiene

for Development 6 (1): 142–50.

decentralization of deep-well provision can be prone to elite capture, in which

wells are preferentially targeted toward political, social, or economic elites in the

community.

Evidence of elite capture of deep-well placement has mounted. In 2017,

Human Rights Watch accumulated anecdotal evidence based on village interviews

that politicians were preferentially placing wells near political supporters (Human

Rights Watch 2016). Van Geen et al. (2016) report that about one-third of deep

wells were placed in inaccessible locations such as inside the compounds of

private households. Madajewicz et al. (2017) find that a community participation

intervention that was designed to limit the influence of elites led to an increase in

clean water access. Finally, Mobarak, van Geen, and Mangoubi (2019) investigated

34 A D R

the extent of elite capture by combining geospatial data on well placement and

newly collected geocoded data on the location of political and economic elites. The

authors find strong evidence that local politicians are more likely to have deep wells

built near them during periods when their political party is in power. This form of

elite capture accounts for about one-fifth of the inefficient spatial allocation of deep

wells.

IV. Complexities of Scaling Policy Interventions

Large-scale public health problems such as arsenic poisoning across

Bangladesh require scalable solutions. Implementing the strategies discussed

in this article—treating contaminated groundwater or switching to low arsenic

groundwater—is challenging and complexities may arise when going from a project

in one district to a nationwide policy. As a program scales, for example, there

may be spillover effects on nonbeneficiaries, friends and neighbors, and markets;

political reactions from voters and governments; macroeconomic, growth, and

welfare impacts; as well as concerns about the external validity of small-scale pilot

results (Davis and Mobarak 2020).

Interventions may have spillover effects onto neighboring households or

communities, interact with social networks, and affect market prices and wages.

For example, the more that people use filters, purchase well test kits, or engage

in well-sharing arrangements, the more attractive these behaviors may become

to other members in a social network. The installation of more deep wells or

community filtration systems in a given area could increase demand for spare parts,

tools, and skilled labor, leading to positive spillovers in maintenance costs and

better functionality. If people are less exposed to arsenic, they may become more

productive employees, leading to more employment opportunities and higher wages

in the community. More research on spillover effects can inform policy makers on

unintended costs and benefits that can remain hidden in small-scale programs. This

can motivate cost-effective intervention designs. For example, subsidies for test kits

or filters may only need to be provided to a subset of households if demand for such

products is interlinked between households, thus lowering the cost of the program

substantially.

People may also adapt and react to policies in ways that can produce

unintended effects. Some of those consequences might be negative. For example,

Field, Glennerster, and Hussam (2011) hypothesize that the widespread switching

to surface water after the discovery of arsenic in 1994 might have led to higher

exposure to fecal–oral pathogens, which in turn increased infant and child mortality.

On the other hand, people may also adapt in ways that produce unintended benefits.

Keskin, Shastry, and Willis (2017) show that mothers react to arsenic exposure

risk by increasing the propensity and the duration of breastfeeding, which provides

infants some measure of protection against arsenic contamination, and this in turn

E S R A P S A 35

reduces infant mortality. The effects of any arsenic mitigation policies at scale will

be inclusive of such adaptation and behavioral responses. For comprehensive policy

evaluation, it is important for social scientists to provide rigorous analysis on these

types of questions.

As an arsenic mitigation program scales up, it may change the behaviors of

politicians and policy makers in response to the program. For example, if politicians

have control of discretionary funding of deep wells, they may choose to install

more in their home areas to gain votes or target placement near other politicians to

gain political supporters. If funding for wells is externally funded by international

NGOs, programs could erode political accountability if leaders claim credit for

successful programs (Deaton 2013). On the other hand, externally funded programs

may elicit political or financial support, as found in the case of externally funded

sanitation programs in Bangladesh (Guiteras and Mobarak 2016). Research has

already shown that political factors have led to inefficient deep-well placement in

Bangladesh through elite capture. More research is needed on how best to address

these political influences. For example, research on community participation in

deep-well placement that imposed rules designed to limit the appropriation of

projects by elites effectively expanded access to clean water sources (Madajewicz

et al. 2017).

Changes in individual behavior induced by a program can, at scale, have

macro-level impacts. Large-scale interventions that reduce arsenic consumption

could boost human capital and labor productivity, which can lead to long-run

growth. However, macroeconomic models often require parameters to properly

predict macro-level impacts. Rigorous evidence from randomized controlled trials

can help calibrate these models to more accurately determine these impacts in the

medium to long run, and even simulate alternative policy scenarios. Net welfare

impacts are also important when evaluating a program but are difficult to measure

without modeling. For example, people may experience nonmonetary disutility by

walking a longer distance to a communal deep well as they may be more vulnerable

to crime if they must walk far and/or at night. Modeling can be used to answer

normative questions about welfare trade-offs that are important for policy decisions.

Social science research aspires to generate evidence that policy makers can

use to scale promising programs. Even if the research discussed above produced

internally valid estimates of the policies studied at pilot scale, there are open

questions about how programs would work outside the context of those evaluations.

Replication studies and subsequent meta-analyses will be useful to aggregate results

from different contexts.

V. Policy Recommendations

There are trade-offs in expanding access to clean water through well testing

versus installing deeper, more expensive low-arsenic wells. Jamil et al. (2019)

36 A D R

Figure 2. Cost-Effectiveness of Different Interventions

Note: This figure shows the number of people whose exposure was reduced in Araihazar, Bangladesh compared to

the cost per person of each type of intervention.

Source: Jamil, Nadia, Huan Feng, Kazi Ahmed, Imtiaz Choudhury, Prabhat Barnwal, and Alexander van Geen. 2019.

“Effectiveness of Different Approaches to Arsenic Mitigation over 18 Years in Araihazar, Bangladesh: Implications

for National Policy.” Environmental Science and Technology 53 (10): 5596–604.

conducted a cost-effectiveness analysis of alternative strategies in a particular area

and found that free nationwide well testing would be the most cost-effective way

of reducing exposure (Figure 2). Well testing alone reduced the exposed population

in their study area of Araihazar in the short term by an estimated 130,000 people.

The next most effective way was installing private intermediate-depth wells, which

lowered exposure for 60,000 people at a cost of $30 per person. In contrast,

installation of deep tube wells and piped-water supply systems by the government

reduced the exposure of little more than 7,000 inhabitants at a cost of $150 per

person (see Table on Comparison of Interventions). These numbers are a strong

argument in favor of free well testing.

Simply providing test results addresses an information failure, which has

been found to be a major impediment to the adoption of preventive health

technologies in a variety of contexts. Informing people about the level of arsenic

they are exposed to bolsters their demand for alternative sources of water. Therefore,

well tests must precede investments in alternative sources in order to maximize the

effectiveness of testing.

If well testing were complemented with interventions that make private

intermediate-depth wells more affordable, such as with subsidies or microcredit,

it could induce adoption and reduce exposure. A national database of well locations

with test results can help policy makers target subsidies to areas with a high density

of contaminated shallow wells. Designing subsidies that encourage sharing private

intermediate-depth wells with neighbors can also increase the coverage.

E S R A P S A 37

Com

pari

son

of

Inte

rven

tion

s

Exp

ose

dE

xp

osu

reE

xp

ose

dC

ost

per

Tota

lC

ost

Cost

per

Tota

lC

ost

Tota

lC

ost

Ara

ihaza

rP

op

ula

tion

Pro

porti

on

Pop

ula

tion

Govt/

NG

OG

ovt/

NG

OH

ou

seh

old

Hou

seh

old

per

Exp

osu

re

Mit

igati

on

Met

hod

Act

ivit

yR

each

edR

edu

ced

Red

uce

d($

)($

)($

)($

)R

edu

ced

($)

Tes

ting

and

swit

chin

g48,8

00

wel

lste

sted

(21,3

00

safe

)

220,0

00

60%

132,0

00

2.5

122,0

00

1

Pri

vat

ein

term

edia

tew

ells

8,4

50

inte

rmed

iate

wel

lsin

stal

led

(7,6

10

safe

)

67,6

00

90%

60,8

00

200

1,6

90,0

00

28

Dee

ptu

be

wel

ls916

dee

pw

ells

inst

alle

d(9

07

safe

)

51,2

00

10%

5,1

20

800

733,0

00

143

Pip

edw

ater

supply

312

connec

tions

inst

alle

d(3

12

safe

)

2,1

80

100%

2,1

80

250,0

00

250,0

00

300

a93,6

00

158

NG

O=

nongov

ernm

ent

org

aniz

atio

n.

Note

s:a10

yea

rsat

$2.5

0/m

onth

.T

his

table

pre

sents

com

par

isons

of

the

effe

ctiv

enes

san

dco

sts

of

var

ious

form

sof

arse

nic

mit

igat

ion

conduct

edin

Ara

ihaz

ar,B

angla

des

h.

Sourc

e:Ja

mil

,N

adia

,H

uan

Fen

g,

Kaz

iA

hm

ed,

Imti

azC

houdhury

,P

rabhat

Bar

nw

al,

and

Ale

xan

der

van

Gee

n.

2019.

“Eff

ecti

ven

ess

of

Dif

fere

nt

Appro

aches

toA

rsen

icM

itig

atio

n

over

18

Yea

rsin

Ara

ihaz

ar,B

angla

des

h:

Impli

cati

ons

for

Nat

ional

Poli

cy.”

Envi

ronm

enta

lSci

ence

and

Tec

hnolo

gy

53

(10):

5596–604.

38 A D R

Although Jamil et al. (2019) find that deep-well construction is much less

cost-effective, the analysis by Mobarak, van Geen, and Mangoubi (2019) suggests

that much of that is due to inefficient placement and elite capture. Deep wells are

often forcibly “privatized” by politicians to use as a personal resource. This prevents

other households from gaining access to clean water, even after expensive deep-

well construction. Institutional reform that limits the discretion of public officials

to site deep wells as they please would increase the efficiency of public funds that

are deployed for well construction. Increasing either voter awareness or national

government supervision of local politicians might put pressure on politicians to

distribute deep wells in a fairer and more efficient way. We think that combining

a national well-testing program with policy interventions that address these market

failures currently preventing deep-well construction is required to properly address

this massive health crisis in Bangladesh.

References

Ahmad, Sk Akhtar, Manzurul Haque Khan, and Mushfiqul Haque. 2018. “Arsenic Contamination

in Groundwater in Bangladesh: Implications and Challenges for Healthcare Policy.” Risk

Management and Healthcare Policy 11: 251–61. doi: 10.2147/RMHP.S153188.

Ahmed, Mushfiq Mobarak, Feroze Ahmed, Satinder Ahuja, Mohammad Alauddin, Stephan J.

Hug, Jon R. Lloyd, Alexander Pfaff, Thomas Pichler, Chad W. Saltikov, Martin Stute, and

Alexander van Geen. 2006. “Ensuring Safe Drinking Water in Bangladesh.” Science 314

(5806): 1687–88. https://doi.org/10.1126/science.1133146.

Ahuja, Amrita, Michael Kremer, and Alix Zwane. 2010. “Providing Clean Water: Evidence from

Randomized Evaluations.” Annual Review of Resource Economics 2: 237–56.

Ashraf, Nava, James Berry, and Jesse Shapiro. 2010. “Can Higher Prices Stimulate Product Use?

Evidence from a Field Experiment in Zambia.” American Economic Review 100 (5): 2383–

413. https://doi.org/10.1257/aer.100.5.2383.

Bagwell, Kyle, and Michael Riordan. 1991. “High and Declining Prices Signal Product Quality.”

The American Economic Review 81 (1): 224–39.

Bakhtiar, Mehrab, Raymond Guiteras, and Ahmed Mushfiq Mobarak. 2019. “Encouraging

Sanitation Adoption Leveraging Social Networks.” Unpublished.

Balasubramanya, Soumya, Alexander Pfaff, Lori Bennear, Alessandro Tarozzi, Kazi Matin

Ahmed, Amy Schoenfeld, and Alexander van Geen. 2014. “Evolution of Households’

Responses to the Groundwater Arsenic Crisis in Bangladesh: Information on

Environmental Health Risks Can Have Increasing Behavioral Impact Over Time.”

Environment and Development Economics 19 (5): 631–47.

Bangladesh Bureau of Statistics and UNICEF. 2011. Bangladesh National Drinking Water

Quality Survey of 2009. Dhaka.

Bardhan, Pranab, and Dilip Mookherjee. 2000. “Capture and Governance at Local and National

Levels.” American Economic Review 90 (2): 135–39.

Barnwal, Prabhat, Alexander van Geen, Janvon der Goltz, and Chander Kumar Singh. 2017.

“Demand for Environmental Quality Information and Household Response: Evidence from

Well-Water Arsenic Testing.” Journal of Environmental Economics and Management 86

(8136): 160–92. https://doi.org/10.1016/j.jeem.2017.08.002.

http://dx.doi.org/10.2147/RMHP.S153188

E S R A P S A 39

Bennear, Lori, Alessandro Tarozzi, Alexander Pfaff, Soumya Balasubramanya, Kazi Matin

Ahmed, and Alexander van Geen. 2013. “Impact of a Randomized Controlled Trial

in Arsenic Risk Communication on Household Water-Source Choices in Bangladesh.”

Journal of Environmental Economics and Management 65: 225–40.

BenYishay, Ariel, Andrew Fraker, Raymond Guiteras, Giordano Palloni, Neil Buddy Shah, Stuart

Shirrell, and Paul Wang. 2017. “Microcredit and Willingness to Pay for Environmental

Quality: Evidence from a Randomized-Controlled Trial of Finance for Sanitation in

Rural Cambodia.” Journal of Environmental Economics and Management. 86: 121–40.

https://doi.org/10.1016/j.jeem.2016.11.004.

BenYishay, Ariel, and Ahmed Mushfiq Mobarak. 2019. “Social Learning and Incentives for

Experimentation and Communication.” Review of Economic Studies 86 (3): 976–1009.

https://doi.org/10.1093/restud/rdy039.

Berry, James, Greg Fischer, and Raymond Guiteras. 2019. “Eliciting and Utilizing Willingness-

to-Pay: Evidence from Field Trials in Northern Ghana.” Journal of Political Economy 128

(4): 1436–73. https://doi.org/10.1086/705374.

Björkman Nyqvist, Andrewa Guariso, Jakob Svensson, and David Yanagizawa-Drott. 2017.

“Reducing Child Mortality in the Last Mile: A Randomized Social Entrepreneurship

Intervention in Uganda.” Paper presented at UNU-WIDER Development Conference,

Maputo. https://yanagizawadrott.com/wp-content/uploads/2017/08/livinggoods_april2017

.pdf.

Brown, Julia, Ahmed Mushfiq Mobarak, and Tetyana Zelenska. 2014. “Barriers to Adoption

of Products and Technologies that Aid Risk Management in Developing Countries.”

Background Paper for World Development Report 2014. Washington, DC: World Bank.

http://siteresources.worldbank.org/EXTNWDR2013/Resources/8258024-1352909193861

/8936935-1356011448215/8986901-1380046989056/WDR-2014_Complete_Report.pdf.

Chakraborty, A.K., and K.C. Saha. 1987. “Arsenical Dermatosis from Tubewell Water in West

Bengal.” Indian Journal of Medical Research 85: 326–34.

Chen, Yu, Faruque Parvez, Vesna Slavkovich, Maria Argos, Muhammad Rakibuz-Zaman, Golam

Sarwar, Alexander van Geen, and Habibul Ahsan. 2011. “Arsenic Exposure from Drinking

Water and Mortality from Cardiovascular Disease in Bangladesh: Prospective Cohort

Study.” British Medical Journal 342 (7806): 1–11. https://doi.org/10.1136/bmj.d2431.

Chen, Yu, Alexander van Geen, Joseph H. Graziano, Alexander Pfaff, Malgosia Madajewicz,

Faruque Parvez, A.Z.M. Iftekhar Hussain, Vesna Slavkovich, Tariqul Islam, and Habibul

Ahsan. 2007. “Reduction in Urinary Arsenic Levels in Response to Arsenic Mitigation

Efforts in Araihazar, Bangladesh.” Environmental Health Perspectives 115 (6): 917–23.

https://doi.org/10.1289/ehp.9833.

Cocciolo, Serena, Ahasan Habib, and Anna Tompsett. 2019. “Group Size and Collective Action.”

Working Paper. https://doi.org/10.1177/0010414001034001003.

Cohen, Jessica, and Pascaline Dupas. 2010. “Free Distribution or Cost-Sharing? Evidence from

a Randomized Malaria Prevention Experiment.” The Quarterly Journal of Economics 125

(1): 1–45.

Davis, C. Austin, and Ahmed Mushfiq Mobarak. 2020. “The Challenges of Scaling Effective

Interventions: A Path Forward for Research and Policy.” World Development 127 (C).

https://doi.10.1016/j.worlddev.2019.104817.

Deaton, Angus. 2013. The Great Escape: Health, Wealth, and the Origins of Inequality. Princeton:

Princeton University Press.

40 A D R

Department of Public Health Engineering and Japan International Cooperation Agency. 2009.

Situation Analysis of Arsenic Mitigation. Dhaka.

Devoto, Florencia, Esther Duflo, Pascaline Dupas, William Parienté, and Vincent Pons. 2012.

“Happiness on Tap: Piped Water Adoption in Urban Morocco.” American Economic

Journal Economic Policy 4 (4): 68–99.

Dhar, R. K., Bhajan Kr. Biswas, Gautam Samanta, Badal Kr. Mandal, D. Chakraborti, Shibtosh

Roy, Abu Jafar, Ariful Islam, Gulshan Ara, Saiful Kabir, A. Wadud Khan, S. Akther

Ahmed, and S. Abdul Hadi. 1997. “Groundwater Arsenic Calamity in Bangladesh.” Current

Science 73 (1): 48–59.

Duflo, Esther, Sebastian Galiani, and Ahmed Mushfiq Mobarak. 2012. “Improving Access to

Urban Services for the Poor.” Abdul Latif Jameel Poverty Action Lab Report (October).

Fendorf, Scott, Holly Michael, and Alexander van Geen. 2010. “Spatial and Temporal Variations

of Groundwater Arsenic in South and Southeast Asia.” Science 328 (5982): 1123–27.

https://doi.org/10.1126/science.1172974.

Field, Erica, Rachel Glennerster, and Reshmaan Hussam. 2011. “Throwing the Baby

Out with the Drinking Water: Unintended Consequences of Arsenic Mitigation

Efforts in Bangladesh.” Harvard Manuscript. http://scholar.harvard.edu/field/files/arsenic

_infantmortality_feb10.pdf.

Flanagan, Sara V., Richard B. Johnston, and Yan Zheng. 2012. “Arsenic in Tube Well Water

in Bangladesh: Health and Economic Impacts and Implications for Arsenic Mitigation.”

Bulletin of the World Health Organization 90 (11): 839–46.

Gelman, Andrew, Hao Ly, Matilde Trevisani, and Alexander van Geen. 2004. “Direct Data

Manipulation for Local Decision Analysis as Applied to the Problem of Arsenic in Drinking

Water from Tube Wells in Bangladesh.” Risk Analysis 24 (6): 1597–612.

George, Christine, Jennifer Inauen, Sheikh Rahman, and Yan Zheng. 2013. “The Effectiveness

of Educational Interventions to Enhance the Adoption of Fee-Based Arsenic Testing

in Bangladesh: A Cluster Randomized Controlled Trial.” American Journal of Tropical

Medicine and Hygiene 89 (1): 138–44. https://doi.org/10.4269/ajtmh.12-0664.

George, Christine, Alexander van Geen, Vesna Slavkovich, Ashit Singha, Diane Levy, Tariqul

Islam, Kazi Matin Ahmed, Joyce Moon-Howard, Alessandro Tarozzi, Xinhua Liu, and

Pam Factor-Litvak. 2012. “A Cluster-Based Randomized Controlled Trial Promoting

Community Participation in Arsenic Mitigation Efforts in Bangladesh.” Environmental

Health: A Global Access Science Source 11 (1): 1–10. https://doi.org/10.1186/1476

-069X-11-41.

German, Michael, Todd Watkins, Minhaj Chowdhury, Prasun Chatterjee, Mizan Rahman, Hul

Seingheng, and Arup SenGupta. 2019. “Evidence of Economically Sustainable Village-

Scale Microenterprises for Arsenic Remediation in Developing Countries.” Environmental

Science and Technology 53: 1078–86.

Guiteras, Raymond, James Levinsohn, and Ahmed Mushfiq Mobarak. 2019. “Demand Estimation

with Strategic Complementarities: Sanitation in Bangladesh.” Working Paper.

Guiteras, Raymond, and Ahmed Mushfiq Mobarak. 2016. “Does Development Aid Undermine

Political Accountability? Leader and Constituent Responses to a Large-Scale Intervention.”

Working Paper.

Hossain, M. Amir, Mrinal Kumar Sengupta, Sad Ahamed, Mohammd Mahmudur Rahman,

Debapriya Mondal, Dilip Lodh, Bhaskar Das, Bishwajit Nayak, Bimal K. Roy, Amitava

Mukherjee, and Dipankar Chakraborti. 2005. Ineffectiveness and Poor Reliability of

Arsenic Removal Plants in West Bengal, India. Environmental Science and Technology 39

(11): 4300–06. https://doi.org/10.1021/es048703u.

E S R A P S A 41

Howard, Guy, Feroze Ahmed, Abu Jafar, Shamsul Gafur Mahmud, and Daniel Deere. 2006. “Risk

Assessment of Arsenic Mitigation Options in Bangladesh.” Journal of Health, Population,

and Nutrition 24 (3): 346–55.

Human Rights Watch. 2016. Nepotism and Neglect: The Failing Response to Arsenic in the

Drinking Water of Bangladesh’s Rural Poor. New York.

Hussam, Abul, and Abul K.M. Munir. 2007. “A Simple and Effective Arsenic Filter Based

on Composite Iron Matrix: Development and Deployment Studies for Groundwater of

Bangladesh.” Journal of Environmental Science and Health–Part A Toxic/Hazardous

Substances and Environmental Engineering 42 (12): 1869–78. https://doi.org/10.1080

/10934520701567122.

Inauen, Jennifer, Mohammad M. Hossain, Richard B. Johnston, and Hans-Joachim Mosler. 2013.

“Acceptance and Use of Eight Arsenic-Safe Drinking Water Options in Bangladesh.” PLoS

ONE 8 (1). https://doi.org/10.1371/journal.pone.0053640.

______. 2014. “The Role of Commitment Strength in Enhancing Safe Water Consumption:

Mediation Analysis of a Cluster-Randomized Trial.” British Journal of Health Psychology

19 (4): 701–19. https://doi.org/10.1111/bjhp.12068.

Jalan, Jyotsna, and E. Somanathan. 2008. “The Importance of Being Informed: Experimental

Evidence on Demand for Environmental Quality.” Journal of Development Economics 87

(1): 14–28. https://doi.org/10.1016/j.jdeveco.2007.10.002.

Jamil, Nadia, Huan Feng, Kazi Ahmed, Imtiaz Choudhury, Prabhat Barnwal, and Alexander van

Geen. 2019. “Effectiveness of Different Approaches to Arsenic Mitigation over 18 Years

in Araihazar, Bangladesh: Implications for National Policy.” Environmental Science and

Technology 53 (10): 5596–604. https://doi.org/10.1021/acs.est.9b01375.

Johnston, Richard, Suzanne Hanchett, and Mohidul Hoque Khan. 2010. “The Socio-Economics

of Arsenic Removal.” Nature Geoscience 3 (1): 2–3. https://doi.org/10.1038/ngeo735.

Johnston, Richard, Stephan Hug, Jennifer Inauen, Nasreen Khan, Hans-Joachim Mosler,

and Hong Yang. 2014. “Enhancing Arsenic Mitigation in Bangladesh: Findings

from Institutional, Psychological, and Technical Investigations.” Science of the Total

Environment 488–489 (1): 477–483. https://doi.org/10.1016/j.scitotenv.2013.11.143.

J-PAL. 2012. “J-PAL Urban Services Review Paper.” Abdul Latif Jameel Poverty Action Lab

Review Paper.

Keskin, Pinar, Gauri Kartini Shastry, and Helen Willis. 2017. “Water Quality Awareness and

Breastfeeding: Evidence of Health Behavior Change in Bangladesh.” Review of Economics

and Statistics 99 (2): 265–80.

Khan, Khalid, Ershad Ahmed, Pam Factor-Litvak, Xinhua Liu, Abu Siddique, Gail Wasserman,

Vesna Slavkovich, Diane Levy, Jacob Mey, Alexander van Geen, and Joseph Graziano.

2015. “Evaluation of An Elementary School-Based Educational Intervention for Reducing

Arsenic Exposure in Bangladesh.” Environmental Health Perspectives 123 (12): 1331–36.

https://doi.org/10.1289/ehp.1409462.

Khan, Nasreen, Roy Brouwer, and Hong Yang. 2014. “Household’s Willingness to Pay for Arsenic

Safe Drinking Water in Bangladesh.” Journal of Environmental Management 143: 151–

61.

Kremer, Michael, Jessica Leino, Edward Miguel, and Alix Peterson Zwane. 2011. “Spring

Cleaning: Rural Water Impacts, Valuation, and Property Rights Institutions.” Quarterly

Journal of Economics 126 (1): 145–205. https://doi.org/10.1093/qje/qjq010.

Kremer, Michael, and Edward Miguel. 2007. “The Illusion of Sustainability.” The Quarterly

Journal of Economics 122 (3): 1007–65.

42 A D R

Lokuge, Kamalini, Wayne Smith, Bruce Caldwell, Keith Dear, and Abul Milton. 2004.

“The Effect of Arsenic Mitigation Interventions on Disease Burden in Bangladesh.”

Environmental Health Perspectives 112 (11): 1172–77. https://doi.org/10.1289/ehp.6866.

Madajewicz, Malgosia, Alexander Pfaff, Alexander van Geen, Joseph Graziano, Iftikhar Hussein,

Hasina Momotaj, Roksana Sylvi, and Habibul Ahsan. 2017. “Community Participation

in Decision-Making: Evidence from An Experiment in Safe Drinking Water Provision

in Bangladesh.” Working Paper. http://www.sv.uio.no/esop/english/research/news-and

-events/events/guest-lectures-seminars/esop-seminar/dokumenter/tompsett-community-pa

rticipation-in-decision-making-preview.pdf.

______. 2007. “Can Information Alone Change Behavior? Response to Arsenic Contamination

of Groundwater in Bangladesh.” Journal of Development Economics 84 (2): 731–54.

https://doi.org/10.1016/j.jdeveco.2006.12.002.

Meredith, Jennifer, Jonathan Robinson, Sarah Walker, and Bruce Wydick. 2013.

“Keeping the Doctor Away: Experimental Evidence on Investment in

Preventative Health Products.” Journal of Development Economics 105: 196–210.

https://doi.org/10.1016/j.jdeveco.2013.08.003.

Miguel, Edward, and Mary Kay Gugerty. 2005. “Ethnic Diversity, Social Sanctions,

and Public Goods in Kenya.” Journal of Public Economics 89 (11–12): 2325–68.

https://doi.org/10.1016/j.jpubeco.2004.09.004.

Miller, Grant, and Ahmed Mushfiq Mobarak. 2015. “Learning about New Technologies

through Social Networks: Experimental Evidence on Nontraditional Stoves in Bangladesh.”

Marketing Science 34 (4): 480–99. https://doi.org/10.1287/mksc.2014.0845.

Mobarak, Ahmed Mushfiq, Alexander van Geen, and Tomer Mangoubi. 2019. “The Politics

of Poison: Elite Capture of Clean Water in Bangladesh.” IGC Working Paper S-31419-

BGD-1.

Mosler, Hans-Joachim, Olivia Blochliger, and Jennifer Inauen. 2010. “Personal, Social, and

Situational Factors Influencing the Consumption of Drinking Water from Arsenic-Safe

Deep Tubewells in Bangladesh.” Journal of Environmental Management 91 (6): 1316–23.

https://doi.org/10.1016/j.jenvman.2010.02.012.

O’Keeffe-O’Donovan, Rossa. 2019. “Water, Spillovers, and Free Riding: Provision of

Local Public Goods in a Spatial Network.” SSRN Electronic Journal, 1–46.

https://doi.org/10.2139/ssrn.3172733.

Opar, Alisa, Alex Pfaff, Ashraf Ali Seddique, Kazi Matin Ahmed, Joseph H. Graziano, and

Alexander van Geen. 2007. “Responses of 6,500 Households to Arsenic Mitigation in

Araihazar, Bangladesh.” Health & Place 13: 164–72.

Parvez, Faruque, Gail A. Wasserman, Pam Factor-Litvak, Xinhua Liu, Vesna Slavkovich, Abu

B. Siddique, Rebeka Sultana, Ruksana Sultana, Tariqul Islam, Diane Levy, Jacob L.

Mey, Alexander van Geen, Khalid Khan, Jennie Kline, Habibul Ahsan, and Joseph H.

Graziano. 2011. “Arsenic Exposure and Motor Function among Children in Bangladesh.”

Environmental Health Perspectives 119 (11): 1665–70.

Pattanayak, Subhrendu, and Alexander Pfaff. 2009. “Behavior, Environment, and Health in

Developing Countries: Evaluation and Valuation.” Annual Review of Resource Economics

1 (1): 183–217. https://doi.org/10.1146/annurev.resource.050708.144053.

Pfaff, Alex, Amy Schoenfeld Walker, Kazi Matin Ahmed, and Alexander van Geen. 2017.

“Reduction in Exposure to Arsenic from Drinking Well-Water in Bangladesh Limited

by Insufficient Testing and Awareness.” Journal of Water Sanitation and Hygiene for

Development 7 (2): 331–39.

E S R A P S A 43

Pitt, Mark, Mark Rosenzweig, and Nazmul Hassan. 2020. “Identifying the Hidden Costs

of a Public Health Success: Arsenic Well Water Contamination and Productivity in

Bangladesh.” Forthcoming in Review of Economic Studies.

Pruss-Ustun, Anette, and Carlos Corvalan. 2006. “Preventing Disease through Healthy

Environments: Towards an Estimate of the Environmental Burden of Disease.”

Bulletin of the World Health Organization 12 (2): 1–106. https://doi.org/10.1590/S1413

-41522007000200001.

Quansah, Reginald, Frederick Ato Armah, David Kofi Essumang, Isaac Luginaah, Edith Clarke,

Kissinger Marfoh, Samuel Jerry Cobbina, Edward Nketiah-Amponsah, Proscovia Bazanya

Namujju, Samuel Obiri, and Mawuli Dzodzomenyo. 2015. “Association of Arsenic with

Adverse Pregnancy Outcomes/Infant Mortality: A Systematic Review and Meta-Analysis.”

Environmental Health Perspectives 123 (5): 412–32.

Ravenscroft, Peter, Hugh Brammer, and Keith Richards. 2009. Arsenic Pollution: A Global

Synthesis. Oxford: Wiley-Blackwell.

Ravenscroft, Peter, Ahammadul Kabir, Syed Adnan Ibn Hakim, A.K.M. Ibrahim, Sudhir Kumar

Ghosh, Md. Saifur Rahman, Firoza Akhter, and Md. Abdus Sattar. 2014. “Review Paper

Effectiveness of Public Rural Waterpoints in Bangladesh with Special Reference to Arsenic

Mitigation.” Journal of Water Sanitation and Hygiene for Development 4 (4): 545–62.

https://doi.org/10.2166/washdev.2014.038.

Sanchez, Tiffany, Diane Levy, Mohammed Hasan Shahriar, Mohammad Nasir Uddin, Abu

B. Siddique, Joseph H. Graziano, Agela Lomax-Luu, Alexander van Geen, and Mary

V. Gamble. 2016. “Provision of Well-Water Treatment Units to 600 Households in

Bangladesh: A Longitudinal Analysis of Urinary Arsenic Indicates Fading Utility.” Science

of the Total Environment 563–564: 131–37.

Sarkar, Sudipta, John Greenleaf, Anirban Gupta, Debabrata Ghosh, Lee Blaney, Prasanta Kumar

Bandyopadhyay, Ranjan K. Biswas, Amal Dutta, and Arup SenGupta. 2010. “Evolution of

Community-Based Arsenic Removal Systems in Remote Villages in West Bengal, India:

Assessment of Decade-Long Operation.” Water Research 44: 5813–22.

Smedley, Pauline L., and David G. Kinniburgh. 2002. “A Review of the Source, Behavior, and

Distribution of Arsenic in Natural Waters.” Applied Geochemistry 17: 517–68.

Smith, Allan, Elena Lingas, and Mahfuzar Rahman. 2000. “Contamination of Drinking Water

by Arsenic in Bangladesh: A Public Health Emergency.” Bulletin of the World Health

Organization 78 (9): 1093–103. https://doi.org/10.1590/S0042-96862000000900005.

Somanathan, E. 2010. “Effects of Information on Environmental Quality in Developing

Countries.” Review of Environmental Economics and Policy 4 (2): 275–92.

https://doi.org/10.1093/reep/req012.

Sutherland, D., P.M. Swash, A. C. Macqueen, L.E. McWilliam, D.J. Roass, S.C. Wood. 2002. “A

Field Based Evaluation of Household Arsenic Removal Technologies for the Treatment of

Drinking Water.” Environmental Technology 23 (12): 1385–403.

Tarozzi, Alessandro, Ricardo Maertens, Kazi Matin Ahmed, and Alexander van Geen. 2020.

“Demand for Information on Environmental Health Risk, Mode of Delivery, and Behavioral

Change: Evidence from Sonargaon, Bangladesh.” Forthcoming in World Bank Economic

Review.

Tarozzi, Alessandro, Aprajit Mahajan, Brian Blackburn, Dan Kopf, Lakshmi Krishnan, and

Joanne Yoong. 2014. “Micro-Loans, Insecticide-Treated Bednets, and Malaria: Evidence

from a Randomized Controlled Trial in Orissa, India.” American Economic Review 104 (7):

1909–41. https://doi.org/10.1257/aer.104.7.1909.

44 A D R

Thaler, Richard. 1980. “Toward A Positive Theory of Consumer Choice.” Journal of Economic

Behavior and Organization 1 (1): 39–60.

Vahter, Marie, Nazmul Sohel, Kim Streatfield, and Lars Ake Persson. 2010. “Arsenic Exposure

from Drinking Water and Mortality in Bangladesh.” The Lancet 376 (9753): 1641–42.

https://doi.org/10.1016/S0140-6736(10)62089-2.

van Geen, Alexander, Ershad Ahmed, Lynnette Pitcher, Jacob Mey, Habibul Ahsan, Joseph

Graziano, and Kazi Ahmed. 2014. “Comparison of Two Blanket Surveys of Arsenic in

Tubewells Conducted 12 Years Apart in a 25 km2 Area of Bangladesh.” Science of the Total

Environment 488–489: 484–92. https://doi.org/10.1016/j.scitotenv.2013.12.049.

van Geen, Alexander, Kazi Matin Ahmed, E.B. Ahmed, and Imtiaz Choudhury. 2016.

“Inequitable Allocation of Deep Community Wells for Reducing Arsenic Exposure in

Bangladesh.” Journal of Water Sanitation and Hygiene for Development 6 (1): 142–50.

https://doi.org/10.2166/washdev.2015.115.

van Geen, Alexander, Habibul Ahsan, Allan Horneman, Ratan Dhar, Yan Zheng, Iftikhhar

Hussain, Kazi Matin Ahmed, Andrew Gelman, Martin Stute, H. James Simpson, Sean

Wallace, Christopher Small, Faruque Parvez, Vesna Slavkovich, Nancy Lolacono, Marck

Becker, Zhongqi Cheng, Hassina Momotaj, Mohammad Shahneqaz, Ashraf Ali Seddique,

and Joseph Graziano. 2002. “Promotion of Well-Switching to Mitigate the Current Arsenic

Crisis in Bangladesh.” Bulletin of the World Health Organization 80 (9): 732–37.

van Geen, Alexander, Yun-jiang Zheng, Roelof Versteeg, Matthias Stute, A. Horneman, Ratan

Dhar, Michael Steckler, Andrew Gelman, Christopher Small, Habibul Ahsan, Joseph H.

Graziano, Ishtiaque Hussein, and Kazi Matin Ahmed. 2003. “Spatial Variability of Arsenic

in 6,000 Tube Wells in a 25 km2 Area of Bangladesh.” Water Resources Research 39 (5):

1281–92.

Wasserman, Gail A., Xinhua Liu, Faruque Parvez, Habibul Ahsan, Pam Factor-Litvak, Alexander

van Geen, Zhongqi Cheng, Vesna Slavkovich, Iftikhar Hussain, Hassina Momotaj, and

Joseph H. Graziano. 2004. “Water Arsenic Exposure and Children’s Intellectual Function

in Araihazar, Bangladesh.” Environmental Health Perspectives 112: 1329–33.

World Bank. 2003. World Development Report 2004: Making Services Work for Poor People.

Washington, DC.

Does the All-China Federation of Industry andCommerce Align Private Firms with the Goals

of the People’s Republic of China’s Beltand Road Initiative?

Jeffrey B. Nugent and Jiaxuan Lu∗

This paper demonstrates that the largest business association of private firmsin the People’s Republic of China (PRC), the All-China Federation of Industryand Commerce (ACFIC), has induced its members to help achieve the goalsof the PRC’s extremely ambitious but risky Belt and Road Initiative (BRI)since its inauguration in 2013. Through its newspaper, the ACFIC has drawnthe attention of its member firms to countries participating in the BRI, whichhas led to increased trade between provinces in the PRC and BRI-participatingcountries emphasized by the ACFIC’s newspaper. The results show that thePRC’s exports have been encouraged substantially more than its imports, whichcould be a cause for concern for the sustainability of the BRI. The results wereobtained through various specially designed versions of the gravity model andhave shown to be robust to the use of various methods for mitigating possibleestimation biases.

Keywords: Belt and Road Initiative, business association, People’s Republic ofChina, private firm, tradeJEL codes: D23, F14, F21, L22, O53

I. Introduction

When the initiatives that countries take to achieve certain objectives aremassive, multinational, and laden with serious challenges, coordination among therelevant parties, both public and private, can be difficult since each participatingfirm or agency has its own objectives. The Belt and Road Initiative (BRI) ofthe People’s Republic of China (PRC), inaugurated by President Xi Jinping inKazakhstan in 2013, includes more than 50 partner countries across the continents

∗Jeffrey B. Nugent (corresponding author): Department of Economics, University of Southern California, UnitedStates. E-mail: [email protected]; Jiaxuan Lu: Department of Economics, University of Southern California, UnitedStates. E-mail: [email protected]. We would like to express our thanks for all the comments received at the AsianDevelopment Bank–International Economic Association Roundtable held in July 2019 in Tokyo and at the AppliedMicroeconomics Conference held in December 2019 at the University of Hawaii, and also to Junjie Xia, TerrieWalmsley, the managing editor and the anonymous referee for helpful comments and suggestions. The AsianDevelopment Bank recognizes “China” as the People’s Republic of China, “Russia” as the Russian Federation, and“Vietnam” as Viet Nam. The usual ADB disclaimer applies.





of Asia, Europe, and Africa. It is one of the largest projects ever attempted andwill remain of great importance to the world for decades to come. Yet, withuncertainties concerning its real objectives—and with participating countries ofdifferent sizes, development levels, and political orientations—achieving sufficientintra-BRI coordination is especially challenging.

Lei and Nugent (2018) made the case that the PRC’s government-controlledbusiness association, the All-China Federation of Industry and Commerce (ACFIC),had served well as a coordinating device between the Government of the PRC andthe country’s private firms from 2007 to 2011. This was when Beijing sharplychanged its economic objectives from “Going Outward” to “Going Inward” toescape the adverse effects of the 2008–2009 global financial crisis on importantexporting countries. Yet, that experience had nothing to do with the BRI and thePRC’s leadership of it.

Given the BRI’s ambitious goals, the many countries involved, anduncertainties about the extent to which coordination among all governments, firms,and agencies can be successful, ongoing analysis of the BRI’s progress and theproblems confronted will be needed and will require a wide variety of researchperspectives. However, given the serious concerns about its political and financialviability for some BRI countries identified in the following section, we deem itcrucial to examine the initiative’s early experience to identify the magnitude ofthe risks involved and how it might be improved, and possibly even to reconsiderwhether the BRI is still worth pursuing. This paper’s objective is, therefore, toundertake an analysis of the extent to which the ACFIC has been successful inaligning its member firms with the government’s BRI objectives and the need forpossible reforms.

The remainder of the paper is organized as follows. Section II providesbackground on business associations in general and the ACFIC in particular,as well as on the BRI and some of its challenges. Section III outlines thesteps to be followed in our overall evaluation of the ACFIC as a coordinationdevice in achieving the BRI’s goals. Section IV develops the econometric models,including the methods designed to deal with potential estimation biases. SectionV describes the data used and displays the results from the regression analysis.Section VI conducts robustness checks to resolve selection, heteroskedasticity,and “confounding” issues. Section VII evaluates the extent to which the ACFIC’strade-promoting effects may differ between BRI and non-BRI countries. Finally,section VIII concludes.

II. Background on Business Associations, the All-China Federation of Industryand Commerce, and the Belt and Road Initiative

Can business associations be counted upon to help guide private firms toexert healthy influences on the economy to achieve the desired objectives of the state

Does the ACFIC Align Private Firms with the Goals of the PRC’s BRI? 47

and society? One cannot help but be skeptical about this since business associationsare often believed to hurt the local economy by offering monopolistic protection,triggering corruption and criminal groups, reducing bureaucratic efficiency, andencouraging cartels that raise prices for consumers and reduce allocative efficiency(Doner and Schneider 2000). Moreover, they may use their collective power toinfluence local politics to attain benefits for themselves and undermine goodgovernance (Bräutigam, Rakner, and Taylor 2002). On the other hand, businessassociations can benefit the general public by introducing regulations favoringmarket development and protecting businesses from default, criminal activity,insolvency, and unreasonable governmental interference (Bennett 1998, Önis andTürem 2001). They may be especially helpful in assisting new and small firmsto adopt innovations and break into value chains. As experiences in EasternEurope showed during and after the 1980s, new, freestanding, bottom-up, private-sector-oriented business associations came to play an essential role in the region’stransition from central planning to free markets (Sukiassyan and Nugent 2011).Indeed, they can serve as a coordination device between governments and theirprivate sectors by sharing information and encouraging mutual understanding andsustainable economic growth (Johnson, McMillan, and Woodruff 2000).

However, the ACFIC is very different from most business associations inthat it is entirely government controlled, being run by the Communist Party of China(CPC), and yet its member firms are private. All private firms are eligible to becomeACFIC members, and large ones are especially encouraged to join by national andlocal governments, and CPC officials. While the membership fees for province-levelACFIC are not high (around $3,000 per year), the largest cost to members is in termsof the time required to attend the association’s meetings, use the services offered,and connect to national and local government offices, and other firms. By the endof 2016, there were 3 million ACFIC member firms, accounting for about 10%of all Chinese private firms. Taking advantage of the ACFIC, member firms havelobbied for more favorable government policies, especially those concerning privateproperty rights. The association has also assisted its members to be better informedof new government policies to facilitate connections between business owners andgovernment officials. Given the controversy concerning business associations ingeneral and the uniqueness of the situation in the PRC, we endeavor to contribute tothe literature by exploring ACFIC’s effectiveness in achieving coordination betweengovernment agencies and firms in this new and especially challenging BRI context.

Not surprisingly, there is disagreement in the economics literature over howhelpful the PRC’s top-down ACFIC has been to private firms and the extent to whichit has succeeded in inducing private firms to attain the government’s economicobjectives. For instance, Jia (2014) and Ma, Rui, and Wu (2015) employed standardeconometric techniques, including propensity score matching, to suggest that theACFIC’s most useful function is to allow owners and managers of private firmsto win positions in the CPC or government, but not to boost the performance of


their firms. Yet, taking advantage of some surveys that compare ACFIC memberfirms with nonmember firms in multiple respects, Lei and Nugent (2018) employedvarious estimation techniques to show quite robustly that the ACFIC did play asignificant role in helping its member firms change their focus rather radicallyfrom the government’s earlier goal of promoting outward-looking exports to thesubsequent goal of prioritizing inward-looking investments between 2007 and 2011(i.e., before the BRI’s inauguration). The reason for the sudden change was toprevent the PRC’s economy from falling victim to the 2008–2009 global financialcrisis that did serious damage to the firms and economies of other exportingcountries. Their study also identified the mechanism behind the success in achievinga sharp change in objectives by providing information to member firms about boththe new government objectives and possible means of attaining them.

However, due to limited information on the geographical destination of firm-level sales and investments available in the Chinese Private Enterprise Survey,which was the dataset utilized in previous studies, and the absence of any recentsurvey results, it was not possible to use that data to examine the role of the ACFICin recent years. As explained below, as an alternative source of relevant data, we usedata published by the PRC’s province-level statistical agencies to see if the ACFIChas succeeded in encouraging its member firms to trade with countries favoredby the association in a way that would be consistent with the government’s BRIobjectives.

As noted above, the BRI is of enormous importance, not only to the PRCbut also to the rest of the world. Announced by President Xi Jinping while visitingKazakhstan in September 2013, the initiative is designed to develop transportation,logistics, and other infrastructure to link the PRC with BRI-identified countriesacross the world. By sharply reducing the cost of exporting and importing goods andservices across this enormous network of countries, the BRI is expected to stimulateindustrial production and technological improvements, not only in the PRC butalso throughout Eurasia and Africa (Dunford and Liu 2019). To help accomplishthis, the Asian Infrastructure Investment Bank, the China Development Bank, andthe Export–Import Bank of China were formed, and they have all been growingrapidly since their establishment (Yu 2017). For instance, by the end of 2018, 152countries had joined the BRI in some capacity and 96 of them had joined the AsianInfrastructure Investment Bank as members. The accumulated investments of theseinstitutions amounted to at least $1 trillion by the end of 2018, and this total isexpected to grow to more than $2 trillion to finance the BRI’s infrastructural needs(Hillman 2018). If the BRI develops as expected, it will perhaps become the largestinternational investment project ever created and serve as an integrating force forEurasia and much of the world.

However, the BRI faces enormous challenges. One is that many developedcountries, especially the United States (US) and the United Kingdom, seemto be moving in directions less friendly to global trade. These trends toward


de-integration could possibly spread to Asia and Africa, negatively affecting theirinitially positive attitudes toward the BRI. Additionally, as large numbers of Chineseworkers have been moving into BRI countries to facilitate local construction andother infrastructural activities, resentment has arisen among the nationals of hostcountries like Pakistan (Solangi 2018) and Viet Nam (Elmer 2018). Newspaperarticles (e.g., Lee 2019) have also called attention to the concerns of the US, India,and Japan over the PRC’s recent establishment of military bases in Djibouti, whereeach of these countries had already established its own base. Other studies haveclaimed that without addressing the different needs of various BRI countries forimporting or exporting labor over time, or facilitating internal labor mobility, theBRI could contribute to rising geographic and income inequalities (Gill, Lall, andLebrand 2019; Bruni 2019). In light of these challenges, the future of the PRC’sinvolvements in these countries is increasingly uncertain. Concerning the allocativeefficiency of different regions within the PRC, Gibson and Li (2018) employeddata envelopment analysis and other statistical tools to demonstrate that distributingtoo much effort and resources to low productivity areas in the western PRC alongtransport routes to other BRI countries could jeopardize the overall efficiency of thePRC’s economy and the sustainability of its remarkable growth.

Given both its great economic potential and substantial political andeconomic risks, multiple studies focusing on the BRI’s trade and investmentfacilitation mechanisms have been conducted. Herrero and Xu (2017); Kohl (2019);and Baniya, Rocha, and Ruta (2019) have employed gravity models to arguethat the initiative has sharply increased trade volumes between most participatingeconomies since 2013. Bird, Lebrand, and Venables (2019) have constructed spatialequilibrium models for BRI regions suggesting that the initiative could substantiallyimprove the real incomes of the participating developing economies. Wiederer(2018) and de Soyres et al. (2018) have collected firsthand data from countriesinvolved in the initiative and find that logistical costs have, as intended, been fallingrapidly since 2013. Yet, since most logistical and infrastructural activities have beenthose of the public sector, these studies have done little to determine whether privatefirms have been participating sufficiently for the BRI to be successful.1

It is the combination of the importance of the private sector’s involvementto make BRI successful and uncertainty about whether the PRC’s growing privatesector will become sufficiently engaged in the prioritized activities that motivatesour primary research question: “Has the ACFIC yet come to play a significant rolein assuring sufficient participation of the PRC’s private firms in alignment with thecountry’s BRI objectives?”

1While some studies have shown that the private sector is involved in trade and investment with BRI countries(Cheng 2018, ACFIC 2018, Zhai 2018), other analysts, such as Hillman (2018), have doubted this.


III. Steps for Evaluating the All-China Federation of Industry and Commerce’sRole in the Belt and Road Initiative

To answer our central question, we break our analysis into two parts. First,we seek to determine if the ACFIC has been successful in increasing exports andimports primarily with the countries it seems to prioritize. Then we determine if,since the BRI’s inauguration in 2013, the ACFIC has increased its priority towardBRI countries in general. In view of the “one-way road” argument raised by USVice President Mike Pence at the 2018 Asia–Pacific Economic Cooperation summit(Tarabay and Choe 2018) and echoed in several other countries, we also investigatewhether the BRI has mainly benefited the PRC’s exports to, rather than its importsfrom, BRI countries.2 This would seem especially important as Hurley, Morris, andPortelance (2018) concluded that the negative outcomes for individual BRI partnersare so large as to raise their debt levels enough to trigger defaults by up to eightparticipating BRI countries.

The focus of the paper is, therefore, on testing the validity of the followingthree hypotheses:

(i) The ACFIC’s promotion of trade activities with any non-PRC country ispositively related with the extent to which the ACFIC calls attention to thatcountry in its newspaper, the China Business Times, which acts as a proxyfor the ACFIC’s policy direction.

(ii) Since the inauguration of the BRI in 2013, the ACFIC has emphasizedBRI countries to a larger extent than non-BRI countries in its mostly trade-encouraging news reports, thus implying that the ACFIC encourages itsmember private firms to trade with BRI countries, though not necessarilyequally.

(iii) The ACFIC promotion of trade with BRI countries has resulted in greaterexports from the PRC than imports to the PRC.

IV. Econometric Models

A. Province-Level Gravity Model of International Trade

Since the pioneering efforts of Jan Tinbergen (1962), gravity models haveserved as the most common means of analyzing bilateral trade patterns, which are

2The analyses to date have been mixed on this. Chen and Lin (2018) and Dunford and Liu (2019) denyit, while tending to confirm it are Huang (2016) for BRI countries in general; Irshad, Xin, and Arshad (2015) forPakistan; Yu (2017) for Myanmar; and Kohl (2019) for Europe.


essential to the BRI. According to his specification, bilateral trade volumes fromregion i to region j, Ti j, could be expressed as

Ti j ≡ KGDPi

aGDPjb

Di jc (1)

where GDPi is a proxy for the economic size (gross domestic product [GDP]) ofregion i, GDPj is a proxy for that of region j, Di j is the distance (a proxy for tradingcost) between i and j, and K is a positive constant.

Anderson (1979) provided a theoretical framework for the model byincorporating a Cobb–Douglas utility function. Anderson and van Wincoop (2003)extended that model by using a constant elasticity of substitution utility function,whereby the exports from region i to region j, xi j, could be expressed as

xi j = yiy j

yW

(ti j

PiPj

)1−σ

(2)

where yW is the economic size of the world, measured by GDP, yi and y j are theGDPs of regions i and j, respectively, ti j is the trading cost between regions i and j,Pi and Pj are the relative consumer prices of regions i and j, and σ is the elasticityof substitution in the constant elasticity of substitution utility function.

Since this paper’s primary concern is the economic influence of the ACFICon cross-border trade, we conduct province-level analyses by treating region i as aPRC province and region j as a country or region outside of the PRC. This allowsus to detect variations in the ACFIC’s influence on both exports and imports acrossdifferent province–country pairs. While the ACFIC can do little to directly affectthe economic size of a province or country, it can reduce the information and othertrading costs (or levels of distrust) between PRC provinces and the countries itprioritizes in its official newspaper. Therefore, unless otherwise noted, we assumethat the ACFIC affects the relationship in equation (2) only by lowering the tradingcost ti j.

Following the insight provided by Maurel and Afman (2010) in theirexamination of the effect of establishing foreign missions on trading activities, wespecify trading cost ti j as

ti j ≡ (ACFICiCBTj

)kdi j

ρ (3)

where ACFICi is the number of ACFIC members in province i, CBTj is thefrequency of the name of country j appearing in the China Business Times, anewspaper entirely controlled by the ACFIC, and di j is the distance betweenprovince i and country j. The larger CBTj, the more favorable country j shouldbe in the ACFIC’s eyes. Accordingly, ACFICiCBTj could be perceived as a proxyfor the magnitude of the ACFIC’s influence on the bilateral trade between theprovince–country pair i j, with ACFICi by itself reflecting ACFIC’s power over


private firms in province i. ρ is expected to be positive as geographical distanceincreases transportation costs, and k should be negative as the ACFIC’s influenceon bilateral economic interactions, weighted by CBTj, should be positive.

Plugging equation (3) into equation (2) and transforming it into logarithmicform, we obtain

ln

(xi j

yiy j

)= (1 − σ ) k ln

(ACFICiCBTj

) + (1 − σ ) ρ ln(di j

) − (1 − σ ) ln(yW )

− (1 − σ ) ln(Pi) − (1 − σ ) ln(Pj

)(4)

Transforming equation (4) into a format suitable for empirical estimation andadding both time subscripts and control variables, we obtain

ln(xi j,t+1

) = β1 ln(ACFICitCBTjt

) + β2 ln(GDPit ) + β3 ln(GDPjt

)+ β4 ln

(Distancei j

)+β5Borderi j +β6Religioni j +β7 ln(Populationit )

+ β8 ln(Population jt

) + β9 ln(Areait ) + β10 ln(Area jt

)+ β11SFIjt + β12ln

(TCPj,t−1

) + β0 + πi jt + εi jt (5)

where the dependent variable, xi j,t+1, could alternatively represent exports from i toj or imports from j to i in year t + 1; ACFICit is the number of ACFIC members inprovince i in year t; CBTjt is the frequency of the name of country j appearing in theChina Business Times in year t; GDPit and GDPjt are the GDPs of province i andcountry j in year t, respectively; Distancei j is the geographical distance betweenprovince i and country j; Borderi j and Religioni j are dummy variables indicatingwhether province i and country j share a common border or a common dominantreligion as suggested by Lewer and Van den Berg (2007) (given that some province-level administrative districts in the PRC are Muslim); Populationit and Population jt

are the populations of province i and country j, respectively, in year t; and Areait

and Area jt are the geographic sizes of provinces i and country j, respectively, inyear t. To broaden the analysis from a traditional version of the gravity model, wealso include (i) SFIjt , the state fragility index of country j in year t, as an indicatorof the country’s level of political instability in that year; and (ii) TCPj,t−1, the totalturnover of Chinese-contracted projects in country j in year t − 1. The remainingterms include β0, the intercept; πi jt , the interacted fixed effect for the region in thePRC in which province i is located (Eastern PRC, Central PRC, or Western PRC),for the continent where country j is located, and for year t, and finally, εi jt , theresidual. From equation (4), β1 ≡ (1 − σ )k and β4 ≡ (1 − σ )ρ.

If the model yields the expected results, the treatment effects quantified byβ1, β2, and β3 should be positive, but β4 should be negative to be consistent withthe gravity model. β5 and β6 should be positive because commonality in border andreligion should reduce trading costs. β7 and β8 can be either positive or negative


since the relevance of population sizes and their effects after controlling for GDPare ambiguous. β9 and β10 are expected to be negative because larger land areasincrease the distance between the centers in the two regions. β11 should be negativebecause the insecurity of a country may serve as a hidden cost (Blomberg andHess 2006), and β12 should be positive since the reduction in trading costs throughlogistics, transportation improvements, and information dissemination has been amajor thrust of the BRI (Rehman and Ding 2019). Additionally, we allow for fixedeffects for year and for both province and country to capture unobserved effects.Finally, the dependent variable, xi j,t+1, is designated to be 1 year after the year inwhich the independent variables are measured so as to mitigate simultaneity and/orreverse causality problems.

B. Two-Stage Least Squares Strategy Based on the Province-LevelGravity Model

While equation (5) provides a suitable econometric model, it may be subjectto endogeneity bias. For instance, it could be possible that CPC membership affectsboth ACFICit and trade but with no direct connection between them. To alleviate thistype of imprecision, we devise a two-stage least squares (2SLS) estimation with aninstrumental variable. Using equation (5) as the second stage, following the methodemployed by Lei and Nugent (2018) for the ACFIC’s coordination effects prior tothe BRI’s inauguration, the first stage for ln(ACFICitCBTjt ) becomes

ln(ACFICitCBTjt

) = α1 ln(PrivateFirmitCBTjt

) + α2 ln(GDPit ) + α3 ln(GDPjt

)+ α4 ln

(Distancei j

) + α5Borderi j + α6Religioni j

+ α7 ln(Populationit ) + α8 ln(Population jt

) + α9 ln(Areait)

+ α10 ln(Area jt

) + α11SFIjt + α12ln(TCPj,t−1

)+ α0 + πi jt + εi jt (6)

where PrivateFirmit is the number of private firms in province i in year t; itcorresponds to ACFICit , the number of private firms that are members of theACFIC in province i. We use ln(PrivateFirmitCBTjt ) as the instrumental variable inequation (6) because the number of private firms in a province, which has no directlink with trade, provides a reasonable proxy for the number of ACFIC memberfirms in that province and, hence, for the ACFIC’s potential influence on the tradeof province i with country j. By including all exogenous variables in the secondstage along with the instrument, any correlation between the error term and otherindependent variables that could bias the estimates can be reduced (Wooldridge2010, 89–90).


C. Exports-versus-Imports Comparison by Incorporatingan Interaction Term

To compare the magnitude of the ACFIC’s effects on exports with thaton imports, following Dunlevy (2006), we incorporate an interaction term in theregression model of equation (5):

ln(Exporti j,t+1

) = γ1 ln(ACFICCBTi jt

) + γ2 ln(ACFICCBTi jt

)Typei jt

+ γ3 ln(GDPit ) + γ4 ln(GDPjt

) + γ5 ln(Distancei j

)+ γ6Borderi j + γ7Religioni j + γ8 ln(Populationit )

+ γ9 ln(Population jt

) + γ10ln (Areait ) + γ11 ln(Area jt

)+ γ12SFIi jt + γ13ln

(TCPi j,t−1

) + γ0 + πi jt + εi jt (7)

where i represents the origin and j the destination. If Typei jt = 1, i is a PRCprovince and j a non-PRC country, and ACFICCBTi jt is the number of ACFICmembers in province i times the frequency of the name of country j appearing inthe China Business Times in year t; if Typei jt = 0, i is a foreign country and j aPRC province, and ACFICCBTi jt is the number of ACFIC members in provincej multiplied by the frequency of the name of country i appearing in the ChinaBusiness Times in year t. Other variables are the same as those in precedingequations. If γ2 > 0 and is statistically significant, this would support the “one-wayroad” argument.

V. Data Sources and Statistical Analysis

A. Data Sources

To carry out econometric analysis for the ACFIC’s effects on trade, we relyon bilateral trade data from province-level statistical yearbooks between 2010 and2017. The values from 2010 to 2017 are used for the dependent variable and thosefrom 2009 to 2016 for the lagged trade variables appearing as explanatory variables.Combining all available trade data for the 8-year interval allows us to construct adataset with more than 20,000 observations.

The greatest data collection challenge is with respect to the measuresof ACFICit and CBTjt . For ACFICit , we utilize the yearbooks published by theACFIC since 2009. In each yearbook, each province-level ACFIC branch has anannual report on its membership, though it does not in every case disclose theprecise number of members in that year. After inspecting these reports, we foundinformation to be missing for 40 out of 248 province–year observations. For CBTjt ,programming techniques were used to identify all country names on the website ofthe China Business Times, the entries were then read, and their numbers recorded


for each country and year from 2009 to 2016. Many of these news notes include thekind of information that should encourage member firms to consider activities inthese countries, and hence, these newspaper-oriented trends could serve as a validproxy for the policy direction of the ACFIC’s influence over its member firms ingeneral. In other words, the ACFIC should affect its member firms in the same wayas its newspaper influences its readers. Appendix 1 presents an English translationof a typical news article in China Business Times. Data sources for the controlvariables are indicated in Table 1, which also reports descriptive statistics for allvariables.

B. Exports and Imports

We report the ordinary least squares (OLS) results of our regression analysisfor xi j,t+1, being the log of exports from province i to country j in year t + 1 incolumns (1) and (2) of Table 2. Similarly, columns (3) and (4) report the OLSresults for imports. The odd-numbered columns contain no fixed effects, while theeven-numbered columns include interacted fixed effects for year, PRC province,and non-PRC country to capture the time-invariant unobserved variables. Missingvalues are omitted because they could represent unrecorded, rather than 0, values.The possible selection bias caused by this truncation will be addressed in section VI.

As shown in the first row, the coefficient of ln(ACFICitCBTjt ) is positiveand statistically significant in all columns, indicating the ability of the ACFICto influence its member firms to increase exports to and imports from the non-PRC countries frequently covered in its mouthpiece, the China Business Times.Additionally, as expected, the parameter estimates of the GDP terms, border andreligion dummies, and the lagged turnover of contracted projects are all positiveand statistically significant, while those for the distance and state fragility index arenegative and statistically significant. The large R-squared values also demonstratethe model’s relatively high explanatory power. Following the method employed byLei and Nugent (2018), but in this quite different context, we test for the stability ofthe coefficient of ln(ACFICitCBTjt ) by calculating the ratio of the R-squared valueobtained with ln(ACFICitCBTjt ) as an independent variable to the R-squared valueobtained in equation (5) without it (i.e., when β1 = 0 for equation [5]). As suggestedby Altonji, Elder, and Taber (2005) and Oster (2017), this ratio, represented by δ,measures how large the impact of unobserved variables must be to invalidate theidentified treatment effect of ln(ACFICitCBTjt ) in Table 2. For this to be so, δ oughtto be at least 1. As shown in the last row of Table 2, all specifications yield greater-than-one values of δ, demonstrating that unobserved variables are unlikely to nullifyour statistical results.

Despite the strong statistical significance of most of the explanatory variablesdisplayed in Table 2, these results could be subject to various endogeneitybiases. Therefore, in Table 3, using equation (6) as the first stage equation and


Table 1. Descriptive Statistics

Dependent No. of StandardVariables Observations Mean Deviation Min Max Source

Log of exports 13,605 10.317 3.061 0.000 19.385 Provincial-levelyearbooks

Log of imports 9,065 10.181 3.504 0.000 17.695 Provincial-levelyearbooks

Independent No. of StandardVariables Observations Mean Deviation Min Max Source

ln(ACFICit ) 203 11.328 0.961 7.688 12.682 ACFIC yearbooksln(CBTjt ) 1,229 2.558 1.800 0.000 7.974 China Business Times

websiteln(GDPit ) 248 19.121 1.013 15.681 20.920 National Bureau of

Statistics of Chinaln(GDPjt ) 1,438 17.320 2.315 10.368 23.552 World Bank World

DevelopmentIndicators

ln(Distancei j ) 1,878 8.832 0.650 4.716 9.899 Google Maps andCEPII

Borderi j 1,861 0.007 0.083 0.000 1.000 Google Maps andCEPII

Religioni j 1,861 0.028 0.166 0.000 1.000 Organization ofIslamic Cooperation

ln(Populationit ) 248 17.319 0.847 14.901 18.516 National Bureau ofStatistics of China

ln(Population jt ) 1,438 15.574 2.116 9.253 21.004 World Bank WorldDevelopmentIndicators

ln(Areait ) 31 12.016 1.225 9.031 14.305 National Bureau ofStatistics of China

ln(Area jt ) 298 11.110 2.682 0.693 16.611 World Bank WorldDevelopmentIndicators

SFIjt 1,325 8.309 6.225 0.000 25.000 Quality of Governmentdatabase

ln(TCPj,t−1) 1,426 7.142 2.509 −1.204 13.445 National Bureau ofStatistics of China

Instrumental No. of StandardVariables Observations Mean Deviation Min Max Source

ln(PrivateFirmit ) 186 11.934 1.244 7.312 14.253 National Bureau ofStatistics of China

ACFIC = All-China Federation of Industry and Commerce, CEPII = Centre d’Etudes Prospectives et d’InformationsInternationales.Notes: The time range of the dependent variables is from 2010 to 2017; those for the independent and instrumentalvariables, except the log of CBTjt , are from 2009 to 2016. The time range of the log of CBTjt is from 2009 to 2017because it becomes a dependent variable in section V. The variables measured by currency values are in thousandcurrent United States dollars. Land area is measured in square kilometers.Source: Authors’ calculations.


Table 2. Ordinary Least Squares Estimates

Exports Imports

Explanatory Variables (1) (2) (3) (4)

ln(ACFICitCBTjt ) 0.263*** 0.265*** 0.280*** 0.257***

(0.029) (0.026) (0.058) (0.054)ln(GDPit ) 0.861*** 1.716*** 1.366*** 1.762***

(0.057) (0.062) (0.110) (0.111)ln(GDPjt ) 0.417*** 0.479*** 0.761*** 0.867***

(0.046) (0.044) (0.098) (0.090)ln(Distancei j ) −0.315*** −0.431*** −0.620*** −0.586***

(0.060) (0.056) (0.111) (0.105)Borderi j 3.572*** 3.451*** 3.155*** 3.686***

(0.606) (0.575) (0.842) (0.810)Religioni j 1.897*** 2.568*** 1.443*** 1.856***

(0.223) (0.223) (0.536) (0.500)ln(Populationit ) 0.591*** −0.448*** −0.084 −0.604***

(0.075) (0.078) (0.154) (0.146)ln(Population jt ) 0.266*** 0.205*** −0.279*** −0.270***

(0.046) (0.046) (0.095) (0.090)ln(Areait ) −0.650*** −0.488*** −0.970*** −0.798***

(0.031) (0.023) (0.060) (0.046)ln(Area jt ) −0.079*** −0.078*** 0.194*** 0.179***

(0.021) (0.020) (0.038) (0.036)SFIjt −0.038*** −0.038*** −0.065*** −0.079***

(0.011) (0.011) (0.024) (0.022)ln(TCPj,t−1) 0.126*** 0.129*** 0.090** 0.092***

(0.020) (0.019) (0.035) (0.032)

Province ProvinceFixed effects No Country Year No Country Year

No. of observations 8,504 8,504 6,599 6,599F-statistic 443.889 961.344 194.546 361.034R-squared 0.717 0.685 0.552 0.514δ 1.012 1.015 1.010 1.011

Notes: Standard errors clustered at the level of province–country pair are included in parentheses.Significance level = *p < 0.1, **p < 0.05, ***p < 0.01.Source: Authors’ calculations.

equation (5) as the second stage, we report 2SLS estimates. Panel A of the tabledisplays the second-stage results, and panel B displays the fist-stage results. Thesetting of fixed effects in Table 3 is the same as in Table 2. The treatment effectof ln(ACFICitCBTjt ) is again statistically significant and slightly larger than inTable 2, while the effects of the other explanatory variables are similar. At thebottom of Table 3, following Stock, Wright, and Yogo (2002), we also reportthe Cragg–Donald statistics, which are the same as the F-statistics testing thesignificance of the instrumental variable in the first-stage equations given thatthere is only one such variable. Since the Cragg–Donald statistics are muchlarger than their corresponding critical values shown in parentheses, our use


Table 3. Two-Stage Least Squares Estimates

Exports Imports


A. Second stage; dependent variable: ln(Exporti j,t+1) and ln(Importi j,t+1)ln(ACFICitCBTjt ) 0.336*** 0.378*** 0.381*** 0.288***

(0.034) (0.032) (0.068) (0.066)ln(GDPit ) 0.827*** 1.721*** 1.581*** 2.389***

(0.072) (0.066) (0.129) (0.125)ln(GDPjt ) 0.372*** 0.399*** 0.672*** 0.837***

(0.050) (0.048) (0.103) (0.104)ln(Distancei j ) −0.236*** −0.271*** −0.543*** −0.740***

(0.063) (0.060) (0.115) (0.115)Borderi j 3.513*** 3.335*** 3.288*** 3.287***

(0.597) (0.586) (0.829) (0.834)Religioni j 1.991*** 2.838*** 1.582*** 1.463***

(0.233) (0.256) (0.541) (0.467)ln(Populationit ) 0.576*** −0.508*** −0.379** −1.333***

(0.091) (0.084) (0.175) (0.154)ln(Population jt ) 0.267*** 0.226*** −0.273*** −0.320***

(0.048) (0.049) (0.098) (0.100)ln(Areait ) −0.645*** −0.507*** −0.953*** −0.732***

(0.033) (0.025) (0.062) (0.046)ln(Area jt ) −0.087*** −0.086*** 0.187*** 0.170***

(0.022) (0.021) (0.038) (0.037)SFIjt −0.033*** −0.026** −0.059** −0.070***

(0.012) (0.012) (0.024) (0.024)ln(TCPj,t−1) 0.117*** 0.113*** 0.093*** 0.108***

(0.021) (0.020) (0.035) (0.035)B. First stage; dependent variable: ln(ACFICitCBT jt )ln(PrivateFirmitCBTjt ) 1.001*** 1.015*** 0.988*** 0.975***

(0.009) (0.009) (0.010) (0.010)ln(GDPit ) −0.896*** −0.976*** −0.898*** −1.031***

(0.018) (0.020) (0.019) (0.024)ln(GDPjt ) 0.030*** 0.008 0.034** 0.072***

(0.011) (0.011) (0.013) (0.014)ln(Distancei j ) −0.025 0.074*** −0.029 −0.145***

(0.016) (0.016) (0.019) (0.019)Borderi j −0.067 −0.029 −0.093 −0.237*

(0.100) (0.097) (0.103) (0.131)Religioni j −0.115*** −0.173*** −0.003 −0.074

(0.043) (0.038) (0.087) (0.083)ln(Populationit ) 0.518*** 0.592*** 0.603*** 0.746***

(0.027) (0.024) (0.031) (0.028)ln(Population jt ) −0.002 0.021* −0.006 −0.040***

(0.012) (0.013) (0.015) (0.015)ln(Areait ) 0.162*** 0.135*** 0.138*** 0.097***

(0.010) (0.008) (0.013) (0.010)ln(Area jt ) 0.000 0.007 0.001 −0.002

(0.006) (0.006) (0.007) (0.007)SFIjt −0.001 −0.008*** −0.002 0.001

(0.003) (0.003) (0.003) (0.003)

Continued.


Table 3. Continued.

Exports Imports


ln(TCPj,t−1) 0.008* 0.004 0.013** 0.030***

(0.005) (0.005) (0.005) (0.006)

Province ProvinceFixed effects No Country Year No Country Year

No. of observations 6,203 6,203 4,870 4,870Cragg–Donald statistic 25,938 23,083 19,117 14,459(Critical value) (16.38) (16.38) (16.38) (16.38)First-stage R-squared 0.957 0.947 0.950 0.932Second-stage R-squared 0.702 0.786 0.549 0.640

Notes: Standard errors clustered at the level of province–country pair are included inparentheses. Significance level = *p < 0.1, **p < 0.05, ***p < 0.01. The critical value ofthe Cragg–Donald statistic in each parentheses corresponds to the 10% maximal IV size.Source: Authors’ calculations.

of ln(PrivateFirmitCBTjt ) as the instrumental variable seems validated and theconclusion from Table 2 confirmed.

In summary, through its official newspaper’s country-specific news coverage,the ACFIC has encouraged its member firms to increase exports and imports withthe countries it prioritizes. According to Tables 2 and 3, if the OLS and 2SLSestimates are unbiased as assumed, a 1% increase in the frequency a country’s nameappearing in the China Business Times would be expected to increase the PRC’strade activities with that country by around 0.3%. In view of the large volumes ofPRC exports and imports across the globe, this level of magnitude of the impacton trade is impressive. Therefore, this result convincingly demonstrates the largeinfluence of the ACFIC on the trading destinations of its member firms.

C. Exports versus Imports

Next, we conduct a comparative analysis between exports and imports basedon the technique articulated in equation (7). These estimates are reported in Table4.3 Those in columns (1) and (2) contain no fixed effects, and those in columns(3) and (4) contain interacted fixed effects for year, province, and country. Forcomparison purposes, columns (1) and (3) are the baseline regressions withoutthe interaction term with Typei jt , while columns (2) and (4) estimate coefficientswith the interaction term included as in equation (7). The primary objective inthis comparison is to examine the likelihood of a positive γ2, the coefficient of theinteraction term, to determine whether the effects of ACFICitCBTjt on exports aregreater than those on imports.

3Since the log of ACFICitCBTjt appears twice in equation (7), greatly complicating matters, we choose notto carry out 2SLS estimation in this case.


Table 4. Estimates of Exports-versus-Imports Comparison


ln(ACFICCBTi jt ) 0.251*** 0.193*** 0.193*** 0.135***

(0.036) (0.036) (0.032) (0.033)ln(ACFICCBTi jt )Typei jt 0.100*** 0.102***

(0.005) (0.005)ln(GDPit ) 0.945*** 1.167*** 1.253*** 1.485***

(0.052) (0.055) (0.050) (0.054)

ln(GDPjt ) 0.792*** 0.666*** 1.082*** 0.954***

(0.053) (0.052) (0.051) (0.050)ln(Distancei j ) −0.218*** −0.209** −0.214*** −0.212***

(0.080) (0.081) (0.070) (0.072)Borderi j 3.268*** 3.540*** 3.626*** 3.884***

(0.716) (0.733) (0.739) (0.757)Religioni j 0.707** 0.563* 1.480*** 1.255***

(0.306) (0.308) (0.284) (0.282)

ln(Populationit ) 0.176*** −0.139** −0.013 −0.322***

(0.068) (0.067) (0.065) (0.065)ln(Population jt ) −0.076 0.261*** −0.229*** 0.119**

(0.053) (0.054) (0.053) (0.054)ln(Areait ) −0.321*** −0.177*** −0.362*** −0.231***

(0.027) (0.027) (0.025) (0.023)ln(Area jt ) −0.138*** −0.323*** −0.185*** −0.381***

(0.022) (0.022) (0.020) (0.021)SFIi jt 0.022* 0.019 0.055*** 0.054***

(0.013) (0.013) (0.013) (0.013)ln(TCPi j,t−1) 0.094*** 0.099*** 0.078*** 0.087***

(0.022) (0.023) (0.022) (0.022)

Province ProvinceFixed effects No No Country Year Country Year

No. of observations 15,103 15,103 15,103 15,103F-statistic 317.004 308.884 678.995 614.109R-squared 0.522 0.562 0.479 0.521δ 1.012 1.090 1.009 1.099

Notes: Standard errors clustered at the level of province–country pair are included in parentheses.Significance level = *p < 0.1, **p < 0.05, ***p < 0.01.Source: Authors’ calculations.

Consistent with hypothesis (3) stated in section III, γ2 is positive andstatistically significant in columns (2) and (4). The coefficient of ln(ACFICitCBTjt )is also positive, indicating that the ACFIC promotes both exports and importswith the specific countries it prefers, albeit exports more than imports, and thatthe coefficients of the economic size variables are positive, while the coefficientof bilateral distance is negative. The coefficients of most control variables are alsoconsistent with those found in the previous tables except that most of the coefficientson SFIi jt are positive and statistically significant.

We also compute the value of δ to test the stability of the coefficient ofln(ACFICitCBTjt ) and that of the interaction term. In columns (1) and (3), the


definition of δ is the same as that displayed in Table 2. δ is slightly different incolumns (2) and (4) because it is defined as the ratio of the R-squared value withγ1 �= 0 and γ2 �= 0 to that with γ1 = γ2 = 0. Simply put, δ compares the R-squaredvalues before and after incorporating any term related to ACFICitCBTjt or Typei jt .Since all values of δ are larger than 1, the estimates of the treatment effects arestatistically valid.

In summary, the estimation results support hypothesis (3) in section III andshow that the ACFIC seems to have the effect of encouraging its member firms toconduct export activities to a greater extent than import activities with the countriesit prioritizes in the China Business Times. Numerically, while a 1% increase inACFICitCBTjt is expected to increase exports from province i to country j by0.24%, it is only expected to increase the imports of province i from country jby 0.14%. Consequently, this 0.1 percentage points difference could lead to a tradesurplus in province i and a trade deficit in country j.

VI. Robustness Checks

As pointed out when discussing our treatment of missing data for exportsand imports in the previous section, omitting observations for trade with zeroescan evade the problem arising when the log of zero is undefined, but it cannotassure econometric validity. To resolve this issue in a rigorous manner, we usetwo alternative methods: Heckman (1979) selection and Poisson pseudo-maximum-likelihood (PPML) models. Furthermore, based on the fixed-effects model and theArellano–Bond estimation, we add the lagged dependent variable as an independentvariable to help deal with confounding and endogeneity biases.

A. Heckman Selection Model

The Heckman selection model is an econometric maneuver to correct forbias caused by nonrandomly selected samples. In the context of trade, independentvariables with missing observations could possess properties different from thosewith nonmissing observations. Consequently, omitting them might have led toconsiderable imprecision. Following the earlier applications of the Heckmanselection model to the gravity model by Bikker and de Vos (1992) and Head andMayer (2010), we construct a Heckman-augmented, two-step gravity model. In thefirst step, the probability of a trade interaction being recorded between province–country pair i j is estimated by using a probit model:

Any_xi j,t+1 = φ

⎛⎝ln

(ACFICitCBTjt

), ln(GDPit ) , ln

(GDPjt

), ln

(Distancei j

),

Borderi j, Religioni j, ln(Populationit ) , ln(Population jt

),

ln(Areait ) , ln(Area jt

), SFIjt, ln

(TCPj,t−1

)⎞⎠

(8)


where Any_xi j,t+1 = 1 if there is a record of the relevant cross-border economicactivity (i.e., exports, imports, or exports-versus-imports, depending on thedefinition of xi j) between province i and country j, and Any_xi j,t+1 = 0 ifsuch a record does not exist. The symbol φ indicates that this is a probit-estimating operation rather than a linear function. In short, the first step could beunderstood as a selection process, detecting the commonalities among those withmissing observations and preparing to correct for the biases resulting from thesecommonalities. Then, in the second step, we use an estimating equation similar tothe combination of equations (5) and (7), but without fixed effects to avoid excessivecomplexity. Also, the inverse Mills ratio, λi jt , computed for each observationbased on the first stage is added as an additional regressor because if β13, thecoefficient of the inverse Mills ratio λi jt , is statistically significant, then the OLSestimations might well be subject to selection biases (Heckman 1979; Helpman,Melitz, and Rubinstein 2008). Thus, the new equation for this second step isexpressed as

ln(xi j,t+1

) = β1 ln(ACFICitCBTCBTjt

) +η ln(ACFICCBTi jt

)Typei jt

+ β2 ln(GDPit) + β3 ln(GDPjt

) + β4 ln(Distancei j

) + β5Borderi j

+ β6Religioni j + β7 ln(Populationit ) + β8 ln(Population jt

)+ β9 ln(Areait ) + β10 ln

(Area jt

) + β11SFIjt + β12 ln(TCPj,t−1

)+ β13λi jt + β0 + εi jt (9)

where η = 0 except when the equation is employed for the exports-versus-importscomparison.

Table 5 presents the regression results based on equation (8) for the firststage and equation (9) for the second stage for each of the different measures ofxi j (exports, imports, or exports-versus-imports). Columns (1) and (2) show theresults for exports and imports, respectively. Columns (3) and (4) show the resultsof the exports-versus-imports comparison without and with the interaction term,respectively. As shown, the inverse Mills ratio is only statistically significant incolumns (1) and (4), implying that our earlier estimates for imports can be trustedat least from the perspective of selection bias. After incorporating the inverseMills ratio as an additional regressor, the implications drawn from the results insection V still hold true for both exports and the exports-versus-imports comparisoneven though there no longer remains strong statistical evidence to support somecomponents of the gravity model, especially in column (1). Since the treatmenteffects represented by the parameters β1 and η remain positive and hover around0.25 in all specifications, the results robustly confirm all findings in section V.


Table 5. Heckman Selection Model

Exports Imports Exports versus Imports


ln(ACFICitCBTjt ) 0.290*** 0.279*** 0.245*** 0.190***

(0.022) (0.031) (0.019) (0.017)ln(ACFICCBTi jt )Typei jt 0.090***

(0.005)ln(GDPit ) 1.420*** 1.367*** 0.917*** 1.116***

(0.096) (0.104) (0.034) (0.036)ln(GDPjt ) 0.084 0.766*** 0.762*** 0.640***

(0.055) (0.082) (0.027) (0.025)ln(Distancei j ) −0.072 −0.618*** −0.151*** −0.120**

(0.054) (0.074) (0.058) (0.053)Borderi j 2.701*** 3.165*** 3.111*** 3.321***

(0.304) (0.401) (0.237) (0.227)Religioni j 2.015*** 1.444*** 0.678*** 0.520***

(0.135) (0.240) (0.125) (0.120)ln(Populationit ) −0.052 −0.087 0.147*** −0.151***

(0.114) (0.163) (0.042) (0.033)ln(Population jt ) 0.154*** −0.264*** −0.096** 0.198***

(0.036) (0.048) (0.038) (0.047)ln(Areait ) −0.878*** −0.968*** −0.347*** −0.226***

(0.038) (0.040) (0.019) (0.024)ln(Area jt ) −0.058*** 0.185*** −0.152*** −0.325***

(0.016) (0.020) (0.016) (0.014)SFIjt 0.007 −0.064*** 0.038** 0.043***

(0.009) (0.015) (0.016) (0.014)ln(TCPj,t−1) 0.082*** 0.085*** 0.086*** 0.087***

(0.015) (0.021) (0.013) (0.012)

Inverse Mills ratio −2.303*** −0.011 −0.449 −0.605**

(0.309) (0.348) (0.332) (0.302)

No. of observations 20607 19819 40426 40426No. of observations (Selected) 8420 6516 14936 14936No. of observations (Nonselected) 12187 13303 25490 25490

Notes: Standard errors are included in parentheses. Significance level = *p < 0.1, **p < 0.05,***p < 0.01.Source: Authors’ calculations.

B. Poisson Pseudo-Maximum-Likelihood Estimation

However, since the coefficients of economic size and bilateral distancevariables in the gravity model augmented by Heckman selection were not alwaysstatistically significant, we implement PPML estimation to reexamine the suitabilityof the gravity model. First introduced by Silva and Tenreyro (2006), the PPMLmethod estimates the gravity equation in its multiplicative form to simultaneouslysolve the problem of zero flows and to mitigate the presence of heteroskedasticity.Mathematically, the estimating equation for PPML is


xi j,t+1 = exp[β1 ln

(ACFICitCBTCBTjt

) +η ln(ACFICCBTi jt

)Typei jt +β2 ln(GDPit)

+ β3 ln(GDPjt

) + β4 ln(Distancei j

) + β5Borderi j + β6Religioni j

+ β7 ln(Populationit ) + β8 ln(Population jt

) + β9 ln(Areait )

+ β10 ln(Area jt

) + β11SFIjt + β12 ln(TCPj,t−1

) + β0 + πi jt + εi jt

](10)

where η = 0 except when the equation is employed for the exports-versus-importscomparison.

Table 6 reports the PPML estimation results. As in Table 5, columns (1) and(2) correspond to exports and imports, respectively. Columns (3) and (4) containthe results for the exports versus imports comparison. Unlike the OLS or 2SLSestimations using R-squared values to quantify the percentage of the varianceexplained by the independent variables, Table 6 employs pseudo R-squared, a proxyfor the regular R-squared, the estimates of which are displayed at the bottom ofthe table. Accordingly, δ becomes the ratio of the pseudo R-squared value withβ1 �= 0 to that with β1 = 0 in columns (1) through (3), and the ratio of the pseudoR-squared values with β1 �= 0 and β2 �= 0 to that with β1 = β2 = 0 in column (4).

According to Table 6, when this somewhat more rigorous variant of thegravity model is used, we find that the corresponding treatment effect of the ACFICand its newspaper on the PRC’s exports is around 30% higher than that in the OLSestimates obtained from Table 2, and the treatment effect on the PRC’s importsremains at roughly the same level. In addition, although the estimated coefficient onthe interaction term, η, in Table 6 is smaller than the OLS estimates from Table 4,it is still positive and statistically significant. The features of the gravity model alsoseem to hold, and the pseudo R-squared values are larger than the R-squared valuesin the previous tables. Thus, despite some changes in magnitudes, the directions ofall the findings regarding the ACFIC’s treatment effects on trade in section V areconfirmed by Table 6.

C. Are Past Economic Interactions Confounders?

Thus far, our statistical analysis has confirmed that the hypothesizedcorrelations between the ACFIC’s pair-wise (province–country) influences exertedby ln(ACFICitCBTjt ) on both exports and imports between the same pairsin the next year are both statistically significant and free of selection andheteroskedasticity biases. We have also dealt with the potential endogeneity ofACFICit with an instrumental variable approach. Yet, these discovered relationshipsmight still not be causal if the assumptions used to eliminate biases are incorrectand/or if there exists any other variable linking the dependent and any of theindependent variables, such as CBTjt , in our models. For example, some previousprovince–country economic interactions might have impacted both the ACFIC’s


Table 6. Poisson Pseudo-Maximum-Likelihood Estimates

Exports Imports Exports versus Imports


ln(ACFICitCBTjt ) 0.388*** 0.294*** 0.191*** 0.184***

(0.055) (0.079) (0.064) (0.066)ln(ACFICCBTi jt )Typei jt 0.015*

(0.008)ln(GDPit ) 1.618*** 1.132*** 0.845*** 0.900***

(0.216) (0.338) (0.099) (0.113)ln(GDPjt ) 0.515*** 0.509*** 0.946*** 0.897***

(0.084) (0.136) (0.110) (0.103)ln(Distancei j ) −0.414*** −0.630*** −0.456*** −0.456***

(0.090) (0.164) (0.154) (0.154)Borderi j 3.448*** 3.529*** 2.711*** 2.713***

(0.339) (0.732) (0.485) (0.486)Religioni j 2.919*** 1.811*** 1.699*** 1.699***

(0.414) (0.621) (0.509) (0.509)ln(Populationit ) 0.302 0.477 0.013 −0.051

(0.261) (0.409) (0.111) (0.121)ln(Population jt ) 0.028 −0.153 −0.081 −0.028

(0.069) (0.122) (0.107) (0.100)ln(Areait ) −0.673*** −0.830*** −0.230*** −0.209***

(0.069) (0.111) (0.051) (0.053)ln(Area jt ) −0.093*** 0.069** −0.176*** −0.195***

(0.029) (0.028) (0.038) (0.039)SFIjt −0.056*** −0.123*** −0.020 −0.020

(0.018) (0.033) (0.022) (0.022)ln(TCPj,t−1) 0.197*** 0.191*** 0.133*** 0.133***

(0.037) (0.060) (0.041) (0.041)

Province Province Province ProvinceCountry Country Country Country

Fixed effects Year Year Year Year

No. of observations 20607 19819 40426 40426Pseudo R-squared 0.864 0.775 0.768 0.769δ 1.049 1.063 1.061 1.062

Notes: Standard errors clustered at the level of province–country pair are included inparentheses. Significance level: *p < 0.1, **p < 0.05, ***p < 0.01.Source: Authors’ calculations.

current influence on that pair and that pair’s future economic interactions. If so,this would imply that the correlations identified above could be spurious.

To address this type of potential threat, we use a fixed-effects model anda dynamic panel data approach by including ln(xi j,t ), the lagged value of thedependent variable, in the set of independent variables. Mathematically, the neweconometric equation can be expressed as

ln(xi j,t+1

) = α1 ln(xi j,t

) + α2 ln(ACFICitCBTjt

) + W ′i jtξ + πt + πi j + εi jt

(11)


where xi j,t+1 represents the economic interactions (exports or imports) betweenprovince i and country j in year t + 1, xi j,t is the lagged value of xi j,t+1 in yeart, Wi jt is the set of other control variables identified in equation (5), πt is the fixedeffect for year t, and πi j is the interacted fixed effect for province i and countryj (not just for their corresponding PRC region and continent). There is only onelagged value of the dependent variable because the timeliness of news is assumed.Based on equation (11), we endeavor to employ various estimation techniques toexamine whether the coefficient of ln(ACFICitCBTjt ), α2, is positive, statisticallysignificant, and not overridden by the presence of ln(xi jt ), so that the results of thisexercise would at least be indicative of the robustness of the conclusions from theprevious sections.

We first use standard fixed-effects models to estimate equation (11), theresults of which are reported in the first four columns of Table 7. Columns (1)and (3) exclude the control variable set Wi jt , and columns (2) and (4) include it.While xi j represents the exports from province i to country j in columns (1) and(2), in columns (3) and (4) it represents the imports by province i from country j.According to the four columns, adding the 1-year lagged value of the dependentvariable does not override the finding that the coefficient of ln(ACFICitCBTjt ) ispositive and statistically significant, implying that the ACFIC has managed to exertsubstantial effects on the trading activities between province i and country j evenafter controlling for the influence of past economic interactions.

Moreover, we truncate our dataset into a balanced panel data and conducta Harris–Tzavalis unit root test designed for samples with short time periodsbut many cross-sectional units (Harris and Tzavalis 1999). This test helps ensurethat the premise of stationarity for the practice of including lagged dependentvariable is not violated as suggested by Keele and Kelly (2006). As shown at thebottom of the table, all the p-values from the Harris–Tzavalis unit root tests for thedependent variables (exports or imports) are far smaller than 5%, thereby rejectingthe null hypothesis of the existence of unit roots. Thus, the dependent variables arestationary, and the use of their lagged values legitimate.

While the results from the first four columns in Table 7 are quite satisfying,they could be exposed to the Nickell bias (Nickell 1981) in that the differencebetween each dependent or independent variable and its mean across years withina cross-sectional unit could create a correlation between the independent variablesand the error term. To mitigate this imprecision, Arellano and Bond (1991) deviseda dynamic panel data approach, which takes the first differences of the dependent,lagged dependent, and independent variables, utilizes the first differences of thelagged dependent and lagged independent variables as instruments, and estimatesthe entire system with the generalized method of moments. Since our panel datasetcontains a small number of time periods and a large number of cross-sectionalunits, which is of the type for which this Arellano–Bond estimator was designed,we believe that its use is appropriate in this context and serves to minimize


Table 7. Dynamic Panel Data Analysis (Dependent Variables: Exports and Importsin Year t + 1)

Arellano–BondFixed-Effects Estimation Estimation

(1) (2) (3) (4) (5) (6)

xi j,t as xi j,t as xi j,t as xi j,t as xi j,t as xi j,t asDependent Variables Export Export Import Import Export Import

ln(xi j,t ) 0.896*** 0.800*** 0.855*** 0.806*** 0.273** 0.440***

(0.007) (0.009) (0.008) (0.009) (0.139) (0.152)ln(ACFICitCBTjt ) 0.109*** 0.069*** 0.116*** 0.074*** 1.140*** 0.992**

(0.009) (0.012) (0.013) (0.020) (0.212) (0.475)ln(GDPit ) 0.422*** 0.242***

(0.052) (0.093)ln(GDPjt ) 0.073*** 0.060*

(0.017) (0.031)ln(Distancei j ) −0.032 −0.074

(0.032) (0.050)Borderi j 0.658*** 0.776***

(0.130) (0.219)Religioni j 0.431*** 0.449***

(0.076) (0.171)ln(Populationit ) −0.132** −0.001

(0.059) (0.118)ln(Population jt ) 0.068*** −0.006

(0.018) (0.031)ln(Areait ) −0.122*** −0.217***

(0.015) (0.034)ln(Area jt ) −0.026*** 0.016

(0.008) (0.013)SFIjt −0.011*** −0.025***

(0.004) (0.008)ln(TCPj,t−1) 0.027*** 0.015

(0.008) (0.014)

Province Province Province Province Province ProvinceCountry Country Country Country Country Country

Fixed effects Year Year Year Year Year Year

No. of observations 6,685 6,685 4,958 4,958 6,229 4,535Harris–Tzavalis

statistics0.159*** 0.159*** 0.083*** 0.083***

Hansen test (p-value) [0.529] [0.004]Serial correlation of

order 1 (p-value)[0.000] [0.000]

Serial correlation oforder 2 (p-value)

[0.274] [0.127]

Notes: Standard errors clustered at the level of province–country pair are included in parentheses. Significance level= *p < 0.1, **p < 0.05, ***p < 0.01.Source: Authors’ calculations.


the endogeneity bias driven by past trading activities in our panel data, therebyenhancing the credibility of the findings of the substantial effects exerted by theACFIC.

The last two columns of Table 7 display the results of the Arellano–Bondestimation on equation (11). Following Arellano and Bond (1991), we performserial correlation tests to determine whether this estimator’s assumption that thedifferenced error term is first-order, but not second-order serially correlated, issatisfied. As shown at the bottom of the table, both p-values for the first-orderserial correlations are smaller than 1%, and both for the second-order are larger than10%, so the serial correlation tests are passed. Moreover, to avoid overidentificationcaused by having too many strong instruments, we collapse the generalized methodof moments style instruments and restrict the lagged periods to year t − 6 andyear t − 7 to be sufficiently far away from year t, as suggested by Wintoki, Linck,and Netter (2012). These procedures help us eventually obtain the greater-than-10% p-value of the Hansen test for exports, but the p-value of the Hansen test forimports is still smaller than 1%. Thus, though the overidentifying condition forexports is satisfied, that for imports is still violated. These statistical tests ensurethe appropriateness of the Arellano–Bond estimation for exports but cast doubton this practice for imports. In summary, although the estimates of α2 are bothpositive and statistically significant for exports and imports, we are only confidentthat the estimates are free of possible bias in the case of exports. Numerically, wefind that the estimates of α2 in both columns are larger than 0.5, which suggeststhat the values of the coefficient on ln(ACFICitCBTjt ) in previous tables might beunderestimated.

As there could still be other estimation biases, the implications drawn fromthis subsection do not guarantee that the ACFIC’s influences on the province–country pair and that pair’s future trade activities are causal. Yet, the results based onthis practice of including the lagged value of the dependent variable and applyingthe Arellano–Bond estimation still increase the credibility of both the estimatespresented and the econometric methods used throughout this study, especially inthe case of exports.

VII. Has the ACFIC Promoted Relations between the PRC and BRI Countries?Difference-in-Differences Analysis

Even after the discussion above of how the ACFIC influences the PRC’sforeign trade with the countries it seems to favor, hypothesis (2) about the BRIremains untested. This section concentrates on whether since 2013 the ACFIC hascome to prioritize BRI countries in the China Business Times. Only if BRI countrieshave indeed become the ACFIC’s increasingly favored targets since 2013 may wesafely conclude that the ACFIC has encouraged its member firms to trade more withBRI countries.


To identify the change in the global distribution of the ACFIC’s frequentlymentioned targets before and after the BRI’s initiation, we combine anautoregressive model of order one (AR[1]) with the difference-in-differences (DiD)method to determine whether the BRI or related geographical information has acausal relationship with CBTj.4 Formally,

CBTjt = ϕ1 CBTj,t−1 + ϕ2Dj + ϕ3Postt + ϕ4DjPostt + ϕ5 ln(GDPjt

)+ ϕ6 ln

(Population jt

) + ϕ7g jt + ϕ8n jt + ϕ0 + εi jt (12)

where Dj is the dummy for country j being a BRI member, Postt = 0 if t ≤ 2013or Postt = 1 if t > 2013, g jt is the GDP growth rate of country j in year t, and n jt isthe population growth rate of country j in year t. As alternatives to the dummy forthe BRI as a whole as the dependent variable, we also create dummy variables forDj for four different subregions of the BRI: Central Asia and the Caucasus, Africa,Eastern Europe, and Southeast Asia. Applying the model to the four subregionsseparately, we can determine whether there is any difference between these regionsin terms of ϕ4, the coefficient on DjPostt . As in Card and Krueger (1994) and otherDiD empirical studies, if ϕ4 is positive, this would indicate that the ACFIC hasincreased its reports about the countries defined by the dummy variable Dj sincethe inauguration of the initiative in 2013, or the opposite if ϕ4 is negative.

The regression estimates based on equation (12) are reported in Table 8.5

Each column reports the results for the dummy Dj and its interaction with Postt fora different set of BRI countries. Column (1) is for BRI membership as a whole,column (2) is for BRI countries in Africa, column (3) for those in Central Asia andthe Caucasus, column (4) for those in Eastern Europe, and column (5) for those inSoutheast Asia.

The entries in the first row of the table represent the effects of the CBTj inthe previous year, which are indeed all positive and strong, revealing considerablepersistence of this AR(1) model. The estimate of the parameter ϕ2 is negative inmost specifications although not always statistically significant, suggesting thatBRI countries have received smaller amounts of attention from the ACFIC thanother large trading partners of the PRC such as the US and Japan. This is areasonable finding because most BRI countries are developing countries. The resultthat the estimated values of ϕ3 are also negative indicates that the ACFIC hasdecreased its overall news reports about non-PRC countries, consistent with its“Going Inward” strategy since 2009 as documented by Lei and Nugent (2018).In none of the columns are the coefficients of GDP growth or population growthstatistically significant, indicating that the ACFIC’s attention to a specific countrydoes not necessarily depend on that country’s economic or demographic status.

4See Wooldridge (2010, 197) for explanations and examples of the AR(1) model and Angrist and Pischke(2008, 227–46) for the DiD model.

5Appendix Figure A2 confirms that our DiD model satisfies the parallel trend assumption.


Table 8. Difference-in-Difference Estimates of the All-China Federation of Industry andCommerce’s Prioritization in the China Business Times

Central Asia Eastern SoutheastBRI Africa and Caucasus Europe Asia

Explanatory Variables (1) (2) (3) (4) (5)

CBTj,t−1 0.940*** 0.940*** 0.940*** 0.939*** 0.940***

(0.003) (0.003) (0.003) (0.003) (0.003)Dj −13.959* −8.959*** −7.371*** −7.317* 4.752

(5.624) (2.406) (2.006) (3.089) (4.281)Postt −49.487*** −23.822*** −20.761*** −21.134*** −19.086***

(14.585) (5.582) (4.716) (4.991) (4.746)DjPostt 40.931** 20.559*** 21.019*** 14.874* −2.503

(14.698) (5.702) (4.846) (6.225) (8.091)ln(GDPjt ) 1.098* 0.549 0.729 0.655 0.688

(0.452) (0.497) (0.447) (0.448) (0.432)ln(Population jt ) 0.234 0.473 0.407 0.464 0.355

(0.319) (0.280) (0.256) (0.255) (0.239)g jt −21.676 −21.035 −23.616 −25.908 −25.899

(22.204) (21.361) (22.109) (20.835) (20.527)n jt −53.893 −51.396 −58.668 −69.335 −51.615

(37.050) (42.369) (39.428) (47.293) (39.036)No. of observations 985 985 985 985 985F-statistic 42,047.26 47,125.74 46,984.37 44,900.18 50,166.27R-squared 0.968 0.967 0.967 0.967 0.967

BRI = Belt and Road Initiative.Notes: Standard errors clustered at the country level are included in parentheses. Significance level = *p < 0.1,**p < 0.05, ***p < 0.01.Source: Authors’ calculations.

Most importantly, however, ϕ4 is positive in all columns except the two forSoutheast Asia, which demonstrates that since 2013 the ACFIC has indeed boostedits relative attention to the BRI in general and to Africa, Central Asia and theCaucasus, and Eastern Europe (but not Southeast Asia) in particular.

Combining this finding with the conclusion drawn from previous sections,it would appear that the ACFIC has induced its member firms to engage inmore trade with BRI countries since 2013. This statistical implication persuasivelydemonstrates that the ACFIC has substantially helped the central government toalign its member firms with the national objective of developing the BRI, at leastbased on the information disseminated by the ACFIC’s newspaper. However, thisimpact has been quite unequal across different groups of BRI countries.

VIII. Conclusion

The results presented in sections V, VI, and VII have demonstrated that theACFIC has managed to induce its member firms from the private sector in the PRC’sdifferent provinces to engage in both exports and imports with the countries thatthe ACFIC has stressed in its newspaper, the China Business Times. On average, a1% increase in the newspaper’s level of dissemination of the positive opportunities


in a non-PRC country has increased the PRC’s trade activities in that country byaround 0.3% (and perhaps more as indicated in Tables 6 and 7). The results havealso been quite robust to different model specifications and means of dealing withpossible econometric problems, although the implications for exports are likely tobe more reliable than those for imports based on the Arellano–Bond estimates. Thelast step in the analysis showed that, although the ACFIC has been posting fewernews articles about other countries in recent years, reflecting the continuation of its“Going Inward” strategy, its focus on news about BRI countries has not decreased.In addition, from the use of the interaction term that compares the ACFIC’s effectson exports with those on imports, we find fairly strong evidence that the ACFIC’sinfluence on the PRC’s exports to BRI countries has been substantially larger thanon its imports from those countries.

Given the vulnerability of such a massive program as the BRI to so manydifferent risks, especially with regard to debt default risks that have been risingin several BRI countries, the Government of the PRC and the ACFIC might dowell to be concerned by the evidence presented here of the unequal balance ofpayment effects between the PRC and many of its BRI partners. The results suggestthat some attention should be given to policies that could increase imports intothe PRC from these BRI countries to prevent them from defaulting on loans orexperiencing other macroeconomic crises. In cases where business associations inother BRI countries appear to have some potential to act as a coordinating entity, itmay also be useful to see if the ACFIC can coordinate with, or even train membersof, such business associations in other BRI countries to increase their ability tocoordinate with member firms. The PRC may also want to increase its imports oflabor-intensive goods and services from other BRI countries so as to focus on itsMade in China 2025 strategic plan.

We admit, however, that this analysis has been conducted based onincomplete data, especially in terms of being limited to the first 5 years of theBRI’s implementation. Hence, regular updates of the present analysis, includingon efficiencies within the PRC, will be needed and preferably also extendedto commodity-specific and/or firm-specific trade and investment among BRIcountries.

References

All-China Federation of Industry and Commerce. 2018. “Report on Top 500 Chinese PrivateFirms (��500��).” Beijing. [In Chinese]

Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005. “Selection on Observedand Unobserved Variables: Assessing the Effectiveness of Catholic Schools.” Journal ofPolitical Economy 113 (1): 151–84.

Anderson, James E. 1979. “A Theoretical Foundation for the Gravity Equation.” AmericanEconomic Review 69 (1): 106–16.

Anderson, James E., and Eric van Wincoop. 2003. “Gravity with Gravitas: A Solution to theBorder Puzzle.” American Economic Review 93 (1): 170–92.


Angrist, Joshua D., and Jörn-Steffen Pischke. 2008. Mostly Harmless Econometrics: AnEmpiricist’s Companion. Princeton: Princeton University Press.

Arellano, Manuel, and Stephen Bond. 1991. “Some Tests of Specification for Panel Data: MonteCarlo Evidence and an Application to Employment Equations.” Review of Economic Studies58 (2): 277–97.

Baniya, Suprabha, Nadia Rocha, and Michele Ruta. 2019. Trade Effects of the New Silk Road: AGravity Analysis. World Bank.

Bennett, Robert J. 1998. “Business Associations and Their Potential Contribution to theCompetitiveness of SMEs.” Entrepreneurship and Regional Development 10 (3): 243–60.

Bikker, Jacob A., and Aart F. de Vos. 1992. “An International Trade Flow Model with ZeroObservations: An Extension of the Tobit Model.” Brussels Economic Review 135: 379–404.

Bird, Julia Helen, Mathilde Sylvie Maria Lebrand, and Anthony J. Venables. 2019. The Belt andRoad Initiative: Reshaping Economic Geography in Central Asia? Washington, DC: WorldBank.

Blomberg, S. Brock, and Gregory D. Hess. 2006. “How Much Does Violence Tax Trade?” Reviewof Economics and Statistics 88 (4): 599–612.

Bräutigam, Deborah, Lise Rakner, and Scott Taylor. 2002. “Business Associations and GrowthCoalitions in Sub-Saharan Africa.” Journal of Modern African Studies 40 (4): 519–47.

Bruni, Michele 2019. “The Belt and Road Initiative: Demographic Trends, Labour Markets andWelfare Systems of Member Countries.” Global Labor Organization Discussion Paper No.300.

Card, David, and Alan B. Krueger. 1994. “Minimum Wages and Employment: A Case Study ofthe Fast-Food Industry in New Jersey and Pennsylvania.” American Economic Review 84(4): 772–93.

Chen, Maggie Xiaoyang, and Chuanhao Lin. 2018. Foreign Investment across the Belt and Road:Patterns, Determinants, and Effects. Washington, DC: World Bank.

Cheng, Xiaobo, ed. 2018. Big Data Report of Trade Cooperation under the Belt and RoadInitiative 2018. Beijing: State Information Center of China. [In Chinese]

de Soyres, François, Alen Mulabdic, Siobhan Murray, Nadia Rocha, and Michele Ruta. 2018.How Much Will the Belt and Road Initiative Reduce Trade Costs? Washington, DC: WorldBank.

Doner, Richard F., and Ben Ross Schneider. 2000. “Business Associations and EconomicDevelopment: Why Some Associations Contribute More Than Others?” Business andPolitics 2 (3): 261–88.

Dunford, Michael, and Weidong Liu. 2019. “Chinese Perspectives on the Belt and RoadInitiative.” Cambridge Journal of Regions, Economy and Society 12 (1): 145–67.

Dunlevy, James A. 2006. “The Influence of Corruption and Language on the Pro-Trade Effect ofImmigrants: Evidence from the American States.” Review of Economics and Statistics 88(1): 182–86.

Elmer, Keegan. 2018. “Anti-China Protests in Vietnam Set to Aggravate Tensions withBeijing.” June 13. https://www.scmp.com/news/china/diplomacy-defence/article/2150653/anti-china-protests-vietnam-set-aggravate-tensions.

Gibson, John, and Chao Li. 2018. “The Belt and Road Initiative and Comparative RegionalProductivity in China.” Asia and the Pacific Policy Studies 5 (2): 168–81.

Gill, Indermit, Somik Y. Lall, and Mathilde Lebrand. 2019. “Winners and Losers alongChina’s Belt and Road.” Brookings Institution, June 21. https://www.brookings.edu/blog/future-development/2019/06/21/winners-and-losers-along-chinas-belt-and-road/.


Harris, Richard D.F., and Elias Tzavalis. 1999. “Inference for Unit Roots in Dynamic PanelsWhere the Time Dimension is Fixed.” Journal of Econometrics 91 (2): 201–26.

Head, Keith, and Thierry Mayer. 2010. “Illusory Border Effects: Distance MismeasurementInflates Estimates of Home Bias in Trade.” In The Gravity Model in International Trade:Advances and Applications. Cambridge: Cambridge University Press.

Heckman, James J. 1979. “Sample Selection Bias as a Specification Error.” Econometrica 47 (1):153–61.

Helpman, Elhanan, Marc Melitz, and Yona Rubinstein. 2008. “Estimating Trade Flows: TradingPartners and Trading Volumes.” Quarterly Journal of Economics 123 (2): 441–87.

Herrero, Alicia Garcia, and Jianwei Xu. 2017. “China’s Belt and Road Initiative: Can EuropeExpect Trade Gains?” China & World Economy 25 (6): 84–99.

Hillman, Jonathan E. 2018. “How Big is China’s Belt and Road?” Center for Strategicand International Studies, April 3. https://www.csis.org/analysis/how-big-chinas-belt-and-road.

Huang, Yiping. 2016. “Understanding China’s Belt & Road Initiative: Motivation, Frameworkand Assessment.” China Economic Review 40: 314–21.

Hurley, John, Scott Morris, and Gailyn Portelance. 2018. “Examining the Debt Implications ofthe Belt and Road Initiative from a Policy Perspective.” Center for Global DevelopmentPolicy Paper 121.

Irshad, Muhammad Saqib, Qi Xin, and Hamza Arshad. 2015. “One Belt and One Road:Does China-Pakistan Economic Corridor Benefit for Pakistan’s Economy?” Journal ofEconomics and Sustainable Development 6 (24): 200–07.

Jia, Nan. 2014. “Are Collective Political Actions and Private Political Actions Substitutes orComplements? Empirical Evidence from China’s Private Sector.” Strategic ManagementJournal 35 (2): 292–315.

Johnson, Simon, John McMillan, and Christopher Woodruff. 2000. “Entrepreneurs and theOrdering of Institutional Reform: Poland, Slovakia, Romania, Russia and UkraineCompared.” Economics of Transition 8 (1): 1–36.

Keele, Luke, and Nathan J. Kelly. 2006. “Dynamic Models for Dynamic Theories: The Ins andOuts of Lagged Dependent Variables.” Political Analysis 14 (2): 186–205.

Kohl, Tristan. 2019. “The Belt and Road Initiative’s Effect on Supply-Chain Trade: Evidencefrom Structural Gravity Equations.” Cambridge Journal of Regions, Economy and Society12 (1): 77–104.

Lee, Jeong-ho. 2019. “How the Tiny African Nation of Djibouti Became the Linchpin inChina’s Belt and Road Plan.” April 28. https://www.scmp.com/news/china/diplomacy/article/3007924/how-tiny-african-nation-djibouti-became-linchpin-chinas-belt.

Lei, Zhenhuan, and Jeffrey B. Nugent. 2018. “Coordinating China’s Economic Growth Strategyvia Its Government-Controlled Association for Private Firms.” Journal of ComparativeEconomics 46 (4): 1273–93.

Lewer, Joshua J., and Hendrik Van den Berg. 2007. “Estimating the Institutional and NetworkEffects of Religious Cultures on International Trade.” Kyklos 60 (2): 255–77.

Ma, Guangrong, Oliver Meng Rui, and Yiping Wu. 2015. “A Springboard into Politics:Do Chinese Entrepreneurs Benefit from Joining the Government-Controlled BusinessAssociations?” China Economic Review 36: 166–83.

Maurel, Mathilde, and Emiel Afman. 2010. “Diplomatic Relations and Trade Reorientationin Transition Countries.” In The Gravity Model in International Trade: Advances andApplications. Cambridge: Cambridge University Press.


Nickell, Stephen. 1981. “Biases in Dynamic Models with Fixed Effects.” Econometrica 49 (6):1417–26.

Önis, Ziya, and Umut Türem. 2001. “Business, Globalization and Democracy: A ComparativeAnalysis of Turkish Business Associations.” Turkish Studies 2 (2): 94–120.

Oster, Emily. 2017. “Unobservable Selection and Coefficient Stability: Theory and Evidence.”Journal of Business & Economic Statistics 37 (2): 1–18.

Rehman, Faheem Ur, and Yibing Ding 2019. “The Nexus between Outward Foreign DirectInvestment and Export Sophistication: New Evidence from China.” Applied EconomicsLetters 1: 1–9.

Silva, J.M.C. Santos, and Silvana Tenreyro. 2006. “The Log of Gravity.” Review of Economicsand Statistics 88 (4): 641–58.

Solangi, Murtaza. 2018. “Why are Separatist Militants Violently Targeting Chinese inPakistan?” November 24. https://www.scmp.com/week-asia/geopolitics/article/2174837/why-are-separatist-militants-violently-targeting-chinese.

Stock, James H., Jonathan H. Wright, and Motohiro Yogo. 2002. “A Survey of Weak Instrumentsand Weak Identification in Generalized Method of Moments.” Journal of Business &Economic Statistics 20 (4): 518–29.

Sukiassyan, Grigor, and Jeffrey B. Nugent. 2011. “Lobbying or Information Provision: WhichFunctions of Associations Matter Most for Member Performance?” Eastern EuropeanEconomics 49 (2): 30–63.

Tarabay, Jamie, and Sang-Hun Choe. 2018. “Xi and Pence Stake Out Trade Positions in DuelingSpeeches at Pacific Rim Forum.” New York Times, November 17.

Tinbergen, Jan. 1962. Shaping the World Economy: Suggestions for an International EconomicPolicy. New York: Twentieth Century Fund.

Wiederer, Christina K. 2018. Logistics Infrastructure along the Belt and Road InitiativeEconomies. Washington, DC: World Bank.

Wintoki, M. Babajide, James S. Linck, and Jeffry M. Netter. 2012. “Endogeneity and theDynamics of Internal Corporate Governance.” Journal of Financial Economics 105 (3):581–606.

Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data. Cambridge:MIT Press.

Yu, Hong. 2017. “Motivation behind China’s “One Belt, One Road” Initiatives and Establishmentof the Asian Infrastructure Investment Bank.” Journal of Contemporary China 26 (105):353–68.

Zhai, Fan. 2018. “China’s Belt and Road Initiative: A Preliminary Quantitative Assessment.”Journal of Asian Economics 55: 84–92.

Appendix 1. English Translation of a Sample Article in the China Business TimesDedicated to the Construction of the Interconnected InformationInfrastructure in Africa

“We believe that more than 150 thousand kilometers of optical cables willbe laid in the next 15 to 20 years, and the consumption of cable-related goods inAfrica will be greater than 100 billion US dollars.” In the eyes of Wang Jianyi,the chairman of Zhejiang’s Federation of Industry and Commerce as well as thechairman of Futong Group’s board of directors, Africa is a continent full of hope. He


is very optimistic about the future of the interconnected information infrastructurein Africa.

Founded in 1987 and headquartered in Hangzhou, Zhejiang, FutongGroup is a Chinese private firm focusing on high-tech manufacturing. Itsindustrial specializations include optical fiber communication and electric powertransmission, and its research specializations include energy storage, high-temperature superconductor and submarine photoelectric composite cable. Today,Futong Group has 1 international headquarters, 3 regional headquarters, 31factories, 15 national high-tech subsidiaries and more than 12,000 registeredemployees.

In recent years, following the Belt and Road Initiative, hundreds of Chinesecompanies have been participating in the construction of foreign interconnectedinformation infrastructure. Futong Group is one of the participants as well as thebeneficiaries.

Futong’s development in Africa exemplifies the company’s recentglobalization. In countries such as Kenya, Nigeria, Seychelles, and Angola,Futong’s products have been widely applied to local telecommunication, electricaltransmission, automobile manufacturing, mobile terminal, and household electricalappliances.

According to chairman Wang, Chinese private firms are very competitivein fields such as optical fiber transmission and terminal equipment. Given theseadvances, Chinese firms are able to lead the construction of the interconnectedinformation infrastructure in Africa.

Futong’s long-term goal is to become an international cable manufacturingconglomerate respected by the society and promoting global sustainabledevelopment. African continent is a wonderful market from chairman Wang’sperspective. Following the “Made in China 2025” strategy, Futong has been activelyparticipating in the construction of information infrastructure in multiple Africancountries to realize the upgrade of local optical communication industry and builda world-class cluster of advanced manufacturing.

“The industrialization in Africa and the manufacturing reform in (theprovince of) Zhejiang are highly complementary, and there is a perfect synergybetween them.” According to chairman Wang, the industrialization in Africashould rely on Zhejiang’s advances in manufacturing, automotive and informationtechnology. As the chairman of Zhejiang’s Federation of Industry and Commerce,he expresses that Zhejiang’s Federation of Industry and Commerce is very willingto advocate the economic cooperation between China and Africa and accelerate theindustrialization of African countries.

Chairman Wang also argues that Chinese private firms need to agglomeratetogether when they are developing their business in Africa. In other words, takingadvantage of constructing industrial parks, Chinese private firms should developorderly industrial chains instead of doing business on their own.


Wang’s proposal is inspired by his conglomerate’s recent experiences ofglobalization. Futong Group, together with other Chinese private firms, hasdeveloped an eight-squared-kilometer high-end industrial park in Mexico. Throughcombining each company’s advantages, they together formed orderly industrialchains and competed with other countries’ firms.

As a Chinese poem goes, the immense sea allows fish to leap at liberty, andthe vast sky allows birds to fly at liberty. 2018 is the fifth anniversary of the Beltand Road Initiative. As the cornerstone of information interconnection, informationinfrastructure is an important component in the development of the Belt and RoadInitiative. Following the Belt and Road Initiative and develop industrial parks inforeign countries, Chinese firms such as Futong Group obtain a greater amount ofopportunities for their business development.

Futong Group’s Official Website (in English): http://www.futonggroup.com.cn/en/

Source: Li, Renping. 2018. “Futong Group Wants to Become an InternationalCable Manufacturing Conglomerate.” China Business Times, September 18. http://epaper.cbt.com.cn/epaper/uniflows/html/2018/09/18/01/01_68.htm. [In Chinese]

Appendix 2

Figure A2. Parallel Trend Test for Difference-in-Difference Estimates of the All-ChinaFederation of Industry and Commerce’s Prioritization in the China Business Times

BRI = Belt and Road Initiative, CBT = China Business Times.Notes: The vertical axis represents CBTj,t , the frequency of the name of country j appearing in the China BusinessTimes in year t. The horizontal axis represents year t. The dashed line represents the average CBTj,t of all non-BRIcountries. The solid line represents the average CBTj,t of all BRI countries. The vertical line represents the thresholdwhen the BRI intervention began to take effect. The dotted line represents the counterfactual average CBTj,t of allBRI countries if the BRI did not exist.Source: Authors’ calculations.

The Borrowing Puzzle: Why Do FilipinoDomestic Workers in Hong Kong, China

Borrow Rather than Dissave?Wooyoung Lim and Sujata Visaria∗

Despite their predictable and regular incomes, Filipino domestic workers inHong Kong, China commonly finance large expenses through interest-bearingloans rather than savings. Our analysis of survey data and records of a creditcooperative for migrant workers suggests that this cannot be explained by theirinability to save, financial illiteracy, short time horizon, or limited liability.Instead, we speculate that the strict schedules and high interest rates of theseloans create a disciplining effect that these individuals find desirable. Thismay help them avoid unnecessary consumption or demands from their socialnetwork. However, interventions should also consider that these workers oftenreceive nonmonetary reciprocal benefits from members of their social network.

Keywords: coholding, loans, migrants, savingsJEL codes: O15, O16

I. Introduction

Domestic workers make up a significant flow of migrants in Asia. Employersfrom several higher-income regions such as Malaysia; Singapore; and HongKong, China recruit live-in domestic help from countries such as the Philippinesand Indonesia. This migration is low-skilled, temporary, and almost entirelymotivated by economic gain. However, our understanding of the benefits of suchmigration on the migrant’s household is sparse. Migrant remittances have beenshown to improve dependents’ contemporaneous living standards and educational

∗Wooyoung Lim: Department of Economics, Lee Shau Kee Business Building, Hong Kong University of Scienceand Technology (HKUST), Clear Water Bay, Hong Kong, China. E-mail: [email protected]; Sujata Visaria(corresponding author): Department of Economics, HKUST, Clear Water Bay, Hong Kong, China. E-mail:[email protected]. A large team of HKUST Undergraduate Research Opportunities Program (UROP) students, ablyled and supported by Arpita Khanna, Sheren Ku, and Jimmy Santiago, helped collect the data for this paper. TheAsian Migrants Credit Union kindly shared their records. Ethics approval was obtained from HKUST. We thank RinaLookman Jio and Ziyi Hong for their terrific help analyzing the data. We also thank Utpal Bhattacharya, ClarenceLee, Dilip Mookherjee, Jane Y. Zhang, and seminar audiences at the 2019 Asian Development Bank-InternationalEconomic Association Round Table and the Indian Institute of Management Ahmedabad for their insightfulconversations, as well as the managing editor and anonymous referees for helpful comments and suggestions. Thisresearch was funded by an HKUST Institute for Emerging Market Studies Research Grant. The Asian DevelopmentBank recognizes “Hong Kong” as Hong Kong, China. The usual ADB disclaimer applies.





investments (Yang 2011), but does this migration enable long-term economicgains and opportunities for upward mobility? In particular, are these migrantseconomically self-sufficient when they retire and return to their home countries?Does this migration successfully lead the next generation to move into higher-wageoccupations?

Definitive answers to these questions require large-scale and long-term datacollection and plausibly exogenous variation in the migration decision. However,even in their absence, we can observe and analyze migrants’ constraints and choicesand make inroads toward an understanding. This paper begins with the observationthat migrants’ financial choices during their tenure in the host country greatlyinfluence their own and their households’ future outcomes. Whether their incomesare spent only on consumption or also saved, and whether they make productiveinvestments for financial gain, will determine whether they will retire comfortablyand whether migration will improve their economic status.

Our study population is Filipino domestic workers in Hong Kong, China. Weexamine how they manage their finances, specifically, their choice between savingsand loans. As we will document, it would appear that their financial choices donot maximize their economic gains. We find that they commonly finance foreseen,discretionary investments through debt rather than savings—taking interest-bearingloans from moneylending companies rather than building up their savings andthen dissaving cheaply. In fact, we find evidence of “coholding,” that is, they holdborrowings and liquid savings at the same time. We will argue that this imposes asignificant financial cost and yet offers no financial benefit: debt contracts are notdesigned to transfer project risk to the lender, interest rates are nonnegligible, and,in fact, loan default imposes a heavy cost, with the real risk of losing their job andcutting off future earnings potential.

To identify and examine this “borrowing puzzle,” we draw on data collectedfrom different sources. In 2017, we interviewed a sample of 136 Filipino domesticworkers and asked about their employment history, wage income, remittances,savings, and loans. Subjects also participated in a lab-in-the-field experiment wherethey allocated a given endowment across a set of options with differing risks andreturns. We also rely on the records of a credit cooperative that caters to migrantworkers.

We document the following facts. First, as expected, most Filipino domesticworkers remit to their home country regularly. These remittances often support notjust their immediate nuclear family but also educational and health expenses fortheir extended family as well. Thus, these migrants take on the responsibility ofsupporting several individuals back home.

Second, although the majority of migrants have bank accounts, they do notappear to use them as a savings device. Bank balances tend to be low and monthlyinflows into the accounts are small. However, this is not to say that their entiremonthly salary is consumed. Anecdotal evidence suggests that many of them invest

The Borrowing Puzzle: Filipino Domestic Workers in Hong Kong, China 79

in “projects” in the Philippines, such as land purchase, house construction, houserenovation and repair, and small businesses.

Third, it is common for these workers to borrow from moneylendingcompanies in Hong Kong, China. On average, these companies charge an interestrate of 25% per annum. Migrants repay these loans from the salaries they earn.Our data suggest that only a small fraction of these loans are used for unforeseenemergency expenses; the majority are remitted home for school fees, consumptionneeds, or investment.

This leads us to the central observation in this paper: Filipino domesticworkers appear to routinely finance their investments through loans rather thansavings. Moneylending companies have standard contracts for loans to overseasworkers, where repayment begins the very next month after the loan is disbursed.Workers’ investments generally do not start generating immediate returns, sorepayment is usually financed from workers’ wages. These wages are contracted andregular and therefore predictable. Default carries heavy penalties; thus, borrowingdoes not transfer risk to the lender.

Our data allow us to examine the plausibility of different explanations forwhy many migrants would rather finance investments through borrowing thandissaving. Although we cannot conclusively accept or reject a particular hypothesis,we discuss possibilities for future research that could shed light on this issue.

The paper is organized as follows. In the next section, we describe theempirical background against which this paper is set. In section III, we describehow we collected our data, and in section IV we use these data to present a pictureof how these domestic workers manage their finances. Sections V and VI discussexplanations for their high indebtedness and overborrowing. Section VII presentsour proposed explanation. In conclusion, section VIII discusses avenues for futureresearch and some broad implications for policy and practice.

II. The Context

Migrant domestic workers made up 9.3% of the workforce in Hong Kong,China in 2016. More than half of these workers were from the Philippines(Government of the Hong Kong SAR 2017). They performed a range of services fortheir employers, including cleaning, cooking, shopping for groceries, babysitting,ferrying children to and from school and extra-curricular activities, and caring forthe employers’ aged parents and pets. Their services facilitated the labor forceparticipation of working-age women in Hong Kong, China especially mothers withyoung children (Cortes and Pan 2013).1

1The benefits from employing foreign domestic workers likely extend beyond the effect on female laborsupply. Tan and Gibson (2013) argue that domestic workers do not increase female labor force participation in


The Philippines was one of the first countries to send workers through theforeign domestic helper program that began in the 1970s. This program grantsmigrants a special “foreign domestic helper” visa, which entitles them to work fora single employer. Employers are required to pay at least a “minimum allowablewage.” In 2020, this minimum wage was HK$4,630. The minimum wage is usuallyrevised once a year to adjust for changes in the cost of living.2 Foreign domestichelpers cannot qualify for permanent residence in Hong Kong, China. They cancontinue to reside there as long as they are gainfully employed as domestic workers,but the employment contract and visa must be renewed every 2 years.3 Thus, theseindividuals are temporary economic migrants: they live in Hong Kong, China onlyfor as long as they can be gainfully legally employed. They are generally aware thatemployers become less likely to employ them as they get older and that they willretire in their home country.

These migrants’ wages are lower than most of the population in Hong Kong,China.4 They do not qualify for pensions or other financial benefits. Commercialbanks usually target higher-income groups; as a result, foreign domestic workershave only limited access to formal banking services. However, the local laws putno restrictions on whether and how much migrants can borrow from moneylendingcompanies.5 This creates a dichotomy where they have only limited access to formalsavings accounts but extensive access to formal loans. This paper examines howthey manage their finances against this backdrop.

III. Data Collection

In 2017, we enrolled 141 Filipino domestic workers to participate in oursurvey and lab-in-the-field experiment. Of these, we have survey data from the 136who successfully completed both parts of our interview.6 Below we describe theprocess by which this sample was created.

Malaysia, but they speculate that Malaysian employers can enjoy increased leisure and can specialize in child-rearing,with possible gains in their children’s human capital.

2Note that by law, the worker must live in the residence of the employer, and so she is not expected to incurany housing costs.

3If they become unemployed, migrant domestic workers must leave within 2 weeks and may only reenter afteran employer has signed a new 2-year contract. Note that when employers fire a domestic worker, they are required topay for the worker’s travel back to her home country.

4There are no regulations about or definitive data on the number of hours that migrants work per day.However, taking a conservative estimate of 10 hours per day for 6 days a week, the monthly minimum allowablewage for domestic helpers in 2019 translates to an hourly rate just below HK$20 or roughly 50% of the minimumwage in Hong Kong, China. Domestic helpers do not fall under the purview of the minimum wage ordinance.

5In contrast, in 2019, the Government of Singapore placed a limit on how much individuals can borrow fromSingaporean moneylenders. Some have even argued that domestic workers’ loan applications should be preapprovedby their employers (Ng and Tan 2019).

6All 141 participated in the first face-to-face interview, but only 136 could be contacted 4–8 weeks later fora follow-up phone interview to report financial transactions that had occurred since the first interview and answeradditional questions that helped to compute loan interest rates.


Migrant domestic workers in Hong Kong, China are required by law to live intheir employer’s house. They usually work 6 days per week. Their hours of work arenot specified in their contract, although the law states they are allowed one rest dayper week. This rest day is generally spent outside their employer’s residence. Thismakes it very difficult for investigators to survey them in their residence. Given thelength of our survey and experimental sessions, we believed it would be difficultto enroll participants by approaching them on the street or in public places.7 Wetherefore advertised our study through WhatsApp and Facebook with certain nodalFilipino domestic workers and asked them to pass the advertisement on.8 Interestedpersons could click on a web link and answer a short enrollment questionnaire. Wethen contacted these enrollees and reserved a study session slot for them.9 In thisway, we created a respondent-driven sample.

A respondent-driven sample may not be representative of the underlyingpopulation of interest. The nodes that we began with were not randomly chosen, andif we had only relied on the nodes to spread the word, we might have only reachedtheir friends. We therefore attempted to create a “snowball” by offering participantsa bonus for each person we recruited through their referral. This created an incentivefor every participant to spread the word to her friends. To avoid swamping thesample with acquaintances of the more popular participants, we offered this bonusfor only four referrals and no more. However, snowball samples only approximaterandom samples in the limit, and our sample of 136 respondents is unlikely to besufficiently large (McKenzie and Mistiaen 2009). We therefore present reweighteddescriptive statistics that better approximate the true population of Filipino domesticworkers living in Hong Kong, China.10

We also draw on the records of the Asian Migrants Credit Union, a creditcooperative in Hong Kong, China that primarily targets migrant domestic workers.All individuals who join the cooperative are provided a savings account. Six monthsafter they enroll, members become eligible to use the credit facility. First-timeborrowers may only borrow up to two times their savings. The entire loan iscollateralized by the savings balance of the borrowing member as well as the savingsof the guarantors, who are other members of the cooperative. The interest rate is setat 1% per month and repayment is on a monthly schedule. We look at the contractsof all loans that were approved by the cooperative in 2017–2018 to examine the

7Relatedly, Barua, Shastry, and Yang (2019) found that domestic workers recruited in public places inSingapore were unlikely to sustain participation in their study.

8These nodal persons were elected officers of the credit cooperative that we also use data from.9Nearly all migrant domestic workers in Hong Kong, China use smartphones, and a very large fraction use

social media, so it was fairly easy for them to view and answer our enrollment questionnaire. Once we received theironline submission, we called them to explain the details of the study session and register them into a time slot.

10We draw the weights from the distribution of Filipino domestic workers’ age, education level, and length ofstay in the 5% microsample of the Hong Kong 2016 Population By-census. Appendix 3 provides further informationabout this microsample and how we use it.


stated purposes of these loans. We also analyze members’ saving and borrowingbalances during the period 2011–2018.11

Finally, we draw on the findings of a lab-in-the-field experiment with all 141subjects who participated in our study. The goal of the experiment was to examinehow participants respond to the rate of return when making savings choices. Allsubjects were given an endowment of 100 tokens (1 token was equivalent to HK$1),and in each round they were asked to allocate these tokens across three accounts:a savings account that generated a sure return, an investment account that wouldgenerate a return of 10% but with uncertainty, and a lottery account where eachtoken would give them a chance of winning a handbag as a prize. Subjects wererandomly assigned to groups and played multiple rounds within each group. Furtherdetails about this study are provided in Appendix 1.

IV. Some Facts

Table 1 presents reweighted descriptive statistics for the respondents in oursurvey. The average Filipino domestic worker was 36.5 years old. She had left thePhilippines for work about 6.5 years prior to our study and had been living in HongKong, China for nearly 5 of those years, three of them with her current employer.Given their profession, domestic workers reported relatively high education levels.A third of the workers had studied further after high school, and a fifth had acquiredan additional academic qualification after their high school diploma.12

As we mentioned before, migrant domestic workers must work for only oneemployer. Contracts are signed for a 2-year duration. Our study took place betweenJanuary and May 2017, thus the workers we interviewed would have signed theircurrent contracts any time in or after January 2015. The minimum allowable wagewas set at HK$4,310 for contracts signed between October 2016 and September2017 and HK$4,210 for contracts signed between October 2015 and September2016. The median worker in our sample received exactly HK$4,210, but the meanwas slightly lower at HK$4,150. Overall, the evidence suggests that employerscomply with the minimum allowable wage regulation.13 More than 80% of workersreceived their wages as cash, which is indicative of the limited use of formal bankingor other financial services.

11Although we have the credit union’s transaction records for the period 2011–2018, we only have access toloan applications during 2017–2018.

12This includes associate degrees, vocational training and professional courses, as well as university degrees.13Of course, it is possible that domestic workers who received lower than the minimum allowable wage did

not participate in our study. However, since the contract is a formal document that is approved by the ImmigrationDepartment, it cannot state a wage lower than the minimum. Most observers believe that domestic workers generallydo receive the contracted wage.


Table 1. Descriptive Statistics of Survey Respondents

Mean Standard Error(1) (2)

Age 36.5 0.09Education

Less than high school 0.49 0.01High school 0.08 0.00Some post-high school 0.24 0.01Post-high school completed 0.19 0.01

Number of children 2.33 0.02Number of dependents 3.72 0.03Years since leaving the Philippines 6.59 0.09Years in Hong Kong, China 4.84 0.08Years with current employer 3.26 0.06Mean salary (HKD per month) 4,150.3 16.6Paid in cash 0.81 0.01Food provided 0.90 0.00Food allowance provided 0.10 0.00Remitted in past 2 months 0.88 0.00Mean remittances (HKD per month) 2,163.5 20.2Fraction of income remitted 0.52 0.00Remittance method

Bank 0.09 0.00Money service operator 0.64 0.01Online 0.01 0.00Other 0.17 0.01

Has bank account 0.83 0.01HKD account 0.32 0.01PHP account 0.91 0.00Has single-holder account 0.97 0.00Has joint account 0.06 0.00Total bank balance (HKD) 5,839.4 206.5Mean savings per month (HKD) −39.8 18.7

ROSCA member 0.11 0.00Uses a money guard 0.00 0.00Has outstanding debt 0.37 0.01If yes:

Total amount borrowed (HKD) 21,445.8 296.4Total repayment amount (HKD) 27,200.3 406.7Monthly repayment amount as fraction of salary 0.55 0.01

Has outstanding credit 0.14 0.03

HKD = Hong Kong dollar, PHP = Philippine peso, ROSCA = Rotating Savings and CreditAssociation.Note: The survey sample is reweighted to match the distribution of Filipino domestic workersin the population of Hong Kong, China, as estimated in the 5% microsample of the Hong Kong2016 Population By-census.Source: Authors’ calculation.

As is to be expected, workers remitted a large part of their salaries to theirhouseholds in the Philippines. Eighty-eight percent of the sample had remittedmoney home within the 2 months prior to the survey. On average, they remittedHK$2,164 or 52% of their monthly salary. These remittances supported on average


3.7 individuals, which included not just their immediate family but their extendedfamily as well.14

A. Savings

Strikingly, 83% of our sample had active bank accounts at the timeof the survey. Of these, 91% of workers had accounts in the Philippines,whereas a much lower 32% had accounts in Hong Kong, China. Nearly allbank accounts were single-holder accounts; only 6% of account holders reportedhaving joint accounts. However, bank balances were low. Across both locations,the average respondent held only about HK$5,840 or 1.4 months’ salary inthe bank. Net inflows were actually negative during the 2 months prior to theinterview. Other saving devices were not very common. Only 11% reportedmembership in a rotating savings and credit association (ROSCA), where theymade an average monthly contribution of HK$340. Nobody reported using amoney-guarding arrangement.15

Migrant workers may have held small bank balances because the accountsoffered low rates of return.16 To examine whether their savings balances wouldrespond to interest rates, our experiment randomly assigned participants to a safe“savings product” with one of two rates of return: a low 3% rate or a high 10%rate. Strikingly, we find no evidence that participants assigned to the high-returncondition allocated more tokens into the safe account. Respondents in thelow-return condition placed HK$53.3 out of HK$100 worth of tokens into the safeaccount, and those in the high-return condition placed a nearly identical HK$51.7(difference = 1.6, p = 0.68). The rates of return on savings also did not affectallocation to the investment and lottery accounts.17

To understand whether this behavior can be explained by migrantcharacteristics, we examine in Table 2 whether respondents with differentcharacteristics respond differently to a change in the rate of return. Our data consistof 324 person-round-level observations across the 141 respondents who participatedin the experiment. Our regressions include dummy variables for the round in whichthe allocation was made. This controls for round-specific effects or learning overtime. In column (1), we include as an explanatory variable a measure of therespondent’s risk aversion.18 As expected, we find that more risk-averse respondents

14Fifty-six percent were supporting their parents, and 34% were supporting other dependents, such asgrandchildren, siblings, nieces and nephews, cousins, or grandparents.

15A money guard is a person who holds money for the subject to help her avoid spending or losing it (Collinset al. 2009).

16Hong Kong, China’s savings interest rates are nearly 0%. At 0.1%, interest rates in the Philippines are onlyslightly higher. Inflation rates in 2017 were 1.48% in Hong Kong, China and 2.85% in the Philippines.

17Investment accounts: HK$32.2 in the low savings return condition vs. HK$30.5 in the high return condition(difference = 1.7, p = 0.64); lottery accounts: HK$23.3 vs. HK$24.4 (difference = 1.1, p = 0.73).

18Risk preferences were elicited using an incentivized Lowry list method where participants were asked tochoose between a safe option and a lottery with a high and low payout, where the probability of a high payout


Table 2. Heterogeneous Treatment Effects of a High Rate of Return on Token Allocation

Risk Aversion Financial Literacy

(1) (2) (3) (4) (5) (6) (7) (8)

Characteristic 1.50*** 1.24** 1.24** 0.14 0.72 −0.20 0.03 −3.81(0.51) (0.52) (0.52) (0.84) (2.05) (2.09) (2.11) (3.31)

High interest rate −2.62 −21.63* −2.54 −17.57*

(3.38) (11.98) (3.45) (10.60)Characteristic × 1.77* 6.34

High interest rate (1.07) (4.23)Age 0.19 0.17 0.22 0.25 0.24 0.23

(0.27) (0.27) (0.27) (0.27) (0.27) (0.27)Length of stay −0.04 −0.02 −0.10 −0.04 −0.03 −0.01

(0.28) (0.28) (0.28) (0.28) (0.28) (0.28)Education

High school −3.26 −3.29 −4.40 −4.02 −3.98 −4.08(5.95) (5.96) (5.98) (6.04) (6.05) (6.04)

Some post-high 6.76 6.75 6.10 6.61 6.63 7.56school (4.85) (4.86) (4.86) (4.90) (4.91) (4.94)

Completed 10.07** 9.95** 9.55** 10.70** 10.56** 11.30**

post-high school (4.81) (4.82) (4.81) (4.85) (4.85) (4.87)Constant 35.18*** 24.27** 26.37** 37.71*** 49.71*** 35.69*** 37.11*** 45.41***

(5.89) (11.09) (11.43) (13.30) (5.36) (11.87) (12.04) (13.23)R-squared 0.03 0.07 0.07 0.08 0.01 0.05 0.05 0.06No. of observations 324 324 324 324 324 324 324 324

Notes: The dependent variable is the number of tokens (out of 100) that the respondent allocated to the safe accountin a round. Respondents were randomly assigned to a group of 4 or 5 members (say n members) and then playedn rounds with that group before being reassigned to a new group. The data contain observations of the first roundthey played with each group they were assigned to. We include round dummy variables to control for round-specificeffects. Standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.Source: Authors’ calculation.

placed a larger number of tokens in the safe return box. In column (2), we includecontrols for age, education, and length of stay in Hong Kong, China; this does notchange the coefficient on risk aversion significantly. In column (3), we add a dummyvariable for whether the respondent faced a 10% return on savings. Controlling forrisk aversion, we do not find that respondents who faced a higher rate of returnplaced more tokens than comparable respondents with a lower rate of return. Finally,in column (4), we interact the risk aversion measure with the dummy variable forthe high-return treatment. There is no evidence that more risk-averse individualsresponded differently to the rate of return than the less risk averse. In columns(5)–(8), we consider heterogeneous effects by the respondent’s financial literacy.19

Again, there is no evidence that financial literacy levels affected how participantsresponded to the rate of return. To the extent that these results can be translated into

successively increased. In line with Yu, Zhang, and Zuo (2019), respondents were encouraged to choose a singleswitching point from safe option to lottery.

19In section V.C, we describe how we measured financial literacy.


Figure 1. Relationship between Average Monthly Savings per Member in the CreditCooperative and Previous Year’s Dividend Rate

Source: Authors’ calculation.

their behavior in daily life, it does not appear that migrants’ disinterest in saving isdriven by low rates of return.20

Note also that the credit cooperative paid considerably higher dividends(1%–3% per annum) than the interest rate of commercial banks in Hong Kong,China during this period. Despite this, we find a low savings rate among membersof the credit cooperative. The average member made a net deposit of only HK$44per month into her account. There is also no evidence that the members’ savingsrates responded to dividends. In Figure 1, we plot the monthly net deposits permember against the dividend rate that the credit union paid in the previous year.21

There is no indication that members saved more per month when dividends werehigher.

B. Credit

Loans allow individuals with small cash inflows to consume or invest in thepresent, instead of having to postpone or forgo these activities. They can also helpsmooth consumption in the face of negative shocks. Below we examine the natureof borrowing by our survey respondents.

20As we see across the table, respondents with more education placed more tokens in the safe return box.However, there is no evidence that they increased their safe token allocation when they faced a higher safe return(results available upon request).

21Each year’s dividends are announced at the end of the year and depend on the credit union’s profits duringthe year. Arguably, members could not have known the dividend when they made the savings decision, but they couldhave used the previous year’s dividends as a predictor.


Table 3. Loan Characteristics

Credit Friend orOverall Moneylender Cooperative Employer Relative

(1) (2) (3) (4) (5)

Fraction of loans 1.00 0.88 0.03 0.03 0.04Principal (HKD) 20,453.5 22,176.2 13,153.9 15,339.0 1,480.8

(272.8) (279.5) (182.44) (943.0) (159.98)Repayment amount (HKD) 31,619.9 34,577.6 19,426.5 17,322.0 1,480.8

(439.6) (449.2) (423.9) (1,239.9) (160.0)Interest rate (%) 23 26 7 − −

(0.00) (0.00) (0.00)With monthly repayment 0.98 1.00 1.00 1.00 0.53

(0.00) (0.00) (0.00) (0.00) (0.06)Loan duration (months) 9.8 11.0 5.1 − −

(0.12) (0.10) (0.20)

− = data are not specified as these are flexible duration loans, HKD = Hong Kong dollar.Notes: The survey sample is reweighted to match the distribution of Filipino domestic workers in the population ofHong Kong, China, as estimated in the 5% microsample of the Hong Kong 2016 Population By-census. The annualinflation rate in Hong Kong, China in 2017 was 1.48%. Standard errors in parentheses.Source: Authors’ calculation.

In Table 3, our summary statistics once again have been reweighted to matchthe distribution of Filipino domestic workers in the Hong Kong 2016 PopulationBy-census (Government of the Hong Kong SAR, Census and Statistics Department2017). Forty-six respondents (or 37% of our reweighted sample) reported having anoutstanding loan from a lender in Hong Kong, China at the time of the survey.

When we computed their monthly repayment obligation we found that theyhad committed to paying on average 55% of their salary in loan installments eachmonth. Table 3 also shows that 88% of the loans were taken from moneylendingcompanies. The cooperative gave out only 3% of the loans. The other sources wereinformal: either informal borrowing from their employers or loans from friendsor relatives in Hong Kong, China. Loans from moneylending companies were thelargest of all: the average principal amount was HK$22,176. They had an 11-monthduration on average, the interest rate was 26% per annum, and payment was due ona monthly basis.

Employers gave zero-interest loans. The size of an average loan given byan employer was HK$15,339, equivalent to just over 3 months’ salary. Employersalso took payments on a monthly basis, usually by garnishing the worker’s salary.Loans from the credit cooperative were similar in size at HK$13,154 on average.At 7% per annum, the cooperative charged less than one-third the interest rate ofthe moneylending companies.22 Loans from friends and relatives were significantlysmaller, and although many subjects expected to repay on a monthly basis, therepayment schedules were more likely to be flexible.

22The cooperative capped the loan principal at two times the borrowing member’s savings balance in thecooperative.


In keeping with our finding that respondents had received loans from friendsand relatives in Hong Kong, China, 14% of our survey respondents told us thateither friends or relatives residing there owed them money at the current time.Of these 27 loans, five represented a sharing arrangement where the respondenthad taken a loan from a moneylending company and then shared it with anotherdomestic worker.23 Another three loans were given out by a single respondent tothree different friends at 10% interest over a 6-month duration or 1.67% per month.

Thus, the survey data suggest that it is common for Filipino domestic workersto borrow. In fact, records of the credit cooperative suggest that migrants need notborrow as much as they do. We present evidence below that a large proportion ofcooperative members “cohold” loans at the same time as they hold liquid savingsthat could be drawn down instead.

Since the interest cost on these loans (1% per month) is considerably higherthan the return on savings (the cooperative’s dividend rates range from 1% to 3%per year), it is in the members’ interest to take the smallest loan necessary to financetheir needs. However, in 17.5% of the 200 loans that the cooperative extendedover the period 2011–2017, the member had a larger amount of savings in thecooperative than the amount she borrowed. Clearly it would have been cheaper toinstead withdraw these savings and avoid the loan altogether. Instead, by taking theloan and securing it with part of her savings, she not only took on an additionalinterest expense, but also rendered part of her savings illiquid.24

Moreover, even when the member’s savings were smaller than her loanamount, the evidence suggests she could have borrowed less than she did.Specifically, she could have financed part of her need by withdrawing her savings,thereby reducing the loan size and interest cost. To see this, note that the cooperativerequires that one-half of the loan is secured by the member’s savings; this amountcannot be withdrawn until the loan is repaid. Call this her illiquid savings, i. Theremaining savings is liquid, which we denote by l. Thus, total savings s = i + l. If theexpense is e then we know that e − l = 2i. We can then calculate her illiquid savingsas i = e − s, so that the remaining l = 2s − e can be withdrawn. By withdrawing theentire l, a member would take the smallest loan necessary to finance the expense.For example, a member who needs to finance an expense of $1,000 and has $700 insavings would minimize costs by maintaining a savings balance of $1,000 − $700= $300 to take a loan of $600 and withdrawing the remaining $400.

23This is an informal arrangement where the individual who is formally listed as the borrower “shares” someof the loan principal with a friend. Often this friend is either the reference person or guarantor for this loan. Bothfriends then pool their money to make the monthly payments. In case of default, the moneylender first contactsthe borrower and then contacts the reference person or guarantor. They may also contact the employer of eitheror both domestic workers and demand payment from them. It is possible that this informally created joint liabilityimproves loan repayment for both the borrower and her friend. Moneylenders offer “VIP status” to borrowers withgood repayment records. VIPs earn in-kind rewards and rebates on their own loan payments in return for referringnew borrowers to the moneylending company.

24The cooperative limits a member’s loan amount to twice her total savings balance at the time of the loanapplication. This savings balance secures the loan and cannot be withdrawn while the loan is outstanding.


Instead, we find that in 62.5% of loans from the credit cooperative, the loanprincipal was smaller than two times the member’s savings balance at the time.25

The data thus suggest that it is common for Filipino domestic workers to financetheir investments and perhaps even their families’ consumption expenses throughloans rather than savings. In what follows, we consider different explanations forwhy they may do so.26

V. Explanations for Indebtedness

We start by examining some common explanations for the high incidence ofdebt.

A. Debt Due to Migration Costs

It is widely reported in Hong Kong, China that domestic workers bear alarge financial cost to get a job placement. This appears to apply both to workerslocated in the Philippines looking to migrate as well as workers located in HongKong, China who are in between employers. The Progressive Labor Union ofDomestic Workers in Hong Kong and the Hong Kong Federation of Asian DomesticWorkers (PLU and FADWU 2016) report on an investigation where researchersmade anonymous phone calls pretending to be domestic workers in search ofemployment. They found that most employment agencies charge workers a sizableillegal fee for the placement service. If the worker is unable to pay the placementfee upfront, the employment agency often refers her to a lending company. Jobapplicants can also take loans in the Philippines and repay them from Hong Kong,China. The report estimates that it takes 6 months to pay off the average loan.

Thus, migrants might be arriving in Hong Kong, China already in debt, andthey may continue to be indebted for a significant duration of their first contract. Ifthey incur a placement fee again when they switch employers, then they may need totake another loan and could be indebted for part of their first contract with the newemployer. If they face any large unexpected consumption or investment expenseduring this period, they may need to take another loan, possibly setting them on apath of repeated indebtedness.27

25Our methodology and findings are similar to those used by Laureti (2018) in her analysis of the clients ofSafeSave, which provides flexible savings-and-loans accounts to slum dwellers in Dhaka, Bangladesh.

26When a cooperative member overborrows, not only does she take a larger loan than necessary but alsosecures a larger fraction of the loan with her own savings, thereby relying on a guarantor to secure a smaller fraction.If she instead withdrew her savings and took a smaller loan, she would still need a guarantor to secure the same dollaramount. Thus, overborrowing does not reduce dependence on a guarantor.

27Recently, three major moneylending companies appear to have started sharing information about theirclients’ loan records; in informal conversations, domestic workers report that they can no longer take multiple loansfrom different moneylenders.


If indebtedness is explained by the placement fee expense, then we should seegreater indebtedness among migrants who are in their first contracts than migrantswho have been with their current employer for a longer time.28 In fact, only 29%of our sample who were in their first contract currently had a loan in Hong Kong,China versus 37% of those who had been with their current employer for a longertime (the difference is statistically nonsignificant).29

B. Unexpected Expenses

We have referred above to the possibility that negative shocks may inducemigrants to borrow to smooth consumption. If migrants are using the bulk of theirincomes to support their families’ regular expenses, then even a migrant who savesregularly may simply not have enough saved to cope with a shock. However, datafrom the credit cooperative suggest that this cannot be a complete explanation.When we analyzed the stated purpose of the 40 loans that migrant domestic workerstook from the credit cooperative during the years 2017–2018, we found that nearlytwo-thirds were for expenses that could have been anticipated: land purchase, homerenovation, or school fees for children back home. Only 21% of the loans werefor medical expenses of relatives.30 Of course, there is the question about whetherwe can trust the stated purpose of the loan. However, the credit cooperative hasan informal policy of providing faster customer service for emergency loans, andso it seems unlikely that borrowers underreport the true incidence of emergencies.Rather, it appears that the bulk of the loans are not being used to smooth shocks.

C. Lack of Financial Knowledge

Domestic workers may not understand the financial costs of borrowing. Inother words, they may not realize that they can lower their financial costs by usingtheir savings instead of borrowing. To assess whether this can explain the observedbehavior, we examine whether indebtedness varies by financial literacy levels. Ourmeasure of financial literacy comes from two questions we asked in our survey.In each question, the respondent was presented with two alternative hypothetical

28Recontracting with the same employer is a relatively easy process and usually does not involve anemployment agency.

29The low incidence of loans among those in their first contract may seem puzzling. Noting that it is illegalto charge placement fees to workers, it is possible that more experienced workers are more aware of this rule or canfind employment more easily through word of mouth or other means, rather than using an agency. We therefore testbut reject the hypothesis that new arrivals are more likely to be in debt. This could be because their loans were takenin the Philippines and therefore not reported as local loans. In any case, this does not suggest that migration-relatedcosts are causing the indebtedness we observe.

30The rest could not be clearly classified into emergency or nonemergency purposes. For example, houserepair could be an urgent expense in response to sudden damage, or it could be a nonurgent expense that was plannedahead.


loans and asked to determine which loan was cheaper.31 The respondent’s financialliteracy score takes a value of 0, 1, or 2 depending on whether she correctlyidentified the cheaper loan in none, one, or both of the questions, respectively.

Our data do not suggest that respondents are generally unable to evaluate thecost of a loan. Fifty-one percent of respondents answered both questions correctly,and 40% answered one correctly. However, 37% of those who answered bothquestions correctly reported having an outstanding loan, compared to 30% of thosewho answered none or one question correctly. This difference is not statisticallysignificant. Thus, it does not appear that their behavior stems from an inability tocompute financial costs.

D. Lack of Self-Control or “Other Control”

Hong Kong, China is a consumerist society, and shopping opportunities areeverywhere. It could be argued that this creates the temptation for Filipino migrantsto purchase goods that may not be strictly necessary. Excessive consumption ofthese goods could prevent them from building up their savings, so they might needto borrow to finance large expenses.

Alternatively, migrants could lack complete property rights over theirearnings. In other words, they could be remitting larger sums than they had plannedto or purchasing items that they did not plan to, not because they lack self-controlbut because others in their social network make demands on their incomes. Forexample, their families back home may demand gifts or ask for larger remittances.Similarly, if they have surplus cash, their local friends may request loans or treatsand these may not be repaid or reciprocated.

Either of these two mechanisms could lead to low savings, making itnecessary to borrow to finance large expenses. Although our current data do notallow us to validate these explanations, we will discuss these mechanisms further insection VII.

VI. Explanations for Overborrowing

As we discussed above, the puzzle is not only that savings tend to be low onaverage but that individuals often choose to borrow, instead of withdrawing theirsavings. We now discuss some explanations for this behavior.

31In the first question, the two loans had an identical duration but differed in the principal and the interest.Thus, respondents would have had to work out which loan was cheaper per dollar of principal. In the second question,the two loans had identical principal and duration but the loan installment size and installment frequency varied.Thus, they would have had to work out which loan required the larger repayment amount. The exact questions arereproduced in Appendix 2. For each pair of loans we asked them two questions: which loan was cheaper and whichloan they would prefer to take.


A. Limited Liability

Recall that the stated purpose for most credit cooperative loans was oftena productive investment such as an educational expense, property purchase orconstruction, or a business investment. These are potentially risky investments. Ifloan contracts offer borrowers limited liability, then by financing the investmentthrough a loan, the migrant worker transfers the downside risk to the lender andprotects her savings.

The fact is, however, that neither moneylending companies nor the creditcooperative offers limited liability. Most loans had rigid repayment schedules andsignificant additional costs in case of default. For example, it is common formoneylending companies to call the borrower, her guarantor or reference person,and/or their employers over the phone to demand payment for an unpaid installment.Employers who receive these phone calls may fire the domestic worker, thus cuttingoff her income. The credit cooperative also does not limit the borrower’s liability;in fact, its loans are completely secured. It recovers unpaid loans by seizing theborrower’s and/or her guarantor’s savings. Short of quitting the job, a worker cannotdefault on a loan from her employer since her payments are deducted from hersalary. If she did quit before she had paid back her loan, she would likely find itimpossible to find alternative employment in Hong Kong, China.32 It thus seemsimplausible that migrant domestic workers borrow in order to avoid the downsiderisk of their investment projects.

B. Short Time Horizon

Although repayment is enforced strongly within Hong Kong, China,moneylenders may be unable to recover their loans once the worker leaves. If amigrant worker is uncertain about how much longer she will stay, she may anticipatenot paying back, effectively lowering the cost of the loan to her.33

Migrants who have been reemployed by the same employer multiple timesmay be more secure about their job. If workers with lower expected tenure intheir job face lower debt costs, then we should find greater indebtedness amongmigrant workers with a shorter tenure in their jobs. Instead, Figure 2 suggests thatindebtedness rises as the number of years with the current employer increases. Thisis likely connected to the fact that lenders reward workers who have a more secure

32Individuals who have breached their previous employment contract are unlikely to be granted a visa for anew contract (Government of the Hong Kong SAR 2020).

33To our knowledge, there is no mechanism to prevent migrants from leaving Hong Kong, China while theyare in debt. Moneylenders are aware of this risk and partly mitigate this by scheduling loans to mature before theworker’s current employment contract ends. Thus, to avoid repaying the loan by leaving the city, the borrower wouldeither have to run away or her contract would have to be terminated prematurely. Both employers and workers havethe right to terminate the contract at any time, with 1 month’s notice or 1 month’s payment in lieu of notice. Employerswho terminate the contract are required to pay for the worker’s travel back to her home.


Figure 2. Indebtedness as a Function of Job Tenure

Source: Authors’ calculation.

job by offering them better loan terms. However, it does not match the pattern wewould expect if workers’ indebtedness were solely caused by the insecurity of theirjobs.

VII. Loans as a Device to Solve Control Problems

We have shown that individuals take loans even when they have sufficientsavings. They then repay these loans from the monthly salaries they earn asdomestic workers. This raises the following question. If they had dissaved instead,then the same monthly salary could have been used to rebuild the depleted savings.One reason individuals may not dissave is that it is difficult to rebuild savings.Indeed, in informal interviews, Filipino domestic workers agree that it would bebetter to save than to borrow. However, they report it is difficult to save, becausethere is always a reason to spend the money instead. This self-reported inabilityto save has also been documented in several other contexts. For example, it hasprovided a rationale for the success of simple savings technologies in Kenya (Dupasand Robinson 2013). However, as we have shown above, nearly all our subjects haveat least one savings account; access to savings products does not seem to facilitatetheir saving.

This, in turn, suggests that loans may serve an additional purpose. Themajority of the loans that we found in our survey data have strict repaymentschedules, and so borrowers are committed to pay monthly installments until theloan is paid off. Since default carries large penalties, this could create a crediblerationale for avoiding other expenses. The prospect of default could help resist


the temptation to buy oneself an unnecessary consumption good or the pressureto purchase such items for one’s family or to gift or treat one’s friends. Migrantswho are sophisticated about their lack of control may then actively choose to takea loan even if it is not financially necessary. Indeed, Baland, Guirkinger, and Mali(2011) argue that members of a credit cooperative in Cameroon borrow more thanthey need to so that they can “pretend to be poor.”

ROSCAs also help overcome the difficulty of saving. Gugerty (2007) reportsthat ROSCA members in western Kenya believe that the collective element givesthem the “strength to save.” In our context, ROSCAs are not very common, probablybecause ROSCAs rely on mutual “trust,” which is more likely to develop whenmembers can monitor each other and enforce promises. These conditions areunlikely to develop organically in a population of transient urban migrants whoonly meet once a week in a public location.34 Even among ROSCA members, theROSCA does not replace loans altogether, most likely because the ROSCA pot islimited by the savings capacity of its members.35

In contrast, moneylending companies offer much larger loans. Domesticworkers have easy access to these loans: to apply, they only need to show theirHong Kong Identity Card and employment contract and provide the phone numberof a reference person or bring along a friend as a guarantor.36 The high interestrates and strict repayment schedules effectively limit future liquidity and flexibilityto smooth consumption shocks and, in extreme situations, can cause the workerto lose her income. It is possible that these features actually make these loansattractive. Morduch (2010) discusses the case of a South Indian woman who tooka high-interest loan that she could have avoided. She believed the high interest rateincentivized her to pay back the loan much more quickly than she could have savingup the same amount.

VIII. Conclusion

Our research has benefited from a large literature that precedes it. Manyscholars have noted that the poor do not save as much as they could (Banerjee andDuflo 2011). However, others have also argued that borrowing remains an attractivechoice for many poor individuals, even when they have the wherewithal to save(Collins et al. 2009, Morduch 2010). This is because the high interest costs or

34Recall that our survey suggested that 11% of Filipino migrant domestic workers belong to ROSCAs. Themajority of the ROSCA members in our survey belonged to the same island in the Philippines and knew each otherwell. However, 2 years after our survey, the ROSCA manager embezzled the pot; as of 2020, the members are stillwaiting to get their savings back.

35Thirty percent of ROSCA members reported they had an outstanding loan from a moneylender. Their meanloan size (HK$25,500) was also similar to the mean loan size of those who did not belong to ROSCAs (HK$23,149).

36Migrant domestic workers who either have a good repayment record or who have been with their currentemployer for longer than 5 years do not even need a reference person or a guarantor.


the penalties for nonrepayment induce the borrower to repay the loan and makeit possible to avoid consumption in a way that voluntary savings mechanisms donot.

An important question in this context is: What are the compulsions thatprevent Filipino domestic workers in Hong Kong, China from saving successfullybut at the same time allow them to repay loans regularly? A possible explanationis the lack of self-control in the face of consumption opportunities. Certainly, thereare abundant shopping opportunities that might test an individual’s self-control, andthere are anecdotes about domestic workers who splurge on consumption goods thatmight seem excessively expensive given their low wages.

A second explanation points to the role of “kin taxes” or insecure propertyrights over one’s earnings and savings. Many of the workers we surveyed wereearning considerably higher wages than their kin in the Philippines and wereremitting money regularly to support their expenses. Anderson and Baland (2011)argue that women participate in ROSCAs in order to wrest control from theirhusbands who might spend on unnecessary items. Ashraf et al. (2015) find thatEl Salvadoran migrants in the United States deposited more into savings accountswhen they had sole control over withdrawals than when they shared control withtheir relatives back home. In a lab-in-the-field experiment, Jakiela and Ozier (2016)find that Kenyan village residents are more unwilling to publicly reveal theirearnings to a room full of fellow residents when a larger proportion of the fellowresidents are their kin. Baland, Guirkinger, and Mali (2011) argue that Camerooniancredit cooperative members take loans instead of withdrawing their savings so thatthey can “pretend to be poor” and avoid gifting or contributing to their friends andrelatives. The Kenyan savers studied by Dupas and Robinson (2013) also say theirsavings boxes help them hide their money from their social network.

In informal interviews, Filipino domestic workers report that their familiesback home sometimes make unreasonable demands for money and haveunrealistically rosy ideas about their financial situation. However, many also statethat the purpose of their migration is to provide for their family, and they view thisas their main responsibility. Thus, although the literature has typically portrayed thedemands made by relatives as “taxes,” these relationships could be more complexin reality. Admittedly, financial support tends to flow in only one direction from themigrant to her family members back home. However, the spouse, siblings, aunts,and cousins in the Philippines are often looking after the migrant’s children orelderly parents, overseeing house construction and repair, or running the smallbusiness that the migrant has invested in. Thus, these may also be reciprocalarrangements, where one side provides financial support while the other provideshuman resources and facilitates peace of mind. Similarly, although friends in HongKong, China may borrow or request gifts, they also lend in return and make giftswhen the individual herself is in need. In future work, we will investigate the


role that these social networks play in shaping the borrowing choices of Filipinodomestic workers.

Ideally, probable interventions should be evaluated in light of these possiblemechanisms. Commitment savings products that restrict the individual fromwithdrawing until a target date or savings balance is reached may be suitablefor individuals with present-biased time preferences (Ashraf, Karlan, and Yin2006). Those who wish to flexibly finance the expenses of their families may bebetter suited to a contractual savings product that requires them to replenish theirsavings after they have drawn them down (Morduch 2010). In a credit cooperative,this could take the form of a combination loan-and-savings product where, witheach installment, the borrower both repays the loan and simultaneously makes asavings deposit. In future research, we hope to investigate the effectiveness of suchalternative products.

References

Anderson, Siwan, and Jean-Marie Baland. 2011. “The Economics of Roscas and Intra-HouseholdResource Allocation.” The Quarterly Journal of Economics 117 (3): 963–95.

Ashraf, Nava, Diego Aycinena, Claudia Martinez, and Dean Yang. 2015. “Savings inTransnational Households: A Field Experiment among Migrants from El Salvador.” TheReview of Economics and Statistics 97 (2): 332–51.

Ashraf, Nava, Dean Karlan, and Wesley Yin. 2006. “Tying Odysseus to the Mast: Evidence froma Commitment Savings Product in the Philippines.” The Quarterly Journal of Economics121 (2): 635–72.

Baland, Jean-Marie, Catherine Guirkinger, and Charlotte Mali. 2011. “Pretending to be Poor:Borrowing to Escape Forced Solidarity in Cameroon.” Economic Development and CulturalChange 60: 1–16.

Banerjee, Abhijit V., and Esther Duflo. 2011. Poor Economics: A Radical Rethinking of the Wayto Fight Global Poverty. New York: Public Affairs Books.

Barua, Rashmi, Gauri Kartini Shastry, and Dean Yang. 2019. “Financial Education for FemaleForeign Domestic Workers in Singapore.” Economics of Education Review (Forthcoming).https://doi.org/10.1016/j.econedurev.2019.101920.

Collins, Daryl, Stuart Rutherford, Jonathan Morduch, and Orlanda Ruthven. 2009. Portfoliosof the Poor: How the World’s Poor Live on $2 a Day . Princeton: Princeton UniversityPress.

Cortes, Patricia, and Jessica Pan. 2013. “Outsourcing Household Production: Foreign DomesticWorkers and Native Labor Supply in Hong Kong.” Journal of Labor Economics 31 (2):327–71.

Dupas, Pascaline, and Jonathan Robinson. 2013. “Why Don’t the Poor Save More? Evidence fromHealth Savings Experiments.” American Economic Review 103 (4): 1138–71.

Government of the Hong Kong Special Administrative Region (SAR). 2017. “Foreign DomesticHelpers and Evolving Care Duties in Hong Kong.” Legislative Council Secretariat OfficeResearch Brief 2016–2017 (4).

______. 2020. Foreign Domestic Helpers Entry Visa for a New Helper. https://www.immd.gov.hk/eng/services/visas/foreign_domestic_helpers.html (accessed February 29, 2020).


______. Census and Statistics Department. 2017. Hong Kong 2016 Population By-census– Technical Report. https://www.censtatd.gov.hk/hkstat/sub/sp459.jsp?productCode=B1120099.

Gugerty, Mary Kay. 2007. “You Can’t Save Alone: Commitment in Rotating Savings and CreditAssociations in Kenya.” Economic Development and Cultural Change 55 (2): 251–82.

Jakiela, Pamela, and Owen Ozier. 2016. “Does Africa Need a Rotten Kin Theorem? ExperimentalEvidence from Village Economies.” Review of Economic Studies 83: 231–68.

Laureti, Carolina. 2018. “Why Do Poor People Co-hold Debt and Liquid Savings?” Journal ofDevelopment Studies 54 (2): 213–34.

McKenzie, David, and Johan Mistiaen. 2009. “Surveying Migrant Households: A Comparisonof Census-Based, Snowball and Intercept Point Surveys.” Journal of the Royal StatisticalSociety Series A (Statistics in Society) 172 (2): 339–60.

Morduch, Jonathan. 2010. “Borrowing to Save.” Journal of Globalization and Development 1 (2):1–11.

Ng, Desmond, and Cheryl Tan. 2019. “Borrowing, Brokering, Lending: Inside the TangledWeb of Maids and Moneylenders.” Channel News Asia, March 2. https://www.channelnewsasia.com/news/cnainsider/maids-domestic-workers-moonlight-brokers-moneylenders-borrowing-11304566.

Progressive Labor Union of Domestic Workers in Hong Kong (PLU) and Hong KongFederation of Asian Domestic Workers (FADWU). 2016. “Between a Rock and a HardPlace: The Charging of Illegal Agency Fees to Filipino Domestic Workers in thePhilippines and Hong Kong.” https://www.hkctu.org.hk/zh-hant/content/between-rock-and-hard-place-english-versionphilippine-and-hong-kong-governments-fail-stop.

Tan, Peck-Leong, and John Gibson. 2013. “Impact of Foreign Maids on Female Labor ForceParticipation in Malaysia.” Asian Economic Journal 27 (2): 163–83.

Yang, Dean. 2011. “Migrant Remittances.” Journal of Economic Perspectives 25 (3): 129–52.Yu, Chi Wai, Y. Jane Zhang, and Xuejing Zuo. 2019. “Multiple Switching and Data Quality in the

Multiple Price List.” Review of Economics and Statistics (Forthcoming). https://doi.org/10.1162/rest_a_00895.

Appendix 1. Lab-in-the-Field Experiment

Each subject participated in a single experimental session. Each sessionconsisted of 8–15 participants who sat at individual computer terminals. Eachsubject was randomly assigned to a group of 4 or 5 members and played 4 or 5rounds of the decision-making experiment with this group before being randomlyassigned to a new group. At no point could they identify their groupmates fromamong the participants in the room.

In each round, participants were given an endowment of 100 tokens and askedto allocate them across three accounts (or “boxes”): a blue safe box that wouldgive a certain return of x percent; a red box where if the “investment” option wereexercised the return would be 40% with probability 0.8 and 0 otherwise; and a green


box that would generate a fixed in-kind return with a probability proportional to thenumber of tokens placed in the box.

We experimentally varied the rate of return in the blue box to either be 3% or10%. Experimental sessions were randomly assigned to one rate or the other.

The decision of whether to invest the amount in the red box was made by adifferent player in the group; we only analyze rounds where the participant was notan investor. Also, to avoid endogenous token allocation in response to what others inthe group did in previous periods, we only analyze the first round that the participantplayed with each group.

Each token placed in the green lottery box gave a 0.5% probability ofsuccess, so that if the participant placed 10 tokens in this box she would have a5% probability of winning a handbag as a prize. The total earnings from each roundwere displayed to the player at the end of the round. At the end of the session, oneround was randomly chosen and implemented, with an exchange rate of one token= $1. Thus, the participant received the cash payment equal to her earnings as wellas the handbag if she had won it in the randomly selected round.

Appendix 2. Financial Literacy Questions

Question 1. Suppose you need to take a loan here in [Hong Kong, China]. Thereare two choices. Loan A: You will get $10,000 for 6 months. You will have to payback $10,500 at the end of 6 months. Loan B: You will get $20,000 for 6 months.You will have to pay back $20,800 at the end of 6 months.

Which loan is cheaper?Which loan would you prefer?

Question 2. Suppose you need to take a loan of $10,000 here in [Hong Kong,China]. There are two choices. Loan A: You can get $10,000 for 6 months. Youhave to repay $2,000 every month for 6 months. Loan B: You can get $10,000 for 6months. You have to repay $600 every week for 24 weeks.

Which loan is cheaper?Which loan would you prefer?

Appendix 3. Reweighting Our Sample Using a Random 5% Microsample fromthe Hong Kong 2016 Population By-census

The Hong Kong 2016 Population By-census sampled about one-tenth of allresidential quarters in Hong Kong, China and collected detailed socioeconomicdata from all households that lived there. We use the 5% sample of the microdatareleased by the Census and Statistics Department and consider the subsample of


individuals who are Filipino and female and whose relationship to the householdhead is reported as “live-in domestic helper.” We check that this subsample plausiblyconsists of Filipino domestic workers—all individuals reported they are currentlyworking, their economic activity as “employees,” their industry as “domesticpersonnel,” and their occupation as one of the following three categories: “cleaners,helpers, and related workers”; “personal care workers”; or “drivers and mobilemachine operators.”

A simple comparison of the summary statistics for variables that are availablein both datasets suggests some differences in age, education levels, and length ofstay in Hong Kong, China. Accordingly, we construct the multivariate frequencydistribution along these three dimensions in the Census dataset and then reweightour survey sample accordingly.

Note that since the lab-in-the-field experiment implemented a randomizedintervention within the sample, unweighted average and heterogeneous treatmenteffects are internally valid.

Impacts of an Information and CommunicationTechnology-Assisted Program on Attitudesand English Communication Abilities: An

Experiment in a Japanese High SchoolYuki Higuchi, Miyuki Sasaki, and Makiko Nakamuro∗

We conducted a randomized experiment targeting 322 Japanese high schoolstudents to examine the impacts of a newly developed English-languagelearning program. The treated students were offered an opportunity tocommunicate for 25 minutes with English-speaking Filipino teachers via Skypeseveral times a week over a 5-month period as an extracurricular activity.The results show that the Skype program increased the interest of the treatedstudents in an international vocation and in foreign affairs. However, thestudents did not improve their English communication abilities, as measuredby standardized tests, probably because of the program’s low utilization rate.Further investigation showed that the utilization rate was particularly low amongstudents demonstrating a tendency to procrastinate. These results suggest theimportance of maintaining students’ motivation to keep using such informationand communication technology-assisted learning programs if they are notalready incorporated into the existing curriculum. Having procrastinatorsself-regulate may be especially crucial.

Keywords: Japan, learning English, procrastination, randomized controlledtrial, SkypeJEL codes: C93, H40, I21

I. Introduction

Providing students with high-quality learning resources is criticallyimportant in improving the quality of education. In recent years, information and

∗Yuki Higuchi (corresponding author): Faculty of Economics, Sophia University, Japan. E-mail:[email protected]; Miyuki Sasaki: Faculty of Education and Integrated Arts and Sciences, Waseda University,Japan. E-mail: [email protected]; Makiko Nakamuro: Faculty of Policy Management, Keio University,Japan. E-mail: [email protected]. This study was conducted as a part of the Measurement of the Qualitiesof Health and Education Services, and Analysis of their Determinants project undertaken at the Research Instituteof Economy, Trade and Industry. We would like to thank Tomohiko Inui, Yukichi Mano, Ryoji Matsuoka, ShinpeiSano, an anonymous referee, and participants of the Asian Development Bank–International Economic AssociationRoundtable for helpful comments and suggestions. We also acknowledge Takeshi Kamimura, Tomohisa Kato, andTomoya Sugiyama for their active research collaboration. This research was financially supported by MEXT/JSPSKAKENHI Grant Number: 18H05314, Grant-in-Aid for Research at Nagoya City University, where the first andsecond authors were affiliated with until March 2020, and Keio University. All errors are our own. The usual ADBdisclaimer applies.




Impacts of an ICT-Assisted Program on Attitudes and English Communication Abilities 101

communication technology (ICT) has increasingly been used as an alternativeto more conventional resources (e.g., Gee and Hayes 2011, Levy 2009). SuchICT-assisted educational resources can be best used to help overcome the limitationsof conventional resources. In particular, because ICT can provide customized andself-paced learning opportunities, the use of ICT in education has huge potential toimprove the effectiveness of home learning.

According to surveys by Bulman and Fairlie (2016) and Snilstveit et al.(2016), the classroom use of ICT generally has positive impacts, especially forstudents in lower grades studying math or science. While earlier observationalstudies found large positive impacts of home use of ICT on students’ academicoutcomes, these studies suffered from the selection bias that students or teacherswith unobserved high ability or motivation tended to introduce the new ICT-assistedresources. More recent experimental studies tended to find smaller or even noimpacts.1 Such mixed results for the home use of ICT partly reflect differencesin the grades of the sampled students, their proficiency levels, sampled countries,and studied or targeted subjects; however, we particularly need evidence on whetherthe home use of ICT can compensate for the weaknesses of conventional educationresources.

To test the usefulness of the home use of ICT in complementing currenteducation programs, we conducted a randomized controlled trial (RCT) thatprovided ICT-assisted resources for Japanese high school students learning English.In contrast to the high internationally normed performance of Japanese studentsin reading, math, and science—as measured by the Organisation for EconomicCo-operation and Development’s Program for International Student Assessment forGrade 9 students—their performance in English has been far from satisfactory.According to a nationwide English test conducted in 2014 by the Ministry ofEducation, Culture, Sports, Science and Technology, Japan (MEXT), a majorityof Grade 12 students ranked at the lowest level (A1) in the Common EuropeanFramework of Reference for Languages, with their speaking performance lowestamong the four skills measured. Based on these results, MEXT recognized that thequality of English education, particularly in nurturing speaking ability, should beimproved (MEXT 2015a). As conventional English education programs in Japanhave been unsuccessful, there is scope for the use of ICT-assisted resources toimprove the quality of such education.

We experimentally introduced a newly developed online English learningprogram as an extracurricular activity to 322 Japanese students in Grade 10.This online program is an individualized, self-paced program in which studentscommunicate with English-speaking Filipino interlocutors, mostly consisting of

1This is reminiscent of Glewwe et al. (2004), who compared an observational study with an experimental oneand found that the large positive impact of the introduction of flipcharts to Kenyan schools found in the observationalstudy was no longer detected in the experimental one.


current students or graduates of the University of the Philippines, the top nationaluniversity in the country. The students can communicate with them at mutuallyconvenient times via Skype using learning materials of their own choice. Thisprogram is an example of human resource arbitrage from developing to developedcountries with the help of modern ICT technology. Although it is beyond thescope of this paper, the program may have positive impacts not only on theJapanese-student side but also on the Filipino-instructor side by creating earningopportunities.

We introduced the Skype English program with a crossover design.2 First, werandomly selected half of our sample (161 students) to be given the opportunity touse the program for 5 months from July to November 2015, while the remaining 161students were given the opportunity to use the program for 5 months from Januaryto May 2016. While all the students had an equal opportunity to use the programby May 2016, only half of them had taken this opportunity as of December 2015,when we conducted the endline survey. We therefore refer to the students exposedto the program in the first round (July–November 2015) as the treatment group andthose exposed to it in the second round (January–May 2016) as the control group.3

Combining program usage records and panel data collected before and afterthe introduction of the program to the treatment group (but not yet to the controlgroup), two main findings emerge. First, the program changed the attitudes of thetreated students positively, especially in terms of their interest in an internationalvocation and in foreign affairs. In particular, our estimates of the local averagetreatment effect (LATE) suggest that the effects were large for students with greaterprogram utilization. This finding is important because past longitudinal studiessuggested that it is difficult to change students’ attitudes toward an internationalvocation and foreign affairs when they study a foreign language (Ortega andIberri-Shea 2005). This may be particularly the case in the Japanese schoolenvironment, which is known to have a monocultural and monolingual orientation.Furthermore, Sasaki (2011); Yashima (2002); and Yashima, Zenuk-Nishide, andShimizu (2004) found that such attitudinal change among Japanese students willeventually lead to improvements in their English communication skills.

Second, despite the positive impacts on the students’ attitudes, there is nomeasured impact on their English communication skills. This may be attributed

2Although an RCT is now recognized as best practice in impact evaluation, it is extremely difficult to runsuch a trial in Japanese public schools, where priority is given to equality of resource allocation within the samecohort of students. Hence, as a second-best strategy, we conducted an RCT with a crossover design, ensuring thatall students received the same treatment within the same academic year, with the only difference being in respect tothe timing of the treatment. A shortcoming of this strategy is that the evaluation period is less than 6 months, but weemphasize that our study is a unique RCT conducted in a public school in Japan.

3A referee suggested to additionally use a difference-in-differences (DiD) “in reverse” approach, exploitingthe change in status of the control group from before-treatment to after-treatment, while the treatment group remainedafter-treatment status (Kim and Lee 2019). We, however, were unaware of this approach and did not conduct a surveyor a standardized English test after the intervention with the control group. We note that DiD “in reverse” is a usefulapproach in a crossover RCT in general.


to the low intensity of the program (25 minutes per lesson) in comparison withthe students’ concurrent regular English classes (50 minutes per lesson on mostweekdays) as well as the program’s low utilization rate. Only 10 of the 161students in the treatment group took 50 or more lessons over the 5-month period,as recommended by the program provider, and 31 students took no lessons overthe same period. In addition, regression analyses show that the utilization ratewas particularly low among students with a tendency to procrastinate, which isconsistent with the emerging literature on self-control problems (e.g., Duckworth,Milkman, and Laibson 2018). These findings warrant further research on how toimprove and maintain students’ motivation, particularly those with a tendency toprocrastinate, to adopt home-use ICT programs such as the one targeted in thisstudy.

The remainder of this paper is organized as follows. Section II describes ourexperiment, including the sample, timeline, and details of the intervention. SectionIII discusses sample balance and program utilization, and section IV presents theestimated program impacts. Finally, section V contains a summary of the findingsand implications for future studies.

II. Experiment

A. Sample

We collaborated with a public high school that is a top-tier school in centralJapan. This school was selected by the Government of Japan in 2015 as oneof the 112 Super Global High Schools among the 4,939 high schools in Japan.Super Global High Schools receive extra budgetary support to nurture globalizedleaders with high levels of interest in societal problems, communication skills,and problem-solving abilities, who will play internationally active roles in thefuture (MEXT 2015b). The school agreed to introduce the online program as anextracurricular activity.

Our sample consisted of all 322 first-year high school students (Grade 10)who were newly admitted to the school a few months before the experiment.4

In Japan, high school admissions, whether public or private, are mostly basedon students’ academic performance on the entrance examination, with studentssubsequently tracked into different high schools of varying quality. After our samplestudents were admitted to our target high school, they were randomly assigned toone of eight classrooms, each consisting of 40 or 41 students. Classroom assignment

4We provided all the parents of the sample students with information on our research and its purpose beforecommencing data collection and intervention. As the parents of one student refused to provide data for our analyses,we excluded the data collected from that student. Thus, the sample size is 321 in our empirical analyses.


Figure 1. Randomization

Source: Authors’ illustration.

was not affected by any preexisting peer groups; we took advantage of this to attainrandomization in our experiment.

Further, each of the four full-time English teachers in the school wererandomly assigned to teach two of these eight classes. To achieve balance in thequality of the English teachers in the classroom, we stratified the sample of studentsat the teacher–classroom level, randomly assigning one of the two classes instructedby each English teacher to the treatment group and the other to the control group(Figure 1). In sum, we have four treatment classes (with 160 students) and fourcontrol classes (with 161 students). Although our experiment may suffer froma small number of clusters (i.e., eight classes), the classroom-level intraclustercorrelation coefficients for outcome variables at the baseline survey are close to 0,indicating that there is little correlation of responses within a cluster, and thus, ourrandomization can be considered as being close to the student-level randomization.5

B. Timeline

Before introducing the program, we conducted a baseline survey designed tocollect information on the students’ characteristics and attitudes toward Englishcommunication. The survey was conducted in June 2015, using a mark-sheetquestionnaire we developed. The timeline of our research is presented in Table 1.

Soon after the baseline survey, the sample students took the Versant speakingtest (Pearson Inc. 2008), a standardized test designed to evaluate the oral English

5The classroom-level randomization will help us mitigate the violation of the Stable Unit TreatmentValue Assumption caused by spillover effects among students in the same classroom. While admitting that it istechnically difficult to separate the direct effect of our intervention from the indirect effect through their peers in theclassroom-level randomization, as pointed out by Imbens and Wooldridge (2009), we think that the degree of suchindirect effect is limited because our outcome variables are individual measures of attitudes and test scores, whichare more likely to be affected by interactions with English teachers than by those with their classmates.


Table 1. Research Timeline

June 2015 Baseline survey and (i) Versant test1 July 2015 Online English program for the treatment group startsJuly 2015 (ii) Benesse testNovember 2015 (ii) Benesse test and (iii) GTEC English test30 November 2015 Online English program for the treatment group endsDecember 2015 Endline survey and (i) Versant testJanuary–May 2016 Online English program for the control group

GTEC = Global Test of English Communication.Source: Authors’ compilation.

skills (integrated listening and speaking) of nonnative English speakers.6 The testwas administrated solely for this research project (although the results were sharedwith the students as feedback) to construct our measure of English communicationability. Following the survey and the Versant test administered in June 2015, wecommenced the intervention on 1 July 2015. The students in the treatment classeswere provided with opportunities to use the online program free of charge, althoughthe market price of the program was ¥5,800 (about $52) per month. This includedone 25-minute lesson for every day of the intervention period.

Soon after our intervention commenced, the students took a nationallyadministrated English test developed and distributed by Benesse Co. The test isa mock university entrance exam designed primarily to measure students’ Englishreading ability. The sample students took a similar test again in November, towardthe end of our intervention. Although the tests were not taken for the purpose of ourstudy, the school agreed to share the results with us to be used as another measureof the students’ English abilities. In addition, in November, the students took theGlobal Test of English Communication (GTEC), a standardized test developed anddistributed by Benesse Co. to evaluate reading, listening, writing, and speakingskills in English.7 The school also agreed to share the results of this test with us.

In December 2015, when only the treated students had received the program,we conducted an endline survey and Versant test. In other words, to investigate theeffects of the online program, the treatment and the control classes were comparedusing a difference-in-differences (DiD) design. To mitigate inequality between thetwo groups (as mentioned above), we provided the same amount of intervention

6We chose this particular test because of its reported high validity and reliability among populations similarto the sample in the present study and because it requires a relatively short time (20 minutes) to conduct comparedwith other English communication tests (e.g., TOEFL iBT). During the Versant test, the students listened to questionsspoken in English and provided verbal answers in English. Their answers were recorded and automatically markedonline. The test was conducted by class in a computer room inside the school, and thus, the test-taking environmentwas essentially the same for all students. The Versant test scores ranged from 20 to 80 and involved four criteria:(i) sentence mastery, (ii) vocabulary, (iii) fluency, and (iv) pronunciation. The scores correspond with the levels ofthe Common European Framework of Reference for Languages: for example, a Versant score of 20–25 is equivalentto the lowest (A1) level, while a score of 79–80 is equivalent to the highest (C2) level.

7The test consists of 30 multiple-choice reading items (24 minutes), 30 multiple-choice listening items(13 minutes), 3 performative writing items (26 minutes), and 4 performative speaking items (12 minutes).


with a time lag, with the program being made available to the control classroomsfrom January to May 2016. By the end of May 2016, all 322 students had beenexposed to the same intensity of intervention (or lack thereof).

C. Intervention

Our intervention consisted of providing the sampled students withopportunities to use the online program. In contrast to conventional face-to-faceEnglish learning methods, in this program, learners and teachers do not have to bepresent in the same space. In addition, learners can be matched with teachers on amore flexible basis because learners can select among available teachers at a time oftheir convenience. Such online English programs have become increasingly popularamong Japanese businesspeople, partly because of time flexibility advantages andpartly because of the low cost of such programs relative to similar face-to-faceEnglish learning programs offered by commercial conversation or crammingschools. However, according to the baseline survey that we conducted before thebeginning of the intervention, 65% of the students had never heard of this typeof online English learning program, only one student was using such a program,and another 10 had used one in the past. In this baseline survey, 30% of thestudents responded that they would be very willing to use the program if giventhe opportunity, and another 50% responded that they were moderately willing touse it. Hence, while the program was new to most of the students, it was favorablyperceived at the beginning of our intervention.

The online program was provided to the students outside of their regularEnglish classes. Each lesson took 25 minutes, and the students were recommendedto take one lesson every 3 days (i.e., 10 lessons a month, or 50 in total) to take fulladvantage of the program. The students could make an appointment for a lesson atany time between 6 a.m. and 1 a.m. on the following day and could choose any ofthe available teachers. If the student’s preferred teacher was not available at the timeof their convenience, they were able to choose another time slot or another availableteacher in the same slot. The pool of teachers consisted mostly of current studentsor graduates of the University of the Philippines. Because English is the language ofinstruction in their home university and also because they were screened on the basisof the company’s strict hiring criteria, we judged that the quality of the teachers wasreasonably guaranteed. While some of the teachers spoke Japanese, participatingstudents had to communicate entirely in English with the help of the chat (texting)function in Skype. Students were free to choose appropriate study materials foreach lesson from a wide range of materials provided by the program, includingdaily conversation, academic talk, grammar and vocabulary, and business English.In other words, the participants’ choice of teachers, time slots, and study materialswere their decision entirely. Most importantly, while we provided the students withopportunities to use the program at home, it was ultimately up to them whether and


Table 2. Balance Check

Differencep-value forEquality ofTreatment Control

Mean N Mean N Means

Procrastination (z-score) −0.020 157 0.021 155 0.72Male (1 = yes) 0.50 159 0.50 161 0.99English since Grade 1 or 2 (1 = yes) 0.42 156 0.40 154 0.63English since Grade 3 or 4 (1 = yes) 0.41 156 0.42 154 0.92English since Grade 5 or later (1 = yes) 0.17 156 0.19 154 0.62Been abroad (1 = yes) 0.39 157 0.37 159 0.66Own room (1 = yes) 0.89 157 0.84 159 0.15Own personal computer (1 = yes) 0.08 152 0.12 154 0.20Own tablet (1 = yes) 0.23 156 0.16 159 0.10Commuting 20 minutes or less (1 = yes) 0.26 156 0.21 155 0.24Commuting 21–40 minutes (1 = yes) 0.38 156 0.42 155 0.53Commuting 41–60 minutes (1 = yes) 0.26 156 0.26 155 0.87Commuting 61 minutes or more (1 = yes) 0.10 156 0.11 155 0.70Belongs to sports club (1 = yes) 0.65 156 0.57 155 0.19Number of books at homea 2.66 154 2.33 155 0.06

N = number of observations.Notes: aNumber of books at home; 0 = none, 1 = approximately 20, 2 = approximately 50, 3 =approximately 100, 4 = approximately 200, and 5 = over 300.Source: Authors’ calculations.

how often to take the lessons, especially because their participation did not affecttheir grades.

One of the biggest advantages of this online program is its cost-effectiveness.The government launched the Japan Exchange and Teaching Program in 1987,which involved providing English-speaking aides known as Assistant EnglishTeachers (AETs) to Japanese English teachers in primary, middle, and high schools(Grades 1–12). This program has expanded since then—a total of 5,163 AETs wereemployed as of 2017. The individual annual cost for an AET is approximately$53,000, including salary, coordination, and transportation, while the market priceof this English program is $600 per year. Based on the program provider’sback-of-the-envelope calculation, the program enables students to devote 15 timesmore minutes to speaking with English-speaking partners than speaking with anAET for every dollar spent.

III. Balance and Program Utilization

A. Balance

Table 2 presents the basic characteristics of students that could potentiallyinfluence the take-up rate and effects of the online program. As the literaturefinds that a lack of self-control, including procrastination, can result in poor


test performance or low grades (e.g., Golsteyn, Grönqvist, and Lindahl 2014;Onji and Kikuchi 2011), we constructed an index of procrastination as a controlvariable based on the six questions to rate students’ perception of themselves, takenfrom Osaka University (2013) and Honda and Nishijima (2007). The questions(originally written in Japanese and translated by the authors) included items such as“Are you a person who postpones plans even when you make them?” and “Are youa person who is happy as long as you are having fun now?” The students answeredall six questions with categorical responses: (i) yes, (ii) moderately yes, (iii) 50/50,(iv) moderately no, or (v) no. We assigned a score of 4 to the answer yes, 3 tomoderately yes, 2 to 50/50, 1 to moderately no, and 0 to no. We then aggregated thescores for all six questions to construct a single index of procrastination, whichranged from 0 to 24 (maximum of 4 multiplied by 6 items). These aggregatedscores were normalized by subtracting the sample mean and then dividing by thestandard deviation. The mean z-score of the procrastination index is −0.02 amongthe treatment group and 0.021 among the control group; importantly, these meansare statistically not different.

Other control variables include gender, past exposure to English (whetherthe student has been abroad and the grade at which they started learning Englishin primary school), and current study environment (having their own room andelectronic device, such as a personal computer connected to the internet or atablet, commuting time to school, and membership of a school sports club), as wellas their family background (number of books at home and parental educationalattainment).8 We also collected information on smartphone ownership, but almostall of the students (96%) owned one so we do not include this variable as a control.The differences in means between the two groups are statistically insignificant atthe 5% level for all the variables, indicating that randomization was performedsuccessfully.

B. Program Utilization

Figure 2 shows daily changes in the number of students who took the lessonsbased on program usage records. Of the 160 students assigned to the treatmentgroup, the average number of students who took lessons each day was 25 in July2015. However, if all students had completed the recommended 10 lessons a month,that number would be 52 (10 lessons multiplied by 160 students and divided by 31days). Thus, the take-up rate in the first month of the intervention was about 50%.Moreover, the number of students taking lessons decreased gradually, presumablybecause the novelty effect faded and peer pressure was muted by the summer

8As a number of students (27 in the treatment group and 21 in the control group) did not report their parentaleducational attainment, we do not use the variables of father’s education and mother’s education. Instead, we usethe variable of number of books at home as a proxy of parental socioeconomic status. Kawaguchi (2016) found acorrelation between the number of books at home and parents’ earnings among Japanese Grade 10 students.


Figure 2. Daily Change in Number of Students Taking Lessons, 2015

Source: Authors’ calculations.

vacation, which started during the last week of July, with the average number fallingto 15 in August, 12 in September, 6 in October, and 5 in November. While Figure2 shows daily changes in program utilization, Figure 3 shows the student-levelnumber of lessons taken during the intervention period. Thirty-one (19%) of the160 students never took any lessons in the 5-month period, and 57 (36%) took fiveor fewer lessons. Only 23 students (14%) completed 25 or more lessons, one-halfof the recommended number, of whom only 10 (6%) completed the recommended50 or more lessons.

To identify the factors associated with program utilization, we estimatedthe ordinary least squares models while controlling for the English teacherdummies. Column 1 shows that the effect of the procrastination index is negativeand significant, illustrating the detrimental effect of procrastination on programutilization. The significance of this variable remains robust and consistent, evenafter the variables listed in Table 2 are controlled (column 2). In terms of the sizeof the effects, a 1 standard deviation increase in the procrastination index reducesthe number of lessons by about 4 times, where the mean was 12.2 times; thus, theinfluence of procrastination seems nonnegligible.

As the program was new to most of the students and the first few trials ofthe program are critical for subsequent utilization, we estimated a linear probabilitymodel, where the dependent variable is coded as a dummy variable that equals 1if a student has ever used this Skype program and 0 otherwise. Indeed, accordingto our informal interviews with some of the students, regular Skype users startedto like the program as they proceeded through the initial few talks with Filipino


Figure 3. Distribution of Lessons Taken by a Student over 5 Months

Source: Authors’ calculations.

interlocutors, whereas nonusers felt hesitant to take the first lesson. Columns 4–6show the results, and the procrastination variable is negative and significant.

Table 3 also shows that the English teacher dummies are large in magnitudeand statistically significant. For instance, a student with English teacher D wasabout 40 percentage points less likely to have ever used the program than a studentwith English teacher A (base category). The degree of in-class encouragementand reminders substantially differed from one teacher to another, with teacherA, who is the most senior and experienced among the four teachers, providingmore encouragement and more frequent reminders to students to participate in theSkype tasks. According to our informal interviews, this teacher frequently askedthe students whether they used the program to put gentle pressure on them as wellas to share their experiences with other classmates. This teacher also posted aneye-catching message in the classroom to regularly use the program. Theseobservations suggest that the frequencies of such promotive acts from teachers maybe critical to the home use of ICT-assisted inputs.

IV. Impacts

A. Descriptive Analyses: Attitudes

We included two sets of outcome measures to evaluate the impactsof the online program: (i) attitudes and (ii) English communication abilities.


Table 3. Correlates of Program Utilization (Ordinary Least Squares Estimation)

(1) (2) (3) (4) (5) (6)Number of lessons taken = 1 if completed at least

in 5 months one lesson in 5 months

Procrastination −3.82** −4.35** −3.88* −0.097*** −0.085** −0.084**

[z-score] (−2.60) (−2.26) (−1.96) (−3.26) (−2.58) (−2.49)Male −1.41 −0.083 −0.12* −0.12*

(1 = yes) (−0.29) (−0.02) (−1.88) (−1.81)English since Grade 3 or 4x 1.09 0.0079 0.052 0.054(1 = yes) (0.29) (0.00) (0.80) (0.83)English since Grade 5 or later −2.72 −3.23 0.039 0.044(1 = yes) (−0.59) (−0.72) (0.44) (0.49)Been abroad −0.70 0.11 −0.11* −0.11*

(1 = yes) (−0.22) (0.03) (−1.67) (−1.75)Own room −3.72 −4.26 −0.15* −0.15*

(1 = yes) (−0.71) (−0.82) (−1.81) (−1.79)Own personal computer 1.17 1.37 −0.15 −0.15(1 = yes) (0.21) (0.25) (−1.24) (−1.25)Own tablet −0.32 0.38 0.067 0.068(1 = yes) (−0.07) (0.09) (1.04) (1.06)Commuting time 21–40 minutes 6.68* 5.51 0.011 0.013(1 = yes) (1.76) (1.49) (0.15) (0.18)Commuting time 41–60 minutes 4.42 4.40 −0.024 −0.026(1 = yes) (0.89) (0.89) (−0.27) (−0.29)Commuting time 61 minutes 1.37 1.35 0.17 0.16or over (1 = yes) (0.30) (0.29) (1.49) (1.37)Sports club −1.53 −2.88 −0.12* −0.12*

(1 = yes) (−0.33) (−0.65) (−1.94) (−1.82)Number of books −0.36 −0.70 0.046** 0.046**

[1–6] (−0.28) (−0.58) (2.43) (2.39)Baseline international posture 0.28 0.013(z-score) (0.20) (0.44)English teacher B −0.14 −0.98 −2.77 −0.21*** −0.24*** −0.24***

(1 = yes) (−0.03) (−0.21) (−0.63) (−3.21) (−3.45) (−3.35)English teacher C 0.038 −1.31 −1.05 −0.19*** −0.17** −0.17**

(1 = yes) (0.01) (−0.22) (−0.18) (−3.18) (−2.40) (−2.43)English teacher D −7.23** −10.2** −9.91** −0.42*** −0.39*** −0.40***

(1 = yes) (−2.09) (−2.37) (−2.27) (−5.28) (−4.97) (−4.97)

Mean of the outcome variable 12.2 0.81

R-squared 0.064 0.107 0.099 0.192 0.352 0.352Adjusted R-squared 0.039 −0.002 −0.021 0.170 0.272 0.266No. of observations 157 147 146 157 147 146

Notes: Estimated coefficients are reported here. ***, **, and * indicate 1%, 5%, and 10% levels of statisticalsignificance, respectively. Numbers in parentheses are t-statistics based on heteroscedasticity-robust standard errors.The base category for the English-since variable is “English since Grade 1 or 2,” for the commuting time variable itis “Commuting time 20 minutes or less,” and for the teacher dummies it is “Teacher A.”Source: Authors’ calculations.

To quantitatively measure any changes in students’ attitudes toward Englishcommunication before and after the intervention, we employed two motivationalattributes that have been found to influence students’ second-language


development: (i) international posture and (ii) willingness to communicate(WTC) (e.g., Yashima, Zenuk-Nishide, and Shimizu 2004). First, the constructof international posture was operationally defined as a composite of foursubconstructs: (i) intercultural orientation; (ii) interest in an international vocation;(iii) reactions to different customs, values, or behaviors; and (iv) interest inforeign affairs. These subcomponents and corresponding items were adapted fromthose made available on the homepage of Professor Tomoko Yashima, who firstintroduced this construct to the field of applied linguistics.9 This construct hasproved to be one of the most distinct and significant factors explaining students’motivation, especially in English-as-a-foreign-language contexts (see, for example,Dörnyei and Ryan 2015). Using all 22 available items (seven for subcomponent 1,six for subcomponent 2, five for subcomponent 3, and four for subcomponent 4),we then created questions requiring either yes or no answers. Although the originalversions of the 22 questions required responses using a six-point Likert scale,we simplified it to yes–no answers to avoid causing excessive fatigue among thestudents, who had to respond to many questions in our survey. We computed a scorefor each of the four subcomponents of international posture and then computedtotal scores, which ranged from 0 to 22, with a higher score indicating a moreinternationally oriented student. Finally, we computed z-scores for the total scoreas well as for the four subcomponents.10

Panel A of Table 4 presents the means of the international posture scoresby group, before and after our intervention with the treatment group (but not yetwith the control group). First, the means of all the scores before the interventionwere not statistically different between the two groups (see the p-values reportedon the right). For instance, the baseline mean z-score for the treatment group was0.042, which was slightly higher than the control group mean of −0.041, but thescores are not statistically different. After the intervention, however, the total scorebecame higher among the treatment group than the control group, and the differenceis statistically significant at the 5% level. If we examine the subcomponents, asignificant difference is observed for subcomponent 2 (interest in an internationalvocation) and subcomponent 4 (interest in foreign affairs).

Interestingly, the total score dropped from the baseline mean of −0.041 toan endline mean of −0.172 among the control group (z-scores were computedusing the means and standard deviations among the baseline samples), which is adecline of 0.13 standard deviations. This declining trend was particularly observablefor subcomponents 1 and 2, which suggests that the motivation of students tolearn English shifted from a more to less internationally oriented one: preparationfor university entrance exams. In the top-tier high school where we conducted

9Tomoko Yashima. Kokusai. http://www2.ipcku.kansai-u.ac.jp/∼yashima/data/kokusai.pdf (accessed April15, 2019).

10Appendix Table A1 presents regression results that analyze the baseline correlates of the internationalposture z-score as well as the baseline correlates of our other outcome variables discussed below.


Table 4. Differences in Attitudes and English Communication Test Scoresby Group

A. International posture and willingness to communicate


Mean N Mean N Means

Total international posture [z-score, 22 criteria]Baseline 0.042 156 −0.041 159 0.47Endline 0.068 155 −0.172 157 0.05

Sub 1. Intercultural approach tendency [z-score, 7 criteria]Baseline 0.024 157 −0.024 159 0.67Endline −0.091 155 −0.162 157 0.55

Sub 2. Interest in international vocation [z-score, 6 criteria]Baseline 0.011 157 −0.011 159 0.84Endline 0.054 155 −0.170 157 0.05

Sub 3. Reaction to different customs [z-score, 5 criteria]Baseline 0.034 156 −0.033 159 0.56Endline 0.010 155 −0.031 157 0.71

Sub 4. Interest in foreign affairs [z-score, 4 criteria]Baseline 0.068 157 −0.067 159 0.23Endline 0.259 155 −0.076 157 0.01

Willingness to communicate [z-score, 8 criteria]Baseline 0.063 156 −0.063 155 0.27Endline −0.082 155 −0.27 156 0.09

Cambodia study tour (1 = yes)Endline 0.101 159 0.068 161 0.30

B. English communication test


Mean N Mean N Means

(i) Versant score [z-score]Baseline 0.095 142 −0.093 146 0.11Endline 0.671 124 0.406 141 0.05

(ii) Benesse score [z-score]Baseline −0.032 156 0.031 158 0.58Endline −0.030 156 0.030 156 0.60

(iii) GTEC overall score [z-score]Endline 0.002 158 −0.001 160 0.98

Sub 1. ReadingEndline −0.012 159 0.012 161 0.83

Sub 2. ListeningEndline 0.034 158 −0.033 161 0.54

Sub 3. WritingEndline 0.024 158 −0.023 160 0.68

Sub 4. SpeakingEndline −0.011 159 0.011 161 0.84

Continued.


Table 4. Continued.

GTEC = Global Test of English Communication.Notes: z-scores are computed using the means and standard deviations among the baseline samples for internationalposture, willingness to communicate, and Versant score. The level of the Benesse test is different from one test toanother, as it is in accordance with the school curriculum; z-score is separately computed for baseline and endlinesamples. For the GTEC score, we only have observations at the endline; z-scores are computed using the meansand standard deviations among the endline samples. ***, **, and * indicate 1%, 5%, and 10% levels of statisticalsignificance, respectively.Source: Authors’ calculations.

the experiment, the curriculum focuses on exam preparation even for first-yearstudents (Sasaki 2018). Hence, panel A appears to suggest that our program helpedmitigate the worsening attitudes among sampled students by stimulating theirinterest in an international vocation and international affairs (subcomponents 2 and4, respectively).

The second motivational variable, WTC, also has significant and complexrelationships with second-language learner confidence, motivation, and actuallanguage use (e.g., MacIntyre 2007). As in the case of international posture, we tookthe eight items that measured WTC from the above-mentioned homepage becausethey have been successfully used in the past with Japanese high school studentslearning English as a second language (e.g., Yashima 2009).11 The questions askedwhether the students would be willing to communicate in English in hypotheticalsituations such as “group discussions on an English course,” “giving a speech inpublic,” and “a chance meeting with a foreign friend in the street.” A six-pointLikert scale offered the following choices: always, usually, sometimes, not veryoften, seldom, and never. We assigned 5 points to the answer always, 4 to usually,3 to sometimes, 2 to not very often, 1 to seldom, and 0 to never, and computed thez-value of the total points.

The means of the z-scores are reported toward the bottom of panel A in Table4. Similar to international posture, the control mean dropped from the baseline tothe endline. However, the drop was smaller among the treatment group, and theinitially nondifferent means became marginally different in the endline. This findingsuggests that although the students’ WTC tended to decline as a result of an Englishcurriculum, such as the one followed in the top-tier high school under study, theSkype program played a role in mitigating the declining WTC.

As an additional variable to examine the attitudes of sample students, weuse the Cambodia study tour dummy variable reported at the bottom of panel A.The school organized a 1-week study tour to Cambodia in December 2017 andthe students had a chance to voluntarily apply for inclusion. The school providedus with a list of students who applied, and we constructed a dummy variable that

11Tomoko Yashima. WTC Scale. http://www2.ipcku.kansai-u.ac.jp/∼yashima/data/wtc_scale.pdf (accessedApril 15, 2019).


equals 1 if a student applied and 0 otherwise. Sixteen (10.1%) of the treated studentsand 11 (6.8%) of the control students applied. Although the difference is notstatistically significant, the application rate was 4.2 percentage points higher amongthe treatment group. Importantly, the correlation between the application dummyand the total endline international posture score was positive with a correlationcoefficient of 0.21 (not reported). Thus, the ICT program may have encouragedmore students to apply by improving their international posture, which we may notbe able to detect because of the weak statistical power.

B. Descriptive Analyses: English Communication Abilities

To quantify the students’ English abilities, we use three sets of Englishtests: Versant, Benesse, and GTEC. We conducted the Versant tests both beforeand after our intervention to measure the development. In addition, the Benessetest was taken soon after our intervention started and toward the end of it, sothe Benesse test score can also be used for the comparison using a DiD design.The GTEC test only measures cross-sectional differences after the intervention. Allthe test scores are presented as standardized z-scores. The scores of the standardizedVersant test are comparable over time, and we computed z-values using the meansand standard deviations among the baseline samples. Thus, we can measure theimprovement in English communication abilities by looking at the changes in thoseabilities. However, the Benesse test score differs from one round to the other, as itis designed in accordance with the school curriculum and the difficulty of the testincreases as students proceed with the curriculum. Thus, the z-scores are computedseparately for the baseline and endline samples, and the changes in the z-scoresbefore and after the treatment do not necessarily indicate changes in students’ levelsof English abilities because the Benesse test is likely to be more difficult in theendline.

Panel B in Table 4 shows the results of the treatment and control groups’respective scores in the international posture and English tests. Although weprimarily intended to use the Versant test as our measure of English communicationabilities, the answers provided by some students were not properly recorded becauseof overburdened internet connections. That is, the test was conducted in a computerroom inside the school in order to provide the same test-taking environment for allstudents, but we ultimately organized a follow-up session for the students whoseanswers were not recorded. Because not all students attended the follow-up session,the problem is that scores were unrecorded for students who were less confidentand more hesitant to retake the test. Appendix Table A2 presents the regressionresults, where the left-hand-side variable is a dummy variable equal to 1 if thestudent took the Versant test. The results show that the Versant take-up was notcorrelated with the observable characteristics at the baseline, but was correlatedwith the baseline Versant score at the endline (column 6). This suggests that poorly


performing students were less likely to have taken the endline Versant test, and weshould therefore interpret the results cautiously.

For the Versant score, there is a slight difference between the two groupsat the baseline, but it is not statistically significant. The score at the endline isstatistically different between the two groups, with the treatment group having ahigher score. However, this difference may be due to the types of students choosingto take the test, particularly among the treated students. Panel B also shows thatthe control mean increased from −0.093 to 0.406, which is a one-half standarddeviation increase over 6 months. This is equivalent to a 2-point increase in theVersant score (out of a full score of 80), which is quite large according to theservice provider. This improvement is most likely the consequence of the regularcurriculum. By contrasting this result with our discussion above, we argue thatwhile the regular school curriculum was unsuccessful in making the students’motivation to learn English more internationally oriented, it did improve theirEnglish communication abilities. The Skype program has the potential to sustainthe students’ intrinsic motivation and therefore supplement the regular curriculum.

The mean scores of the Benesse test, reported in the middle of panel B,were balanced at the baseline and there was no significant difference at the endline.One possible reason for this null result is that the Benesse test primarily measuresreading abilities, whose improvement was not the main focus of the Skype program.The same logic applies to the overall GTEC score, which comprehensively measuresfour English-language skills. Yet, even when we look at the subcomponents of theGTEC, there was no statistical difference in subcomponent 2 (listening ability) orin subcomponent 4 (speaking ability). Taken together, the results shown in Panel Bsuggest that our intervention did not improve the English communication abilitiesof the treated students.

C. Econometric Specification

To rigorously analyze the impacts of the online program by controllingthe baseline level of outcome variables or other characteristics, we applied twoeconometric specifications: analysis of covariance (ANCOVA) and DiD regression.Let yijkt be an outcome variable of student i in classroom j with English teacher k attime t. The ANCOVA specification is written as

yi jkt = α + βTreatment j + γ yi jkt−1 + ηk + εi jkt (1)

where Treatmentj is a dummy variable equal to 1 for the student in treated class j,yijkt−1 is an outcome variable at t − 1 (since we have only two time periods, t − 1represents the baseline and t the endline), ηk is a set of English teacher dummies, andεijkt is a heteroscedasticity-robust standard error. The standard error is not clusteredbecause the number of clusters is much smaller than the rule-of-thumb number of


42 (Angrist and Pischke 2009). To control for possible intracluster correlations,together with correcting for the small number of clusters, we report the 95%confidence intervals (CIs) based on the wild cluster bootstrap method suggested inCameron, Gelbach, and Miller (2008). We used boottest Stata command developedby Roodman et al. (2019) for the computation of the bootstrapped CIs.

In equation (1), β is the parameter of interest, which captures the intention-to-treat (ITT) impacts of the program. In addition to the ANCOVA specification,we also estimate a standard DiD model to control for unobserved, time-invariant,student-level heterogeneity, υ i, using the following specification:

yi jkt = α + βTreatment j ∗ Endlinet + δEndlinet + υi + εi jkt (2)

where Endlinet is a dummy variable equal to 1 if the data are collected in the endline(i.e., after the intervention). β in equation (2) is the parameter of interest, whereasδ measures the changes in the outcome variable from the baseline to the endline,which are mainly consequences of the regular school curriculum, as well as otherchanges that are common to all students.12

To analyze the different impacts of the online program by level of utilization,we use an instrumental approach to estimate the LATE (Imbens and Angrist 1994).Specifically, we replace Treatmentj in equations (1) and (2) with Lessonsk

i , whichequals 1 if student i took at least k lessons during the intervention period. Weuse Treatmentj as an instrument for Lessonsk

i to estimate the program impact forstudents in compliance by changing the threshold number of lessons. Since theassignment of treatment was randomized and the control students could not takeany lessons, Treatmentj works as a valid instrument. We, however, suffer from theweak instrument problem since the take-up rate was not high. To correct for thisproblem, we report the 95% CIs based on the wild cluster bootstrap because italso corrects for weak instruments (Roodman et al. 2019). In addition, we performthe conditional likelihood ratio tests developed by Moreira (2003), using condivregStata command by Moreira and Poi (2003) for robustness check.

D. Econometric Analyses: Intention to Treat

Table 5 shows the ITT estimates of the program impacts. Odd-numberedcolumns present the ANCOVA estimation results based on equation (1), whileeven-numbered columns present the DiD results based on equation (2). Panel Apresents the estimated impacts on the attitude measures. Column 2 shows thepositive and significant coefficients of the treatment on the total internationalposture score and the wild cluster bootstrap CI excludes 0, supporting our

12According to McKenzie (2012), ANCOVA analysis would be beneficial in power rather than DiD analysiswhen autocorrelations are low. The autocorrelation in our analysis ranged from 0.4 to 0.8, which is neither high norlow. We thus provide the results from both the ANCOVA and DiD analyses in Table 5.


Tabl

e5.

Impa

cts

ofO

nlin

eP

rogr

am(I

nten

tion

-to-

Tre

atE

stim

atio

n)

A.A

ttit

udes

(1)

(2)

(3)

(4)

(5)

(6)

Tot

alIn

tern

atio

nalP

ostu

reW

illin

gnes

sto

Com

mun

icat

eC

ambo

dia

Tou

r(1

=ye

s)

AN

CO

VAD

iDA

NC

OVA

DiD

OL

SO

LS

(wit

hco

ntro

l)a

Tre

atm

ent(

1=

yes)

0.15

*0.

130.

033

0.03

9(1

.88)

(1.4

5)(1

.05)

(1.1

4)T

reat

men

t×E

ndli

ne(1

=ye

s)0.

120.

078

(1.4

1)(0

.81)

Bas

elin

eou

tcom

e0.

78**

*0.

64**

*

(22.

61)

(13.

96)

End

line

(1=

yes)

−0.1

1**−0

.23**

*

(−2.

09)

(−3.

16)

Eng

lish

teac

her

B(1

=ye

s)0.

110.

014

−0.0

28−0

.049

(0.9

4)(0

.11)

(−0.

59)

(−0.

95)

Eng

lish

teac

her

C(1

=ye

s)−0

.13

−0.1

6−0

.053

−0.0

70(−

1.24

)(−

1.22

)(−

1.17

)(−

1.39

)E

ngli

shte

ache

rD

(1=

yes)

−0.2

1*−0

.14

−0.0

43−0

.071

(−1.

80)

(−1.

13)

(−0.

92)

(−1.

40)

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[0.0

90.

20]

[−0.

090.

34]

[−0.

050.

32]

[−0.

140.

31]

[−0.

010.

08]

[0.0

00.

08]

No.

ofob

serv

atio

ns30

862

730

362

232

029

2


B.I

nter

nati

onal

post

ure

(sub

com

pone

nt)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Sub

1.In

terc

ultu

ral

Sub

2.In

tern

atio

nal

Sub

3.Su

b4.

Ori

enta

tion

Voc

atio

nD

iffe

rent

Cus

tom

sFo

reig

nA

ffai

rs

AN

CO

VAD

iDA

NC

OVA

DiD

AN

CO

VAD

iDA

NC

OVA

DiD

Tre

atm

ent(

1=

yes)

0.01

90.

17**

0.02

00.

25**

(0.2

1)(2

.21)

(0.1

9)(2

.56)

Tre

atm

ent×

End

line

(1=

yes)

−0.0

100.

16*

−0.0

087

0.18

*

(−0.

11)

(1.8

7)(−

0.07

)(1

.71)

Bas

elin

eou

tcom

e0.

70**

*0.

73**

*0.

41**

*0.

57**

*

(16.

10)

(19.

61)

(7.4

4)(1

2.01

)E

ndli

ne(1

=ye

s)−0

.12*

−0.1

2**−0

.012

−0.0

069

(−1.

82)

(−2.

08)

(−0.

13)

(−0.

10)

Eng

lish

teac

her

B(1

=ye

s)0.

043

0.05

10.

220.

095

(0.3

5)(0

.47)

(1.4

9)(0

.69)

Eng

lish

teac

her

C(1

=ye

s)−0

.11

−0.0

21−0

.077

−0.1

3(−

0.89

)(−

0.19

)(−

0.55

)(−

1.00

)E

ngli

shte

ache

rD

(1=

yes)

−0.1

5−0

.080

−0.1

8−0

.18

(−1.

19)

(−0.

68)

(−1.

22)

(−1.

31)

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[−0.

310.

34]

[−0.

260.

21]

[0.1

30.

22]

[0.0

10.

33]

[−0.

180.

21]

[−0.

250.

25]

[0.0

40.

48]

[−0.

010.

39]

No.

ofob

serv

atio

ns30

962

830

962

830

862

730

962

8


C.E

nglis

hco

mm

unic

atio

nab

iliti

es

(1)

(2)

(3)

(4)

(5)

(6)

(i)

Tot

alV

ersa

nt(i

i)B

enes

se(i

ii)T

otal

GT

EC

AN

CO

VAD

iDA

NC

OVA

DiD

OL

SO

LS

(wit

hco

ntro

l)a

Tre

atm

ent(

1=

yes)

0.09

9−0

.005

10.

0034

−0.0

035

(1.0

9)(−

0.06

)(0

.03)

(−0.

03)

Tre

atm

ent×

End

line

(1=

yes)

0.04

20.

018

(0.4

4)(0

.20)

Bas

elin

eou

tcom

e0.

78**

*0.

71**

*

(17.

38)

(16.

72)

End

line

(1=

yes)

0.52

***

−0.0

13(8

.41)

(−0.

21)

Eng

lish

teac

her

B(1

=ye

s)−0

.27**

0.00

900.

052

0.02

0(−

2.13

)(0

.08)

(0.3

1)(0

.12)

Eng

lish

teac

her

C(1

=ye

s)−0

.31**

−0.0

056

0.00

97−0

.081

(−2.

55)

(−0.

05)

(0.0

6)(−

0.50

)E

ngli

shte

ache

rD

(1=

yes)

−0.4

4***

−0.2

4*−0

.19

−0.2

6(−

3.37

)(−

1.96

)(−

1.16

)(−

1.50

)

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[−0.

090.

32]

[−0.

290.

36]

[−0.

220.

19]

[−0.

210.

25]

[−0.

120.

12]

[−0.

150.

14]

No.

ofob

serv

atio

ns24

355

331

262

731

829

1


D.G

TE

Cte

st(s

ubco

mpo

nent

)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Sub

1.R

eadi

ngSu

b2.

Lis

teni

ngSu

b3.

Wri

ting

Sub

4.Sp

eaki

ng

OL

SO

LS

OL

SO

LS

(wit

h(w

ith

(wit

h(w

ith

OL

Sco

ntro

l)a

OL

Sco

ntro

l)a

OL

Sco

ntro

l)a

OL

Sco

ntro

l)a

Tre

atm

ent(

1=

yes)

−0.0

23−0

.011

0.06

70.

038

0.04

80.

061

−0.0

21−0

.012

(−0.

20)

(−0.

10)

(0.6

0)(0

.32)

(0.4

3)(0

.50)

(−0.

19)

(−0.

10)

Eng

lish

teac

her

B(1

=ye

s)0.

0040

−0.0

100.

200.

17−0

.10

−0.1

40.

020

−0.0

037

(0.0

2)(−

0.06

)(1

.19)

(0.9

5)(−

0.71

)(−

0.87

)(0

.12)

(−0.

02)

Eng

lish

teac

her

C(1

=ye

s)0.

053

−0.0

490.

026

−0.0

20−0

.15

−0.1

30.

051

−0.0

53(0

.35)

(−0.

31)

(0.1

6)(−

0.12

)(−

0.93

)(−

0.74

)(0

.33)

(−0.

33)

Eng

lish

teac

her

D(1

=ye

s)−0

.20

−0.2

8*−0

.031

−0.0

77−0

.12

−0.1

1−0

.22

−0.2

9*

(−1.

28)

(−1.

69)

(−0.

19)

(−0.

44)

(−0.

78)

(−0.

66)

(−1.

34)

(−1.

71)

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[−0.

190.

17]

[−0.

120.

10]

[−0.

110.

27]

[−0.

110.

20]

[−0.

480.

57]

[−0.

480.

60]

[−0.

180.

18]

[−0.

140.

11]

No.

ofob

serv

atio

ns26

524

355

326

524

355

326

524

3

AN

CO

VA

=an

alys

isof

cova

rian

ce,C

I=

confi

denc

ein

terv

al,D

iD=

diff

eren

ce-i

n-di

ffer

ence

s,G

TE

C=

Glo

bal

Test

ofE

ngli

shC

omm

unic

atio

n,IT

T=

inte

ntio

nto

trea

t,O

LS

=or

dina

ryle

asts

quar

es.

Not

es:

Est

imat

edco

effi

cien

tsar

ere

port

ed.

***,

**,

and

*in

dica

te1%

,5%

,an

d10

%le

vels

ofst

atis

tica

lsi

gnifi

canc

e,re

spec

tivel

y.N

umbe

rsin

pare

nthe

ses

are

t-st

atis

tics

base

don

hete

rosc

edas

tici

ty-r

obus

tsta

ndar

der

rors

.a In

OL

S(w

ith

cont

rol)

,th

eco

ntro

lva

riab

les

inco

lum

n2

ofTa

ble

4ar

ead

ded.

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

isfo

rth

etr

eatm

ent

orth

etr

eatm

ent×

endl

ine

vari

able

.U

sing

boot

test

Sta

taco

mm

and

deve

lope

dby

Roo

dman

etal

.(20

19),

we

impl

emen

ted

wil

dcl

uste

rbo

otst

rapp

ing

wit

h1,

000

repl

icat

ions

.In

sodo

ing,

we

used

the

gam

ma

dist

ribu

tion

wit

hth

esh

ape

para

met

erof

4an

dth

esc

ale

para

met

erof

0.5

asw

eigh

tfor

cons

truc

ting

the

boot

stra

psa

mpl

es.

Sou

rce:

Aut

hors

’ca

lcul

atio

ns.


discussion in the previous section. In the DiD estimation reported in column 2, theimpact is positive but insignificant although the t-statistic is as large as 1.41, withthe corresponding p-value of 0.148 (not reported). The point estimate is 0.12 andthat of Endline is −0.11, which is statistically significant; these coefficients suggestthat the overall international posture score declined from the baseline survey inJune 2015 to the endline survey in December of the same year, but the Skypeprogram offset the declining international posture score among the treated students.Furthermore, the significant teacher dummy suggests the presence of substantialteacher heterogeneity, as discussed in section III.B.

We report our results on WTC in columns 3 and 4. While not statisticallysignificant, the point estimate is positive in both the ANCOVA and DiD estimations.In columns 5 and 6, we report results on the Cambodia tour. The point estimate isnot significant, but the CI barely includes 0 in column 5 and excludes 0 in column6. Hence, the treated students were more likely to have voluntarily applied for theopportunity to study abroad.

Panel B shows positive and significant impacts on subcomponents 2(columns 3 and 4) and 4 (columns 7 and 8). The CIs for these two subcomponentsexclude 0 (except for column 8, where the CI barely includes 0). With the pointestimates for subcomponents 1 and 3 being close to 0, the impact on internationalposture comes from the changes in subcomponents 2 and 4. In particular, we findthat while the Grade 10 students tended to become less interested in an internationalvocation—the size of the effect being 0.12 standard deviations (see column 4)—such a tendency was compensated for by our intervention.

Panel C of Table 5 shows the ITT estimates of the program impacts onstudents’ English communication abilities in the same manner as panel A. The pointestimates are small or even negative, particularly for the Benesse (columns 3 and4) and GTEC tests (columns 5 and 6), and the corresponding t-statistics are closeto 0. In addition, all the CIs include zero. Even if we look at the subcomponentsof the GTEC shown in panel D, particularly subcomponents 2 (listening) and 4(speaking), we find similar patterns of small coefficients with small t-statistics andCIs including zero. Hence, our regression analyses show that the Skype programhad limited impacts on the students’ English communication abilities.

However, attitudinal attributes have been reported to lead to eventualimprovement in students’ second-language skills (e.g., Sasaki 2011, Yashima 2002);therefore, the Skype program may have significant impacts over the long term.Unfortunately, all of the sample students had received the same amount of onlineintervention by the end of May 2016, and thus, we do not have variation to evaluatesuch long-term impacts. In addition, we may possibly have detected an effect ifour intervention had been implemented for a longer period because Ross (2000),among others, finds that the duration is a major determinant of the effectiveness ofsecond-language learning. Another important point to note from panel C is thesignificant coefficient of the endline dummy in column 2. As the scores of


the standardized Versant test are intertemporally comparable, the positive andsignificant coefficients suggest that students’ communication abilities significantlyimproved over time, most likely due to the regular school curriculum in this top-tierhigh school.

E. Econometric Analyses: Local Average Treatment Effect

Table 6 reports the LATE estimates of program impacts on attitudes in panelA and on English communication abilities in panel B. In columns 1, 4, and 7 (wherek = 5), the lesson dummy equals 1 if a student took at least five lessons in theintervention period; thus, the coefficient captures the impacts of the online programfor students who completed at least five lessons.

In panel A, the size of the coefficient increases with k, indicating that thestudents who took more lessons benefited more from the program. For instance,the students who took 25 or more lessons (half of the recommended number bythe service provider) have an international posture z-score that is 1.01 standarddeviation higher than the average of the control students (column 3). However, thefirst-stage F-statistics decrease and the CIs widen as k increases because only 23students (14%) completed 25 or more lessons, and the standard errors increase withk. This is one of the reasons why we do not find statistically significant coefficientsfor WTC (columns 4–6). In columns 7–9, although the coefficient is insignificant,CIs exclude or barely include zero, indicating the positive impact on students’participation in the overseas study.

In panel B, we find a similar increasing pattern for the Versant test (columns1–3), but not for the Benesse test (columns 4–6) or the overall GTEC scores(columns 7–9). Unfortunately, none of the three indicators are a perfect measure ofEnglish communication abilities: (i) the Versant test with the nonrandom attrition,(ii) the Benesse test with the primary focus on reading skills, and (iii) the GTECwith the cross-sectional nature. Our tentative conclusion is that the impacts of ourintervention on English communication abilities were at most limited.

F. Additional Analyses

We conducted two sets of additional analyses. First, we analyzed theheterogeneous treatment effects by interacting the treatment dummy with thecontrol variables, including procrastination, gender, past exposure to English,family background, and baseline levels of the outcome variable. Panel A of Table 7reports results for the international posture score; no interaction term is statisticallysignificant, including those not reported (Table 7 only reports the results for thevariables that were found to be correlated with some outcome variables in AppendixTable A1.) This may be because of the moderate size of the average treatmenteffects. Panel B reports the results for the Benesse test score. We found that


Tabl

e6.

Impa

cts

ofO

nlin

eP

rogr

am(L

ocal

Ave

rage

Tre

atm

ent

Eff

ect

Est

imat

ion)

A.A

ttit

udes

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Tot

alIn

tern

atio

nalP

ostu

reW

illin

gnes

sto

Com

mun

icat

eC

ambo

dia

Tou

r(1

=ye

s)

IVIV

IVIV

IVIV

IVIV

IVk

=5

k=

10k

=25

k=

5k

=10

k=

25k

=5

k=

10k

=25

Les

son

atle

ast

kti

mes

0.29

*0.

45*

1.01

*0.

240.

380.

840.

063

0.10

0.23

(1.8

9)(1

.88)

(1.8

4)(1

.46)

(1.4

5)(1

.43)

(1.0

6)(1

.05)

(1.0

4)B

asel

ine

outc

ome

0.77

***

0.77

***

0.79

***

0.65

***

0.65

***

0.64

***

(22.

34)

(22.

30)

(22.

07)

(14.

34)

(13.

98)

(13.

78)

Teac

her

(str

ata)

dum

mie

sY

YY

YY

YY

YY

Firs

t-st

age

F-s

tati

stic

s37

.215

.15.

140

.715

.55.

447

.518

.96.

7

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[0.0

70.

47]

[0.0

40.

74]

[0.2

61.

70]

[−0.

030.

56]

[−0.

190.

96]

[−0.

071.

95]

[0.0

10.

10]

[−0.

000.

22]

[0.0

20.

47]

Con

diti

onal

LR

test

(95%

CI)

[−0.

010.

60]

[−0.

020.

96]

[−0.

052.

34]

[−0.

080.

57]

[−0.

140.

92]

[−0.

312.

24]

[−0.

060.

18]

[−0.

090.

31]

[−0.

210.

72]

No.

ofob

serv

atio

ns30

830

830

830

330

330

332

032

032

0


B.E

nglis

hco

mm

unic

atio

nab

iliti

es

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(i)

Ver

sant

(ii)

Ben

esse

(iii)

Tot

alG

TE

C

IVIV

IVIV

IVIV

IVIV

IVk

=5

k=

10k

=25

k=

5k

=10

k=

25k

=5

k=

10k

=25

Les

son

atle

ast

kti

mes

0.18

0.32

0.79

−0.0

096

−0.0

16−0

.034

0.00

660.

011

0.02

3(1

.10)

(1.0

9)(1

.06)

(−0.

06)

(−0.

06)

(−0.

06)

(0.0

3)(0

.03)

(0.0

3)B

asel

ine

outc

ome

0.77

***

0.77

***

0.78

***

0.70

***

0.70

***

0.70

***

(16.

61)

(16.

85)

(17.

56)

(17.

08)

(17.

06)

(16.

94)

Teac

her

(str

ata)

dum

mie

sY

YY

YY

YY

YY

Firs

t-st

age

F-s

tati

stic

s31

.210

.83.

137

.314

.45.

448

.719

.06.

7

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[−0.

200.

50]

[−0.

620.

98]

[−0.

922.

35]

[−0.

370.

40]

[−0.

640.

60]

[−1.

261.

24]

[−0.

230.

39]

[−0.

400.

58]

[−0.

851.

41]

Con

diti

onal

LR

test

(95%

CI)

[−0.

140.

51]

[−0.

240.

93]

[−0.

622.

67]

[−0.

320.

29]

[−0.

530.

49]

[−1.

191.

10]

[−0.

430.

43]

[−0.

700.

70]

[−1.

641.

63]

No.

ofob

serv

atio

ns24

324

324

331

231

231

231

831

831

8

CI=

confi

denc

ein

terv

al,G

TE

C=

Glo

balT

esto

fE

ngli

shC

omm

unic

atio

n,IV

=in

stru

men

talv

aria

ble,

LR

=li

keli

hood

rati

o.N

otes

:E

stim

ated

coef

fici

ents

are

repo

rted

.**

*,

**,

and

*in

dica

te1%

,5%

,an

d10

%le

vels

ofst

atis

tica

lsi

gnifi

canc

e,re

spec

tivel

y.N

umbe

rsin

pare

nthe

ses

are

t-st

atis

tics

base

don

hete

rosc

edas

tici

ty-r

obus

tst

anda

rder

rors

.In

the

IVes

tim

atio

ns,

the

less

ondu

mm

yeq

uals

1if

ast

uden

tto

okat

leas

tk

less

ons,

and

this

dum

my

vari

able

isin

stru

men

ted

wit

hth

etr

eatm

ent

dum

my.

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

isfo

rth

ele

sson

atle

ast

kti

mes

vari

able

.Usi

ngbo

otte

stS

tata

com

man

dde

velo

ped

byR

oodm

anet

al.(

2019

),w

eim

plem

ente

dw

ild

clus

terb

oots

trap

ping

wit

h1,

000

repl

icat

ions

.In

sodo

ing,

we

used

the

gam

ma

dist

ribu

tion

wit

hth

esh

ape

para

met

erof

4an

dth

esc

ale

para

met

erof

0.5

asw

eigh

tfor

cons

truc

ting

the

boot

stra

psa

mpl

es.C

ondi

tion

alL

Rte

stis

for

the

less

onat

leas

tkti

mes

vari

able

,com

pute

dus

ing

cond

ivre

gS

tata

com

man

dde

velo

ped

byM

orei

raan

dPo

i(20

03).

Sou

rce:

Aut

hors

’ca

lcul

atio

ns.


Tabl

e7.

Het

erog

eneo

usT

reat

men

tE

ffec

t

A.I

nter

nati

onal

post

ure

(1)

(2)

(3)

(4)

(5)

(6)

Bas

elin

eM

ale

Bee

nA

broa

dSp

orts

Clu

bN

umbe

rIn

tern

atio

nal

XP

rocr

asti

nati

on(1

=ye

s)(1

=ye

s)(1

=ye

s)of

Boo

ksP

ostu

reSc

ore

Tre

atm

ent(

1=

yes)

0.14

*0.

21*

0.24

**0.

120.

120.

15*

(1.8

3)(1

.95)

(2.4

1)(0

.95)

(0.7

7)(1

.87)

Tre

atm

ent×

X−0

.053

−0.1

2−0

.24

0.04

1−0

.002

2−0

.11

(−0.

63)

(−0.

72)

(−1.

46)

(0.2

4)(−

0.04

)(−

1.52

)X

−0.0

70−0

.053

0.32

***

−0.0

290.

047

N.A

.(sa

me

as(−

1.31

)(−

0.50

)(2

.75)

(−0.

28)

(1.4

1)ba

seli

neou

tcom

e)B

asel

ine

outc

ome

0.78

***

0.77

***

0.75

***

0.79

***

0.80

***

0.83

***

(22.

53)

(21.

73)

(19.

62)

(22.

37)

(23.

50)

(17.

58)

Eng

lish

teac

her

B0.

130.

110.

091

0.13

0.09

70.

11(1

=ye

s)(1

.15)

(0.9

7)(0

.81)

(1.1

1)(0

.88)

(0.9

4)E

ngli

shte

ache

rC

−0.1

3−0

.14

−0.1

2−0

.10

−0.0

92−0

.12

(1=

yes)

(−1.

18)

(−1.

24)

(−1.

13)

(−0.

97)

(−0.

86)

(−1.

15)

Eng

lish

teac

her

D−0

.20*

−0.2

1*−0

.22**

−0.1

9*−0

.18

−0.2

1*

(1=

yes)

(−1.

78)

(−1.

83)

(−1.

98)

(−1.

69)

(−1.

55)

(−1.

85)

No.

ofob

serv

atio

ns30

330

830

830

330

130

8


B.B

enes

sesc

ore

(1)

(2)

(3)

(4)

(5)

(6)

Mal

eB

een

Abr

oad

Spor

tsC

lub

Num

ber

Bas

elin

eX

Pro

cras

tina

tion

(1=

yes)

(1=

yes)

(1=

yes)

ofB

ooks

Ben

esse

Scor

e

Tre

atm

ent(

1=

yes)

−0.0

26−0

.055

−0.1

3−0

.054

−0.1

7−0

.005

0(−

0.32

)(−

0.52

)(−

1.29

)(−

0.39

)(−

1.12

)(−

0.06

)T

reat

men

t×X

−0.1

00.

099

0.31

*0.

044

0.05

1−0

.019

(−1.

27)

(0.6

2)(1

.89)

(0.2

6)(0

.94)

(−0.

23)

X−0

.006

7−0

.068

−0.1

0−0

.016

0.00

67N

.A.(

sam

eas

(−0.

11)

(−0.

60)

(−0.

82)

(−0.

13)

(0.1

6)ba

seli

neou

tcom

e)B

asel

ine

outc

ome

0.69

***

0.70

***

0.70

***

0.69

***

0.68

***

0.72

***

(15.

75)

(16.

47)

(16.

62)

(15.

74)

(15.

18)

(11.

82)

Eng

lish

teac

her

B−0

.033

0.00

890.

022

−0.0

21−0

.025

0.00

95(1

=ye

s)(−

0.28

)(0

.08)

(0.1

8)(−

0.17

)(−

0.20

)(0

.08)

Eng

lish

teac

her

C−0

.067

−0.0

045

0.00

15−0

.036

−0.0

23−0

.007

0(1

=ye

s)(−

0.55

)(−

0.04

)(0

.01)

(−0.

30)

(−0.

20)

(−0.

06)

Eng

lish

teac

her

D−0

.29**

−0.2

4*−0

.22*

−0.2

7**−0

.26**

−0.2

5**

(1=

yes)

(−2.

26)

(−1.

93)

(−1.

74)

(−2.

14)

(−2.

08)

(−1.

98)

No.

ofob

serv

atio

ns30

631

230

930

530

331

2

X=

cont

rolv

aria

bles

(i.e

.,pr

ocra

stin

atio

n,m

ale,

been

abro

ad,s

port

scl

ub,n

umbe

rof

book

s,an

dba

seli

neB

enes

sesc

ore)

.N

otes

:E

stim

ated

coef

fici

ents

are

repo

rted

.***,**

,and

*in

dica

te1%

,5%

,and

10%

leve

lsof

stat

isti

cal

sign

ifica

nce,

resp

ectiv

ely.

Num

bers

inpa

rent

hese

sar

et-

stat

isti

csba

sed

onhe

tero

sced

asti

city

-rob

usts

tand

ard

erro

rs.

Sou

rce:

Aut

hors

’ca

lcul

atio

ns.


only the interaction with the abroad dummy is positive and marginally significant,suggesting that the program may have widened the gap between strongly performingstudents with greater degrees of international exposure and those showing nosuch orientation because the former is more likely to take advantage of learningopportunities to further improve their English communication abilities.

The second set of analyses is the impact of the Skype program on thestudents’ school performance based on their self-reported information. Whileadmitting that we do not have more objective data based on assessments by theirteachers, the treated students were more likely to work hard and actively participatein English classes at school (Table 8, columns 1–4). In addition, the treated studentsmay be more likely to work hard in classes other than English classes (columns5–6). Therefore, the program had positive impacts on overall school performance.In addition, the possibility of a crowding-out effect, where the students spend moretime studying English while spending less time on other subjects, seems limited.

IV. Conclusion

We conducted a unique and rare field experiment in collaboration with aJapanese public high school to provide students with a home-use, ICT-assistedprogram for English. Through the examination of program usage records and paneldata, we analyzed the factors associated with program utilization and estimatedthe program impacts. In our descriptive and econometric analyses, we found thatthe program significantly changed the internationally oriented attitudes of thetreated students but not their English communication abilities. We could justifiablyspeculate that the insignificant improvement in their communication abilities wasdue to the low take-up rate of the targeted program. As we found that studentsshowing a tendency to procrastinate were less likely to start and continue using theprogram, more research is warranted on how to improve and maintain students’motivation, particularly those with a tendency to procrastinate, and encourage themto use ICT-assisted programs such as the one targeted in this study. In addition, asimproved internationally oriented attitudes could have a positive impact on students’English development on a long-term basis, future studies need to evaluate thelong-term impacts of such programs.

We also found that although the entrance-exam-oriented regular schoolcurriculum did improve the students’ English (oral) communication abilities, itseemed to have negative effects on their international orientation. As we identifiedthe positive causal effects of the online English learning program on the students’attitudes, given that it supplemented the weaknesses of the regular curriculum,future research should consider how to combine regular English lessons andsuch ICT-based programs in a complementary manner. In addition to encouraginginterventions designed to encourage home use, using such programs during regularEnglish lessons also might be an option.


Tabl

e8.

Impa

cts

onSe

lf-R

epor

ted

Scho

olP

erfo

rman

ce(I

nten

tion

-to-

Tre

atE

stim

atio

n)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Iw

ork

hard

Iex

pres

sm

yop

inio

nI

wor

kha

rdI

expr

ess

my

opin

ion

inE

nglis

hcl

asse

sin

Eng

lish

clas

ses

inot

her

clas

ses

inot

her

clas

ses

OL

SO

LS

OL

SO

LS

OL

SO

LS

OL

SO

LS

Tre

atm

ent(

1=

yes)

0.29

***

0.24

***

0.22

*0.

25**

0.27

***

0.25

***

0.02

50.

13(3

.10)

(2.9

7)(1

.93)

(2.5

6)(2

.73)

(2.6

9)(0

.22)

(1.2

7)B

asel

ine

outc

ome

0.53

***

0.53

***

0.41

***

0.48

***

(11.

40)

(10.

83)

(8.1

5)(8

.89)

Eng

lish

teac

her

B(1

=ye

s)0.

140.

077

0.05

30.

017

−0.0

130.

016

0.17

0.15

(1.0

2)(0

.69)

(0.3

0)(0

.12)

(−0.

09)

(0.1

3)(0

.97)

(1.0

4)E

ngli

shte

ache

rC

(1=

yes)

−0.0

51−0

.098

−0.1

4−0

.071

−0.1

3−0

.13

−0.0

770.

0066

(−0.

38)

(−0.

88)

(−0.

78)

(−0.

48)

(−0.

93)

(−1.

03)

(−0.

44)

(0.0

5)E

ngli

shte

ache

rD

(1=

yes)

−0.0

30−0

.056

−0.2

4−0

.28**

−0.0

94−0

.081

−0.0

77−0

.21

(−0.

21)

(−0.

47)

(−1.

44)

(−1.

99)

(−0.

66)

(−0.

61)

(−0.

45)

(−1.

42)

Con

trol

mea

nat

base

line

4.0

3.0

3.9

3.2

Wil

dcl

uste

rbo

otst

rap

(95%

CI)

[0.1

10.

48]

[0.0

80.

40]

[−0.

000.

44]

[0.0

60.

44]

[0.0

70.

48]

[0.0

70.

43]

[−0.

190.

24]

[−0.

070.

34]

No.

ofob

serv

atio

ns31

030

230

929

931

030

230

829

8

CI=

confi

denc

ein

terv

al,O

LS

=or

dina

ryle

asts

quar

es.

Not

es:

All

outc

omes

are

mea

sure

dus

ing

afiv

e-po

int

Lik

ert

scal

ew

ith

1=

not

atal

l,2

=no

,3=

neut

ral,

4=

yes,

and

5=

defi

nite

lyye

s.E

stim

ated

coef

fici

ents

repo

rted

.***,**

,an

d*

indi

cate

1%,5

%,a

nd10

%le

vels

ofst

atis

tica

lsig

nifi

canc

e,re

spec

tivel

y.N

umbe

rsin

pare

nthe

ses

are

t-st

atis

tics

base

don

hete

rosc

edas

tici

ty-r

obus

tsta

ndar

der

rors

.Wil

dcl

uste

rbo

otst

rap

(95%

CI)

isfo

rth

etr

eatm

ent

vari

able

.Usi

ngbo

otte

stS

tata

com

man

dde

velo

ped

byR

oodm

anet

al.(

2019

),w

ere

peat

edw

ild

clus

ter

boot

stra

ppin

gfo

r1,

000

tim

es.I

nso

doin

g,w

eus

edth

ega

mm

adi

stri

buti

onw

ith

the

shap

epa

ram

eter

of4

and

the

scal

epa

ram

eter

of0.

5fo

rw

eigh

ts.

Sou

rce:

Aut

hors

’ca

lcul

atio

ns.


References

Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: AnEmpiricist’s Companion . Princeton: Princeton University Press.

Bulman, George, and Robert W. Fairlie. 2016. “Technology and Education: Computers, Software,and the Internet.” In Handbook of the Economics of Education, Volume 5, edited by Eric A.Hanushek, Stephen J. Machin, and Ludger Woessmann, 239–80. Amsterdam: Elsevier.

Cameron, Colin A., Jonah B. Gelbach, and Douglas L. Miller. 2008. “Bootstrap-BasedImprovements for Inference with Clustered Errors.” Review of Economics and Statistics90 (3): 414–27.

Dörnyei, Zoltán, and Stephen Ryan. 2015. The Psychology of the Language Learner Revisited.London: Routledge.

Duckworth, Angela L., Katherine L. Milkman, and David Laibson. 2018. “Beyond Willpower:Strategies for Reducing Failures of Self-Control.” Psychological Science in the PublicInterest 19 (3): 102–29.

Gee, James P., and Elisabeth R. Hayes. 2011. Language and Learning in the Digital Age. London:Routledge.

Glewwe, Paul, Michael Kremer, Sylvie Moulin, and Eric Zitzewitz. 2004. “Retrospective vs.Prospective Analyses of School Inputs: The Case of Flip Charts in Kenya.” Journal ofDevelopment Economics 74 (1): 251–68.

Golsteyn, Bart H. H., Hans Grönqvist, and Lena Lindahl. 2014. “Adolescent Time PreferencesPredict Lifetime Outcomes.” Economic Journal 124 (580): F739–F761.

Honda, Yuki, and Hiroshi Nishijima. 2007. “Survey of Lifestyles, Behaviors, and Attitudesof High-Schoolers in Tokyo.” (Japanese). http://berd.benesse.jp/berd/center/open/report/toritsu_kousei/2009/pdf/siryou_01.pdf (accessed April 15, 2019).

Imbens, Guido M., and Angrist, Joshua D. 1994. “Identification and Estimation of Local AverageTreatment Effects.” Econometrica 62 (2): 467–76.

Imbens, Guido M., and Jeffrey M. Wooldridge. 2009. “Recent Developments in the Econometricsof Program Evaluation.” Journal of Economic Literature 47 (1): 5–86.

Kawaguchi, Daiji. 2016. “Fewer School Days, More Inequality.” Journal of the Japanese andInternational Economies 39: 35–52.

Kim, Kimin, and Myoung-jae Lee. 2019. “Difference in Differences in Reverse.” EmpiricalEconomics 57 (3): 705–25.

Levy, Mike. 2009. “Technologies in Use for Second Language Learning.” Modern LanguageJournal 93 (s1): 769–82.

MacIntyre, Peter D. 2007. “Willingness to Communicate in the Second Language: Understandingthe Decision to Speak as a Volitional Process.” Modern Language Journal 91 (4): 564–76.

McKenzie, David. 2012. “Beyond Baseline and Follow-up: The Case for More T in Experiments.”Journal of Development Economics 99 (2): 210–21.

Ministry of Education, Culture, Sports, Science and Technology, Government of Japan (MEXT).2015a. “Results of the English Test Conducted to Improve English Education inJapan in 2015.” (Japanese). http://www.mext.go.jp/a_menu/kokusai/gaikokugo/1358258.htm (accessed April 15, 2019).

______. 2015b. “Selection of 2015 Super-Global High Schools.” (Japanese). http://www.mext.go.jp/a_menu/kokusai/sgh/1356366.htm (accessed April 21, 2019).

Moreira, Marcelo J. 2003. “A Conditional Likelihood Ratio Test for Structural Models.”Econometrica 71 (4): 1027–48.


Moreira, Marcelo J., and Brian P. Poi. 2003. “Implementing Tests with Correct Size in theSimultaneous Equations Model.” Stata Journal 3 (1): 57–70.

Onji, Kazuki, and Rina Kikuchi. 2011. “Procrastination, Prompts, and Preferences: Evidencefrom Daily Records of Self-Directed Learning Activities.” Journal of Socio-Economics 40(6): 929–41.

Ortega, Lourdes, and Gina Iberri-Shea. 2005. “Longitudinal Research in Second LanguageAcquisition: Recent Trends and Future Direction.” Annual Review of Applied Linguistics25: 26–45.

Osaka University. 2013. “Survey of Preference Parameters at Osaka University.” (Japanese).http://www.iser.osakau.ac.jp/survey_data/doc/japan/questionnaire/japanese/2013QuestionnaireJAPAN.pdf (accessed April 15, 2019).

Pearson Inc. 2008. “Consistency of Versant English Test Scores Over Multiple Administrators.”Unpublished.

Roodman, David, James G. MacKinnon, Morten Ørregaard Nielsen, and Matthew D. Webb. 2019.“Fast and Wild: Bootstrap Inference in Stata Using Boottest.” Stata Journal 19 (1): 4–60.

Ross, Steven. 2000. “Individual Differences and Learning Outcomes on the Certificate of Spokenand Written English.” In Studies in Immigrant English Language Assessment, edited byGeoff Brindley, 191–214. Sydney: NCELTR.

Sasaki, Miyuki. 2011. “Effects of Varying Lengths of Study-Abroad Experiences on JapaneseEFL Students’ L2 Writing Ability and Motivation: A Longitudinal Study.” TESOLQuarterly 45 (1): 81–105.

______. 2018. “Application of Diffusion of Innovation Theory to Educational Accountability:The Case of EFL Education in Japan.” Language Testing in Asia 8 (1): 1–18.

Snilstveit, Birte, Jennifer Stevenson, Radhika Menon, Daniel Phillips, Emma Gallagher, MaisieGeleen, Hannah Jobse, Tanja Schmidt, and Emmanuel Jimenez. 2016. The Impact ofEducation Programmes on Learning and School Participation in Low- and Middle-IncomeCountries: A Systematic Review Summary Report. 3ie Systematic Review Summary 7.London: International Initiative for Impact Evaluation (3ie).

Yashima, Tomoko. 2002. “Willingness to Communicate in a Second Language: The JapaneseEFL Context.” Modern Language Journal 86 (1): 54–66.

______. 2009. “International Posture and the Ideal L2 Self in the Japanese EFL Context.” InMotivation, Language Identity, and the L2 Self, edited by Zoltán Dörnyei and Ema Ushioda,144–63. Clevedon: Multilingual Matters.

______. 2009. Kokusai. http://www2.ipcku.kansai-u.ac.jp/∼yashima/data/kokusai.pdf (accessedApril 15, 2019).

______. 2009. WTC Scale. http://www2.ipcku.kansai-u.ac.jp/∼yashima/data/wtc_scale.pdf(accessed April 15, 2019).

Yashima, Tomoko, Lori Zenuk-Nshide, and Kazuaki Shimizu. 2004. “The Influence of Attitudesand Affect on Willingness to Communicate and Second Language Communication.”Language Learning 54 (1): 119–52.


Appendix

Table A1. Baseline Correlates of Outcome Variables (Ordinary Least Squares Estimation)

(1) (2) (3) (4)Total Willingness

International to Versant BenessePosture Communicate Score Score

Treatment 0.074 0.017 0.12 −0.10(1 = yes) (0.65) (0.14) (1.08) (−0.88)Procrastination −0.096 −0.15*** −0.077 −0.067[z-score] (−1.58) (−2.67) (−1.17) (−1.25)Male −0.22* 0.030 0.28 0.20(1 = yes) (−1.85) (0.25) (1.54) (1.37)English since Grade 3 or 4 0.051 −0.10 −0.013 −0.10(1 = yes) (0.41) (−0.84) (−0.09) (−0.82)English since Grade 5 or later 0.012 −0.19 −0.095 −0.23(1 = yes) (0.08) (−1.11) (−0.62) (−1.45)Been abroad 0.62*** 0.43*** 0.29** 0.0054(1 = yes) (5.50) (3.62) (2.11) (0.04)Own room 0.13 0.36** 0.15 −0.13(1 = yes) (0.65) (2.25) (0.97) (−1.00)Own personal computer 0.35* 0.094 0.62 0.34(1 = yes) (1.79) (0.46) (1.55) (1.42)Own tablet 0.10 0.19 0.085 0.16(1 = yes) (0.69) (1.34) (0.41) (1.11)Commuting time 21–40 minutes −0.20 −0.092 −0.042 0.13(1 = yes) (−1.36) (−0.66) (−0.19) (0.74)Commuting time 41–60 minutes −0.12 −0.058 −0.086 −0.15(1 = yes) (−0.69) (−0.36) (−0.42) (−0.86)Commuting time 61 minutes 0.24 0.26 0.11 −0.15or over (1 = yes) (1.21) (1.32) (0.51) (−0.70)Sports club −0.0061 0.27** 0.10 −0.080(1 = yes) (−0.05) (2.12) (0.55) (−0.54)Number of books 0.021 0.054 0.081** 0.052[1–6] (0.56) (1.32) (2.26) (1.28)English teacher B 0.11 0.12 0.088 −0.069(1 = yes) (0.70) (0.74) (0.44) (−0.39)English teacher C −0.041 0.11 0.072 −0.17(1 = yes) (−0.25) (0.73) (0.41) (−0.90)English teacher D 0.16 0.19 0.12 −0.21(1 = yes) (0.99) (1.09) (0.65) (−1.25)

R-squared 0.149 0.141 0.122 0.078Adjusted R-squared 0.096 0.087 0.061 0.020No. of observations 291 292 262 289

Notes: Estimated coefficients are reported. ***, **, and * indicate 1%, 5%, and 10% levels of statistical significance,respectively. Numbers in parentheses are t-statistics based on heteroscedasticity-robust standard errors. The basecategory for the English-since variable is “English since Grade 1 or 2,” for the commuting time variable it is“Commuting time 20 minutes or less,” and for the teacher dummies it is “Teacher A.”Source: Authors’ calculations.


Appendix A2. Versant Take-Up (Ordinary Least Squares Estimation)

(1) (2) (3) (4) (5) (6)= 1 if scored in Versant test

Baseline Endline

Treatment −0.020 −0.012 −0.089*** −0.093** −0.077** −0.081**

(1 = yes) (−0.60) (−0.34) (−2.65) (−2.50) (−2.21) (−2.06)Procrastination 0.0018 0.0044 0.0077[z-score] (0.09) (0.27) (0.45)Male 0.052 −0.016 −0.012(1 = yes) (1.15) (−0.39) (−0.29)English since Grade 3 or 4 −0.037 −0.0020 −0.018(1 = yes) (−0.99) (−0.05) (−0.45)English since Grade 5 or later −0.015 0.0020 −0.0050(1 = yes) (−0.26) (0.04) (−0.10)Been abroad 0.023 −0.037 −0.047(1 = yes) (0.63) (−0.95) (−1.12)Own room 0.042 −0.013 −0.024(1 = yes) (0.77) (−0.28) (−0.51)Own personal computer −0.095 0.0024 −0.0011(1 = yes) (−1.21) (0.05) (−0.02)Own tablet −0.079 0.063* 0.050(1 = yes) (−1.46) (1.75) (1.35)Commuting time 21–40 minutes 0.036 0.067 0.067(1 = yes) (0.73) (1.33) (1.31)Commuting time 41–60 minutes 0.087 0.014 0.022(1 = yes) (1.63) (0.24) (0.40)Commuting time 61 minutes 0.048 0.059 0.054or over (1 = yes) (0.72) (0.90) (0.78)Sports club −0.064 −0.0098 −0.0069(1 = yes) (−1.37) (−0.23) (−0.16)Number of books −0.0033 0.015 0.020[1–6] (−0.24) (1.11) (1.46)English teacher B 0.015 −0.0050 0.016 0.0077 −0.012 −0.028(1 = yes) (0.36) (−0.11) (0.37) (0.17) (−0.29) (−0.64)English teacher C −0.010 −0.022 −0.022 −0.046 −0.042 −0.070(1 = yes) (−0.24) (−0.49) (−0.45) (−0.90) (−0.90) (−1.47)English teacher D −0.081 −0.091* −0.031 −0.041 −0.030 −0.039(1 = yes) (−1.60) (−1.67) (−0.64) (−0.82) (−0.62) (−0.76)Versant score in the baseline 0.030** 0.026

(2.16) (1.51)

R-squared 0.017 0.057 0.026 0.058 0.030 0.067Adjusted R-squared 0.005 −0.001 0.013 −0.000 0.012 −0.002No. of observations 320 292 320 292 288 262

Notes: Estimated coefficients reported. ***, **, and * indicate 1%, 5%, and 10% levels of statistical significance,respectively. Numbers in parentheses are t-statistics based on heteroscedasticity-robust standard errors. The basecategory for the English-since variable is “English since Grade 1 or 2,” for the commuting time variable it is“Commuting time 20 minutes or less,” and for the teacher dummies it is “Teacher A.”Source: Authors’ calculations.

Analyzing the Sources of Misallocation inIndian Manufacturing: A Gross-Output

ApproachSujana Kabiraj∗

It is well established that misallocation of factor resources lowers productivity.In this paper, I use data from both formal and informal firms to studydistortions in input and output markets as sources of misallocation in theIndian manufacturing sector. My work extends the seminal work of Hsieh andKlenow (2009). I consider output, capital, raw material, energy, and servicesector distortions in a monopolistically competitive framework to measurethe aggregate dispersion in total factor revenue productivity (TFPR). I alsodecompose the variance in TFPR and show that raw material and outputdistortions play a major role in defining aggregate misallocation.

Keywords: distortion, Indian manufacturing, misallocation, productivityJEL codes: E10, O41, O47

I. Introduction

According to the World Bank, the per capita income of the United States(US) was 30 times that of India in 2017. Explaining such differences is one of thefundamental problems in growth economics. Klenow and Rodriguez-Clare (1997)and Hall and Jones (1999) demonstrate that the disparity in total factor productivity(TFP) is the primary source of cross-country income differences. In this context,another debate is about the sources of TFP differences among rich and poor nations.Banerjee and Duflo (2005), Restuccia and Rogerson (2008), and Hsieh and Klenow(2009) argue that in poor countries, some TFP differences are generated from amisallocation of resources across firms. In this paper, I follow the aforementionednotion that resource misallocation is a primary source of variation in TFP. I includeintermediate inputs such as raw materials, energy, and services into the model ofHsieh and Klenow (2009) to obtain the extent of misallocation that originates fromfactor market distortions in a developing country such as India.

∗Sujana Kabiraj: University of Wisconsin-Stevens Point, United States. E-mail: [email protected]. I am grateful toJenny Minier of the University of Kentucky for providing the data used for the empirical analysis. I thank AreendamChanda and other members of the faculty at Louisiana State University for their valuable comments as well as themanaging editor and anonymous referees for helpful suggestions. The Asian Development Bank recognizes “China”as the People’s Republic of China and “Bangalore” as Bengaluru. The usual ADB disclaimer applies.




Analyzing the Sources of Misallocation in Indian Manufacturing 135

When measuring physical TFP, one can adopt either of the two knownapproaches to measuring a firm’s output: value added or gross output. The formerexcludes intermediate inputs, whereas the latter includes them. The differencebetween the two measures of TFP is more pronounced at the firm or industrylevel rather than in aggregate output. Gullickson and Harper (1999); van der Wiel(1999); Hulten, Dean, and Harper (2001); and Cobbold (2003) have demonstratedthe benefits of the gross-output approach over the value-added method. Theproductivity manual published by the Organisation for Economic Co-operation andDevelopment (2001) concludes that the gross-output approach is more appropriatefor productivity measurement because it reduces productivity measurement bias.Based on these findings, I extend the Hsieh–Klenow model to measure productivityusing the gross-output approach by including raw material, energy, and servicesector intermediate inputs as factors of production.1 The inclusion of these factorsseparately into the production process enables a more detailed representation offactor misallocation. Furthermore, the decomposition of factor market distortionsby considering each factor input distortion separately provides a way to distinguishthe level of misallocation in each factor market and to identify the correspondingpotential gain from reallocation. I find that distortions in the output market and rawmaterial market explain the lion’s share of the variation in productivity.

TFP is a residual in the production process and is not observed directly.Moreover, it is difficult to measure firm-level TFP as the unit of production variesacross firms. Therefore, I measure the variation in total factor revenue productivity(TFPR), which by definition is the product of output price and the physical TFP ofa firm. In the absence of any factor market misallocation, TFPR should be equalfor all firms within an industry. The intuition behind this claim is as follows: if afirm has a high TFP, the marginal cost as well as the output price for that firmwill be proportionally lower compared to a low-TFP firm in a particular industry,thus equalizing TFPR. Based on this intuition from Restuccia and Rogerson (2008)and Hsieh and Klenow (2009), I build my empirical results by using data fromboth formal and informal manufacturing sector firms in India for the survey year2005–2006. In such a developing country, the informal sector plays an extensiverole in shaping the economy. The informal manufacturing sector in India consists ofaround 17 million firms that provide 82% of total employment in that sector. Hence,it seems appropriate to include informal sector data in the empirical analysis.

My work has the closest resemblance to the paper by Chatterjee (2011). Iextend the paper by including service sector inputs and energy inputs in the modelseparately. There exists a severe distortion in tariff rates in India’s energy sector. Forexample, during 1999–2000, the industrial sector paid a tariff on electricity almost

1The KLEMS gross-output approach decomposes the factors of production into capital (K), labor (L), energy(E), materials (M), and services (S). The use of the KLEMS approach facilitates a decomposition of sources of growthin the production of firms as well as the inputs used in such production.


15 times higher than that paid by the agriculture sector and 2.1 times higher thanthat paid by the domestic sector (Thakur et al. 2005). Although the Electricity Act of2003 worked toward the reduction and gradual elimination of cross subsidies, suchdistortions may have some impact on the cost of energy usage for small and largefirms as well as in formal and informal sectors. Besides, small firms often have touse other electricity sources (such as generators), which in turn may impact resourceallocation differently in smaller firms as compared to their bigger counterparts.Additionally, liberalization in the service sector in the early 1990s has resulted insignificant growth in the sector. According to Chanda and Gupta (2011), servicesector reforms along with external market linkages led to substantial growth inthe most liberalized service sectors such as business services, banking, insurance,education, medical and health, and others. There is evidence in the literature thatcan link service sector reform to productivity in the manufacturing sector. Forexample, Arnold, Javorcik, and Mattoo (2011) demonstrate a positive relationshipbetween service sector reform and the performance of manufacturing firms in theCzech Republic. In India, the cost share of service inputs is around 10% and thatof energy is around 7% for the manufacturing sector. Exclusion of these factorinputs might lead to misleading measurements of output and productivity. I alsoinclude distortions in the energy and service sectors to verify whether some of thevariation in firm-level TFPR is attributed to these factors. I find that there is verylittle variation in TFPR due to energy input distortions and that misallocation inservice inputs is more pronounced in the dispersion of TFPR. On the other hand, Ifind output and raw material distortions are the primary sources of misallocation inthe manufacturing sector. Another interesting result is that the distortions, whentaken from several factor markets, together reduce the variation in TFPR. Thissurprising result will be the subject of further research.

The rest of the paper is organized as follows. Section II discusses the relevantliterature. In section III, I present a theoretical model to show how TFPR is affectedby firm-level distortions. Section IV describes the data, and section V analyzes theempirical results and the decomposition of the variance of TFPR. Section VI shedssome light on the misallocation among different groups of industries within themanufacturing sector. In section VII, I construe some relationship between firmsize and misallocation in factor markets. Finally, I conclude in section VIII.

II. Literature Review

My work is related to a large body of literature that has accumulated overthe last few decades. Hsieh and Klenow (2009) argue that in a monopolisticallycompetitive framework, misallocation in factor markets can result in largedifferences in TFP and in output among firms within an industry. For example, acapital market distortion caused by the disparity in access to cheap credit will resultin differences in the marginal product of capital among firms. Hsieh and Klenow


argue that in such a situation, the aggregate economy will be better off by allocatingmore capital to the firm with the higher marginal product of capital. Using firm-level data from India and the People’s Republic of China (PRC), they calculatethe TFP gain from reallocating capital, equalizing TFPR within the industry, to be30%–50% in the PRC and 40%–60% in India. I follow the same intuition inthis paper. I include raw materials, energy, and service sector inputs as factors ofproduction and find the effect of distortions in all those inputs on firm-level TFPR.The goal is to find the empirical measurement of distortions in individual factormarkets on aggregate TFPR.

Restuccia and Rogerson (2008) demonstrate the effect of factor distortionon TFP. They state that different taxes and policies across firms create disparitiesin prices and lead to a 30%–50% decrease in output and TFP in developingcountries. Midrigan and Xu (2014) argue that financial frictions cause variationsin TFP across firms through two channels. In particular, financial frictions distortentry decisions and technological adoption of producers. Furthermore, they createdisparities in return to capital among producers. Fernald and Neiman (2011)deviate from the standard setup of monopolistic competition. They show that, in atwo-sector economy with heterogeneous financial policies and monopoly power,TFP measured in terms of quantities and real factor prices can diverge.

There is a body of literature based on Hsieh and Klenow’s framework.Camacho and Conover (2010) use Hsieh and Klenow’s methodology to measureproductivity differences through misallocation in resources for Colombianindustries. Taking the US as the benchmark economy, they find a wide TFPRdistribution for Colombia, which implies large resource misallocation across firms.They also calculate that the reallocation of labor and capital among firms willimprove aggregate TFP by 47%–55%. Another paper by Kalemli-Ozcan andSørensen (2014) measures TFP dispersion through capital misallocation for 10African countries using the World Bank enterprise survey data. They argue thataccess to finance is one of the main sources of substantial capital misallocation.Dias, Marques, and Richmond (2016) extend the Hsieh–Klenow model to includeintermediate inputs and measure TFP disparity using firm-level data from Portugal.They consider data from all sectors of the economy. Consequently, the endogenousintermediate inputs in their model take into account goods produced by all sectors.They find huge misallocation across industries. According to their results, in theabsence of misallocation within industries, there would have been a 48%–79%gain in value-added output during 1996–2011. In India, it is rather difficult to findfirm-level data for sectors other than manufacturing; hence, I take aggregate inputproduced by other sectors as exogenously given in the model.

The most closely related work to my research is the paper by Chatterjee(2011), which extends the Hsieh–Klenow framework for both formal and informalmanufacturing sectors in India. Chatterjee also includes intermediate input marketdistortions in the model as a source of variation in TFP. She assumes that theeconomy has an intermediate input, aggregated from a fraction of total production


by each existing firm. The data used in Chatterjee’s paper was obtained fromthe Annual Survey of Industries (ASI) for formal firms and the National SampleSurvey Office (NSSO) for informal sector firms, as in the case in my paper.Since both of these surveys primarily focus on manufacturing sector firms, theaggregated intermediate input produced from these firms will take into account onlymanufacturing sector products. Consequently, Chatterjee ignores inputs from othersectors such as energy and services in her model. On the other hand, I consideraggregated energy and service inputs as exogenously given in the model apart fromthe combined raw materials produced by the existing firms. In the next section, Iextend Hsieh and Klenow’s model to measure the degree of misallocation in theeconomy.

III. Model

I consider a static one-period model without uncertainty, used by Hsieh andKlenow (2009). I assume that the economy consists of J manufacturing industriesindexed as j = 1, 2, … , J. Each industry consists of Nj monopolistically competitivefirms indexed as i = 1, 2, … , Nj. Each firm produces differentiated products andthus has substantial market power. The firms have heterogeneous productivity Aij

as exogenously given and an endowment of capital Kij, labor Lij, raw material Mij,energy Eij, and service sector input Zij. Firms combine the factors to produce agood using a Cobb–Douglas production function. The firm’s production function isas follows:

Yi j = Ai jKαKj

i j LαL j

i j MαM j

i j EαE j

i j ZαZ j

i j

where∑

S αS j = 1 and S ∈ {K, L, M, E, Z}.I consider only manufacturing sector firms in the model because I could

find data only for the manufacturing sector in India, which I use for the empiricalanalysis. For the sake of simplicity, I assume that all raw materials coming fromthe manufacturing sector are aggregated into a single raw material M , whereas allenergy inputs and service sector inputs are aggregated into factor inputs E and Z,

respectively. A fraction of the output produced by manufacturing firms consideredin the model is aggregated as the manufacturing input M and used by the samefirms; hence, the price of M is taken as endogenously determined. On the other hand,since service and energy inputs that are produced by firms in their respective sectorsare not considered in the model, I take the output prices of such firms as exogenous.Some manufacturing products may also be used by service and energy sector firmsas intermediate inputs; these products are considered as part of consumption goodsin the model. I further assume that all firms in an industry have the same cost shareof factor inputs αS j , but there is a variation in factor shares between industries.

In this paper, I measure the misallocation in resources that affects firm-levelTFPR. Distortion in an input or output market does not always uniformly increase


(or decrease) the marginal product of the factors of production (MPF) for all firms.As firms equalize price with the marginal product of factor inputs, a firm that facestaxes will have higher MPF for service inputs than the firms facing subsidies. Theintuition behind the entire literature based on Hsieh and Klenow (2009) originatesfrom the hypothesis that aggregate productivity will be higher if factors can bereallocated from lower MPF firms to higher MPF firms.

I assume several types of factor market distortions in the model. Someelements that change the MPF for all inputs by the same proportion are calledoutput distortions (τYi j ). Tax on the output of a firm affects all inputs proportionallyand can be identified as an example of an output distortion. Moreover, if thedistortion creates a discrepancy in only the marginal product of capital, I call itcapital distortion (τKi j ) in accordance with Hsieh and Klenow. Similar remarkshold for raw material distortion (τMi j ), energy distortion (τEi j ), and service sectorinput distortion (τZi j ). For example, differentiation in electricity price between smalland large businesses is perceived as an energy distortion as it affects only themarginal product of energy. Note that labor distortion is not considered separatelybut that every other distortion affects the respective MPF, relative to the marginalproductivity of labor.

Each firm produces a single good Yi j that is used both as a final consumptiongood and as an intermediate raw material. Ci j and Xi j denote final consumptiongood and intermediate raw material, respectively, which are produced by the ith firmfrom the jth industry. Firms face a downward sloping demand schedule that resultedfrom the assumption of a differentiated product environment in a monopolisticallycompetitive market. Hence, the industry’s final good appears to be a constantelasticity of substitution aggregation of all firms’ final goods represented as

Yj =⎛⎝ Nj∑

i=1

Yi jρ−1ρ

⎞⎠

ρ

ρ−1

where ρ > 1 is the elasticity of substitution. For simplicity, I assume the elasticityof substitution is the same for all industries. This assumption follows from theliterature. Each industry’s output is sold as consumption good Cj and intermediateraw material Xj as was the case with firm-level output. I further assume that themarkets for consumption goods and raw materials, produced by each industry,is perfectly competitive. Hence, the final consumption good is aggregated fromindustry-level consumption goods by a Cobb–Douglas production function:

C =J∏

j=1

Cθ j

j

whereJ∑

j=1

θ j = 1


The intermediate raw material is produced endogenously by aggregating eachindustry’s production of raw materials, again using a Cobb–Douglas productionfunction:

M =J∏

j=1

Xλ j

j

whereJ∑

j=1

λ j = 1

In the above two equations, θ j and λ j are the factor shares of each industryin total consumption and total intermediate raw materials production, respectively.Each firm chooses intermediate raw materials from the aggregated M according totheir productivity. The aggregate quantity of other inputs such as energy E andservices Z are exogenous in the model. Hence, each firm chooses the optimalamount Ei j and Zi j based on its production function. The industry aggregates Ej

and Zj are given by the sum over each firm’s use of energy and services in thatindustry.

I will now solve the model for optimal factor resources and output bymaximizing profit for the firm, industry, and economy. I assume that total factorresources are limited in the manufacturing sector by the aggregate use of factorresources of the firms in the sector.

For each S ∈ {K, L, M, E, Z}, we can write the aggregate factor resources as

S =J∑

j=1

Nj∑i=1

Si j

and solve for the equilibrium to identify the effects of distortion on productivity.

A. Equilibrium Analysis

In this section, I present a comprehensive equilibrium structure for firms,industries, and the economy. The equilibrium consists of the quantities of theconsumption good and the intermediate raw materials produced at the level of thefirm, industry, and aggregate economy. It also takes into account the optimal amountof capital, labor, raw materials, energy, and services used by each firm. The inputmarkets and final goods markets clear at equilibrium. I now solve the optimizationproblems for each market.

1. Final Goods Problem

I assume a representative firm produces a final good Y that is used inconsumption C and in raw material M for further production. C is produced using


consumption good Cj produced by the industries. I assume C to be a numerairecommodity with unit price P. Likewise, Pj represents the price of the fraction ofoutput or final good produced by each industry Yj. I do not distinguish between priceof final good Cj and raw material Xj, produced by each industry, on the assumptionthat both are of the same good and are subject to the same cost and marketstructure. Hence, the optimization problem for the final consumption good is givenby

maxCj

PC −J∑

j=1

PjCj (1)

subject to

C =J∏

j=1

Cθ j

j (2)

2. Intermediate Raw Materials Problem

The fraction of output used as the intermediate raw material (M) isconstructed by the representative firms aggregating the produced raw materials (Xj)from each industry. The price of the aggregated intermediate raw material M isgiven by pm. The representative firm optimizes the production of M as follows:

maxM j

pmM −J∑

j=1

PjXj (3)

subject to

M =J∏

j=1

Xλ j

j (4)

We can solve the final goods problem from equations (1) and (2) and theintermediate raw materials problem from equations (3) and (4) to find the prices setby representative firms. The market clearing price of the final good is

P =J∏

j=1

(Pj

θ j

)θ j

= 1 (5)

and the intermediate raw material’s price is

pm =J∏

j=1

(Pj

λ j

)λ j

(6)


The second equality in equation (5) follows from the assumption that C is anumeraire good. Both prices are functions of the industry price (Pj) and the shareof each industry in producing the same good (θ j and λ j, respectively).

3. The Industry’s Problem

The final goods produced by each industry Yj are used as both finalconsumption good Cj and intermediate raw material Xj. I assume that Cj and Xj

are fractions of the same good, hence they face the same optimization problem.Furthermore, Ci j and Xi j are fractions of a firm’s output Yi j; therefore, I assume thatthey are produced using the same production function and that they also incur thesame marginal cost. It is safe to assume that the firms charge the same price Pi j forboth parts of their output. I represent the industry’s problem as

maxYj

PjYj −J∑

j=1

Pi jYi j (7)

subject to

Yj =⎛⎝ Nj∑

i=1

Yi jρ−1ρ

⎞⎠

ρ

ρ−1

(8)

The market clearing industry price is

Pj =⎛⎝ Nj∑

i=1

Pi j1−ρ

⎞⎠

ρ

ρ−1

(9)

4. The Firm’s Problem

To allow for factor misallocation in the input and output markets, I considerseveral types of distortions. I assume that there exists an output distortion (τYi j ) thataffects the marginal product of each factor of production by the same proportion.I also consider capital distortion (τKi j ), raw material distortion (τMi j ), energydistortion (τEi j ), and service sector input distortion (τZi j ) that affect the marginalproduct of capital, raw materials, energy, and service inputs, respectively, relative tothe marginal product of labor.2 Each firm solves the following profit maximization

2Since all distortions are measured relative to the labor market, I do not explicitly use labor distortion(τLi j = 0).


problem to choose the optimized level of capital, labor, raw materials, energy, andservice inputs:3

maxYj

Pi jYi j

(1 − τYi j

) − wLi j − r(1 + τKi j

)Ki j − pm

(1 + τMi j

)Mi j

− pe

(1 + τEi j

)Ei j − pz

(1 + τZi j

)Zi j (10)

subject to

Yi j = Ai jKαKj

i j LαL j

i j MαM j

i j EαE j

i j ZαZ j

i j (11)

Solving firm i’s problem yields

K∗i j =

(ρ − 1

ρ

)αKj

(1 − τYi j

)Pi jYi j(

1 + τKi j

)r

(12a)

L∗i j =

(ρ − 1

ρ

)αLj

(1 − τYi j

)Pi jYi j

w(12b)

M∗i j =

(ρ − 1

ρ

)αM j

(1 − τYi j

)Pi jYi j(

1 + τMi j

)pm

(12c)

E∗i j =

(ρ − 1

ρ

)αEj

(1 − τYi j

)Pi jYi j(

1 + τEi j

)pe

(12d)

Z∗i j =

(ρ − 1

ρ

)αZj

(1 − τYi j

)Pi jYi j(

1 + τZi j

)pz

(12e)

Optimal quantities of factor inputs contain both output distortion anddistortion in their respective factor markets. By combining equations (12a)–(12e)with the firm’s objective function in equation (10), we can find the market clearingprice for each firm:

Pi j =(

ρ − 1

ρ

)(MC

ε

) (1 + τKi j

)αKj(1 + τMi j

)αM j(1 + τEi j

)αE j(1 + τZi j

)αZ j(1 − τYi j

)Ai j

(13)

where

ε =∏

S

ααS j

S j

MC = rαKj wαL j pmαM j pe

αE j pzαZ j

3It is important to note that, although the price of aggregated raw materials (M) produced in themanufacturing sector is determined using the profit maximization problem in section III.A.2, firms may consumedifferent combinations (Mi j) of such raw materials for their production; hence, they take the price (pm) as exogenouslygiven.


Note that the firm-level price in equation (13) comprises the marginal cost ofproduction, markup, distortions, and the reciprocal of firm-level productivity. Giventhe assumption that firms in an industry have the same factor shares and input costs,I can infer that in the absence of distortions, the price of each firm in an industrywould have been inversely proportional to the TFP of the firm. This inference isin line with my conjecture that all firms in an industry will have the same revenueproductivity in the absence of any misallocation in factor resources.

I define firm-level total revenue productivity as TFPRi j = Pi jAi j. SolvingTFPRi j from equation (13) yields

TFPRi j =(

ρ − 1

ρ

)(MC

ε

) (1 + τKi j

)αKj(1 + τMi j

)αM j(1 + τEi j

)αE j(1 + τZi j

)αZ j(1 − τYi j

)(14)

Revenue productivity given by equation (14) is a measure of firm-leveldistortion. Variation in TFPRi j gives us the degree of misallocation in input andoutput markets. I build my empirical findings on this intuition and try to measure theextent of variation in firm-level revenue productivity in the presence of distortions.

I define the marginal revenue products of factor inputs for an industry as theweighted average of the value of firm-level marginal revenue products, where theweights are calculated as a share of a firm’s output in the industry:

MRPSj = PS

∑Nj

i=1

(1−τYi j

)Pi jYi j(

1+τSi j

)PjYj

(15)

Recall that S consists of all factor inputs such as K, L, M , E, and Z. PS

denotes the corresponding factor prices r, w, pm, pe, and pz, respectively, and τSi j

indicates the corresponding factor distortions relative to labor.I define industry-level total factor revenue productivity (TFPRj) to be

proportional to the geometric average of the average marginal revenue productsof factor inputs in the industry (given in equation [15]):

TFPRj =(

ρ − 1

ρ

)(MC

ε

)⎡⎢⎢⎢⎣

1

∑Nj

i=1

(1−τYi j

)Pi jYi j(

1+τKi j

)PjYj

⎤⎥⎥⎥⎦

αKj ⎡⎢⎣ 1

∑Nj

i=1

(1−τYi j

)Pi jYi j

PjYj

⎤⎥⎦

αL j

⎡⎢⎢⎢⎣

1

∑Nj

i=1

(1−τYi j

)Pi jYi j(

1+τMi j

)PjYj

⎤⎥⎥⎥⎦

αM j⎡⎢⎢⎢⎣

1

∑Nj

i=1

(1−τYi j

)Pi jYi j(

1+τEi j

)PjYj

⎤⎥⎥⎥⎦

αE j⎡⎢⎢⎢⎣

1

∑Nj

i=1

(1−τYi j

)Pi jYi j(

1+τZi j

)PjYj

⎤⎥⎥⎥⎦

αZ j

.

(16)


B. Allocation of Factors in the Industry

I now solve for the allocation of factor resources for each industry. Iaggregate factor resources used by all firms in an industry using their marginalproducts to get the following:

S j =Nj∑

i=1

Si j = SαS jθ j/MRPSj∑Jj=1 αS jθ j/MRPSj

(17)

Recall that S ∈ {K, L, M, E, Z}. and S = ∑Jj=1 S j are aggregate supplies of

factor inputs in the economy. Also recall that θ j is the share of each industryin producing the final consumption good. Note that factor accumulations in eachindustry are affected by factor distortions only through the corresponding marginalrevenue products. This result is due to the Cobb–Douglas aggregation at theindustry level. Combining industry-level factor inputs (17) and revenue productivity(16), we can derive

PjYj = TFPRjKαKj

j LαL j

j MαM j

j EαE j

j ZαZ j

j (18)

Combining industry price Pj from (9) and firm’s price Pi j from (13) togetherwith firm-level revenue productivity from (14), we can simplify

Pj =⎡⎣ Nj∑

i=1

(TFPRi j

Ai j

)(1−ρ )⎤⎦

11−ρ

(19)

Equating (18) and (19), we get

Yj = TFPjKαKj

j LαL j

j MαM j

j EαE j

j ZαZ j

j (20)

where

TFPj =⎡⎣ Nj∑

i=1

(Ai jTFPRj

TFPRi j

)(ρ−1)⎤⎦

11−ρ

(21)

Hence, the total factor productivity of each firm is a function of thefirm-level TFP, TFPR, and industry-level revenue productivity. Now, we can writethe final consumption outcome of the economy as follows:

C∗ =J∏

j=1

(TFPjK

αKj

j LαL j

j MαM j

j EαE j

j ZαZ j

j

)θS

(22)

and the intermediate good of the economy as

M∗ =J∏

j=1

(TFPjK

αKj

j LαL j

j MαM j

j EαE j

j ZαZ j

j

)λS

(23)


Following Hsieh and Klenow (2009), I now assume that TFP (Ai j) andrevenue productivity (TFPRi j) are jointly log-normally distributed to depict theeffect of firm-level distortion on the productivity of an industry. By this assumption,the logarithm of industry-level TFP can be expressed as

log TFPj = 1

1 − ρlog

⎛⎝ Nj∑

i=1

Ai j(ρ−1)

⎞⎠ − ρ

2Var

(log TFPRi j

)(24)

Equation (24) shows that factor distortions reduce overall productivity ofan industry through the variance of firm-level TFPR. On the basis of this finding,I will now proceed to show how factor distortions contribute to firm-level TFPRvariation. Note that I assume that the number of firms are unaffected by factormarket distortions. This assumption is elaborated in more detail in Hsieh andKlenow (2009).

IV. Data

This study uses data on the formal manufacturing sector from the AnnualSurvey of Industries (ASI) collected by the Central Statistical Office of India. ASIis the primary source of industrial statistics in India, which covers all factoriesas defined in the Factories Act of 1948. ASI data is an annual survey of formalmanufacturing firms with more than 50 workers and a random one-third samplesurvey of firms with more than 10 workers (with electricity) or firms with morethan 20 workers (without electricity). I use the 62nd round of ASI data collected inthe survey year 2005–2006.

I also take into account data for the unorganized manufacturing sectorcollected by the National Sample Survey Office (NSSO) of India for the surveyyear 2005–2006. The NSSO collects firm-level data for the informal manufacturingsector in India every 5 years. The dataset includes small manufacturing firms alongwith some service sector firms and some unincorporated proprietary firms. Thesefirms are not registered under the Factories Act of 1948; hence, they are not includedin the ASI data. The data for the informal sector consists of a large number offirms that use one or two workers. These firms have missing values for most of thevariables I take into consideration. Also, they contribute a very small percentage oftotal value added.

Table 1 summarizes the distribution of informal firms and the correspondingcumulative percentages of contributions to total value added, according to thenumber of employees. There are over 30,000 one-employee firms that contributeonly 1.6% of total value added and almost none of them have data for labor andcapital in the corresponding dataset. In my analysis, I do not include such firms. Ionly consider informal firms that have at least six employees, a cutoff set on thebasis of these firms’ substantial market share. To keep the two datasets comparable,


Table 1. Informal Firms Distribution

Cumulative % ofNo. of Employees No. of Firms Value Added

1 31,874 1.62 23,734 4.43 9,468 6.94 4,658 9.15 2,601 11.66 1,948 17.87 1,349 21.88 1,054 26.09 732 31.010 648 37.011–20 1,712 57.421–30 342 62.831– 50 190 73.051–100 92 95.0More than 100 61 100.0

Source: National Sample Survey Office (NSSO) survey ofunorganized manufacturing enterprises, 2005–2006.

Table 2. Distribution of Firms: Annual Survey of Industriesversus National Sample Survey Office

ASI Data NSSO Data

No. of Employees No. of Firms No. of Employees No. of Firms

1–10 4,663 6–10 3,51211–20 6,818 11–20 1,19421–50 6,771 21–50 39751–100 3,719 51–100 77101–500 7,648 101–500 31More than 500 2,254 More than 500 5

ASI = Annual Survey of Industries, NSSO = National Sample Survey Office.Sources: Government of India, Ministry of Statistics and Programme Implementation(2005–2006a and 2005–2006b).

I only consider manufacturing industries that are covered in both the ASI and NSSOdatasets. The literature in the field (La Porta and Shleifer 2008) argues that informalsector firms are small and highly unproductive compared to formal sector firms. Theassumption of monopolistic competition among firms in the model allows for firmswith different levels of productivity to coexist in the market.

Table 2 shows the distribution of firms in the analysis. There are around31,000 formal sector firms taken from the ASI data, whereas the number of informalsector firms from the NSSO data is around 5,000. For this analysis, I had to dropsome observations from both sectors due to missing data. Formal firms consist ofall sizes, while informal firms are mostly small. To simplify the analysis, I use2-digit industry-level data developed in the National Industrial Classification (NIC)


system. I consider 23 different industries, including food and beverage, machineryand equipment, wood, paper, publishing, computing machinery, and others (seeTable 3).

Hsieh and Klenow (2009) use the value-added method to measureproductivity and distortion in capital and output. They did not incorporate rawmaterials, services, or energy inputs in the production function. I first replicate theirresults using the value-added method and then extend the model to incorporateintermediate inputs as factors of production. This extension will adopt thegross-output method. I use nominal revenue of the firm as the output variable.

Aside from firm revenue, the variables that I use for this analysis are thefirm’s industry (2-digit NIC), labor compensation, net book value of fixed capitalstock, rent on capital, intermediate input costs, and fuel and energy costs. I assumethat service input cost is the same as the residual cost. I use labor compensationincluding wages, bonuses, and benefits as proxy for labor input. Capital is measuredby the average of net book value of capital at the beginning and end of the year.I deviate from Hsieh and Klenow (2009) and Chatterjee (2011), as well as otherprevious works based on the measurement of the rental cost of capital. Existingliterature in this field uses an exogenous percentage of capital as rental cost, whereasI measure the same by variables such as rent on machinery, building, and land,interest paid on loans, and other miscellaneous capital cost, which are taken fromthe ASI data for the formal sector.

However, for informal firms, the NSSO data do not explicitly provide renton capital. I measured rental cost from the residual of value added after subtractingtotal labor cost. The costs of raw materials and energy are calculated explicitlyfrom the cost of inputs of production. Service input costs consist of transportand communication costs, insurance charges, license costs, and other operativeexpenses.

The elasticity of substitution (ρ) is assumed to be constant in the model.Based on the literature in this field, I assume the value of ρ to be 3. In most of myempirical analysis, I use factor shares from industries in the US as a benchmarkto identify the effect of distortion on productivity. The factor shares data are fromthe KLEMS measures found in the National Income and Product Accounts industrydatabase (2005) provided by the US Bureau of Labor Statistics.

V. Empirical Analysis

My identification strategy is similar to that of Hsieh and Klenow (2009)and Chatterjee (2011). I established identification of distortions based on therationale that, in the absence of distortions, revenue factor shares of output willbe proportional to the parameters αK j, αL j, αM j, αE j, and αZ j in a market withmonopolistic competition. Because I assume distortions in factor markets, the

Analyzing the Sources of Misallocation in Indian Manufacturing 149Ta

ble

3.Fa

ctor

Shar

es:

Indi

ave

rsus

Uni

ted

Stat

es

Cap

ital

Shar

eL

abor

Shar

eR

awM

ater

ialS

hare

Ene

rgy

Shar

eSe

rvic

esSh

are

NIC

2-D

igit

Cod

eM

anuf

actu

ring

Indu

stry

Indi

aU

SIn

dia

US

Indi

aU

SIn

dia

US

Indi

aU

S

15Fo

odan

dbe

vera

ge0.

030.

120.

050.

150.

830.

570.

030.

030.

060.

1316

Toba

cco

0.02

0.12

0.17

0.15

0.62

0.57

0.02

0.03

0.18

0.13

17Te

xtil

es0.

040.

100.

100.

230.

680.

530.

090.

030.

090.

1118

Wea

ring

appa

rel

0.03

0.11

0.15

0.34

0.55

0.32

0.03

0.02

0.24

0.21

19L

eath

er0.

020.

110.

080.

340.

760.

320.

030.

020.

110.

2120

Woo

d0.

040.

070.

080.

240.

730.

530.

050.

030.

090.

1421

Pape

r0.

040.

140.

090.

210.

670.

460.

130.

080.

080.

1222

Pub

lish

ing

0.04

0.05

0.15

0.32

0.58

0.39

0.03

0.02

0.20

0.23

23Pe

trol

eum

prod

ucts

0.00

0.28

0.01

0.03

0.95

0.67

0.01

0.01

0.03

0.01

24C

hem

ical

0.03

0.19

0.07

0.17

0.77

0.40

0.06

0.06

0.08

0.19

25R

ubbe

ran

dpl

asti

cs0.

030.

130.

060.

220.

790.

490.

050.

020.

070.

1426

Non

met

alli

cm

iner

als

0.05

0.16

0.09

0.29

0.50

0.31

0.26

0.08

0.11

0.16

27B

asic

met

als

0.04

0.14

0.06

0.20

0.74

0.39

0.11

0.10

0.06

0.17

28Fa

bric

ated

met

als

0.03

0.14

0.09

0.31

0.74

0.37

0.03

0.02

0.12

0.17

29M

achi

nery

and

equi

pmen

t0.

020.

110.

100.

280.

730.

450.

020.

010.

130.

1530

Com

puti

ngm

achi

nery

0.02

0.17

0.08

0.37

0.85

0.24

0.01

0.01

0.06

0.20

31E

lect

rica

lmac

hine

ry0.

020.

110.

090.

270.

780.

470.

020.

010.

090.

1432

Com

mun

icat

ion

mac

hine

s0.

040.

110.

080.

270.

760.

470.

020.

010.

110.

1433

Med

ical

inst

rum

ents

0.04

0.17

0.13

0.33

0.66

0.31

0.02

0.01

0.16

0.18

34M

otor

vehi

cles

0.01

0.10

0.07

0.25

0.82

0.51

0.02

0.01

0.08

0.14

35T

rans

port

equi

pmen

t0.

020.

100.

090.

250.

780.

510.

030.

010.

100.

1436

Fur

nitu

re0.

020.

110.

050.

310.

850.

400.

010.

010.

070.

1737

Rec

ycli

ng0.

020.

170.

060.

330.

740.

310.

070.

010.

100.

18

NIC

=N

atio

nalI

ndus

tria

lCla

ssifi

cati

on,U

S=

Uni

ted

Sta

tes.

Not

es:F

orU

Sfa

ctor

shar

es,m

ulti

fact

orpr

oduc

tivit

yan

dre

late

dK

LE

MS

mea

sure

sar

efr

omth

eN

atio

nalI

ncom

ean

dP

rodu

ctA

ccou

nts

indu

stry

data

base

(usi

ngba

seye

ar20

05).

For

Indi

afa

ctor

shar

es,d

ata

are

from

the

Inpu

t–O

utpu

tTra

nsac

tion

sTa

ble

2003

–04.

Sou

rces

:US

Bur

eau

ofL

abor

Sta

tist

ics

(200

5);G

over

nmen

tof

Indi

a,M

inis

try

ofS

tati

stic

san

dP

rogr

amm

eIm

plem

enta

tion

(200

8);G

over

nmen

tof

Indi

a,M

inis

try

ofS

tati

stic

san

dP

rogr

amm

eIm

plem

enta

tion

.htt

p://

mos

pi.n

ic.in

/sit

es/d

efau

lt/fi

les/

mai

n_m

enu/

nati

onal

_ind

ustr

ial_

clas

sifi

cati

on/n

ic_2

004_

stru

c_2d

igit

.pdf

.


revenue shares will give a biased estimation of the parameters. We can validatethis from the first-order conditions of the firms:

αS j =(

ρ

ρ − 1

)PSSi j

(1 + τSi j

)Pi jYi j

(1 − τYi j

) (25)

where αS j = {αK j, αL j, αM j, αE j, αZ j} and the respective PS = {r, w, pm, pe, pz}.4Recall that S consists of all factor inputs and τSi j denotes corresponding distortionsrelative to the labor market. In the presence of distortions, I cannot distinguishthe misallocation in resources from the bias in the parameters. Following Hsiehand Klenow (2009), I take into account US factor shares. The strategy is basedon the assumption that US factor markets are less distorted than in India and thetechnology used in the industries is the same for both countries. A more detaileddiscussion on the assumptions are presented in Chatterjee (2011). Factor shares forboth countries, described in Table 3, represent the average of the cost share for eachfactor in each industry.

Figure 1 illustrates the bias in factor shares in Indian industries withrespect to benchmark US industries. Any deviation from the 45-degree line showsmisallocation in the corresponding factor markets in India. I find a similar patternin capital, labor, and raw material shares presented in Chatterjee (2011). It isevident from the figure that cost shares of capital, labor, and service inputs aresignificantly higher in the US than in India, whereas shares of raw materials andenergy are higher in India. Next, I analyze within-industry variation in averagerevenue product of labor. Figure 2 illustrates the distribution of the logarithm offirm-level average revenue product of labor (APRL) relative to the industry mean,log (APRLi j/APRLj ). I trim 1 percentile from both ends to avoid outliers. Thehorizontal axis shows log (APRLi j/APRLj ), whereas the vertical axis measures thedensity of firms. There is a substantial variation in average revenue product of laborwithin an industry with a variance of 3.76.

A. Value-Added versus Gross-Output Approach

The goal in this section is to measure the variation in firm-level TFPR asan indicator of misallocation in factor markets. The variable of interest is thelogarithm of firm-level TFPR relative to the industry TFPR, log (TFPRi j/TFPRj ).I depict both value-added and gross-output approaches to measure TFPR. First, Ireplicate the results from Hsieh and Klenow (2009) using the value-added approach.They estimate the distribution of TFPR using formal manufacturing sector data for

4The assumption of common factor prices is a hypothetical frictionless situation, which allows me to identifythe factor allocation distortion apart from any exogenous variation that may affect factor prices. As long as distortionsexist, firms pay different factor prices, and the dispersion in distortions provides us with a sense of how much firmsactually deviate from the equilibrium.


Figure 1. Factor Shares for the United States and India

US = United States.Source: Author’s calculation based on the Annual Survey of Industries and the National Sample Survey Office ofthe Government of India, Ministry of Statistics and Programme Implementation (2005–2006a, 2005–2006b); andKLEMS measures from the National Income and Product Accounts of the US Bureau of Labor Statistics (2005).

1987–1988 and 1994–1995. I repeat their method using 2005–2006 data for bothformal and informal sectors. I also illustrate the TFPR distribution using thegross-output method using the same data. Cobbold (2003) presented the formalrelationship between value-added and gross-output TFP as

TFPVA = G

VA× TFPGO


Figure 2. Distribution of the Logarithm of Average Revenue Product of Labor

Source: Author’s calculation based on the Annual Survey of Industries and the National Sample Survey Office of theGovernment of India, Ministry of Statistics and Programme Implementation (2005–2006a, 2005–2006b).

where G and VA represent nominal values of total revenue and total value added,respectively.

Several studies, such as Oulton and O’Mahony (1994) and van der Wiel(1999), show that productivity growth measured using value added is much higherthan the measurement considering all inputs. It naturally follows from the aboveequation that given G and VA, TFP as well as TFPR measured using the value-addedapproach will be larger than if measured by the gross-output approach.

Before calculating the variance, I trim 1% tails of log (TFPRi j/TFPRj ) toget rid of outliers. Figure 3 plots the distributions of the logarithm of TFPR relativeto the industry mean. The dashed line shows the value-added TFPR distributionwhereas the solid line shows the distribution using the gross-output approach. Thevariation in value-added TFPR is much higher than the variation in gross-outputTFPR. Table 4 presents the TFPR dispersion statistics in firm-level TFPR. Standarddeviation (SD) in value-added TFPR is around 0.99 compared to 0.47 using thegross-output approach. The difference in both methods is more pronounced whenthe variation in TFPR is estimated at higher percentiles.

Table 5 shows the dispersion in the logarithm of TFPR in Hsieh and Klenow(2009) using the value-added approach and the same variable in Chatterjee (2011)using the gross-output approach. The results display a larger value-added SD than


Figure 3. Distribution of the Logarithm of Total Factor Revenue Productivity

kdensity = kernel density, TFPR = total factor revenue productivity.Source: Author’s calculation based on the Annual Survey of Industries and the National Sample Survey Office of theGovernment of India, Ministry of Statistics and Programme Implementation (2005–2006a, 2005–2006b).

Table 4. Dispersion of the Logarithm of Total FactorRevenue Productivity

Statistics Value Added Gross Output

Standard deviation 0.99 0.4775th to 25th percentiles 1.23 0.5190th to 10th percentiles 2.45 1.08

Note: The variable is log (TFPRij/TFPRj).Source: Author’s calculation based on the Annual Survey of Industries andthe National Sample Survey Office of the Government of India, Ministryof Statistics and Programme Implementation (2005–2006a, 2005–2006b);and KLEMS measures from the National Income and Product Accountsof the US Bureau of Labor Statistics (2005).

Table 5. Dispersion of the Logarithm of Total Factor Revenue Productivityin the Literature

Statistics Hsieh–Klenow (1994–1995) Chatterjee (2004–2005)

Standard deviation 0.67 0.4975th to 25th percentiles 0.81 0.5690th to 10th percentiles 1.60 1.19

Notes: Column 2 shows dispersion of total factor revenue productivity estimated by Hsieh andKlenow (2009) for 1994–1995 data using the value-added approach. Column 3 depicts the samevariable estimated by Chatterjee (2011) for 2004–2005 data using the gross-output approach.Sources: Hsieh and Klenow (2009) and Chatterjee (2011).


in Hsieh and Klenow (2009), who use the same approach with formal sector datafrom 1994 to 1995. This might be the consequence of an increase in overall level ofmisallocation in the last decade or inclusion of the informal sector in my analysis.

Furthermore, I find comparable results (shown in Table 4) with those ofChatterjee (2011) in the dispersion of gross-output TFPR. After including energyand service sector distortions, the SD in firm-level TFPR in my study dropped by0.02 from an overall 0.49 as shown by Chatterjee using 2004–2005 data. The gapbetween the results is more conspicuous in the 75th to 25th percentiles and 90th to10th percentiles.

B. Decomposing the Misallocation in Factor Markets

I now turn to separating the effect of each component attributed to thevariance of firm-level TFPR. Moving forward, only the gross-output approach willbe considered. I take into account several types of distortions in input and outputmarkets. The calculation for each type as a function of total revenue, cost of inputs,and factor shares is derived from first-order conditions of a firm as

1 − τYi j =(

ρ

ρ − 1

)wLi j

αLj Pi jYi j(26a)

1 + τKi j = αKjwLi j

αLj rKi j(26b)

1 + τMi j = αM jwLi j

αLj pmMi j(26c)

1 + τEi j = αEjwLi j

αLj peEi j(26d)

1 + τZi j = αZjwLi j

αLj pzZi j(26e)

where all input market distortions are measured relative to the labor market. Theintuition behind equations (26b)–(26e) is that, in the presence of distortions, inputcosts relative to labor compensation will be lower than given by the output elasticity.Equation (26a) demonstrates that a deviation of labor share from output elasticitywith respect to labor will result in an output distortion.

To give a more elaborate presentation of the above result, I now findthe variance of log (TFPRi j/TFPRj ). The total misallocation is measured by thefollowing variance:

Var[log(TFPRi j/TFPRj

)] = Var (DK + DL + DM + DE + DZ − DY ) (27)


Table 6. Variance Decomposition

Component Variance or Covariance

Var(DK ) 0.0472Var(DL) 0.0119Var(DM ) 0.3490Var(DE ) 0.0048Var(DZ ) 0.0345Var(DY ) 1.1203Cov(DK , DL) 0.0030Cov(DK , DM ) 0.0346Cov(DK , DE ) 0.0029Cov(DK , DZ ) 0.0082Cov(DK , DY ) 0.0727Cov(DL, DM ) –0.0026Cov(DL, DE ) 0.0001Cov(DL, DZ ) 0.0001Cov(DL, DY ) 0.0067Cov(DM , DE ) 0.0129Cov(DM , DZ ) 0.0533Cov(DM , DY ) 0.5716Cov(DE , DZ ) 0.0034Cov(DE , DY ) 0.0302Cov(DZ, DY ) 0.1110Var[log(TFPRi j/TFPRj )] 0.2194

TFPR = total factor revenue productivity.Note: The table shows variances and covariances of thecomponents of log TFPR, where DS (S ∈ {K, L, M, E, Z}) andDY are given by equations (28) and (29).Source: Author’s calculation based on the Annual Survey ofIndustries of the Government of India, Ministry of Statisticsand Programme Implementation (2005–2006a); and KLEMSmeasures from the National Income and Product Accounts ofthe US Bureau of Labor Statistics (2005).

where

DS = αS j log

⎡⎣(

1 + τSi j

) Nj∑i=1

(1 − τYi j

)Pi jYi j(

1 + τSi j

)PjYj

⎤⎦ (28)

DY = log(1 − τYi j

)(29)

Recall that S consists of all factor inputs such as K, L, M, E, and Z. αS j

denotes the corresponding factor shares, and τSi j indicates the corresponding factordistortions. Also recall that I measure factor distortions relative to the labor market,implying τLi j to be 0. In the above equations, DS can be inferred as componentsof each factor input in the variance of TFPR. Table 6 describes the variance andcovariances of each of the above components.


The variance of the components of equation (16) depict the contribution offactor distortions in explaining the variation in firm-level TFPR. Since I measuredistortions in factor markets relative to the labor market in my analysis, thevariance of DL measures the variation in industry TFPR in the presence of onlyoutput distortions, multiplied by the cost share of labor. Moreover, the varianceof DY determines the variation in firm TFPR attributed to only output distortion.Dispersions in DY and DM are very high compared to the overall variance oflog(TFPRi j/TFPRj ), implying that misallocation is highest in output and rawmaterials.

Overall variance in log (TFPRi j/TFPRj ) includes the pairwise covariancebetween the components of equation (16) as well. It is interesting to note that thecovariance between output and raw material distortions is the highest (0.5716). Thisresult may follow from the fact that in my framework, raw materials are endogenous,thus the output of one firm is used as raw materials in another.

Next, I examine the distinct effect of each distortion on the logarithm ofTFPR relative to the industry mean. In Figure 4, the solid lines illustrate thedistribution of the variable of interest, taking one factor distortion at a time. Thedashed line represents the actual firm-level TFPR distribution taking all distortionstogether. The top panels of Figure 4 show TFPR distributions taking either outputor capital distortion. Similarly, the middle panels and bottom panel depict scenarioswith only raw material, energy, or service input distortions, respectively.

In the absence of any distortion, I expect the TFPR of all firms to equalize,which should reflect in a distribution that shows a vertical line in a graph centeredat 0. Any deviation from such a line shows signs of distortion. Taking one factordistortion at a time facilitates the comparison between the contribution of eachfactor input distortion toward the overall distortion in the market.

Since higher dispersion in the distribution shows higher distortion, it isperceptible from Figure 4 that output and raw material distortions play the mainrole in the overall distortion within an industry. Capital and service input distortionscontribute a modest share in the measurement of misallocation. Energy distortionis almost negligible. These results emphasize the findings in Table 6.5

The intriguing observation from Figure 4 is that the misallocation in TFPR islower when all factor market distortions are considered than when considering onlyoutput distortion or only raw material distortion. Such findings imply that factorinput distortions offset each other’s effects in describing total misallocation. This isan area I would like to work on in the future in order to understand the underlyingintuition.

5Since the distortion in other markets are measured relative to the labor market, I cannot estimate themagnitude of the distortion in the labor market. The Appendix shows factor market distortions in different sectorssimilar to Figure 4, relative to the energy input market.


Figure 4. Distribution of Firm-Level Total Factor Revenue Productivity Taking OneDistortion at a Time (relative to the labor market)

TFPR = total factor revenue productivity.Note: All TFPR distributions are in logarithm and relative to the industry average.Source: Author’s calculation based on the Annual Survey of Industries and the National Sample Survey Office of theGovernment of India, Ministry of Statistics and Programme Implementation (2005–2006a, 2005–2006b).

C. Formal versus Informal Sectors

Since I use data from both formal and informal sector firms in my empiricalanalysis, it is important to examine if there is any inherent difference in the patternof input and output distortions between these two sectors. Figure 5 illustrates the


Figure 5. Distribution of Firm-Level Total Factor Revenue Productivity in Formal andInformal Sectors


distribution of TFPR relative to the industry mean for firms in the formal andinformal sectors. The solid line represents formal sector firms, whereas the dottedline represents firms in the informal sector. The dispersions in the distributions aresimilar in both sectors while the mean of the distribution is higher in the formalsector than in the informal sector.

Next, I examine the distribution of factor distortions separately for bothsectors. Figure 6 shows the distribution of output and input distortions for formaland informal firms. The distributions show that in the informal sector, output, rawmaterial, energy, and service input distortions are a little higher than in the formalsector. Capital distortion on the other hand is more dispersed and higher in theformal sector.

VI. Misallocation within the Manufacturing Sector

Misallocation may vary between industries within the manufacturing sector.Therefore, it is useful to look further into the distribution of TFPR in separate


Figure 6. Distribution of Factor Market Distortions in Formal and Informal Sectors

Source: Author’s calculation based on the Annual Survey of Industries and the National Sample Survey Office of theGovernment of India, Ministry of Statistics and Programme Implementation (2005–2006a, 2005–2006b).

groups of industries to infer any inherent pattern of misallocation that may existwithin the manufacturing sector.

A. According to Use of Service Inputs

Services include a vast range of inputs used by manufacturing sectorfirms. Also, manufacturing industries vary widely in their use of such inputs.Using manufacturing firm data from the Czech Republic, Arnold, Javorcik, andMattoo (2011) show that the productivity of manufacturing industries that relyextensively on service inputs is affected more by reforms in the service sector. To


Figure 7. Distribution of Firm-Level Total Factor Revenue Productivity according toIntensity of Use of Service Inputs


investigate if such a connection exists in Indian manufacturing, I use the 2003–2004input–output table provided by the Ministry of Statistics and ProgrammeImplementation of the Government of India to rank the industries according totheir use of services (Government of India, Ministry of Statistics and ProgrammeImplementation 2008). Five industries that used more than 55% of the total serviceinputs used by the manufacturing sector are food and beverage, basic metals,chemical products, wearing apparel, and electric machinery. Figure 7 shows theTFPR distribution of these five industries compared to others. The dispersion inTFPR is lower in the industries that use service inputs more intensively, reflecting alower misallocation in these industries.

B. According to Raw Materials Contribution

Variance decomposition of the misallocation in factor markets in sectionV.B. revealed that not only does raw material distortion play a vital role inexplaining overall misallocation in TFPR, it is also highly correlated with output


Figure 8. Distribution of Firm-Level Total Factor Revenue Productivity for All Industries(only raw material distortion) and Distribution of Total Factor Revenue Productivity for

High Raw Material Contributing Industries (only output distortion)


distortion. Since raw materials in the model are produced by the manufacturingfirms themselves, raw material distortion may partly reflect output distortion in theindustries supplying raw materials. To investigate the source of such a distortion,I categorize industries according to their raw materials contribution, using theinput–output table mentioned previously. Five manufacturing sector industries thatcontributed almost 63% of raw materials to the same sector are food and beverage,textile, petroleum products, basic metals, and chemical and chemical products.

Next, I compare the output distortion in industries that contribute a largeshare of raw materials with the raw material distortion for all manufacturing sectorfirms. The solid line in Figure 8 shows the TFPR distribution with only outputdistortion using firms in the top five raw materials contributing industries, while thedashed line shows the TFPR distribution with only raw material distortion usingall firms in the manufacturing sector. The former distribution is more skewed thanthe latter, which suggests that raw material distortion in the manufacturing sectormay be partly reflecting output distortions in the industries contributing most to


Table 7. Regression of Firm Size on Distortion

Log of Distortions

Output Capital Raw Materials Energy Services(1) (2) (3) (4) (5)

Log (labor) 0.075** 0.274** 0.044** 0.190** 0.076**

(0.005) (0.006) (0.008) (0.005) (0.005)Industry effect Yes Yes Yes Yes YesOwnership effect Yes Yes Yes Yes YesOrganization effect Yes Yes Yes Yes YesRegion effect Yes Yes Yes Yes YesNo. of observations 41,237 44,726 45,589 47,755 47,829

Notes: The dependent variables in the regressions are the logarithm of output and input (capital, rawmaterials, energy, and services, respectively) distortions. Standard errors in parentheses. ** showsp-value < .01.Source: Author’s calculation based on the Annual Survey of Industries and the National SampleSurvey Office of the Government of India, Ministry of Statistics and Programme Implementation(2005–2006a, 2005–2006b); and KLEMS measures from the National Income and ProductAccounts of the US Bureau of Labor Statistics.

manufacturing raw materials. This finding suggests that policies that can reduceoutput distortion in industries that supply the lion’s share of manufacturing rawmaterials should result in lower raw material distortion in the overall manufacturingsector.

VII. Misallocation and Firm Size

There is a body of literature on the sources of factor distortions. Banerjeeand Duflo (2005) discovered that capital market distortions might be originatingfrom disparities in credit policy. Chatterjee (2011) mentions unavailability of rawmaterials as a reason behind intermediate input distortions. Bhidé (2008) showsthat in a developing country such as India, electricity connection from private andpublic enterprises might cause a distortion in energy prices. Hsieh and Klenow(2009) argue that government policy, especially size restrictions, might preventfirms from achieving an optimal scale, thereby creating an output distortion. Theyalso considered firm size as an explanation for TFPR dispersion within an industry.Ha, Kiyota, and Yamanouchi (2016) show a nonlinear relationship between firmemployment size and factor market distortions in the context of manufacturingfirms in Viet Nam. I now proceed to examine the relationship between firm sizeand distortion in factor markets.

Table 7 presents regression coefficients from estimating this relationship. Iuse the logarithm of total labor employed as a measure of firm size. Column (1) usesthe logarithm of firm-level output distortion as the dependent variable. Similarly, thedependent variables for columns (2), (3), (4), and (5) are the logarithms of firm-levelcapital, raw material, energy, and service input distortions, respectively. I controlfor industry fixed effects, ownership type (private, central government owned,


state government owned, etc.), type of organization (individual proprietorship,partnership, co-operative society, etc.), and location of the firms.

I find a positive relationship between firm size and each type of distortion.6

Smaller firms in the formal or informal sector might be able to avoid somepolicy restrictions, unlike their larger counterparts. An assumption of monopolisticcompetition includes the provision of a markup in the model. Though I assumeall firms in an industry have the same markup, larger firms might have greatermarket power and larger markups, which in turn will create more output distortionas well as raw material distortion. It will be fascinating to see the effect of firm sizeon distortion once we relax the assumption of a constant elasticity of substitutionwithin an industry.

VIII. Conclusion

I measure the aggregate misallocation in resources using firm-level datafrom both formal and informal manufacturing sectors in India for the survey year2005–2006. I include energy distortion and service input distortion to extendexisting research such as those by Hsieh and Klenow (2009) and Chatterjee (2011).The dispersion in TFPR within each industry is substantial, implying misallocationcaused by distortion of factor resources. While energy distortion does not contributemuch to aggregate misallocation, the effect of service sector input distortion is morepronounced. I further decompose the variance of TFPR to find the effect of eachfactor market distortion separately. I discover that output distortion and raw materialdistortion contribute the largest share in aggregate misallocation. Reallocation ofsuch factors within industries should result in the largest gain in TFP. Moreover, Ifind a high level of covariance between output and raw material distortion which,along with a further exploration within the manufacturing sector, suggests that someof the distortion in raw materials may reflect the output distortion in industriesproducing a larger share of the raw materials. I also uncover a puzzling result that theinclusion of many factor distortions together offset each other’s effects and resultsin a lower aggregate misallocation. Although unexpected, this result may inspirefurther research in this field.

References

Arnold, Jens, Beata Javorcik, and Aaditya Mattoo. 2011. “The Productivity Effects of ServicesLiberalization: Evidence from the Czech Republic.” Journal of International Economics 85(1): 136–46.

6Note that the distortions are calculated using equations (26a)–(26e). Since the sign of output distortion isinverse in equation (26a), a higher numerical value of output distortion shows a lower magnitude of the same.


Banerjee, Abhijit V., and Esther Duflo. 2005. “Growth Theory through the Lens of DevelopmentEconomics.” In Handbook of Economic Growth Volume 1, Part A, edited by PhilippeAghion and Steven N. Durlauf, 473–552. Amsterdam: North Holland-Elsevier.

Bhidé, Amar. 2008. “What Holds Back Bangalore Businesses?” Asian Economic Papers 7 (1):120–53.

Camacho, Adriana, and Emily Conover. 2010. “Misallocation and Productivity in Colombia’sManufacturing Industries.” Inter-American Development Bank Working Paper Series.

Chanda, Rupa, and Pralok Gupta. 2011. “Service Sector Liberalisation in India: Key Lessons andChallenges.” South African Institute of International Affairs (SAIIA) Occasional Paper No.88.

Chatterjee, Urmila. 2011. “Resource Allocation and Efficiency in Developing Countries.” PhDDissertation, University of California Berkeley. https://escholarship.org/uc/item/6458x094.

Cobbold, Trevor. 2003. “A Comparison of Gross Output and Value-Added Methods ofProductivity Estimation.” Productivity Commission Research Memorandum. Canberra,Australia.

Dias, Daniel A., Carlos Robalo Marques, and Christine Richmond. 2016. “Misallocation andProductivity in the Lead Up to the Eurozone Crisis.” Journal of Macroeconomics 49: 46–70.

Fernald, John, and Brent Neiman. 2011. “Growth Accounting with Misallocation: Or, Doing Lesswith More in Singapore.” American Economic Journal: Macroeconomics 3 (2): 29–74.

Government of India, Ministry of Statistics and Programme Implementation. http://mospi.nic.in/sites/default/files/main_menu/national_industrial_classification/nic_2004_struc_2digit.pdf.

______. 2005–2006a. Annual Survey of Industries. http://www.csoisw.gov.in/cms/en/1023-annual-survey-of-industries.aspx (accessed June 1, 2014).

______. 2005–2006b. National Sample Survey Office. http://mospi.nic.in/NSSOa (accessed June1, 2014).

______. 2008. Input–Output Transactions Table 2003–04. http://mospi.nic.in/publication/input-output-transactions-table-2003-04 (accessed August 1, 2014).

Gullickson, William, and Michael J. Harper. 1999. “Possible Measurement Bias in AggregateProductivity Growth.” Monthly Labor Review 122 (2): 47–67.

Ha, Doan Thi Thanh, Kozo Kiyota, and Kenta Yamanouchi. 2016. “Misallocation andProductivity: The Case of Vietnamese Manufacturing.” Asian Development Review 33 (2):94–118.

Hall, Robert E., and Charles I. Jones. 1999. “Why Do Some Countries Produce So Much MoreOutput per Worker than Others?” The Quarterly Journal of Economics 114 (1): 83–116.

Hsieh, Chang-Tai, and Peter J. Klenow. 2009. “Misallocation and Manufacturing TFP in Chinaand India.” The Quarterly Journal of Economics 124 (4): 1403–48.

Hulten, Charles R., Edwin R. Dean, and Michael J. Harper. 2001. “Total Factor Productivity:A Short Biography.” In New Developments in Productivity Analysis, 1–54. Chicago:University of Chicago Press.

Kalemli-Ozcan, Sebnem, and Bent E. Sørensen. 2014. “Misallocation, Property Rights, andAccess to Finance: Evidence from within and across Africa.” In African Successes, VolumeIII: Modernization and Development, edited by Sebastian Edwards, Simon Johnson, andDavid N. Weil, 183–211. Chicago: University of Chicago Press.

Klenow, Peter J., and Andres Rodriguez-Clare. 1997. “The Neoclassical Revival in GrowthEconomics: Has It Gone Too Far?” NBER Macroeconomics Annual, Volume 12, edited byBen S. Bernanke and Julio Rotemberg, 73–103. Cambridge, MA: MIT Press.


La Porta, Rafael, and Andrei Shleifer. 2008. “The Unofficial Economy and EconomicDevelopment.” Brookings Papers on Economic Activity 47 (1): 123–35.

Midrigan, Virgiliu, and Daniel Y. Xu, 2014. “Finance and Misallocation: Evidence from Plant-Level Data.” American Economic Review 104 (2): 422–58.

Organisation for Economic Co-operation and Development. 2001. Measuring Productivity OECDManual: Measurement of Aggregate and Industry-Level Productivity Growth. http://www.oecd.org/sdd/productivity-stats/2352458.pdf.

Oulton, Nicholas, and Mary O’Mahony. 1994. Productivity and Growth: A Study of BritishIndustry 1954–86. National Institute of Economic and Social Research Occasional Papers,Volume 46. Cambridge, UK: Cambridge University Press.

Restuccia, Diego, and Richard Rogerson. 2008. “Policy Distortions and Aggregate Productivitywith Heterogeneous Establishments.” Review of Economic Dynamics 11 (4): 707–20.

Thakur, Tripta, S.G. Deshmukh, S.C. Kaushik, and Mukul Kulshrestha. 2005. “ImpactAssessment of the Electricity Act 2003 on the Indian Power Sector.” Energy Policy 33(9): 1187–98.

US Bureau of Labor Statistics. 2005. National Income and Product Accounts Industry Database.https://www.bls.gov/mfp/mprdload.htm (accessed June 1, 2014).

van der Wiel, Henry P. 1999. “Sectoral Labour Productivity Growth.” Research MemorandumNo. 159, CPB Netherlands Bureau for Economic Policy Analysis.


Appendix

Distribution of Firm-Level Total Factor Revenue Productivity Taking One Distortion at aTime (relative to the energy sector)


Trade Volatility in the Association of SoutheastAsian Nations Plus Three: Impacts

and DeterminantsThi Nguyet Anh Nguyen, Thi Hong Hanh Pham, and

Thomas Vallée∗

This paper investigates trade volatility in the Association of Southeast AsianNations Plus Three (ASEAN+3) and its links with output volatility, exportdiversification, and free trade agreements. To achieve this research objective, weapply several econometric estimators to data from all ASEAN+3 member statesover the period 1990–2016. We first find evidence of a positive relationshipbetween output volatility and trade volatility. Second, we reveal that the wayexport diversification is measured can influence its impacts on bilateral exportvolatility. Moreover, the relationship between income volatility, trade volatility,and export diversification seems to depend on country size and the level ofeconomic development.

Keywords: ASEAN+3, export diversification, FTA, output volatility, tradevolatilityJEL codes: F02, F14, F15, F21, F40

I. Introduction

International trade is considered one of the most volatile components of grossdomestic product (GDP). According to Bennett et al. (2016), trade volatility andGDP volatility have tended to move together over the past 20 years—from slightlyfalling in the mid-1990s until 2008 to sharply increasing after the global crisis of2008–2009. Obviously, trade has become a transmission mechanism of externalshocks throughout the world economy. In addition, researchers have found a positivecorrelation between trade openness and volatility (Di Giovanni and Levchenko2009), which suggests that trade openness has also played a role in explainingthe volatility of economic growth (Rodrik 1997). Caselli et al. (2015) find that

∗Thi Nguyet Anh Nguyen (corresponding author): Faculty of Business Management, National Economics University,Hanoi, Viet Nam. Email: [email protected]; Thi Hong Hanh Pham: LEMNA, Institute of Economics andManagement, University of Nantes, France. Email: thi-hong-hanh.pham@univ-nantes; Thomas Vallée: LEMNA,Institute of Economics and Management, University of Nantes, France. Email: [email protected]. Wewould like to thank the managing editor and the anonymous referee for helpful comments and suggestions. TheAsian Development Bank recognizes “China” as the People’s Republic of China and “Korea” as the Republic ofKorea. The usual ADB disclaimer applies.





countries that suffer from big country-specific shocks experience a reduction involatility as a result of opening up to international trade, because trade is a sourceof diversification.

Trade volatility also varies among countries at different levels of economicdevelopment. Developing countries are substantially more volatile than developedones (Hakura 2009) because developing countries usually concentrate their exportsin a small number of sectors that are particularly sensitive to external shocks.Therefore, diversification in trade should diminish volatility effects. Moreover,developing countries are characterized by underdeveloped financial markets andweak monetary and fiscal policies. Thus, diversification of the sectoral compositioncan enhance the development of financial markets and alleviate uncertainty in theeconomy.

ASEAN+3 is known as a successful model of economic cooperationbetween the Association of Southeast Asian Nations (ASEAN) and three East Asiancountries—Japan, the People’s Republic of China (PRC), and the Republic of Korea.Although trade integration and growth prospects from ASEAN enlargement overthe last 2 decades have been investigated in the literature, trade volatility in theASEAN+3 has attracted little attention. This paper contributes to empirical tradestudies by analyzing the volatility of trade flows and its links with output volatility,international institutions, and trade diversification in ASEAN+3. In other words,our study aims to address the following three questions:

(i) Does output volatility move together with trade volatility in ASEAN+3?

(ii) What are the links between output volatility and export diversification inASEAN+3?

(iii) How do export diversification and free trade agreements (FTAs) impactASEAN+3’s trade volatility?

To answer these questions, we conduct multivariate statistical tests usingdata from 13 ASEAN+3 countries from 1990 to 2016. First, we investigate theimpacts of trade variables, notably trade volatility and export diversification, onASEAN+3’s output volatility by applying the generalized method of moments(GMM) estimator. Second, using a cross-sectional dataset covering ASEAN+3bilateral exports, we explore the potential determinants of ASEAN+3’s bilateralexport volatility by applying fixed effect (FE) and instrumental variable (IV)techniques.

The rest of the paper is organized as follows. Section II summarizes theliterature on the link between trade volatility and economic growth. Section IIIcharacterizes the trend in trade growth volatility, openness, and diversification over

Trade Volatility in ASEAN+3: Impacts and Determinants 169

the past 20 years in ASEAN+3. Section IV explains the impact of trade volatilityon output volatility. Section V explores the determinants of trade volatility. Theconcluding remarks are in section VI.

II. Literature Framework

The theoretical and empirical links between international trade and economicgrowth have largely been investigated. Helpman and Krugman (1985) and Bhagwati(1988) find that export growth promotes economic growth and stimulates boththe supply and demand sides of the economy. Easterly (2007) also argues thatexports boost economic efficiency with better allocation of resources and raiseeconomic growth in the long run. Our study relates to the literature that predicts therelationship between trade volatility, diversification, and output volatility. In an earlypaper, Rodrik (1997) considers whether trade openness increases macroeconomicvolatility but finds that the effects of openness are still ambiguous. Bejan (2006)analyzes the relationship between trade openness and output volatility and findsthat higher trade openness is associated with higher output volatility. Over the lastdecades, several empirical studies have investigated this trade openness–volatilityrelationship (Easterly, Islam, and Stiglitz 2001; Calderón, Loayza, and Schmidt-Hebbel 2005; Cavallo 2008; Jansen et al. 2009; Balavac and Pugh 2016), butopenness remains the most controversial determinant of volatility. For this reason,we investigate the relationship between trade volatility and economic growthvolatility in this study.

The existing literature has also examined the role of export diversificationin controlling risks arising from trade volatility. On the one hand, several empiricalstudies find a negative but not always a significant effect of export diversificationon volatility (e.g., Cavallo 2008 and Cavalcanti, Mohaddes, and Raissi 2012).Calderón and Schmidt-Hebbel (2008) discover a negative relationship betweenopenness and volatility only when exports are diversified. Similarly, Haddad et al.(2013) find strong evidence that export diversification plays an important role inreducing the vulnerability of countries to global shocks. They also find that productdiversification clearly moderates the effect of trade openness on growth volatility,while market concentration measures yield much more mixed results. Meanwhile,Bejan (2006) shows that the interaction term between openness and export productconcentration is significant only in advanced economies. On the other hand,several scholars have found evidence of a positive effect of diversification ongrowth. According to Melitz (2003), an increase in export diversification can boostproductivity given that exporters are more productive than nonexporters. Moreover,Melitz (2003) suggests that export diversification can reduce exposure to externalshocks, thus reducing macroeconomic volatility and increasing economic growth.

Together with the link between trade volatility and export diversification,we also analyze the impact of international trade institutions on trade volatility,


particularly the impact of FTAs. Initial work by Keohane (1983) points out thatvarious international regimes are explicitly designed to promote “orderly patternsof behavior among members,” implying that they would enhance stability andreduce volatility in interactions among members. Martin and Simmons (1998)state that international institutions “lock in a particular equilibrium, providingstability.” According to Abbott (2000), trade agreements are adopted to reducethe risk for private traders and investors. Bagwell and Staiger (2002) maintain thatthe purpose of trade agreements is to allow countries with market power to reachmore efficient market access exchanges. Most recently, Mansfield and Reinhardt(2008) conducted the first large-scale, multivariate statistical tests and stronglyconfirmed that preferential trade agreements and the World Trade Organizationregime significantly reduced export volatility and also increased export levels.

In sum, based on the theoretical background discussed above, we firstinvestigate the impacts of trade volatility and export diversification on ASEAN+3’seconomic growth volatility. Second, we focus on trade volatility in ASEAN+3 byproviding an empirical analysis on its determinants.

III. Trade Volatility, Openness, and Diversification in ASEAN+3

Over the last 2 decades, ASEAN+3 has witnessed fluctuating trade growth.Figure 1 plots the trend of ASEAN+3 trade growth, which is measured by theaverage growth rate of intra-ASEAN+3 exports and imports. As shown in Figure1, ASEAN+3 experienced a significant decrease in intraregional trade during the1997 Asian financial crisis and the 2008 global financial crisis. Trade flows in theregion increased by around 20% after recovering from the 1997 Asian crisis butdropped again in 2002, before climbing by more than 30% in 2005. After a difficultperiod, due to the 2008 global financial crisis, intra-ASEAN+3 trade attained itshighest growth rate in 2010, but has experienced a decreasing trend since 2011.

We now turn our attention to the volatility of ASEAN+3 trade integrationover the period 1990–2016. Figure 2 plots the volatility of total trade growthand of intra-ASEAN+3 and extra-ASEAN+3 trade growth for each membercountry. As seen in Figure 2, ASEAN+3 member states can be divided into twogroups according to trade volatility levels. The first group includes the PRC, theLao People’s Democratic Republic (Lao PDR), Cambodia, Indonesia, and VietNam, which are characterized by a higher level of trade volatility. Moreover, inthese countries, intra-ASEAN+3 and extra-ASEAN+3 trade volatilities do notmove together. For instance, the volatility of trade flows between Viet Nam andextra-ASEAN+3 countries has an increasing trend, while a decreasing trend involatility is observed for Viet Nam’s intra-ASEAN+3 trade. The second groupincludes Brunei Darussalam, Japan, the Republic of Korea, the Lao PDR, Myanmar,


Figure 1. Average Trade Growth Intra-ASEAN+3

ASEAN+3 = Association of Southeast Asian Nations plus Japan, the People’s Republic of China, and the Republicof Korea.Source: Authors’ illustration from UN COMTRADE data.

Malaysia, the Philippines, Singapore, and Thailand. These countries have a fairlystable evolution of trade volatility at different levels. Among all countries, Malaysiaand Japan have the lowest trade volatility.

Together with a significant change in trade growth, ASEAN+3 has alsoexperienced an important evolution in trade openness, which is measured by theratio of total exports and imports to GDP. As reported in Table 1, the average ofthis ratio over the period 1990–2016 ranges from 7% for Japan to over 100% forSingapore in terms of intra-ASEAN+3 trade. In terms of total world trade, this ratiois much higher for all countries, in particular for Viet Nam, Malaysia, and Singapore(mostly over 100%).

Table 1 also shows that the intra-ASEAN+3 openness indicator increasedin most countries in the 2000s, reflecting the growing importance of intraregionaltrade. This ratio continued to rise in the 2010s in many countries such as Cambodia,Japan, the Lao PDR, the Republic of Korea, Thailand, and Viet Nam, but slightlydeclined in other countries. Japan and the Republic of Korea, in particular,experienced a significant increase in this ratio over the study period, indicatinggreater trade integration with the region. We also observe a similar trend inASEAN+3 trade integration in terms of total world trade.


Figure 2. Volatility of Trade Growth by Country

ASEAN+3 = Association of Southeast Asian Nations plus Japan, the People’s Republic of China, and the Republicof Korea; BRU = Brunei Darussalam; CAM = Cambodia; INO = Indonesia; JPN = Japan; KOR = Republic ofKorea; LAO = Lao People’s Democratic Republic; MAL = Malaysia; MYA = Myanmar; PHI = Philippines; PRC =People’s Republic of China; SIN = Singapore; THA = Thailand; VIE = Viet Nam.Source: Authors’ illustration.


Table 1. ASEAN+3 Openness Indicator

Intra-ASEAN+3 Trade Total World Trade

Full 1990– 2000– 2010– Full 1990– 2000– 2010–Country Period 1999 2009 2016 Period 1999 2009 2016

Brunei Darussalam 0.67 0.73 0.64 0.61 0.87 0.94 0.84 0.81Cambodia 0.34 0.31 0.28 0.50 0.73 0.44 0.84 1.00Indonesia 0.24 0.23 0.25 0.22 0.45 0.48 0.47 0.38Lao PDR 0.36 0.35 0.35 0.38 0.47 0.49 0.47 0.45Malaysia 0.79 0.76 0.86 0.73 1.56 1.56 1.71 1.33Myanmar 0.27 0.21 0.33 0.29 0.39 0.30 0.50 0.35Philippines 0.29 0.23 0.37 0.26 0.66 0.63 0.82 0.48Singapore 1.30 1.17 1.48 1.22 2.78 2.69 3.05 2.53Thailand 0.46 0.32 0.52 0.57 0.97 0.73 1.09 1.12Viet Nam 0.57 0.34 0.61 0.85 1.08 0.66 1.15 1.56People’s Republic of China 0.14 0.13 0.17 0.12 0.43 0.36 0.52 0.42Japan 0.07 0.04 0.08 0.12 0.21 0.15 0.23 0.28Republic of Korea 0.24 0.17 0.25 0.34 0.62 0.49 0.63 0.79

ASEAN+3 = Association of Southeast Asian Nations plus Japan, the People’s Republic of China, and the Republicof Korea; Lao PDR = Lao People’s Democratic Republic.Note: The openness indicator refers to the ratio of total exports and imports to gross domestic product.Source: Authors’ calculation.

Regarding export diversification, Figure 3 shows how export diversificationhas evolved in the last 2 decades in the ASEAN+3 market.1 Figure 3 also illustratesthe difference in export concentration between ASEAN and the Plus Three countries(Japan, the PRC, and the Republic of Korea). As displayed in Figure 3, the evolutionof both indicators, the Herfindahl–Hirschman Index (HHI) and the Theil index, issimilar in each country group.2 ASEAN has a higher concentration in exports withan average HHI of up to 0.3 and an average Theil index of up to 3.3, comparedwith 0.1 and 1.9, respectively, for the Plus Three countries. ASEAN experiencedtwo significant upward trends in export concentration after 1997 and 2009. First,between 1997 and 2000, the HHI climbed from 0.17 to 0.25 and the Theil indexincreased from 2.1 to 3. This trend could be explained by an important change intrade policy of ASEAN countries due to the financial crisis. However, since 2010,these indicators have dramatically decreased, corresponding with the trend towardexport diversification in ASEAN. For the Plus Three countries, both the HHI andthe Theil index varied significantly over the study period. Compared to ASEANcountries, the Plus Three countries experienced a significant increase in both theHHI and the Theil index. During 1990–2016, the HHI ranged from 0.08 to 0.11,and the Theil index ranged from 1.62 to 1.83.

Figure 4 illustrates the evolution of export concentration in each ASEAN+3country. A high value of HHI indicates a high level of export concentration.

1In line with our research objective, Figure 3 only displays trends in export concentration. For importconcentration, see Figure A1.1 in Appendix 1.

2See section IV.A for details on computing the HHI and the Theil index.


Figure 3. Export Concentration: ASEAN versus Plus Three

ASEAN = Association of Southeast Asian Nations, HHI = Herfindahl–Hirschman Index.Note: Plus Three refers to Japan, the People’s Republic of China, and the Republic of Korea.Source: Authors’ illustration.

Accordingly, Brunei Darussalam, which mainly specializes in oil and naturalresources, shows the highest concentration in exports with a value of 0.6. Low-and middle-income economies such as Cambodia, the Lao PDR, Myanmar, andViet Nam maintained a high level of export concentration until 2010. However,since 2010, the HHIs in these countries have fallen, suggesting the start of atrade diversification policy. Compared to other countries, the PRC and Japan havemaintained a low level of HHI throughout the period 1990–2016, clearly affirmingtheir trade diversification policy.

In short, these stylized facts on trade volatility, openness, and exportdiversification provide more details about trade patterns in ASEAN+3, which partlysupport the research objectives of this paper.

IV. Impacts of Trade Volatility

This section addresses the question of whether trade volatility contributesto economic growth volatility in ASEAN+3. Moreover, we also investigate thepotential role of export diversification in explaining output volatility.


Figure 4. Export Diversification Trends by Country

BRU = Brunei Darussalam, CAM = Cambodia, HHI = Herfindahl–Hirschman Index, INO = Indonesia, JPN =Japan, KOR = Republic of Korea, LAO = Lao People’s Democratic Republic, MAL = Malaysia, MYA = Myanmar,PHI = Philippines, PRC = People’s Republic of China, SIN = Singapore, THA = Thailand, VIE = Viet Nam.Source: Authors’ illustration.

A. Methodology

This paper aims to explain the pace in output volatility and its variationacross ASEAN+3 by employing an empirical model, which will allow us to test themain hypothesis of interest. Given this objective, we try to make maximum use ofboth time and cross-country dimensions of our panel dataset, including data at anannual frequency. Based on an annual dataset, our empirical model will considerthe possible slow adjustment of output volatility to changes in other variablesin any given year. Therefore, we specify a dynamic equation, which includes alagged dependent variable. The potential impacts of trade volatility and exportdiversification on output volatility are formulated as follows:

GDPvoli,t = α0 +β1Tvoli,t−1 +β2DIVi,t−1 +β3CONi,t−1 +β4FTAi,t +β5Zi,t +ui,t

(1)

where the dependent variable, GDPvoli,t , is the volatility of the growth of realGDP per capita for country i in period t. As argued in many studies (Bejan 2006,Cavallo 2008, Jansen et al. 2009, and Haddad et al. 2013), we use the growth rateinstead of output level, since it is of greater interest to policy makers to measureeconomic growth stability. Tvoli,t−1 is the volatility of total trade (sum of exports


and imports) of country i in period t − 1. In this paper, we use the 5-year rollingstandard deviation to measure the volatility of output and trade.

In equation (1), uit is an error term that contains country and time-specificfixed effects:

uit = μi + εt + ϑit (2)

where ϑit is independent and identically distributed with mean 0 and variance σ 2ϑ .

DIVi,t−1 is the level of export diversification in country i in period t − 1, whichis proxied by different alternative concentration indicators including the HHI, theTheil index, and the similarity index (SI). Cadot, Carrère, and Strauss-Kahn (2013)decompose the Theil index into two indicators for the intensive and extensiveproduct margins. The extensive Theil index captures concentration in the numberof products (extensive product margin), whereas the intensive Theil index measuresconcentration in the sales volume of products (intensive product margin).

The Theil index is thus T = T ext + T int , where the intensive Theil index isgiven by

T int = 1

Nx

∑k∈Nx

Rk

Rxln

(Rk

Rx

)(3)

and the extensive Theil index is

T ext = ln

(N

Nx

)(4)

where Nx denotes the number of exported products, Rk is the value of exports ofproduct k, and Rx represents the mean value of exported products. Like the HHI, ahigher value means higher export concentration.

The HHI, by construction, measures the changes in the distribution ofexport shares. In other words, this indicator is a measure of the degree of productconcentration. The following normalized HHI is used in order to obtain valuesbetween 0 and 1:

Hj =

√∑ni=1

(xi j

Xj

)2− √

1/n

1 − √1/n

Xj =n∑

i=1

xi j (5)

where xi j is the value of exports for country j and product i, and n is the numberof products at the level of classification. An index value closer to 1 indicates thata country’s exports or imports are highly concentrated on a few products. On thecontrary, values closer to 0 mean that exports or imports are more homogeneouslydistributed among a series of products.


Different from the HHI and Theil index, the similarity index SI is consideredthe simplest measure for comparing export content across countries or across time.To calculate export SI, we use the original export similarity index first introduced byFinger and Kreinin (1979) and largely applied to trade data. The Finger and Kreininsimilarity index between two countries c and d is given as follows:

SIFKc,d =

n∑i=1

min

(xc,i

Xc,

xd,i

Xd

)(6)

where xc,i

Xcand xd,i

Xdare the shares of good i in the total exports of country c and

d, respectively. This index ranges from 0 to 1—from completely different exportshares to identical export shares. According to Finger and Kreinin (1979), thismeasure should not be affected by the relative sizes or scales of total exportsas it is intended to compare only patterns of exports across product categories.All three concentration indexes are calculated at the 4-digit level of the StandardIndustry Trade Classification Revision 3 from the United Nations Commodity TradeStatistics Database.

Apart from trade, Easterly, Islam, and Stiglitz (2000) find evidence ofbivariate correlation between per capita growth volatility and financial sectordevelopment, and price policies. Following this work, we also introduce in equation(1) a set of control variables CONit , which includes inflation volatility and thelevel of financial development, which is measured by the ratio of private sectorbanking credit to GDP and a financial openness indicator. The quality of domesticinstitutions has also been recognized as a cause of macroeconomic instability in theliterature. For instance, research by Rodrik (1999) and Acemoglu et al. (2003) hasfound that corruption and conflict can lead to misleading trade policy. Similarly,Mobarak (2005) and Klomp and de Haan (2009) have found that a democracy ora stable political regime supports macroeconomic stability. Thus, we introduce inequation (1) an institution variable which is proxied by the index of political rightsfrom Freedom House.3 The index ranges from 1 to 7, with 1 indicating the most“free” country. We also include the variable FTAi,t to see the impact of a free tradeagreement in ASEAN on the volatility of output. In addition, time dummies areincluded to control common output fluctuations during financial crises, notably theAsian economic crisis (1997–1999) and the global financial crisis (2007–2009).

The data are summarized in Table 2, which provides definitions and sourcesfor all variables and their units of measurement, mean, standard deviations,and minimum and maximum values. This table also reports the correlationcoefficients between output volatility, trade volatility, and independent variables.The signs, magnitudes, and significance of correlation coefficients help in modelingand confirming the choice of variables. However, the values of the correlationcoefficients are diverse, ranging from negative to positive and small to large. This

3See Balavac and Pugh (2016).

178 Asian Development ReviewTa

ble

2.O

utpu

tV

olat

ility

Mod

el:

Sum

mar

ySt

atis

tics

Cor

rela

tion

wit

h

Des

crip

tive

Stat

isti

csIn

com

eT

rade

Var

iabl

eM

easu

rem

ent

Sour

ceO

bsM

ean

SDM

inM

axV

olat

ility

Vol

atili

ty

Inco

me

vola

tili

ty5-

year

roll

ing

stan

dard

devi

atio

nof

GD

Ppe

rca

pita

grow

thW

DI

325

0.10

50.

077

0.00

00.

459

1.00

0

Tra

devo

lati

lity

5-ye

arro

llin

gst

anda

rdde

viat

ion

ofto

tale

xpor

tsan

dim

port

sU

NC

OM

TR

AD

E32

50.

235

0.27

60.

023

1.76

60.

405*

1.00

0

HH

IA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

274

0.21

20.

145

0.05

50.

707

0.21

1*0.

118

Inte

nsiv

eT

heil

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

274

2.40

00.

838

1.20

35.

566

0.24

9*0.

091

Ext

ensi

veT

heil

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

274

0.30

70.

536

0.03

54.

301

−0.0

360.

118

The

ilin

dex

Aut

hors

’co

mpu

tati

onU

NC

OM

TR

AD

E27

42.

707

1.12

01.

276

6.13

20.

168*

0.12

4*

Sim

ilar

ity

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

274

7.34

82.

787

0.77

014

.616

−0.0

04−0

.093

Infl

atio

nvo

lati

lity

5-ye

arro

llin

gst

anda

rdde

viat

ion

ofin

flat

ion

rate

WD

I33

45.

521

7.84

90.

019

53.6

290.

513*

0.75

6*

Fina

ncia

lde

velo

pmen

tR

atio

ofpr

ivat

ese

ctor

bank

ing

cred

itto

GD

PW

DI

321

6.46

95.

615

0.06

031

.081

0.01

1−0

.092

Fina

ncia

lop

enne

ssK

AO

PE

NC

hinn

and

Ito

(200

8)31

90.

032

1.43

4−1

.904

2.37

4−0

.235

*−0

.215

*

Poli

tica

lrig

hts

Sco

reof

1–7,

wit

h1

repr

esen

ting

the

grea

test

degr

eeof

free

dom

and

7th

esm

alle

stde

gree

offr

eedo

m

Free

dom

Hou

se35

14.

749

2.21

41.

000

7.00

00.

081

0.20

7*

Free

trad

eag

reem

ent

Dum

my

AS

EA

NS

ecre

tari

at35

10.

684

0.46

60.

000

1.00

00.

137*

0.11

1*

AS

EA

N=

Ass

ocia

tion

ofS

outh

east

Asi

anN

atio

ns,G

DP

=gr

oss

dom

esti

cpr

oduc

t,H

HI=

Her

find

ahl–

Hir

schm

anIn

dex,

KA

OP

EN

=C

hinn

–Ito

inde

x,S

D=

stan

dard

devi

atio

n,U

NC

OM

TR

AD

E=

Uni

ted

Nat

ions

Com

mod

ity

Tra

deS

tati

stic

sD

atab

ase,

WD

I=

Wor

ldD

evel

opm

entI

ndic

ator

s.N

ote:

*in

dica

tes

sign

ifica

nce

atth

e5%

leve

l.S

ourc

e:A

utho

rs’

com

pila

tion

base

don

vari

ous

sour

ces.


means we should expect different potential impacts of the independent variables onoutput volatility.

Due to the inclusion of lagged dependent variables in equation (1), thecombination of fixed effects and lagged dependent variables introduces seriouseconometric bias. According to Nerlove (1967) and Nickell (1981), the ordinaryleast squares estimates of the lagged dependent variable’s coefficient in a dynamicpanel model are biased due to the correlation between the fixed effects and thelagged dependent variable.4 Judson and Owen (1999) also suggest that the bias isinversely related to panel length (“T”), but potentially severe biases remain even atT = 30. In this case, the preferred estimator is the GMM suggested by Arellano andBond (1991), which was then developed and extended by Arellano and Bover (1995)and Blundell and Bond (1998) for two reasons. First, the GMM differences themodel of interest to remove country-specific effects or any time-invariant country-specific variable. Second, this method eliminates possible correlation betweenthe country-specific effects and the regressors. In the GMM estimator, momentconditions utilize the orthogonality conditions between differenced errors andthe dependent variable’s lagged values. This means that the disturbances ϑit areserially uncorrelated and that the differenced error is a moving average with 1lag (MA[1]). Therefore, two diagnostics are computed to test for first-order andsecond-order serial correlation in the disturbances. We expect to reject the null ofthe absence of first-order serial correlation and not reject the absence of second-order serial correlation. In the GMM procedure, the number of moment conditionsincreases with time span t. As a result, the Sargan test is performed to test foroveridentification restrictions.

B. Results and Discussion

Our empirical estimations are organized in two steps. First, the benchmarkdynamic GMM estimation treats all right-hand side variables other than the laggeddependent variable as if they were exogenous. In this benchmark estimation, welag all independent variables by one period. Second, we run the dynamic GMMestimator in which the export volatility is treated as endogenous by using anadditional instrument suggested by related literature. We follow Frankel and Romer(1999) who developed “natural openness” instruments for openness. According toWei (2000), natural openness is found by estimating the level of trade opennessdepending on a country’s size and geographic and linguistic characteristics. Inparticular, we estimate the following equation:

log (Openness) = β21Remoteness + β22 log (Population) + β23Lang

+ β24Geo + ε21 (7)

4See further Baltagi (2008).


where Remoteness = ∑j

distancei j

GDPj/GDPw, a formula that measures a country’s average

weighted distance from its trading partners (Head 2003), where the weights arethe partner countries’ shares of total GDP (denoted by GDPw). “Lang” is proxiedby two dummies of language, “English” and “Chinese,” each of which takesthe value of 1 if the country speaks the respective language and 0 otherwise.“Geo” is constructed from the land area and a dummy called “landlock” if thecountry is landlocked. All data are adopted from Centre d’Etudes Prospectiveset d’Informations Internationales (CEPII) data. The method to construct an IV isbased on the same concept as the gravity model. Appendix 2 reports a successionof regressions on ASEAN+3 openness over the period 1990–2016. The predictedvalue of the log of openness (trade-to-GDP ratio) from equation (7) is used as an IVfor trade volatility to deal with the biases.

All empirical results are summarized in Table 3. We estimate equation (1)separately for each of our five export diversification indexes. It is also worth notingthat in all the output volatility regressions reported in Table 3, all diagnosticsare satisfactory, irrespective of whether the trade volatility terms are treated asexogenous or endogenous. First, the presence of first-order serial correlation is notrejected, while the presence of second-order serial correlation is rejected. Second,the Sargan test does not reject the overidentification restrictions. In addition, thelagged dependent variable in all regressions is positive and statistically significant.These results allow us to conclude that the dynamic GMM is an appropriateestimator.

Going straight to the hypothesis of interest, we note that the estimatedcoefficients of trade volatility have the expected sign with a high level ofsignificance. This result suggests that an increase in intra-ASEAN+3 tradevolatility leads to an increase in income volatility of member countries. In addition,when the trade volatility terms are treated as endogenous, the sign and significanceof their estimates are qualitatively similar to those obtained in the benchmark GMMmodels. However, due to the endogeneity treatment, the empirically estimatedcoefficients are somewhat smaller. The positive link between trade and incomevolatility is also illustrated in Figure 5.

In sum, our empirical results confirm that trade has a quantitatively largeand robust positive effect on income (Frankel and Romer 1999). This result is alsoconsistent with that of Cavallo (2008) who argues that output volatility naturallyrelates to the frequency and size of the shocks affecting an economy and to themanner in which the economy handles the shocks. Accordingly, trade openness isassociated with greater output volatility. In other words, the more exposed a countryis to trade, the more vulnerable it is to shocks coming from abroad.

Examining the estimated coefficients associated with export diversificationvariables, we first note that their signs depend on the way export diversificationis measured. For instance, estimates of the extensive Theil index have a negativevalue, while those of the intensive Theil index have a positive value. Second, in

Trade Volatility in ASEAN+3: Impacts and Determinants 181Ta

ble

3.G

ener

aliz

edM

etho

dof

Mom

ents

Est

imat

orR

esul

tsof

the

Out

put

Vol

atili

tyM

odel

Mea

sure

ofE

xpor

tD

iver

sific

atio

n

HH

IE

xten

sive

Inte

nsiv

eT

heil

SI

Lag

ged

depe

nden

tvar

iabl

e0.

942**

0.51

9**0.

873*

0.46

1*0.

886**

0.51

2**0.

922**

0.51

2**0.

781*

0.46

4**

(0.3

94)

(0.2

08)

(0.4

04)

(0.2

17)

(0.3

95)

(0.1

88)

(0.3

72)

(0.1

95)

(0.3

69)

(0.1

66)

Tra

devo

lati

lity

(lag

ged)

0.11

2*0.

073*

0.10

1*0.

069*

0.10

8*0.

078**

0.11

0*0.

075*

0.12

6*0.

081**

(0.0

58)

(0.0

35)

(0.0

57)

(0.0

38)

(0.0

60)

(0.0

36)

(0.0

59)

(0.0

35)

(0.0

63)

(0.0

37)

Exp

ortd

iver

sifi

cati

on(l

agge

d)−0

.031

−0.0

71−0

.035

−0.0

500.

016

0.00

10.

007

−0.0

040.

011*

0.00

4(0

.124

)(0

.130

)(0

.027

)(0

.035

)(0

.019

)(0

.012

)(0

.014

)(0

.012

)(0

.005

)(0

.006

)In

flat

ion

vola

tili

ty(l

agge

d)0.

005

0.01

2**0.

007

0.01

1**0.

007

0.01

1**0.

006

0.01

0**0.

008

0.01

1**

(0.0

09)

(0.0

04)

(0.0

10)

(0.0

05)

(0.0

10)

(0.0

04)

(0.0

09)

(0.0

04)

(0.0

10)

(0.0

05)

Fina

ncia

ldev

elop

men

t(l

agge

d)−0

.001

0.00

0−0

.001

0.00

0−0

.001

0.00

0−0

.001

0.00

0−0

.001

0.00

0(0

.000

)(0

.001

)(0

.001

)(0

.001

)(0

.000

)(0

.000

)(0

.000

)(0

.000

)(0

.001

)(0

.000

)Fi

nanc

ialo

penn

ess

(lag

ged)

0.03

00.

009

0.02

80.

005

0.02

90.

007

0.02

80.

007

0.01

90.

005

(0.0

25)

(0.0

21)

(0.0

26)

(0.0

18)

(0.0

24)

(0.0

15)

(0.0

23)

(0.0

17)

(0.0

23)

(0.0

13)

Poli

tica

lrig

hts

(lag

ged)

0.01

4**0.

020*

0.01

70.

021*

0.01

10.

018

0.01

4**0.

019

0.00

60.

016*

(0.0

07)

(0.0

11)

(0.0

10)

(0.0

11)

(0.0

07)

(0.0

11)

(0.0

07)

(0.0

13)

(0.0

08)

(0.0

08)

Free

trad

eag

reem

ent

−0.0

20−0

.040

−0.0

23−0

.046

−0.0

52−0

.053

−0.0

41−0

.045

−0.0

46−0

.058

(0.0

40)

(0.0

37)

(0.0

29)

(0.0

34)

(0.0

37)

(0.0

46)

(0.0

35)

(0.0

43)

(0.0

29)

(0.0

43)

Cri

sis

0.02

9***

0.02

6**0.

029**

0.02

5**0.

026**

0.02

4**0.

027**

*0.

025**

0.01

70.

021**

(0.0

08)

(0.0

11)

(0.0

11)

(0.0

11)

(0.0

09)

(0.0

10)

(0.0

08)

(0.0

10)

(0.0

11)

(0.0

09)

Tre

atm

ento

ftr

ade

vola

tili

tyE

xoge

neou

sE

ndog

enou

sE

xoge

neou

sE

ndog

enou

sE

xoge

neou

sE

ndog

enou

sE

xoge

neou

sE

ndog

enou

sE

xoge

neou

sE

ndog

enou

s

Firs

t-or

der

seri

alco

rrel

atio

nte

st(p

-val

ue)

−2.2

−1.6

5−2

.27

−1.6

2−2

.18

−1.7

2−2

.2−1

.7−1

.97

−1.6

5(0

.028

)(0

.099

)(0

.023

)(0

.105

)(0

.029

)(0

.085

)(0

.028

)(0

.089

)(0

.049

)(0

.099

)S

econ

d-or

der

seri

alco

rrel

atio

nte

st(p

-val

ue)

−1.4

6−1

.25

−1.5

3−1

.27

−1.5

7−1

.19

−1.4

7−1

.2−1

.24

−1.0

9(0

.143

)(0

.211

)(0

.126

)(0

.204

)(0

.117

)(0

.236

)(0

.142

)(0

.231

)(0

.215

)(0

.277

)S

arga

nte

stof

over

iden

tifi

cati

onre

stri

ctio

ns(p

-val

ue)

15.3

540

.68

14.7

538

.03

14.7

439

.56

15.4

040

.15

14.0

538

.03

(0.7

00)

(0.6

15)

(0.7

39(0

.724

)(0

.739

)(0

.662

)(0

.697

)(0

.637

)(0

.781

)(0

.724

)

HH

I=

Her

find

ahl–

Hir

schm

anIn

dex,

SI=

sim

ilar

ity

inde

x.N

otes

:Fig

ures

inpa

rent

hese

sar

est

anda

rder

rors

.***,**

,and

*in

dica

test

atis

tica

lsig

nifi

canc

eat

the

1%,5

%,a

nd10

%le

vels

,res

pect

ivel

y.S

ourc

e:A

utho

rs’

esti

mat

es.


Figure 5. Output Volatility versus Trade Volatility

ASEAN+3 = Association of Southeast Asian Nations plus Japan, the People’s Republic of China, and the Republicof Korea; GDP = gross domestic product.Note: Out-of-sample predictions are from the generalized method of moments estimator with the Herfindahl–Hirschman Index.Source: Authors’ illustration.

most regressions, the impact of export diversification on output volatility is notstatistically significant. Exceptionally, we find a significant and positive estimatedcoefficient on the similarity index in the GMM model that considers trade volatilityas an exogenous regressor. However, this coefficient loses its significance whentrade volatility is treated as an endogenous variable. Third, the empirical resultsrelating to export diversification are not sensitive to whether the trade volatilityterms are treated as endogenous. Overall, we do not find evidence of a significantand direct link between export diversification and output volatility. Despite theabsence of a direct impact, export diversification could still affect output volatilitythrough its potential link with export volatility. This mechanism will be studied inthe next section.

We now turn our attention to the impact of institutional variables, representedby political rights and FTAs. First, as expected, we find that the estimatedcoefficients for political rights all have a positive value. These coefficients aregenerally significant, except in the intensive Theil index model. This finding meansthat a “freer” country experiences less volatile output growth. In other words,ASEAN+3 countries with a higher level of political rights and democracy havea lower level of output volatility. Second, the estimates associated with FTAs arenegative but statistically insignificant. In fact, it is noteworthy that there are twoeconomic aspects of FTA. On the one hand, an FTA could divert trade away frommore efficient suppliers outside the area toward less efficient ones within the area.On the other hand, an FTA could create trade that may not have otherwise existedand thus raise a country’s national welfare. Reducing the output volatility of acountry should also be considered a trade creation effect of an FTA. According


to our empirical results, this effect potentially exists, but ASEAN+3 countries havenot yet realized it. In other words, the potential trade creation effects of FTAs arenot working in ASEAN+3 in precisely the way that is described by the relevantliterature, suggesting that more nuanced political economy explanations may beneeded. This result can motivate policy makers in ASEAN+3 to conduct furtheranalysis to find out exactly what makes FTAs beneficial, whether by enhancinggrowth or lowering income volatility. The resulting inferences could then be appliedto future FTA negotiations in order to effectively achieve an expected long-runoutcome from FTAs.

On the set of control variables, we find that inflation volatility plays asignificant role in explaining output volatility when trade volatility is treatedas endogenous. This means that a stable macroeconomic environment, whichis captured by the level of inflation volatility, could promote output stability.By contrast, financial variables in all regressions have insignificant estimatedcoefficients. For instance, as expected, the sign of the coefficients on the level offinancial development is negative, meaning a more developed financial system isassociated with more stable output growth, although the results are not statisticallysignificant. Looking at the financial openness variable, the estimates of this variableare positive in all regressions. This positive sign supports the theoretical relationshipbetween financial openness and income volatility—the more integrated a financialsystem the more sensitive a country is to external shocks and the more volatile itsincome growth. Notwithstanding, this potential effect is not statistically significant.

V. Determinants of Trade Volatility

In this section, we use a cross-sectional dataset of intra-ASEAN+3 bilateralexports over the period 1990–2016 to investigate the potential determinants ofASEAN+3 trade volatility.

A. Methodology

The potential determinants of ASEAN+3 trade volatility are modeled in thefollowing equation:

log (Xvol)i, j,t = α2 + θ0 log (INCOMEvol)i,t + θ1log(INCOMEvol) j,t + θ2DIVXi, j,t

+ θ3TARIFFi, j,t + θ4FTAi, j,t + θ5Wi, j,t + μi, j + δt + ε3i,t (8)

where we use the 5-year rolling standard deviation of bilateral exports Xvoli, j,t

between countries i and j as a proxy of trade volatility. Export diversification,DIVXi, j,t , is proxied by five alternative indicators: the HHI, the SI, the Theil index,and the intensive and extensive Theil indexes. INCOMEvoli,t and INCOMEvol j,t arethe income volatilities of the reporter country and its trading partner, respectively.


FTAi, j,t is a dummy variable indicating whether countries i and j are members ofthe same FTA in year t. TARIFFi, j,t is the tariff applied on imported products fromcountry i to country j in year t, which is measured by the most favored nation (MFN)and the effectively applied (AHS) weighted averages. Wi, j,t is a vector of controlvariables, which are proxied by the bilateral exchange rate and political rights.5 μi, j

is a country-specific fixed effect, and δt is a year-specific fixed effect. The statisticalsummary of all variables in equation (8) is reported in Table 4.

According to the Hausman test results, the fixed-effects estimator is theappropriate technique to estimate equation (8).6 In addition, we also reestimateequation (6) by using the IV technique, which allows us to tackle the potentialendogeneity of the income volatility variable. We use country’s infant mortalityrate and the corruption perception index as two external instrumental variables ofthe income volatility variable.7 According to Kalemli-Ozcan (2002), birth rate isindeed a determinant of economic growth. Lower child mortality results in highereducational investment and lower fecundity by parents, which in turn causes lowerpopulation growth and higher economic growth. The other variable, corruptioncontrol, is an essential instrument of governance, which has a sizable long-run effecton economic growth (Kaufmann, Kraay, and Mastruzzi 2007).

Furthermore, we distinguish the impact of each ASEAN+1 FTA (betweenASEAN and the PRC, Japan, or the Republic of Korea) on trade volatility byestimating equation (8) with the presence of three dummies—ACFTA, AJFTA, andAKFTA—which represent ASEAN’s FTA with the PRC, Japan, and the Republicof Korea, respectively.

B. Results and Discussion

We summarize the empirical results of FE and IV estimators in Table 5. Inparticular, the IV results that treat income volatility as endogenous indicate thevalidity of the instruments, which is shown by several diagnostic tests. First, thep-values for Anderson’s canonical correlation test confirm the adequate explanatorypower of our instruments. Second, the Cragg–Donald Wald F-statistics allow usto reject the null hypothesis of a weak instrument. Third, the overidentificationrestriction test (the Sargan statistics) accepts the null hypothesis of the instruments’validity.

The estimated results in Table 5 allow us to confirm a positive and significantimpact of income volatility of the reporter country on its bilateral export volatility.

5The bilateral real exchange rate is calculated as the product of the nominal exchange rate and the relativeGDP deflator in each country: RERi j= ei jt *(p jt /pit ), where ei jt is the nominal exchange rate (IMF, InternationalFinancial Statistics), p jt is the GDP deflator of the exporter, and pit is the GDP deflator of the importer.

6The Hausman test results are not reported here to save space.7Infant mortality rate is extracted from World Bank data. Corruption perception index is adopted from

Transparency International.


ble

4.E

xpor

tV

olat

ility

Mod

el:

Sum

mar

ySt

atis

tics

Cor

rela

tion

wit

hE

xpor

tV

aria

ble

Mea

sure

men

tSo

urce

Obs

Mea

nSD

Min

Max

Vol

atili

ty

Exp

ortv

olat

ilit

y5-

year

roll

ing

stan

dard

devi

atio

nof

AS

EA

N+3

’sbi

late

rale

xpor

tvol

atil

ity

UN

CO

MT

RA

DE

2,54

40.

352

0.36

20.

001

5.23

61.

000

Inco

me

vola

tili

ty(r

epor

ter)

5-ye

arro

llin

gst

anda

rdde

viat

ion

ofG

DP

per

capi

tagr

owth

(rep

orte

rco

untr

y)W

DI

2,86

00.

160

0.11

70.

000

0.97

10.

112*

Inco

me

vola

tili

ty(p

artn

er)

5-ye

arro

llin

gst

anda

rdde

viat

ion

ofG

DP

per

capi

tagr

owth

(par

tner

coun

try)

WD

I2,

860

0.16

10.

115

0.00

00.

971

0.06

6*

HH

IA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

2,54

20.

267

0.19

30.

042

1.00

00.

301*

Inte

nsiv

eT

heil

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

2,55

42.

331

0.78

20.

000

5.31

1−0

.037

Ext

ensi

veT

heil

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

2,55

40.

963

1.12

80.

044

6.94

00.

460*

The

ilin

dex

Aut

hors

’co

mpu

tati

onU

NC

OM

TR

AD

E2,

554

3.29

51.

264

1.24

06.

940

0.37

1*

Sim

ilar

ity

inde

xA

utho

rs’

com

puta

tion

UN

CO

MT

RA

DE

2,55

434

.420

19.1

100.

001

75.9

86−0

.406

*

MF

Nta

riff

Wei

ghte

dm

ostf

avor

edna

tion

tari

ffs

WIT

S2,

155

7.82

88.

714

0.00

088

.941

0.09

1*

AH

Sta

riff

Wei

ghte

def

fect

ivel

yap

plie

dta

riff

sW

ITS

2,15

56.

108

7.95

80.

000

65.9

150.

068*

Exc

hang

era

tevo

lati

lity

Aut

hors

’co

mpu

tati

onIM

F2,

860

0.11

70.

137

0.00

00.

936

0.18

7*

Free

trad

eag

reem

ent

Dum

my

AS

EA

NS

ecre

tari

at2,

970

0.62

00.

486

0.00

01.

000

0.09

8*

AC

FTA

Dum

my

AS

EA

NS

ecre

tari

at2,

970

0.29

10.

454

0.00

01.

000

0.00

8A

JFTA

Dum

my

AS

EA

NS

ecre

tari

at2,

970

0.21

50.

411

0.00

01.

000

−0.0

77*

AK

FTA

Dum

my

AS

EA

NS

ecre

tari

at2,

970

0.24

20.

429

0.00

01.

000

−0.0

18Po

liti

calr

ight

s(r

epor

ter)

Sco

reof

1–7,

wit

h1

repr

esen

ting

the

grea

test

degr

eeof

free

dom

and

7th

esm

alle

stde

gree

offr

eedo

m

Free

dom

Hou

se2,

970

4.35

42.

179

1.00

07.

000

0.19

7*

Poli

tica

lrig

hts

(par

tner

)S

core

of1–

7,w

ith

1re

pres

enti

ngth

egr

eate

stde

gree

offr

eedo

man

d7

the

smal

lest

degr

eeof

free

dom

Free

dom

Hou

se2,

970

4.35

42.

179

1.00

07.

000

0.18

3*

AC

FTA

=A

SE

AN

–Chi

naFr

eeT

rade

Are

a;A

HS

=ef

fect

ivel

yap

plie

dta

riff

;AJF

TA=

AS

EA

N–J

apan

Free

Tra

deA

rea;

AK

FTA

=A

SE

AN

–Kor

eaFr

eeT

rade

Are

a;A

SE

AN

+3=

Ass

ocia

tion

ofS

outh

east

Asi

anN

atio

nspl

usJa

pan,

the

Peop

le’s

Rep

ubli

cof

Chi

na,a

ndth

eR

epub

lic

ofK

orea

;HH

I=

Her

find

ahl–

Hir

schm

anIn

dex;

IMF

=In

tern

atio

nalM

onet

ary

Fun

d;M

FN

=m

ost

favo

red

nati

on;

SD

=st

anda

rdde

viat

ion;

UN

CO

MT

RA

DE

=U

nite

dN

atio

nsC

omm

odit

yT

rade

Sta

tist

ics

Dat

abas

e;W

DI=

Wor

ldD

evel

opm

ent

Indi

cato

rs;

WIT

S=

Wor

ldIn

tegr

ated

Tra

deS

yste

m.

Not

e:*

indi

cate

ssi

gnifi

canc

eat

the

5%le

vel.

Sou

rce:

Aut

hors

’co

mpi

lati

onba

sed

onva

riou

sso

urce

s.


ble

5.E

stim

atio

nR

esul

tsof

the

Exp

ort

Vol

atili

tyM

odel

Lin

ear

Reg

ress

ion

Qua

ntile

Reg

ress

ion

HH

ISI

The

ilH

HI

SIT

heil

FE

IVF

EIV

FE

IV1s

t5t

h1s

t5t

h1s

t5t

h

Inco

me

vola

tili

ty(r

epor

ter)

0.54

8***

0.35

5*0.

548**

*0.

509**

0.56

6***

0.47

7**0.

558**

*0.

539**

*0.

553**

*0.

543**

*0.

573**

*0.

559**

*

(0.0

57)

(0.2

16)

(0.0

58)

(0.2

04)

(0.0

56)

(0.2

22)

(0.0

58)

(0.1

11)

(0.0

65)

(0.0

77)

(0.0

64)

(0.0

83)

Inco

me

vola

tili

ty(p

artn

er)

0.08

60.

093

0.09

10.

182

0.07

50.

154

0.09

90.

072

0.10

20.

080

0.09

10.

058

(0.0

63)

(0.1

89)

(0.0

62)

(0.1

92)

(0.0

64)

(0.1

98)

(0.0

70)

(0.1

32)

(0.0

79)

(0.0

93)

(0.0

75)

(0.0

98)

Bil

ater

alex

chan

gera

tevo

lati

lity

−0.1

75**

−0.0

85−0

.207

**−0

.255

−0.1

80**

−0.2

19−0

.187

**−0

.161

−0.2

22**

−0.1

92−0

.195

**−0

.163

(0.0

80)

(0.2

01)

(0.0

79)

(0.1

96)

(0.0

79)

(0.2

05)

(0.0

90)

(0.1

71)

(0.1

01)

(0.1

20)

(0.0

96)

(0.1

25)

Exp

ortd

iver

sifi

cati

onin

dex

0.40

0***

0.35

8***

0.00

10.

000

0.08

2***

0.03

2***

0.20

0*0.

617**

0.00

10.

000

0.04

7*0.

120*

(0.1

07)

(0.0

42)

(0.0

01)

(0.0

01)

(0.0

20)

(0.0

10)

(0.1

04)

(0.2

94)

(0.0

01)

(0.0

02)

(0.0

26)

(0.0

34)

AH

Sta

riff

0.00

20.

004**

*0.

002

0.00

4***

0.00

10.

004**

*−0

.001

0.00

5*−0

.001

0.00

6**−0

.002

0.00

5*

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

03)

(0.0

02)

(0.0

03)

(0.0

02)

(0.0

03)

MF

Nta

riff

0.00

0−0

.003

***

−0.0

01−0

.003

**0.

000

−0.0

03**

*0.

002

−0.0

02**

0.00

2−0

.004

*0.

002

−0.0

05*

(0.0

02)

(0.0

01)

(0.0

01)

(0.0

01)

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

03)

Free

trad

eag

reem

ent

−0.0

07−0

.010

−0.0

14−0

.009

−0.0

05−0

.008

−0.0

05−0

.021

0.00

0−0

.029

0.00

7−0

.018

(0.0

20)

(0.0

11)

(0.0

21)

(0.0

11)

(0.0

20)

(0.0

12)

(0.0

22)

(0.0

43)

(0.0

26)

(0.0

31)

(0.0

24)

(0.0

31)

AC

FTA

−0.0

71**

*−0

.038

*−0

.078

**−0

.059

***

−0.0

68**

−0.0

49**

−0.0

55*

−0.0

90*

−0.0

65*

−0.0

91**

−0.0

53*

−0.0

84*

(0.0

27)

(0.0

21)

(0.0

30)

(0.0

21)

(0.0

27)

(0.0

22)

(0.0

32)

(0.0

50)

(0.0

38)

(0.0

45)

(0.0

34)

(0.0

44)

AJF

TA−0

.025

−0.0

14−0

.025

−0.0

18−0

.022

−0.0

14−0

.018

−0.0

33−0

.018

−0.0

33−0

.016

−0.0

28(0

.022

)(0

.014

)(0

.024

)(0

.014

)(0

.022

)(0

.014

)(0

.020

)(0

.039

)(0

.024

)(0

.028

)(0

.022

)(0

.028

)A

KF

TA0.

034

0.00

90.

033

0.00

90.

038

0.01

00.

031

0.03

70.

036

0.02

90.

033

0.04

3(0

.024

)(0

.018

)(0

.024

)(0

.018

)(0

.024

)(0

.019

)(0

.023

)(0

.044

)(0

.026

)(0

.031

)(0

.025

)(0

.032

)Po

liti

calr

ight

s(r

epor

ter)

0.00

0−0

.009

**−0

.003

−0.0

12**

−0.0

03−0

.012

**−0

.004

0.00

4−0

.005

0.00

0−0

.006

0.00

1(0

.008

)(0

.005

)(0

.007

)(0

.005

)(0

.008

)(0

.005

)(0

.006

)(0

.012

)(0

.007

)(0

.008

)(0

.007

)(0

.008

)Po

liti

calr

ight

s(p

artn

er)

−0.0

05−0

.001

−0.0

04−0

.002

−0.0

05−0

.001

−0.0

07−0

.002

−0.0

07−0

.002

−0.0

07−0

.003

(0.0

05)

(0.0

05)

(0.0

06)

(0.0

05)

(0.0

05)

(0.0

05)

(0.0

06)

(0.0

12)

(0.0

08)

(0.0

09)

(0.0

07)

(0.0

09)

Con

stan

t0.

158**

*0.

261**

*0.

007

––

––

––

(0.0

44)

(0.0

61(0

.072

)O

bser

vati

ons

1,89

11,

630

1,89

31,

630

1,89

31,

643

1,89

11,

630

1,89

31,

630

1,89

31,

643

R-s

quar

ed0.

116

0.18

500.

002

––

––

––

Con

tinu

ed.


Tabl

e5.

Con

tinu

ed.

Lin

ear

Reg

ress

ion

Qua

ntile

Reg

ress

ion

HH

ISI

The

ilH

HI

SIT

heil

FE

IVF

EIV

FE

IV1s

t5t

h1s

t5t

h1s

t5t

h

And

erso

nca

noni

calc

orre

lati

on48

.784

55.3

9352

.389

––

––

––

LM

stat

isti

c(p

-val

ue)

(0.0

00)

(0.0

00)

(0.0

00)

Cra

gg–D

onal

dW

ald

F-s

tati

stic

9.97

911

.382

10.7

42–

––

––

–(1

0%m

axim

alIV

rela

tive

bias

)(8

.78)

(8.7

8)(8

.78)

Sar

gan

stat

isti

c(p

-val

ue)

6.08

92.

764

3.18

0–

––

––

–(0

.107

4)(0

.429

5)(0

.364

7)

AC

FTA

=A

SE

AN

–Chi

naFr

eeT

rade

Are

a,A

HS

=ef

fect

ivel

yap

plie

dta

riff

,AJF

TA=

AS

EA

N–J

apan

Free

Tra

deA

rea,

AK

FTA

=A

SE

AN

–Kor

eaFr

eeT

rade

Are

a,A

SE

AN

=A

ssoc

iati

onof

Sou

thea

stA

sian

Nat

ions

,F

E=

fixe

def

fect

,H

HI=

Her

find

ahl–

Hir

schm

anin

dex,

IV=

inst

rum

enta

lva

riab

le,

LM

=L

agra

nge

mul

tipl

ier,

MF

N=

mos

tfa

vore

dna

tion

,SI=

sim

ilar

ity

inde

x.N

otes

:We

repo

rton

lyre

sult

sfo

rre

gres

sion

sus

ing

the

The

ilin

dex

tosa

vesp

ace.

Figu

res

inpa

rent

hese

sar

est

anda

rder

rors

.***,**

,and

*in

dica

test

atis

tica

lsig

nifi

canc

eat

the

1%,

5%,a

nd10

%le

vels

,res

pect

ivel

y.S

ourc

e:A

utho

rs’

esti

mat

es.


This impact seems to be slightly sensitive to the endogeneity treatment. Inparticular, when the reporter’s income volatility is treated as endogenous, the levelof its significance drops from 1% to 5% and 10%. In general, a percentage point risein the reporter country’s income volatility is likely to increase its export volatilityfrom 0.355% to 0.509%. By contrast, the reporter’s export volatility does not dependon the income volatility of its trading partner. This means that the intraregionalexport volatility of an ASEAN+3 country seems to be insensitive to a domesticeconomic shock of its partners in the same region. This may be good news forpolicy makers in low-income countries that are relatively closed in ASEAN+3, suchas the Lao PDR, Myanmar, and the Philippines, since opening up their economiesto regional trade will not necessarily make their exports more volatile nor weakentheir domestic economy due to external turmoil.

Examining now the estimates of three export diversification indicators, weobtain two different findings. First, the estimated coefficients of the HHI andthe Theil index in all regressions are positive and significant. For instance, afterresolving the endogeneity problem of the dependent variable, a percentage pointrise in export concentration (the HHI) would, all things being equal, increaseexport volatility by approximatively 0.358%. Accordingly, an ASEAN+3 countrywith a higher level of export concentration could experience a higher level ofbilateral export volatility. In other words, to reduce intraregional export volatility,ASEAN+3 countries should diversify their exports to other ASEAN+3 memberstates. This finding is of major relevance for low-income and less open ASEAN+3countries, notably the Lao PDR, Myanmar, and the Philippines, which do nothave very diversified export baskets. Therefore, we recommend that policy makersof these countries increase export diversification. Higher export diversificationcorresponds with lower export volatility, which in turn fosters stability of economicgrowth. This argument is also supported by Haddad et al. (2013) who arguethat irrespective of whether the effect of trade openness on output volatility ispositive or negative on average, openness lowers output volatility in sufficientlydiversified economies, while it increases volatility in those with more concentratedexport baskets. Second, we find no evidence of a link between the similarity indexand bilateral export volatility. We can thus conclude that the nature of exportdiversification’s impact on ASEAN+3 bilateral export volatility strongly dependson the way export diversification is measured.

We now turn our attention to the potential effect of bilateral tariffs onASEAN+3’s bilateral export volatility. On the one hand, estimates of both tariffindicators, AHS and MFN, are statistically significant only in the model in whichthe export volatility term is treated as endogenous. On the other hand, these twotariff indicators experience conflicting effects on export volatility. Specifically, theimpact of an AHS tariff on bilateral export volatility is positive, while that of anMFN tariff is negative: a percentage rise in the AHS tariff would create a 0.004%increase in export volatility, while the same increase in the MFN tariff would


reduce export volatility by around 0.003%. According to these results, grantingthe most favored nation status reduces the volatility of bilateral exports. Inversely,the bilateral intraregional exports of ASEAN+3 member states can become morevolatile if their trading partners apply protective tariffs. However, these two impactsare quite small.

Looking now at the impact of FTAs, we first note that estimates of the FTAvariable in all regressions are negative as expected but statistically insignificant.This suggests that sharing a common FTA does not allow two ASEAN+3 countrypartners to decrease their export volatility. Second, on the three FTAs signedbetween ASEAN and one of the Plus Three countries, we find that only theASEAN–China Free Trade Area (ACFTA) plays a significant role in reducingthe level of bilateral export volatility. This can be explained by the fact that thePRC’s important position in world trade can help its ASEAN trading partnersto maintain a low volatility of bilateral exports between them and the PRC. Inother words, joining a regional trade block with the PRC reduces the variability oftrade flows among ASEAN member states and the PRC. This finding is consistentwith the hypothesis that “joint membership in a reciprocal trade agreement—apreferential trade agreement or the General Agreement on Tariffs and Trade/WorldTrade Organization—should decrease the volatility of a country’s exports to atrade partner,” according to Mansfield and Reinhardt (2008) in their study of 162countries over the period 1951–2001. The authors also argue that institutions couldprecipitate fluctuations in trade due to commitments and protectionist barriers andreduce the volatility of cross-border transactions.

Similar to equation (1), we also consider political rights as a control variablein equation (8). After treating the export volatility terms, estimates of a reportercountry’s political rights become negative and significant. This suggests thatintra-ASEAN+3 exports of an economy with a higher level of political rightsappear to be more volatile. However, this negative impact of political rights isquite small. The second control variable in equation (8), bilateral exchange rate’svolatility, also has a significantly negative impact on bilateral export volatility.However, this unexpected effect should not be considered since it becomesstatistically insignificant due to the export volatility’s treatment in the IV estimation.Indeed, the exchange rate variable plays a key role in international trade butthe nature of its impact on trade has been a controversial issue in the literature.Therefore, we leave this issue for further research.

All empirical results mentioned above show how the main determinantsaffect export volatility on average. While these results allow us to address thequestion about the role of each main determinant in explaining export volatility,they do not allow us to answer another important question—does each determinantinfluence export volatility differently for exports with low volatility than forexports with average volatility? A more comprehensive picture of the effect of thepredictors on export volatility can be obtained by using quantile regression analysis,


which allows us to model the relationship between the potential determinants andspecific quantiles of export volatility. In other words, quantile regression specifiesthe change in the quantiles of export volatility’s reaction to the change in itsdeterminants. Therefore, to complete our empirical results reported in Table 5, wereestimate equation (8) by applying the regression quantiles approach developed byMachado and Santos Silva (2019). The authors consider two settings: panel datamodels with individual effects and models with endogenous explanatory variables.Coefficient estimates for the 1st and the 5th quintiles are presented in Table 5.

As reported in Table 5, the empirical results provided by the quantileregressions largely support those obtained in the linear regression. However, weobserve that the effect of the main determinants on export volatility slightly changesfrom the 5th quintile to the 1st quintile. First, the effect of an increase in thereporter country’s income on its export volatility is likely larger on exports withlow volatility and smaller on exports with high volatility, while we find the oppositeresult for export diversification’s impact on export volatility. Second, the positiveimpact of an AHS tariff and the negative impact of an MFN tariff on export volatilityare not maintained in the 1st quintile, the group with the lowest export volatility. Bycontrast, the negative effect of bilateral exchange rate volatility is not significantin the 5th quintile, the group with the highest export volatility. Third, we also findthat the role of the ACFTA in reducing export volatility is more important in the5th quintile. In sum, the effect of each determinant is not the same for all levels ofexport volatility. This heterogeneity should be considered as an important factor inpublic policy.

C. Bilateral Export Volatility in the Association of Southeast Asian Nationsand Plus Three Countries: A Comparison

In this subsection, we reestimate equation (8) by applying the FE and IVestimators again for two separate data subsamples—ASEAN countries and PlusThree countries. We aim to address the question of whether differences in sizeand levels of economic development between ASEAN countries and Plus Threecountries can influence the relationship between bilateral export volatility and itspotential determinants. We report the empirical results in Table 6.

First, looking at the impact of income volatility on bilateral exportvolatility, the empirical results are significantly altered. Specifically, in the ASEANsubsample, we fail to maintain the significant impact of the reporter country’sincome volatility on its bilateral export volatility after treating income volatilityas an endogenous variable. By contrast, in the Plus Three subsample, estimates ofthe reporter country’s income volatility are quite high and statistically significant inboth FE and IV estimations, suggesting a considerable effect of income volatility onbilateral export volatility. For instance, on the estimated coefficients of the reporter’sincome volatility in the HHI, SI, and the Theil index models after treating the


ble

6.E

xpor

tV

olat

ility

Est

imat

ion:

ASE

AN

vers

usP

lus

Thr

ee

ASE

AN

Pan

elP

lus

Thr

eeP

anel

HH

ISI

The

ilH

HI

SIT

heil

FE

IVF

EIV

FE

IVF

EIV

FE

IVF

EIV

Inco

me

vola

tili

ty(r

epor

ter)

0.55

7***−0

.032

0.56

6***

0.07

20.

577**

*−0

.034

0.40

0***

1.34

3***

0.45

5***

1.62

8***

0.41

9***

1.41

5***

(0.0

66)

(0.2

02)

(0.0

64)

(0.1

96)

(0.0

65)

(0.1

98)

(0.1

00(0

.412

)(0

.107

)(0

.414

)(0

.099

)(0

.407

)In

com

evo

lati

lity

(par

tner

)0.

051

−0.2

670.

056

−0.1

650.

037

−0.2

630.

126

0.36

90.

162**

0.33

50.

129

0.36

3(0

.083

)(0

.257

)(0

.077

)(0

.249

)(0

.084

)(0

.261

)(0

.077

(0.2

33)

(0.0

78)

(0.2

45)

(0.0

78)

(0.2

38)

Exc

hang

era

tevo

lati

lity

−0.2

04*

0.36

6−0

.271

***

0.23

1−0

.210

**0.

364

−0.0

85−0

.351

*−0

.174

−0.4

25*

−0.1

02−0

.361

*

(0.1

04)

(0.2

30)

(0.1

01)

(0.2

32)

(0.1

01)

(0.2

29)

(0.1

20(0

.210

)(0

.129

)(0

.232

)(0

.121

)(0

.214

)E

xpor

tdiv

ersi

fica

tion

inde

x0.

393**

*0.

213**

*0.

002

0.00

00.

081**

*0.

035**

*0.

488**

*0.

463**

*−0

.002

*−0

.004

***

0.09

3***

0.07

5***

(0.1

19)

(0.0

56)

(0.0

02)

(0.0

01)

(0.0

23)

(0.0

12)

(0.1

31(0

.087

)(0

.002

)(0

.001

)(0

.029

)(0

.014

)A

HS

tari

ff0.

001

0.00

3**0.

003

0.00

4***

0.00

10.

003**

0.00

50.

000

0.00

4−0

.002

0.00

40.

000

(0.0

02)

(0.0

02)

(0.0

02)

(0.0

01)

(0.0

02)

(0.0

02)

(0.0

07(0

.005

)(0

.006

)(0

.006

)(0

.007

)(0

.005

)M

FN

tari

ff0.

000

−0.0

04**

*0.

001

−0.0

04**

*0.

000

−0.0

04**

*−0

.003

0.00

4−0

.003

0.00

4−0

.002

0.00

4(0

.002

)(0

.001

)(0

.001

)(0

.001

)(0

.002

)(0

.001

)(0

.007

(0.0

05)

(0.0

07)

(0.0

06)

(0.0

07)

(0.0

06)

Free

trad

eag

reem

ent

−0.0

36−0

.006

−0.0

42−0

.008

−0.0

33−0

.007

0.01

6−0

.007

0.00

5−0

.019

*0.

021

−0.0

05(0

.034

)(0

.006

)(0

.035

)(0

.005

)(0

.034

)(0

.006

)(0

.017

(0.0

10)

(0.0

17)

(0.0

10)

(0.0

18)

(0.0

10)

AC

FTA

−0.0

75**

0.00

1**−0

.083

**0.

000*

−0.0

74**

0.00

1**–

––

––

–(0

.034

)(0

.000

)(0

.036

)(0

.000

)(0

.034

)(0

.000

)A

JFTA

−0.0

14−0

.031

*−0

.018

−0.0

32*

−0.0

08−0

.028

––

––

––

(0.0

33)

(0.0

19)

(0.0

33)

(0.0

18)

(0.0

32)

(0.0

19)

AK

FTA

0.02

4−0

.021

0.02

5−0

.037

0.02

7−0

.021

––

––

––

(0.0

33)

(0.0

26)

(0.0

33)

(0.0

25)

(0.0

33)

(0.0

26)

Poli

tica

lrig

hts

(rep

orte

r)0.

001

0.00

6−0

.002

0.00

2−0

.002

0.00

7−0

.022

0.00

2−0

.024

−0.0

02−0

.022

0.00

0(0

.008

)(0

.023

)(0

.008

)(0

.021

)(0

.008

)(0

.023

)(0

.015

(0.0

05)

(0.0

16)

(0.0

06)

(0.0

16)

(0.0

05)

Poli

tica

lrig

hts

(par

tner

)−0

.008

−0.0

29−0

.005

−0.0

24−0

.008

−0.0

290.

000

0.00

9−0

.004

−0.0

110.

000

0.00

7(0

.009

)(0

.026

)(0

.010

)(0

.025

)(0

.009

)(0

.026

)(0

.004

(0.0

15)

(0.0

05)

(0.0

17)

(0.0

04)

(0.0

15)

Con

stan

t0.

212**

*0.

283**

*0.

053

0.15

3***

–0.

406**

*–

0.00

9–

(0.0

50)

–0.

066

–(0

.085

)–

(0.0

47)

(0.0

78)

(0.0

69)

Obs

erva

tion

s1,

320

1,11

61,

322

1,11

61,

322

1,10

352

752

757

1–

571

–R

-squ

ared

0.08

9–

0.00

72–

0.12

4–

0.12

90–

0.09

36–

0.12

61–

Con

tinu

ed.


Tabl

e6.

Con

tinu

ed.

ASE

AN

Pan

elP

lus

Thr

eeP

anel

HH

ISI

The

ilH

HI

SIT

heil

FE

IVF

EIV

FE

IVF

EIV

FE

IVF

EIV

And

erso

nca

noni

calc

orre

lati

on65

.541

63.9

3761

.200

19.9

8322

.682

20.5

52L

Mst

atis

tic

(p-v

alue

)–

(0.0

00)

–(0

.000

)–

(0.0

00)

–(0

.000

)–

(0.0

00)

–(0

.000

)C

ragg

–Don

ald

Wal

dF

-sta

tist

ic13

.788

–13

.429

–12

.822

–6.

779a

–7.

736a

–6.

980a

(10%

max

imal

IVre

lativ

ebi

as)

––

––

––

Sar

gan

stat

isti

c(p

-val

ue)

5.58

75.

370

7.11

90.

308

0.06

50.

409

–(0

.133

)–

(0.1

46)

–(0

.106

)–

(0.5

79)

–(0

.798

)–

(0.5

22)

AC

FTA

=A

SE

AN

–Chi

naFr

eeT

rade

Are

a,A

HS

=ef

fect

ivel

yap

plie

dta

riff

,AJF

TA=

AS

EA

N–J

apan

Free

Tra

deA

rea,

AK

FTA

=A

SE

AN

–Kor

eaFr

eeT

rade

Are

a,A

SE

AN

=A

ssoc

iati

onof

Sou

thea

stA

sian

Nat

ions

,F

E=

fixe

def

fect

,H

HI=

Her

find

ahl–

Hir

schm

anIn

dex,

IV=

inst

rum

enta

lva

riab

le,

LM

=L

agra

nge

mul

tipl

ier,

MF

N=

mos

tfa

vore

dna

tion

,SI=

sim

ilar

ity

inde

x.N

otes

:Fig

ures

inpa

rent

hese

sar

est

anda

rder

rors

.***,**

,and

*in

dica

test

atis

tica

lsig

nifi

canc

eat

the

1%,5

%,a

nd10

%le

vels

,res

pect

ivel

y.a 20

%m

axim

alIV

rela

tive

size

.S

ourc

e:A

utho

rs’

esti

mat

es.


endogeneity problem, the results reveal that a percentage point increase in incomevolatility would, all things being equal, raise export volatility by approximately1.343%, 1.628%, and 1.415%, respectively. We therefore conclude that the roleof income volatility in explaining the volatility of bilateral exports betweenASEAN+3 countries depends on the size and level of economic development ofeach member country. Precisely, in strong economies, particularly the Plus Threecountries, the positive impacts of income volatility on export volatility are muchmore pronounced.

Second, our empirical results on the effects of export diversification arenot affected very much after splitting the data. It is worth noting that the effectsof export diversification on export volatility appear to be qualitatively similarto those obtained for the ASEAN+3 as a whole. Exceptionally, the impact ofthe similarity index becomes negative and significant but only in the Plus Threesample. In addition, the estimates of the HHI and Theil index in the Plus Threesample are substantially higher than those in the ASEAN sample. For instance,the IV estimation results show that a percentage point rise in the degree of exportconcentration (the HHI), would increase the export volatility of the Plus Threecountries by approximately 0.463%, while export volatility in ASEAN countrieswould go up by only 0.213%. This means that the Plus Three countries seemto be more cautious than the ASEAN members in terms of considering exportdiversification as a tool for controlling export volatility.

Third, the empirical results on the impact of tariffs on ASEAN+3 as a wholecan only be found in the results for the ASEAN panel. This means that in smallerand less powerful economies, export volatility depends on the variation of tariffsapplied by their trading partners. However, this impact is quite small. A percentagepoint increase in the AHS tariff would increase ASEAN export volatility by around0.003%, while a percentage point increase in the MFN tariff would reduce ASEANexport volatility by approximately 0.004%. Thus, negotiating preferential tariffswith trading partners should be considered an essential tool for controlling exportvolatility. On the other hand, the empirical results reported in Table 6 also confirmthat becoming a member of ACFTA allows an ASEAN country to slightly lowerits bilateral export volatility. By contrast, due to their meaningful position in worldtrade, the bilateral export volatility of the Plus Three countries is not significantlyaffected by tariffs applied by ASEAN countries.

Similar to the previous section, we once again apply the quantile regression’sestimator to investigate the possible heterogeneity of each determinant’s impact onexport volatility in ASEAN and Plus Three countries, as reported in Table 7. Asexpected, the link between each determinant and export volatility changes whenexport volatility moves from the highest to the lowest levels. First, in ASEAN,the change in export volatility from the 5th quintile to the 1st quintile results ina slight increase in the positive impact of the reporter country’s income volatilityby approximatively 0.03%. However, the reverse result is found for Plus Three


ble

7.E

xpor

tV

olat

ility

Qua

ntile

Reg

ress

ion:

ASE

AN

vers

usP

lus

Thr

ee

ASE

AN

Pan

elP

lus

Thr

eeP

anel

HH

ISI

The

ilH

HI

SIT

heil

1st

5th

1st

5th

1st

5th

1st

5th

1st

5th

1st

5th

Inco

me

vola

tili

ty(r

epor

ter)

0.57

1***

0.54

3***

0.58

1***

0.55

1***

0.59

0***

0.56

2***

0.39

4***

0.40

6**0.

446*

0.46

4*0.

396**

*0.

445*

(0.0

00)

(0.1

26)

(0.0

77)

(0.1

11)

(0.0

72)

(0.1

16)

(0.0

02)

(0.1

86)

(0.2

83)

(0.2

72)

(0.1

06)

(0.2

54)

Inco

me

vola

tili

ty(p

artn

er)

0.08

60.

013

0.07

80.

033

0.08

0−0

.009

0.11

40.

141

0.14

00.

187

0.10

50.

155

(0.3

73)

(0.1

66)

(0.1

03)

(0.1

49)

(0.0

94)

(0.1

51)

(0.1

98)

(0.1

31)

(1.0

39)

(0.2

60)

(0.0

76)

(0.2

54)

Exc

hang

era

tevo

lati

lity

−0.1

89*

−0.2

22−0

.270

**−0

.271

−0.2

01*

−0.2

22−0

.104

−0.0

65−0

.190

−0.1

57−0

.114

−0.0

89(0

.100

)(0

.214

)(0

.134

)(0

.195

)(0

.120

)(0

.193

)(0

.389

)(0

.179

)(1

.376

)(0

.362

)(0

.103

)(0

.346

)E

xpor

tdiv

ersi

fica

tion

inde

x0.

211*

0.59

1**0.

003

0.00

10.

051*

0.11

5***

0.17

60.

839**

−0.0

03−0

.003

*0.

039

0.25

5**

(0.1

12)

(0.2

89)

(0.0

02)

(0.0

02)

(0.0

26)

(0.0

43)

(0.4

54)

(0.3

47)

(0.1

84)

(0.0

02)

(0.0

31)

(0.1

04)

AH

Sta

riff

−0.0

010.

004

0.00

00.

006

−0.0

020.

003

0.00

30.

007

0.00

30.

005

0.00

40.

005

(0.6

61)

(0.0

05)

(0.0

03)

(0.0

04)

(0.0

03)

(0.0

04)

(0.5

82)

(0.0

08)

(0.7

64)

(0.0

18)

(0.0

05)

(0.0

16)

MF

Nta

riff

0.00

2−0

.002

0.00

2−0

.004

0.00

2−0

.003

−0.0

03−0

.003

−0.0

03−0

.002

−0.0

03−0

.001

(0.3

67)

(0.0

04)

(0.0

02)

(0.0

03)

(0.0

02)

(0.0

03)

(0.6

41)

(0.0

09)

(0.7

86)

(0.0

18)

(0.0

05)

(0.0

17)

Free

trad

eag

reem

ent

0.00

2−0

.078

−0.0

08−0

.078

0.00

2−0

.071

0.00

70.

027

0.00

00.

011

0.00

90.

033

(0.9

59)

(0.0

67)

(0.0

41)

(0.0

59)

(0.0

37)

(0.0

60)

(0.7

40)

(0.0

29)

(2.4

47)

(0.0

58)

(0.0

17)

(0.0

56)

AC

FTA

−0.0

48−0

.105

−0.0

55−0

.112

*−0

.048

−0.1

03*

––

––

––

(0.2

60)

(0.0

73)

(0.0

45)

(0.0

65)

(0.0

41)

(0.0

66)

AJF

TA−0

.018

−0.0

10−0

.023

−0.0

12−0

.016

0.00

0–

––

––

–(0

.554

)(0

.052

)(0

.030

)(0

.044

)(0

.029

)(0

.047

)A

KF

TA0.

026

0.02

20.

032

0.01

80.

029

0.02

3–

––

––

–(0

.433

)(0

.057

)(0

.034

)(0

.049

)(0

.032

)(0

.051

)Po

liti

calr

ight

s(r

epor

ter)

−0.0

040.

005

−0.0

050.

001

−0.0

060.

002

−0.0

27−0

.016

−0.0

34−0

.013

−0.0

28−0

.015

(0.5

58)

(0.0

11)

(0.0

06)

(0.0

09)

(0.0

06)

(0.0

10)

(0.3

65)

(0.0

44)

(3.5

50)

(0.0

84)

(0.0

25)

(0.0

84)

Poli

tica

lrig

hts

(par

tner

)−0

.012

−0.0

03−0

.011

0.00

0−0

.012

−0.0

030.

002

−0.0

020.

001

−0.0

090.

003

−0.0

04(0

.209

(0.0

17)

(0.0

11)

(0.0

15)

(0.0

09)

(0.0

15)

(0.7

55)

(0.0

10)

(0.8

89)

(0.0

21)

(0.0

06)

(0.0

20)

Obs

erva

tion

s1,

320

1,11

61,

322

1,11

61,

322

1,10

352

752

757

157

157

157

1

AC

FTA

=A

SE

AN

–Chi

naFr

eeT

rade

Are

a,A

HS

=ef

fect

ivel

yap

plie

dta

riff

,AJF

TA=

AS

EA

N–J

apan

Free

Tra

deA

rea,

AK

FTA

=A

SE

AN

–Kor

eaFr

eeT

rade

Are

a,A

SE

AN

=A

ssoc

iati

onof

Sou

thea

stA

sian

Nat

ions

,HH

I=

Her

find

ahl–

Hir

schm

anIn

dex,

MF

N=

mos

tfav

ored

nati

on,S

I=

sim

ilar

ity

inde

x.N

otes

:Fig

ures

inpa

rent

hese

sar

est

anda

rder

rors

.***,**

,and

*in

dica

test

atis

tica

lsig

nifi

canc

eat

the

1%,5

%,a

nd10

%le

vels

,res

pect

ivel

y.S

ourc

e:A

utho

rs’

esti

mat

es.


countries. Second, the negative impact of exchange rate volatility on ASEAN exportvolatility is only significant in the lowest quintile (1st), while this impact is notsignificant for Plus Three countries. Third, the empirical results at the quintilelevel support the significant and strong effects of export diversification on exportvolatility, particularly in the highest quintile (5th), in both ASEAN and Plus Threecountries. Lastly, the very small impact of AHS and MFN tariffs on ASEAN’sexport volatility obtained by using linear regression (Table 6) is not found in thequantile regression. This means that ASEAN’s exports, on average, weakly dependon AHS and MFN tariffs. However, this fragile dependence could be lost at thequintile level.

All the empirical results listed above also allow us to provide rankingproperties among the determinants of interest. We reveal that income volatility ofthe reporter country plays the most important role in explaining ASEAN+3 exportvolatility (for both the full sample and the two separate samples). Following incomevolatility, export diversification and exchange rate volatility play the next importantroles in explaining export volatility in ASEAN+3. By contrast, changes in tradingtariffs, unexpectedly, have a very weak impact on export volatility.

VI. Conclusion

The objective of this paper is twofold. First, we investigate the potential roleof trade variables, notably trade volatility and export diversification, on incomevolatility of ASEAN+3 member states. Second, we provide an empirical analysis onthe potential determinants of ASEAN+3’s bilateral export volatility. To this end, weapply a set of panel and cross-sectional econometric techniques including the GMMand the FE and IV estimators. We carry out our empirical tests for a panel datasetcovering aggregate income and trade volatility of each ASEAN+3 member stateand a cross-sectional data sample of intra-ASEAN+3 bilateral export volatility.A set of important findings on the relationship between income volatility, tradevolatility, and trade diversification can be drawn from our paper.

First, we reveal that the volatility of trade has a positive effect on outputvolatility in ASEAN+3. This result confirms trade’s key role in explainingeconomic growth. Second, we find no evidence that export diversification reducesoutput volatility in ASEAN+3. Similarly, becoming a member of an FTAdoes not allow ASEAN+3 member states to better control economic growthvolatility. Exceptionally, the ACFTA seems to be creating stability in terms ofintra-ASEAN+3 exports. Third, income volatility, in turn, positively influencesintra-ASEAN+3 bilateral export volatility. Fourth, the nature of the link betweenexport diversification and export volatility strongly depends on the measurementof export diversification. Fifth, the type of tariffs applied by a trading partner isalso a considerable factor in determining bilateral export volatility in ASEAN+3,


particularly in ASEAN countries. Lastly, by separating data for ASEAN and PlusThree countries, we find that the nature of the relationship between export volatility,income volatility, and trade diversification could depend on country size and thelevel of economic development.

Our findings also provide a set of important policy implications. First, thetrade–economic growth volatility nexus supports the compensation hypothesis aseconomic growth becomes susceptible to external shocks from trading partners. Inthe case of ASEAN+3, particularly in ASEAN countries in which trade opennessis still considered an economic growth promoter, policy makers should implementefficient tools for controlling trade volatility that, in turn, could shield an economyagainst the detrimental impact of idiosyncratic global shocks on volatility. Anotherimportant finding of this paper concerns the role of export diversification inreducing bilateral export volatility in ASEAN+3. This finding supports the factthat trade openness can reduce economic growth volatility in ASEAN+3 whencountries are well diversified. The relation between export concentration and tradeopenness is also illustrated in Figure A1.2 in Appendix 1. Lastly, we cannot confirmthe theoretical impact of an FTA on export volatility in our empirical analysissince the FTA estimates we find are negative but statistically insignificant. Thisfinding suggests that ASEAN+3 policy makers have to review the implementationof existing intraregional FTAs in order to better benefit from these agreements interms of lessening bilateral export volatility.

References

Abbott, Frederick M. 2000. “NAFTA and the Legalization of World Politics: A Case Study.”International Organization 54 (3): 5–19.

Acemoglu, Daron, Simon Johnson, James Robinson, and Thaicharoen Yunyong. 2003.“Institutional Causes, Macroeconomic Symptoms: Volatility, Crises, and Growth.” Journalof Monetary Economics 50 (1): 49–123.

Arellano, Manuel, and Stephen Bond. 1991. “Some Tests of Specification for Panel Data: MonteCarlo Evidence and an Application to Employment Equations.” The Review of EconomicStudies 58 (2): 277–97.

Arellano, Manuel, and Olympia Bover. 1995. “Another Look at the Instrumental VariableEstimation of Error-Components Models.” Journal of Econometrics 68 (1): 29–51.

Bagwell, Kyle, and Robert Staiger. 2002. The Economics of the World Trading System.Cambridge, MA: MIT Press.

Balavac, Merima, and Geoffrey T. Pugh. 2016. “The Link between Trade Openness, ExportDiversification, Institutions and Output Volatility in Transition Countries.” EconomicSystems 40 (2): 273–87.

Baltagi, Badi H. 2008. Econometric Analysis of Panel Data (4th edition). Chichester, UK;Hoboken, NJ: John Wiley & Sons.

Bejan, Maria. 2006. “Trade Openness and Output Volatility.” Munich Personal RePEc ArchivePaper No. 2759, University Library of Munich, Germany.


Bennett, Federico, Daniel Lederman, Samuel J. Pienknagura, and Roja Diego. 2016. “TheVolatility of International Trade Flows in the 21st Century: Whose Fault Is It Anyway?”World Bank Policy Research Working Paper No. 7781.

Bhagwati, Jagdish N. 1988. “Poverty and Public Policy.” World Development 16 (5): 539–55.Blundell, Richard, and Stephen Bond. 1998. “Initial Conditions and Moment Restrictions in

Dynamic Panel Data Models.” Journal of Econometrics 87 (1): 115–43.Cadot, Olivier, Céline Carrère, and Vanessa Strauss-Kahn. 2013. “Trade Diversification,

Income, and Growth: What Do We Know?” Journal of Economic Surveys 27 (4): 790–812.

Calderón, César, Norman Loayza, and Klaus Schmidt-Hebbel. 2005. “Does Openness ImplyGreater Exposure?” World Bank Policy Research Working Paper No. 3733.

Calderón, César, and Klaus Schmidt-Hebbel. 2008. “Openness and Growth Volatility.” CentralBank of Chile Working Paper No. 483.

Caselli, Francesco, Miklós Koren, Milan Lisicky, and Silvana Tenreyro. 2015. “Diversificationthrough Trade.” NBER Working Paper No. 21498.

Cavalcanti, Tiago V. de V., Kamiar Mohaddes, and Mehdi Raissi. 2012. “Commodity PriceVolatility and the Sources of Growth.” IMF Working Paper WP/12/12.

Cavallo, Eduardo A. 2008. “Output Volatility and Openness to Trade: A Reassessment.”Economia 9 (1): 105–38.

Chinn, Menzie D., and Ito Hiro. 2008. “A New Measure of Financial Openness.” Journal ofComparative Policy Analysis: Research and Practice 10 (3): 309–22.

Di Giovanni, Julian, and Andrei Levchenko. 2009. “Trade Openness and Volatility.” The Reviewof Economics and Statistics 91 (3): 558–85.

Easterly, William. 2007. “Inequality Does Cause Underdevelopment: Insights from a NewInstrument.” Journal of Development Economics 84 (2): 755–76.

Easterly, William, Roumeen Islam, and Joseph E. Stiglitz. 2001. “Shaken and Stirred: ExplainingGrowth Volatility.” In Annual World Bank Conference on Development Economics, editedby Boris Pleskovic and Nicholas Stern, 191–211. Washington, DC: World Bank.

Finger, Joseph, and Mordechai Kreinin. 1979. “A Measure of ‘Export Similarity’ and Its PossibleUses.” Economic Journal 89 (356): 905–12.

Frankel, Jeffrey A., and David Romer. 1999. “Does Trade Cause Growth?” American EconomicReview 89 (3): 379–99.

Haddad, Mona, Jamus J. Lim, Cosimo Pancaro, and Christian Saborowski. 2013. “TradeOpenness Reduces Growth Volatility When Countries Are Well Diversified.” CanadianJournal of Economics 46 (2): 765–90.

Hakura, Dalia S. 2009. “Output Volatility in Emerging Market and Developing Countries: WhatExplains the Great Moderation of 1970–2003?” Czech Journal of Economics and Finance(Finance a úver) 59 (3): 229–54.

Head, Keith. 2003. “Gravity for Beginners.” Mimeo, University of British Columbia. https://vi.unctad.org/tda/background/Introduction%20to%20Gravity%20Models/gravity.pdf.

Helpman, Elhanan, and Paul R. Krugman. 1985. Market Structure and Foreign Trade: IncreasingReturns, Imperfect Competition, and the International Economy. Cambridge, MA: MITPress.

Jansen, Marion, Marc Bacchetta, Carolina Lennon, and Roberta Piermartini. 2009. “Exposureto External Country-Specific Shocks and Income Volatility in Small and DevelopingCountries: Export Concentration Matters.” World Trade Organization Discussion PaperNo. 3.


Judson, Ruth A., and Ann L. Owen. 1999. “Estimating Dynamic Panel Data Models: A Guide forMacroeconomists.” Economics Letters 65 (1): 9–15.

Kalemli-Ozcan, Sebnem. 2002. “Does the Mortality Decline Promote Economic Growth?”Journal of Economic Growth 7 (4): 411–39.

Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi. 2007. “Growth and Governance: AReply.” The Journal of Politics 69 (2): 555–62.

Keohane, Robert O. 1983. “The Demand for International Regimes.” In International Regimes,edited by Stephen D. Krasner, 147–71. Ithaca, New York: Cornell University Press.

Klomp, Jeroen, and Jakob de Haan. 2009. “Political Institutions and Economic Volatility.”European Journal of Political Economy 25 (3): 311–26.

Machado, Jose A.F., and João Santos Silva. 2019. “Quantiles via Moments.” Journal ofEconometrics 213 (1): 145–73.

Mansfield, Edward D., and Eric Reinhardt. 2008. “International Institutions and the Volatility ofInternational Trade.” International Organization 62 (4): 621–52.

Martin, Lisa L., and Beth A. Simmons. 1998. “Theories and Empirical Studies of InternationalInstitutions.” International Organization 52 (4): 729–57.

Melitz, Marc. J. 2003. “The Impact of Trade on Intra-industry Reallocations and AggregateIndustry Productivity.” Econometrica 71 (6): 1695–725.

Mobarak, Ahmed M. 2005. “Democracy, Volatility, and Economic Development.” The Review ofEconomics and Statistics 87 (2): 348–61.

Nerlove, Marc. 1967. “Experimental Evidence on the Estimation of Dynamic Economic Relationsfrom a Time Series of Cross-Sections.” Economic Studies Quarterly 18 (3): 42–74.

Nickell, Stephen. 1981. “Biases in Dynamic Models with Fixed Effects.” Econometrica 49 (6):1417–26.

Rodrik, Dani. 1997. Has Globalization Gone Too Far? Washington, DC: Institute for InternationalEconomics.

______. 1999. “Where Did All the Growth Go? External Shocks, Social Conflict, and GrowthCollapse.” Journal of Economic Growth 4 (4): 385–412.

Wei, Shang-Jin. 2000. “Natural Openness and Good Government.” NBER Working Paper No.7765.


Appendix 1

Figure A1.1. Import Concentration, 1990–2016

ASEAN = Association of Southeast Asian Nations, HHI = Herfindahl–Hirschman Index.Note: Plus Three refers to Japan, the People’s Republic of China, and the Republic of Korea.Source: Authors’ illustration from calculated indexes.

Figure A1.2. Export Concentration and Trade Openness

HHI = Herfindahl–Hirschman Index.Source: Authors’ illustration from calculated indexes.


Appendix 2

Estimation of Trade Openness Model (equation 7)

Dependent Variable: Trade Openness ([Exports + Imports]/GDP)

Regressors Fixed Effects

Log (Population) −0.129***

(0.015)Landlocked −0.451***

(0.086)Area −0.000***

(0.000)Language: English 0.156**

(0.065)Language: Chinese 0.870***

(0.066)Remoteness −0.000***

(0.000)Constant 1.649***

(0.255)Observations 351R-squared 0.76Number of years 27

GDP = gross domestic product.Notes: Robust standard errors in parentheses. *** = p < 0.01, ** = p < 0.05,* = p < 0.10.Source: Authors’ estimates.

Could Weather Fluctuations Affect LocalEconomic Growth? Evidence from Counties

in the People’s Republic of ChinaChengzheng Li, Jiajia Cong, and Haiying Gu∗

This paper uses historical fluctuations of weather variables within counties inthe People’s Republic of China to identify their effects on economic growthfrom 1996 to 2012. We find three primary results. First, higher temperaturessignificantly reduce the growth rate of county-level gross domestic productper capita: an increase in the annual average temperature of 1°C lowers thegrowth rate by 1.05%–1.25%. The effect of higher temperatures is nonlinear.Second, fluctuations in temperature and precipitation not only have a leveleffect, they also have a substantial cumulative effect. Third, weather fluctuationshave wide-ranging effects. Beyond their substantial effects on the growth rateof agricultural output, they also affect nonagriculture sectors, labor productivity,and investment. Our findings provide new evidence for the impact of weatherchanges on economic development and have major implications for adaptationpolicies.

Keywords: climate change, economic growth, precipitation, temperature,weather shocksJEL codes: O13, O44, Q54, Q56

I. Introduction

It is a controversial question whether climate conditions are central toeconomic development. Cross-country analysis shows that hot countries tend tobe poor. The average growth rate of tropical countries was 0.9% lower than that ofnontropical countries from 1965 to 1990 (Gallup, Sachs, and Mellinger 1999). Theincome per capita of Africa in 1992 was equivalent to the income level of WesternEurope in 1820 (Maddison 1995). The prevalence of tropical climate diseases and

∗Chengzheng Li: Institute for Economics and Social Research, Jinan University, Guangzhou, People’s Republic ofChina (PRC). E-mail: [email protected]; Jiajia Cong (corresponding author): School of Management,Fudan University, Shanghai, PRC. E-mail: [email protected]; Haiying Gu: Antai College of Economics andManagement, Shanghai Jiao Tong University, Shanghai, PRC. E-mail: [email protected]. This research is supportedby the National Natural Science Foundation of China (71333010), National Social Science Foundation of China(16ZDA019), Young Scientists Fund of the National Natural Science Foundation of China (71903074), YoungScientists Fund of Natural Science Foundation of Guangdong Province (2018A030310658), and Shanghai PujiangProgram (2019PJC007). We thank Benjamin Jones, Baohua Zhu, Xi Zhu, Mingwang Cheng, Qinghua Shi, and DeyuZhao, as well as the managing editor and the anonymous referees for helpful comments and suggestions. The AsianDevelopment Bank recognizes “China” as the People’s Republic of China. The usual ADB disclaimer applies.





the lack of suitable weather conditions for agriculture are the major reasons for theunderperformance of African countries (Sachs 2001, Masters and McMillan 2001,Sachs 2003). However, many other studies cast doubt on these results. They findthat the impacts of geography and climate on economic development are negligibleonce national characteristics (e.g., institutions and trade policy) are controlled for incross-sectional regressions (Acemoglu, Johnson, and Robinson 2002; Sachs 2003;Rodrik, Subramanian, and Trebbi 2004). Recent studies using a global databasewith resolution of 1° latitude by 1° longitude (Nordhaus 2006) and subnational dataat the municipal level (Dell, Jones, and Olken 2009) find that the negative effect oftemperature on income remains, but that its magnitude is attenuated.1

In addition to possible omitted variable bias, the cross-sectional analysisutilized in the above studies cannot reflect the contemporaneous effect of weathersince cross-sectional regression describes the long-run equilibrium relationshipbetween climate variables and the economy. Recently, a growing number ofempirical studies exploit country-level panel data to estimate the effect of weatherfluctuations on national income (e.g., Dell, Jones, and Olken 2012; Heal and Park2013; Burke, Hsiang, and Miguel 2015). The panel approach identifies the effectsof weather variables by exploiting their variations within an economy over time.Since variations in weather variables are strictly exogenous and stochastic, thisapproach can easily yield causative identification (Deschenes and Greenstone 2007,Deschenes 2014, Barreca et al. 2016) and can clearly isolate the effects of weatherfrom time-invariant country characteristics (Dell, Jones, and Olken 2012).

This paper uses county-level panel data on temperature and precipitation inthe People’s Republic of China (PRC) from 1996 to 2012 to examine their effects onthe growth rate of county-level gross domestic product (GDP) per capita.2 Our panelestimation shows that the growth rate of the county-level economy is negativelyrelated to the average temperature: increasing the annual average temperature by1°C reduces the growth rate of county-level GDP per capita by 1.05%–1.25%.Precipitation fluctuations have a negative effect on the growth rate, especially inagricultural counties and poor counties. Panel-distributed lag models show that theimpacts of weather variations are persistent over time since the cumulative effectsof weather variations are larger than their instantaneous effects. Increasing annualaverage temperature by 1°C reduces the cumulative growth rate of county-levelGDP per capita by 2.03%–3.84%. When 5 or 10 lags of weather variables areincorporated into the regression, the cumulative effect of precipitation becomes

1For example, when temperature rises by 1°C, the average income per capita of 12 countries in the Americasis reduced by 1.2%–1.9% at the municipal level (Dell, Jones, and Olken 2009). The results from cross-sectional datashow that a 1°C increase in temperature is associated with an 8.5% decrease in national income per capita (Horowitz2009; Dell, Jones, and Olken 2009).

2A longer version of the weather dataset from 1980 to 2012 is employed when lags of weather variables areneeded.

Could Weather Fluctuations Affect Local Economic Growth? 203

significantly negative. In other words, the cumulative effect of precipitation onlyshows up in the medium run.

By examining the relationship between weather and the growth rate of threesectors, we find that high temperatures have a substantial negative effect on thegrowth rate of the value added of primary industry and tertiary industry, whilethe negative effect on the growth rate of secondary industry is insignificant. Withseasonal average temperature, we find that only high temperatures in spring andsummer have significantly negative effects on the growth rate of GDP per capita.The negative effect of seasonal temperature is mainly caused by the effect onprimary industry. We construct temperature bins to detect the nonlinear effect ofdaily average temperature on growth. Compared with the reference bin [15, 20)°C,temperatures above 20°C have significantly negative effects on the growth rate ofGDP per capita, while extremely low temperatures and temperatures within [0,15)°C have no significant effect. The negative effect of temperatures above 20°Cis mainly levied on primary industry. Consistent with Colmer (2018) and Emerick(2018), we find that high temperatures have significantly positive effects on thegrowth rate of secondary industry, which indicates evidence of resource reallocationamong economic sectors. Most temperature bins do not show a significant effect onthe growth rate of tertiary industry. Regarding other possible effects of weather oncounty-level economies, we find that high temperatures have a significantly negativeeffect on the growth rate of labor productivity and fixed asset investment.

The effects identified through short-run weather fluctuations may differ fromthe long-run effects. For example, counties may adapt to climate change in the longrun and mitigate the short-run effects of weather variations. However, our data spanonly 17 years, and the long-run effects cannot be fully investigated. Following Dell,Jones, and Olken (2012), we make an initial attempt in this direction to explorethe effects of changes in weather in the medium run. The medium-run estimatesshow that temperature change and precipitation change each have a significantlynegative effect on the growth rate, and the coefficients are larger in magnitude thantheir counterparts in the panel analysis. This result implies that counties in the PRCadapt poorly in our sample period and the negative impacts of changes in weatheraccumulate over time.

Most existing studies on weather fluctuations and growth use cross-countryanalysis (Dell, Jones, and Olken 2012; Burke, Hsiang, and Miguel 2015; Healand Park 2013). A notable exception is Burke and Tanutama (2019) who usedistrict-level data from 37 countries. But their study does not control for any weathervariable other than temperature, which could cause omitted variable bias. In contrastto the obvious shortcomings in cross-country analysis (Burke and Tanutama 2019),our study enjoys several advantages. First, our cross-county analysis substantiallyincreases the number of cross-sectional observations. Although there are morethan 200 countries (regions) in the world, fewer than 140 of them are applicablein cross-country analysis. In contrast, our data contain 1,800 counties. Second,


our cross-county analysis reduces the risk of omitted variable bias. Possibleomitted variables—such as institutions, industrial policy, trade policy, and otherunobserved time-invariant factors—are similar for different counties within acountry. Moreover, our empirical analysis contains all major climatic variables,which further reduces the omitted variable bias. Third, our county-level paneldata have rich statistics, and we can explore many possible channels throughwhich weather fluctuations affect economic outcomes. In contrast, studies usingcountry-level data can only test very limited channels due to data deficiencies. Tothe best of our knowledge, this paper is the first to study the effect of weatherfluctuations on the economic growth of counties in the PRC and provide novelevidence on potential channels for weather–economy relationships.3

The remainder of this paper is organized as follows. Section II introducesthe data and provides descriptive statistics. In section III, we establish a theoreticalframework and describe our estimation strategy. Section IV presents the mainresults and various robustness checks. Section V examines potential channelsthrough which weather affects the growth of aggregate economic outcomes. Insection VI, we estimate the impacts of weather changes in the medium run. Thediscussion and conclusion are presented in section VII.

II. Data and Summary Statistics

A. Data

Our weather data come from the China Meteorological Data Sharing ServiceSystem, which is directed by the National Meteorological Information Center.This grid dataset provides nationwide terrestrial daily average temperature anddaily total precipitation data at 0.5° × 0.5° degree resolution, spanning from 1January 1980 to 31 December 2012. Zhang, Zhang, and Chen (2017) highlight theimportance of weather variables other than temperature and precipitation. Thus,other weather variables—including atmospheric pressure, wind speed, sunshinehours, and relative humidity—are introduced into our empirical analysis. Otherweather variables are drawn from the United States’ National Oceanic andAtmospheric Administration (NOAA). Relative humidity is not reported directly inthe data, but is constructed based on the standard meteorological formula providedby NOAA using temperature and dew point temperature.

We use geospatial software ArcGIS to aggregate the grid weather data tothe county-day level and eliminate counties with observations that are omitted or

3Deryugina and Hsiang (2014) estimate the effect of temperature on the income per capita of counties inthe United States, but they do not examine its effect on GDP growth, which has attracted more attention in thedevelopment literature.


have fatal errors.4 Then, we calculate the annual average temperature, days in eachtemperature bin, annual average precipitation, annual average relative humidity,annual average atmospheric pressure, and total sunshine hours for counties. Thefinal balanced panel of weather data contains weather information for 2,376counties from 1980 to 2012.

The economic data come from the Support System for China StatisticsApplication. The county-level dataset includes various annual statistics of GDP,population, employment, wage, investment, banking, public finance, trade, andsocial welfare, among others. The economic dataset includes data on 1,800 countiesfrom 1996 to 2012.

B. Summary Statistics

We merge a county’s weather data and economic data based on the nameand administrative code. The final combined panel covers 1,657 counties andspans from 1996 to 2012, containing each county’s weather variables and annualeconomic statistics. All monetary values are expressed in constant 2013 Chineseyuan. Summary statistics of key variables are presented in Table A.1 of the onlineAppendix.5

Figure 1 depicts the evolution of the average temperature of sample countiesfrom 1980 to 2012. The average temperature has risen gradually over the past 3decades. The peak annual average temperature of 13.97°C appears in 2012; thelowest annual average temperature of 11.23°C appears in 1984. Figure 2 describesthe evolution of the average precipitation of sample counties from 1980 to 2012.The year of maximum precipitation is 2012, with daily average precipitation of up to2.8 millimeters (annual precipitation is 1,024.8 millimeters). The year of minimumprecipitation is 2011, with an average daily precipitation of 2.1 millimeters (annualprecipitation is 766.5 millimeters).

Figure 3 depicts the relationship between counties’ average temperature andthe growth rate of GDP per capita from 1996 to 2012. Figure 4 presents therelationship between counties’ average precipitation and the growth rate of GDPper capita from 1996 to 2012. The growth rate of county-level GDP per capita isnegatively correlated with average temperature and average precipitation.

4We use a geographic-weighted approach to calculate the average temperature and average precipitationat the county-day level, where the weights are proportions of a county’s area within a specific grid. Counties thatconsistently have 0-value observations for temperature and precipitation are dropped. Weather variables other thantemperature and precipitation are constructed following Zhang et al. (2018).

5The online Appendix can be found at the corresponding author’s homepage: https://sites.google.com/site/jiajiacong/research.


Figure 1. Annual Average Temperature, 1980–2012

Note: Temperature is measured in degrees Celsius.Source: Authors’ calculation based on weather data from the China Meteorological Data Sharing Service System.

III. Theoretical Framework

In this section, we develop a theoretical framework for how weather variablesaffect economic growth and present the estimation strategy used for the empiricalanalysis.

A. Theoretical Framework

Our theoretical framework is based on Bond, Leblebicioglu, and Schiantarelli(2010) and Dell, Jones, and Olken (2012). Consider the production function ofcounty i in year t:

Y (Cit ) = eβCit AitK(Cit )αL(Cit )

1−α (1)

Ait

Ait= gi + γCit (2)

Y is aggregate output; K measures the capital stock; L measures population; Arepresents total factor productivity; and C measures the weather conditions of thiscounty. Equation (1) captures the level effect of weather on economic production


Figure 2. Annual Average Precipitation, 1980–2012

Note: Precipitation is measured in millimeters.Source: Authors’ calculation based on weather data from the China Meteorological Data Sharing Service System.

(e.g., the effect of current weather on aggregate output). Equation (2) captures thegrowth effect of weather (e.g., the effect of weather variables that affect the growthof total factor productivity).

Dividing both sides of equation (1) by population L, we have

y(Cit ) = eβCit Aitk(Cit )α (3)

where y is output per capita and k measures capital stock per capita. Taking logsof equation (3) and differentiating with respect to time, we have a dynamic growthequation as follows:

gy(Cit ) = gi + αgk (Cit ) + (β + γ )Cit − βCit−1 (4)

where gy and gk represent the growth rates of output per capita and capital stockper capita, respectively. Equation (4) indicates two features of weather shocks oneconomic growth. First, there is a lagged effect of weather on growth. Weatherconditions in the previous year affect the growth rate of the current year. Second,weather conditions affect the growth rate through the level effect β, which comesfrom equation (1) and the growth effect γ from equation (2). Equation (4) clearlyidentifies these two effects: (i) the level effect β is the coefficient of Cit−1, and


Figure 3. Growth of County-Level Gross Domestic Product per Capita and AverageTemperature, 1996–2012

GDP = gross domestic product.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing Service System andthe National Oceanic and Atmospheric Administration, and on economic data from the Support System for ChinaStatistics Application.

(ii) the growth effect γ can be derived by summing the coefficients of Cit andCit−1. The level and growth effects can still be clearly identified for more generalmodel structures, such as dynamic models including lagged dependent variablesand lagged weather variables, as we demonstrate in the online Appendix.

B. Model Specification

To estimate weather effects, we adopt the following regression specification:

git =P∑

p=0

λpCit−p + X ′φ + μi + δt + εit (5)

C is a vector of annual average temperature and average precipitation with up to Plags included; X is a vector of control variables containing other weather factors; μi

are county fixed effects; δt are year fixed effects; and εit are error terms.6 The error

6In the robustness check, we use alternative fixed effects and other specifications to cluster error terms. Thegrowth rate of the capital stock per capita gk is not controlled for in the regression because, based on our theory, gk


Figure 4. Growth of County-Level Gross Domestic Product per Capita and AveragePrecipitation, 1996–2012

GDP = gross domestic product, mm = millimeter.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing Service System andthe National Oceanic and Atmospheric Administration, and on economic data from the Support System for ChinaStatistics Application.

terms are simultaneously clustered by county and province-year to allow arbitraryserial correlation within counties and arbitrary spatial correlation within provincesin a year.

Our estimation proceeds as follows. First, we estimate equation (5) with nolags, focusing on the null hypothesis that weather does not affect growth:

H0(P = 0): λ0 = 0 (6)

Failing to reject this null hypothesis indicates the absence of both the level effectand growth effect. Second, we estimate equation (5) with lags and test the nullhypothesis that the weather variables have no instantaneous effect on the growthrate:

H ∗0 (P > 0): λ0 = 0 (7)

is a function of Cit . Including both gk and Cit in the regression would generate the “over-control problem,” whichresults in an underestimation of the effects of weather variables (Dell, Jones, and Olken 2014).


and the null hypothesis that the weather variables have no cumulative effect on thegrowth rate:

H ∗∗0 (P > 0):

P∑p=0

λp = 0 (8)

The value of∑P

p=0 λp corresponds to the growth effect γ in equation (4) as well asthe more general concept of growth effects in models with longer lag structures, asdemonstrated in the online Appendix.

IV. Results

A. Level Effect

We estimate equation (5) without including any lagged weather variables;that is, we test whether fluctuations in the weather variables have a level effecton growth. The null hypothesis is presented in equation (6). Column 1 of Table1 shows that there is a significantly negative relationship between the growth rate ofcounty-level GDP per capita and average temperature: the growth rate decreasesby 1.05% when annual average temperature increases by 1°C. Column 2 showsa negative and significant relationship between the growth rate and averageprecipitation. These results are robust to controlling for other weather variablesincluding atmospheric pressure, wind speed, and sunshine hours, as shown incolumn 3. Other weather variables are controlled for in all the following regressions.

Next, we investigate the heterogeneous effects of weather fluctuations ondifferent counties. We define a county’s agriculture ratio as its sum of gross outputvalue of agriculture from 1996 to 2012 divided by the sum of its GDP from1996 to 2012. A county is defined as an agricultural county if its agricultureratio exceeds the median agriculture ratio of the sample counties. As shown incolumn 4, the coefficient of the interaction between average temperature andthe agricultural county dummy is not statistically significant, indicating thatthe effects of temperature on agricultural and nonagricultural counties do notdiffer substantially. However, agricultural counties are more adversely affectedby average precipitation than are nonagricultural counties. We define a county asa hot county if its average temperature from 1996 to 2012 exceeds the medianof the sample counties. Column 5 shows that the coefficient of the interactionbetween average temperature and the hot county dummy is insignificant, indicatingthat the level effects of temperature on hot counties and cold counties are notsignificantly different. However, hot counties are more negatively affected byaverage precipitation. We define a county as a poor county if its average wage andaverage net income per capita of rural residents from 1996 to 2012 are smallerthan the corresponding medians of the sample counties. Column 6 shows that

Could Weather Fluctuations Affect Local Economic Growth? 211Ta

ble

1.B

asel

ine

Reg

ress

ion

Var

iabl

e(1

)(2

)(3

)(4

)(5

)(6

)(7

)

Ave

rage

tem

pera

ture

−0.0

105**

*−0

.012

6***

−0.0

122**

*−0

.012

5***

−0.0

123**

*−0

.012

1***

−0.0

124**

*

(0.0

037)

(0.0

039)

(0.0

039)

(0.0

039)

(0.0

039)

(0.0

039)

(0.0

039)

Ave

rage

prec

ipit

atio

n−0

.005

4***

−0.0

054**

−0.0

002

0.00

27−0

.001

20.

0113

**

(0.0

020)

(0.0

024)

(0.0

032)

(0.0

050)

(0.0

025)

(0.0

055)

Ave

rage

tem

pera

ture

×A

gric

ultu

ralc

ount

yaav

erag

e0.

0001

0.00

00(0

.000

6)(0

.000

6)P

reci

pita

tion

×ag

ricu

ltur

alco

unty

aver

age

−0.0

075**

−0.0

072**

(0.0

032)

(0.0

032)

Tem

pera

ture

×ho

tcou

ntry

bav

erag

e0.

0012

0.00

11(0

.000

8)(0

.000

8)P

reci

pita

tion

×ho

tcou

ntry

aver

age

−0.0

093*

−0.0

087*

(0.0

048)

(0.0

048)

Tem

pera

ture

×po

orco

unty

cav

erag

e0.

0025

***

0.00

25**

*

(0.0

005)

(0.0

005)

Pre

cipi

tati

on×

poor

coun

try

−0.0

171**

*−0

.017

1***

(0.0

028)

(0.0

028)

Ave

rage

atm

osph

eric

pres

sure

0.00

200.

0020

0.00

200.

0018

0.00

18(0

.001

3)(0

.001

3)(0

.001

3)(0

.001

3)(0

.001

3)A

vera

gew

ind

spee

d0.

0139

*0.

0120

0.01

44*

0.00

950.

0080

(0.0

073)

(0.0

074)

(0.0

073)

(0.0

074)

(0.0

075)

Tota

lsun

shin

eho

urs

−0.0

000

−0.0

000

−0.0

000

−0.0

000

−0.0

000

(0.0

000)

(0.0

000)

(0.0

000)

(0.0

000)

(0.0

000)

Obs

erva

tion

s25

,363

25,3

6325

,318

25,3

1825

,318

25,3

1825

,318

R-s

quar

ed0.

0795

0.07

970.

0801

0.08

170.

0803

0.08

170.

0834

Not

es:T

hede

pend

entv

aria

ble

for

allc

olum

nsis

the

grow

thra

teof

coun

ty-l

evel

gros

sdo

mes

tic

prod

uctp

erca

pita

.All

colu

mns

incl

ude

coun

tyfi

xed

effe

cts

and

year

fixe

def

fect

s.S

tand

ard

erro

rsar

ecl

uste

red

inpr

ovin

ce-b

y-ye

ars

and

coun

ties

.*p

<0.

1,**

p<

0.05

,and

***p

<0.

01.

a Aco

unty

isde

fine

das

anag

ricu

ltur

alco

unty

ifit

sag

ricu

ltur

era

tio

exce

eds

the

med

ian

agri

cult

ure

rati

oof

sam

ple

coun

ties

.bA

coun

tyis

defi

ned

asa

hotc

ount

yif

its

aver

age

tem

pera

ture

from

1996

to20

12ex

ceed

sth

em

edia

nof

sam

ple

coun

ties

.c A

coun

tyis

defi

ned

asa

poor

coun

tyif

its

aver

age

wag

ean

dav

erag

ene

tin

com

epe

rca

pita

ofru

ral

resi

dent

sfr

om19

96to

2012

are

smal

ler

than

the

corr

espo

ndin

gm

edia

nsof

sam

ple

coun

ties

.S

ourc

es:

Aut

hors

’es

tim

atio

nba

sed

onw

eath

erda

tafr

omth

eC

hina

Met

eoro

logi

cal

Dat

aS

hari

ngS

ervi

ceS

yste

man

dth

eN

atio

nal

Oce

anic

and

Atm

osph

eric

Adm

inis

trat

ion,

and

onec

onom

icda

tafr

omth

eS

uppo

rtS

yste

mfo

rC

hina

Sta

tist

ics

App

lica

tion

.


an increase in average temperature has a significantly negative effect on growthand that the effect is mitigated in poor counties. However, average precipitationhas a significantly negative effect only on poor counties. In column 7, we add allinteraction terms into the regression. The results show that average temperature stillhas a significantly negative effect and average precipitation has more adverse effectson agricultural counties and poor counties.

Therefore, we can clearly reject the null hypothesis that weather fluctuationshave no level (instantaneous) effect on the growth rate. Increasing the annualaverage temperature by 1°C would lower the growth rate of county-level GDP percapita by 1.05%–1.25%; average precipitation has a significant and negative effecton the growth rate of agricultural counties and poor counties.

B. Robustness of the Level Effect

In this section, we check whether the results of the level effect are robust toalternative regression specifications and other measures of temperature.

1. Alternative Regression Specifications

Table 2 reports the results under different regression specifications. Manystudies have found that temperature may have a joint impact with humidity (Zhanget al. 2018). In column 1 of Table 2, we use the heat index to measure the jointinfluence of temperature and humidity as a robustness check. The constructionof the heat index follows the standard formula provided by NOAA.7 The heatindex and average precipitation still have significantly negative effects on growth.In column 2, we cluster error terms by counties instead of province-by-years andcounties in the main specification. Column 2 shows that the effects of averagetemperature and average precipitation are significantly negative, consistent withthe baseline result. In column 3, we replace the year fixed effect in the baselineregression with a region × year fixed effect because the eastern, middle, andwestern regions of the PRC have remarkable development gaps and may havedifferent growth patterns.8 The result shows that the significant negative effect ofaverage temperature remains. In column 4, we replace the year fixed effect with1–4 order time trends because the development of county economies may exhibitnonlinear trends. Column 4 shows that the negative level effect of temperature isstill significant. Therefore, our main results, especially the negative level effect oftemperature, are robust to different regression specifications.

7The formula is available at https://www.wpc.ncep.noaa.gov/html/heatindex_equation.shtml; alternatively,please refer to Zhang et al. (2018).

8Following Deryugina and Hsiang (2014), we do not control for province-by-year fixed effects since it mayremove too much identification variation, which may produce attenuation bias that would overwhelm the results(Fisher et al. 2012).


Table 2. Robustness Checks with Alternative Regression Specifications

(1) (2) (3) (4) (5) (6)Heat Alternative Region 1–4 Order Hot Counties Hottest Year

Variable Index Cluster by Year Time Trends Excludeda Excludedb

Heat index −0.0045**

(0.0020)Average

temperature−0.0122*** −0.0146*** −0.0070*** −0.0117*** −0.0139***

(0.0040) (0.0040) (0.0025) (0.0042) (0.0047)Average

precipitation−0.0047* −0.0054** −0.0017 −0.0014 −0.0079** −0.0042(0.0024) (0.0026) (0.0027) (0.0026) (0.0031) (0.0028)

Other weathervariables

Yes Yes Yes Yes Yes Yes

Country fixedeffect

Yes Yes Yes Yes Yes Yes

Year fixed effect Yes Yes No No Yes YesRegion × year

fixed effectNo No Yes No No No

1–4 order timetrends

No No No Yes No No

Observations 25,318 25,318 25,318 25,318 23,800 23,665R-squared 0.0799 0.0801 0.0909 0.0720 0.0786 0.0846

Notes: The dependent variable for all columns is the growth rate of county-level gross domestic product per capita.Standard errors are clustered in province-by-years and counties. Temperature is measured in degrees Celsius andprecipitation is measured in millimeters. *p < 0.1, **p < 0.05, and ***p < 0.01.aIn column 5, counties with an average temperature above 20°C are excluded.bIn column 6, the hottest year 2012 is excluded.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing Service Systemand the National Oceanic and Atmospheric Administration, and on economic data from Support System for ChinaStatistics Application.

In column 5 of Table 2, we exclude counties whose average temperaturesfrom 1996 to 2012 are higher than 20°C to test whether the negative effect oftemperature is solely driven by hot counties. The result shows that temperature alsohas a significantly negative effect on cool counties. In column 6 of Table 2, weexclude the year 2012, which has the highest annual average temperature, from thesample, and test whether the main results are driven by this particularly hot year.The negative effect of temperature on the growth rate remains.

2. Different Measures of Temperature

We introduce different measures of temperature by constructing county-levelseasonal average temperature and temperature bins. A cross-country analysis of28 Caribbean countries indicates that high-temperature shocks have a negativeeffect on income only when they occur during the hottest season (Hsiang 2010).Table A.2 in the online Appendix reports how seasonal average temperature affectsthe growth rates of county-level GDP per capita and the value added of differentsectors. Column 1 shows that only high temperatures in spring and summer have


significantly negative effects on the growth rate of GDP per capita. The negativeeffect of seasonal temperature is mainly caused by the effect on primary industry.

Many empirical studies have found that temperatures have a nonlinear effecton economic activities (Burke, Hsiang, and Miguel 2015; Zhang et al. 2018; Burkeand Tanutama 2019). To check for a possible nonlinear effect of weather variableson growth, we construct temperature bins to measure county-level temperatureconditions. These new results are reported in Table A.3 of the online Appendix.Consistent with the literature, our study finds that temperatures have a nonlineareffect on the growth rate of county-level economies. Column 1 shows that comparedwith the reference bin [15, 20)°C, high temperatures above 20°C have significantlynegative effects on the growth rate of GDP per capita. Temperatures within [–10,15)°C and extremely low temperatures have no significant effect. Column 2 showsthat temperatures within [–20, –10)°C and high temperatures above 20°C havenegative effects on primary industry. In contrast, high temperatures above 20°Chave significantly positive effects on the growth rate of secondary industry. Thisfinding is consistent with micro evidence from Colmer (2018) and Emerick (2018),which indicates that the nonagriculture sector (mainly manufacturing) could benefitfrom weather shocks through labor reallocation between the agriculture sector andnonagriculture sector.9 For the growth rate of tertiary industry, most temperaturebins show no significant effect.

C. Cumulative Effect

This section uses panel-distributed lag models with up to 10 lags of theweather variables to explore the dynamics of weather effects.10 The distributedlag models nest both the level effect (instantaneous effect) and the growth effect(cumulative effect). Table 3 reports the results of estimating equation (5) with 0lags, 1 lag, 3 lags, 5 lags, and 10 lags of the weather variables. The first and secondrows present the level effect of the weather variables, and the bottom two rows reportthe sum of all weather lags (i.e., the cumulative effect).

Columns 2–5 of Table 3 show that a 1°C increase in temperature cumulativelyreduces the growth rate of county-level GDP per capita by 2.03%–3.84%. Themagnitude of temperature’s negative cumulative effect increases as more laggedweather variables are included. The cumulative effect of precipitation is significant

9Chen and Yang (2019) and Zhang et al. (2018) use micro-level data from manufacturing industries and finda significant negative effect of high temperatures on output and total factor productivity. Our results are differentbut not necessarily inconsistent with their findings. First, our focus is the growth rate of output rather than output orproductivity per se. Second, the secondary industry here includes many industries other than manufacturing. Third,we study the growth rate of all firms, including small and medium-sized firms, while Chen and Yang (2019) andZhang et al. (2018) focus on large firms with annual sales above 5 million Chinese yuan.

10Our panel of economic variables spans from 1996 to 2012, and the panel of weather variables spans from1980 to 2012. Even with 10 lags of the weather variables, the balanced panel of weather and economies still has 17years of observations.


Table 3. Regression with Lags

(1) (2) (3) (4) (5)Variables 0 Lags 1 Lag 3 Lags 5 Lags 10 Lags

Average temperature −0.0122*** −0.0116*** −0.0104** −0.0112** −0.0084*

(0.0039) (0.0039) (0.0043) (0.0045) (0.0046)Average precipitation −0.0054** −0.0051** −0.0047* −0.0052** −0.0067***

(0.0024) (0.0024) (0.0025) (0.0026) (0.0026)Other weather variables Yes Yes Yes Yes YesObservations 25,318 25,318 25,318 25,318 25,318R-squared 0.0801 0.0805 0.0814 0.0816 0.0833Cumulative effect of average −0.0203*** −0.0244** −0.0201 −0.0384***

temperature (0.0062) (0.0111) (0.0123) (0.0148)Cumulative effect of average −0.0024 −0.0067 −0.0135** −0.0541***

precipitation (0.0032) (0.0052) (0.0066) (0.0109)

Notes: The dependent variable of all columns is the growth rate of county-level gross domestic product per capita.All columns include county fixed effects and year fixed effects. Standard errors are clustered in province-by-yearsand counties. Given space constraints, the table only reports the sum of coefficients of all lagged weather variables(cumulative effect). Temperature is measured in degrees Celsius and precipitation is measured in millimeters. *p <

0.1, **p < 0.05, and ***p < 0.01.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing Service System andthe National Oceanic and Atmospheric Administration, and on economic data from the Support System for ChinaStatistics Application.

only when 5 lags or 10 lags are included. In other words, the cumulative effect ofprecipitation only shows up in the medium run. Increasing average precipitation by1 millimeter (i.e., increasing annual precipitation by 365 millimeters) cumulativelyreduces the growth rate of county-level GDP per capita by 1.35%–5.41%.

Table 3 demonstrates that the cumulative effect of average temperature islarger in magnitude than its level effect; average precipitation also generates a largercumulative effect than the level effect when 5 lags and 10 lags are introduced. Theseresults imply that the level effect of weather fluctuations in each period accumulates,rather than reverses, and that counties do not fully adapt to weather fluctuations.

V. Channels

In this section, we explore the channels through which the weather variablesexert their effects on growth. Many macroeconomic studies of weather effects focuson limited channels, especially agriculture and income levels, due to data limitations(Dell, Jones, and Olken 2012). Our data contain various economic statistics, so wecan explore whether there are other channels for weather to influence economicactivities.

A. Primary, Secondary, and Tertiary Industries

We investigate the level effect and cumulative effect of average temperatureand average precipitation on the growth rates of primary, secondary, and tertiary


Table 4. Channels: Primary, Secondary, and Tertiary Industry

(1) (2) (3)Value Added Value Added Value Addedof Primary of Secondary of Tertiary

Growth Rate of Industry Industry Industry

A. Models with no lags

Average temperature −0.0253*** −0.0110* −0.0034(0.0051) (0.0061) (0.0047)

Average precipitation −0.0091*** −0.0027 −0.0049(0.0027) (0.0038) (0.0033)

Observations 25,760 25,711 24,796R-squared 0.0673 0.0679 0.0505

B. Models with 5 lagsa

Average temperature −0.0313*** −0.0118* 0.0010(0.0058) (0.0067) (0.0051)


Cumulative effect of average temperature −0.0338** 0.0005 −0.0328**

(0.0158) (0.0199) (0.0154)Cumulative effect of average precipitation −0.0111 −0.0103 −0.0077

(0.0073) (0.0105) (0.0088)Observations 25,760 25,711 24,796R-squared 0.0703 0.0683 0.0520

C. Models with 10 lags

Average temperature −0.0276*** −0.0103 0.0031(0.0058) (0.0068) (0.0052)


Cumulative effect of −0.0316* −0.0292 −0.0534***

average temperature (0.0187) (0.0230) (0.0183)Cumulative effect of −0.0389*** −0.0465*** −0.0310**

average precipitation (0.0118) (0.0179) (0.0141)Observations 25,760 25,711 24,796R-squared 0.0729 0.0695 0.0539

Notes: All columns include other weather variables, county fixed effects, and year fixed effects. Standarderrors are clustered in province-by-years and counties. Given space constraints, part B and part C onlyreport the sum of coefficients of all lagged weather variables (cumulative effect). Temperature is measuredin degrees Celsius and precipitation is measured in millimeters. *p < 0.1, **p < 0.05, and ***p < 0.01.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing ServiceSystem and the National Oceanic and Atmospheric Administration, and on economic data from the SupportSystem for China Statistics Application.

industries. Part A of Table 4 begins with the model with no lagged weathervariables and shows the level effect. Column 1 of part A shows that averagetemperature and precipitation have significantly negative effects on the growth rateof primary industry. Columns 2–3 of part A show that the negative effects of average


temperature and precipitation on secondary and tertiary industry are negative butinsignificant.

Parts B and C introduce 5 lags and 10 lags of the weather variables,respectively, to examine the cumulative effects of average temperature and averageprecipitation. In part B, temperature has significantly negative cumulative effectson the growth of primary industry and tertiary industry. Precipitation has negativebut insignificant cumulative effects on all industries. When 10 lags are introduced,as shown in part C, temperature still has significantly negative cumulative effectson the growth of primary industry and tertiary industry. The cumulative effects ofprecipitation on all industries become significantly negative. This is consistent withour previous finding that the cumulative effect of precipitation emerges only in themedium run.

B. Average Wage, Investment, and Output of Agriculture and Large Firms

We investigate how average temperature and precipitation affect the growthrate of average wage, fixed asset investment, the gross output value of agriculture,and the gross output value of enterprises above a designated size. Part A of TableA.4 in the online Appendix begins with models without lags and reports the leveleffect. Column 1 shows that an increase in temperature and precipitation reduces thegrowth rate of the average wage. Since the average wage represents the productivityof labor, this result implies that the productivity of labor is affected by weatherfluctuations. Column 2 shows that average temperature does not have a significanteffect on the growth rate of fixed asset investment, while precipitation has asignificantly positive effect on it. Column 3 shows that an increase in temperatureand precipitation lowers the growth rate of the gross output value of agriculture,which is consistent with our finding that the agriculture sector is substantiallyinfluenced by weather fluctuations. Column 4 shows that both average temperatureand precipitation have negative effects on the growth of large firms, but only theeffect of precipitation is significant.

Parts B and C introduce 5 lags and 10 lags of the weather variables,respectively, to examine the cumulative effects of average temperature andprecipitation. The first column of parts B and C show that only temperature hasa significantly negative cumulative effect on average wages. The second columnshows that the cumulative effect of temperature on fixed asset investment issignificantly negative and that the cumulative effect of precipitation is significantlypositive.11 As shown by the third column of part C, temperature has a significantcumulative effect on agriculture. The fourth column of parts B and C show that

11One possible reason for this positive cumulative effect is that increasing precipitation accelerates thedepreciation of fixed assets. Thus, the growth rate of fixed asset investment has to be increased to compensate for thedepreciated fixed assets.


average temperature does not have a significant cumulative effect on the grossoutput of large firms.

In summary, in addition to the well-studied channel of agriculture, we findthat average temperature has significantly negative cumulative effects on the growthrates of average wages and fixed asset investment, and that average precipitationhas a significantly positive cumulative effect on the growth rate of fixed assetinvestment.

VI. Medium-Run Estimates

The comparison of the level effect and cumulative effect implies that thelevel effect of weather fluctuations in each period accumulates and that countiesdo not fully adapt to weather fluctuations. To further verify whether counties adaptto weather fluctuations in the medium run, we use a long difference approach toexplore the medium-run relationship between the growth rate and weather variables.Our specification is similar to Dell, Jones, and Olken (2012) and Burke and Emerick(2016). Specifically, given two periods, a and b, each period contains n years. Wedefine the average growth rate of county i in period a as gia = 1

n

∑t∈a git . We define

the vector of average weather variables as Cia. It contains the average temperature,average precipitation, and other weather variables in period a. The relationshipbetween the average growth rate and average weather variables in period a can bedescribed as

gia = ψ + κCia + μi + εia

This relationship is derived by taking averages on both sides of equation (5) with0 lags. The relationship between the average growth rate and average weathervariables in period b can be derived similarly. Then, we can have the followingregression specification:

gib − gia = c + κ (Cib − Cia) + (εib − εia) (9)

The unobservable county fixed effects μi are eliminated. Compared with the cross-sectional models, the long difference approach is free of omitted variable problemscaused by heterogeneity μi. Compared with panel-data models that investigate theshort-run effects, the long difference model investigates the effects of temperatureand precipitation on the growth rate in the medium run. If the estimated κ issmaller than the estimated level effect λ0 in magnitude, the county’s economy showsadaptation in the medium run; if not, the effects of weather fluctuations accumulateover time.

In the following analysis, we set each period at 4 years: period a is1996–1999, and period b is 2009–2012. Figure 5 shows the changes in the averagetemperature and growth rate across periods a and b. This demonstrates a clear


Figure 5. Changes in Growth Rate and Average Temperature, 1996–1999 and 2009–2012

Notes: Temperature is measured in degrees Celsius.Sources: Authors’ calculation based on weather data from the China Meteorological Data Sharing Service Systemand the National Oceanic and Atmospheric Administration, and on economic data from the Support System for ChinaStatistics Application.

negative relationship between the change in growth rate and average temperaturechange. Figure 6 shows the changes in average precipitation and the growth rateacross these two periods. To facilitate comparison, part A of Table 5 reports thepanel results. Column 1 of part B shows that a 1°C increase in temperature wouldlower the growth rate of county-level GDP per capita by 6.17%; the averageprecipitation change also has a significantly negative effect on the growth rate.Columns 2–4 of part B show that temperature has a significantly cumulativenegative effect on all industries, while precipitation has a significantly negativeeffect on secondary industry and tertiary industry.

Table A.5 in the online Appendix reports the effects of temperature andprecipitation on the growth rates of the average wage, fixed asset investment, thegross output value of agriculture, and the gross output value of enterprises abovea designated size across period a and period b. To facilitate comparison, part Arepeats the panel results. Part B shows that temperature change has a significantlynegative effect on most variables except the average wage, which is positivelyaffected; precipitation change has a significantly negative effect on the growth rateof gross output of enterprises above a designated size.


Figure 6. Changes in Growth Rate and Average Precipitation, 1996–1999 and 2009–2012

Notes: Precipitation is measured in millimeters.Sources: Authors’ calculation based on weather data from the China Meteorological Data Sharing Service Systemand the National Oceanic and Atmospheric Administration, and on economic data from the Support System for ChinaStatistics Application.

We introduce region fixed effects and alternative time periods for a and b inequation (9) to check the robustness of our results under a long difference approach.The results are reported in Tables A.6–A.9 of the online Appendix; most resultspersist. We also use the long difference method in Burke and Tanutama (2019) andfind that the negative effect of temperature is larger in magnitude than the leveleffect, consistent with our results in column 1 of Table 5.

In summary, comparing the coefficients of average temperature in part Aand part B of Table 5, the medium-run effect of temperature is larger than its leveleffect in magnitude, which indicates evidence of intensification. Therefore, we canconclude that counties in the PRC adapt poorly to temperature changes during oursample period. The effects of temperature accumulate over time and lead to a largerloss in the medium run.

VII. Conclusion

This paper exploits weather data and economic data from counties in the PRCduring the period 1996–2012 to examine the relationship between weather variablesand economic growth. We find a significantly negative relationship between the


Table 5. Long Difference Regression (I)

A. Panel results

(1) (2) (3) (4)County-Level Value Added Value Added Value Added

GDP per of Primary of Secondary of TertiaryGrowth rate of Capita Industry Industry Industry

Average temperature −0.0122*** −0.0253*** −0.0110* −0.0034(0.0039) (0.0051) (0.0061) (0.0047)

Average precipitation −0.0054** −0.0091*** −0.0027 −0.0049(0.0024) (0.0027) (0.0038) (0.0033)

Observations 25,318 25,760 25,711 24,796R-squared 0.0801 0.0673 0.0679 0.0505

B. Long difference regression

(1) (2) (3) (4)County-Level Value Added Value Added Value Added

GDP per of Primary of Secondary of TertiaryChange in growth rate of Capita Industry Industry Industry

Average temperature change −0.0617*** −0.0629*** −0.1102** −0.0642**

(0.0165) (0.0184) (0.0506) (0.0321)Average precipitation change −0.0388*** 0.0092 −0.0693*** −0.0347**

(0.0103) (0.0111) (0.0232) (0.0161)Early period 1996–1999 1996–1999 1996–1999 1996–1999Late period 2009–2012 2009–2012 2009–2012 2009–2012Observations 1,652 1,600 1,574 1,573R-squared 0.0314 0.0331 0.0262 0.0226

GDP = gross domestic product.Notes: Other weather variables are controlled for in all columns. The robust standard errors are reported inparentheses. *p < 0.1, **p < 0.05, and ***p < 0.01.Sources: Authors’ estimation based on weather data from the China Meteorological Data Sharing Service Systemand the National Oceanic and Atmospheric Administration, and on economic data from the Support System forChina Statistics Application.

growth rate and average temperature. Increasing average temperature by 1°C lowersthe growth rate of county-level GDP per capita by 1.05%–1.25%. The negativeeffect of precipitation mainly occurs in agricultural counties and poor counties.Using models with lags, we find that the cumulative effects of temperature are fargreater than its level effects. Models with 1–10 lags indicate that a 1°C increase inaverage temperature cumulatively lowers the growth rate of county-level GDP percapita by 2.03%–3.84%. When 5 lags and 10 lags are introduced, the cumulativeeffects of precipitation become significantly negative, implying that the cumulativeeffect of precipitation emerges only in the medium run. In addition to the well-studied agriculture channel, we find that temperature has considerable impacts ontertiary industry, labor productivity, and fixed asset investment. The long differenceapproach finds that counties in the PRC adapt poorly to weather changes in oursample period.


Our study has significant policy implications. First, we verify that weatherfluctuations have a considerable negative impact on economic growth. Theunignorable threat of weather changes demands appropriate government responsessuch as introducing crop diversity to help farmers insulate yields and incomeagainst weather extremes (Auffhammer and Carleton 2018). Second, we find thatin addition to the agriculture sector, weather fluctuations can influence economicgrowth by influencing the productivity of labor, fixed asset investment, and theproduction of nonagriculture sectors. These nonagricultural channels deserve moreattention because the share of agriculture in the PRC’s GDP continues to decline,and nonagricultural channels will become key channels in the future. Third, we findthat counties in the PRC adapt poorly to weather changes. This is the main reasonthat the cumulative effects of weather fluctuations outweigh their level effects.Adaptability to weather changes requires additional investment and technologicalinnovation.

References

Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2002. “Reversal of Fortune:Geography and Institutions in the Making of the Modern World Income Distribution.” TheQuarterly Journal of Economics 117 (4): 1231–94.

Auffhammer, Maximilian, and Tamma A. Carleton. 2018. “Regional Crop Diversity and WeatherShocks in India.” Asian Development Review 35 (2): 113–30.

Barreca, Alan, Karen Clay, Olivier Deschenes, Michael Greenstone, and Joseph S. Shapiro. 2016.“Adapting to Climate Change: The Remarkable Decline in the US Temperature MortalityRelationship Over the Twentieth Century.” Journal of Political Economy 124 (1): 105–59.

Bond, Steve, Asli Leblebicioglu, and Fabio Schiantarelli. 2010. “Capital Accumulation andGrowth: A New Look at the Empirical Evidence.” Journal of Applied Econometrics 25(7): 1073–99.

Burke, Marshall, and Kyle Emerick. 2016. “Adaptation to Climate Change: Evidence from USAgriculture.” American Economic Journal: Economic Policy 8 (3): 106–40.

Burke, Marshall, Solomon M. Hsiang, and Edward Miguel. 2015. “Global Non-Linear Effect ofTemperature on Economic Production.” Nature 527: 235–39.

Burke, Marshall, and Vincent Tanutama. 2019. “Climatic Constraints on Aggregate EconomicOutput.” NBER Working Paper No. 25779.

Chen, Xiaoguang, and Lu Yang. 2019. “Temperature and Industrial Output: Firm-Level Evidencefrom China.” Journal of Environmental Economics and Management 95: 257–74.

Colmer, Jonathan. 2018. “Weather, Labor Reallocation and Industrial Production: Evidence fromIndia.” CEP Discussion Paper No. 1544.

Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken. 2009. “Temperature and Income:Reconciling New Cross-Sectional and Panel Estimates.” American Economic Review 99(2): 198–204.

______. 2012. “Temperature Shocks and Economic Growth: Evidence from the Last HalfCentury.” American Economic Journal: Macroeconomics 4 (3): 66–95.


______. 2014. “What Do We Learn from the Weather? The New Climate–Economy Literature.”Journal of Economic Literature 52 (3): 740–98.

Deryugina, Tatyana, and Solomon M. Hsiang. 2014. “Does the Environment Still Matter?Daily Temperature and Income in the United States.” NBER Working Paper No.20750.

Deschenes, Olivier. 2014. “Temperature, Human Health, and Adaptation: A Review of theEmpirical Literature.” Energy Economics 46: 606–19.

Deschenes, Olivier, and Michael Greenstone. 2007. “The Economic Impacts of Climate Change:Evidence from Agricultural Output and Random Fluctuations in Weather.” AmericanEconomic Review 97 (1): 354–85.

Emerick, Kyle. 2018. “Agricultural Productivity and the Sectoral Reallocation of Labor in RuralIndia.” Journal of Development Economics 135: 488–503.

Fisher, Anthony C., W. Michael Hanemann, Michael J. Roberts, and Wolfram Schlenker. 2012.“The Economic Impacts of Climate Change: Evidence from Agricultural Output andRandom Fluctuations in Weather: Comment.” American Economic Review 102 (7): 3749–60.

Gallup, John Luke, Jeffrey D. Sachs, and Andrew D. Mellinger. 1999. “Geography and EconomicDevelopment.” International Regional Science Review 22 (2): 179–232.

Heal, Geoffrey, and Jisung Park. 2013. “Feeling the Heat: Temperature, Physiology & the Wealthof Nations.” NBER Working Paper No. 19725.

Horowitz, John K. 2009. “The Income–Temperature Relationship in a Cross-Section of Countriesand Its Implications for Predicting the Effects of Global Warming.” Environmental andResource Economics 44 (4): 475–93.

Hsiang, Solomon M. 2010. “Temperatures and Cyclones Strongly Associated with EconomicProduction in the Caribbean and Central America.” Proceedings of the National Academyof Sciences of the United States of America 107 (35): 15367–72.

Maddison, Angus. 1995. Monitoring the World Economy, 1820–1992 . Paris: Organisation forEconomic Co-operation and Development.

Masters, William A., and Margaret S. McMillan. 2001. “Climate and Scale in Economic Growth.”Journal of Economic Growth 6 (3): 167–86.

Nordhaus, William D. 2006. “Geography and Macroeconomics: New Data and New Findings.”Proceedings of the National Academy of Sciences of the United States of America 103 (10):3510–17.

Rodrik, Dani, Arvind Subramanian, and Francesco Trebbi. 2004. “Institutions Rule: The Primacyof Institutions over Geography and Integration in Economic Development.” Journal ofEconomic Growth 9 (2): 131–65.

Sachs, Jeffrey D. 2001. “Tropical Underdevelopment.” NBER Working Paper No. 8119.______. 2003. “Institutions Don’t Rule: Direct Effects of Geography on Per Capita Income.”

NBER Working Paper No. 9490.Zhang, Peng, Olivier Deschenes, Kyle Meng, and Junjie Zhang. 2018. “Temperature Effects on

Productivity and Factor Reallocation: Evidence from a Half Million Chinese ManufacturingPlants.” Journal of Environmental Economics and Management 88: 1–17.

Zhang, Peng, Junjie Zhang, and Minpeng Chen. 2017. “Economic Impacts of Climate Changeon Agriculture: The Importance of Additional Climatic Variables Other Than Temperatureand Precipitation.” Journal of Environmental Economics and Management 83: 8–31.


Online Appendix

The online Appendix is posted at https://sites.google.com/site/jiajiacong/research. It contains the following contents:

Table A.1. Summary Statistics of Key Variables

Table A.2. Seasonal Temperature Effects

Table A.3. Nonlinear Effects with Temperature Bins

Table A.4. Channels: Average Wage, Investment, and Output Values of Agricultureand Large Firms

Table A.5. Long Difference Regression (II)

Table A.6. Robustness of Long Difference Regression: Including Region FixedEffect (I)

Table A.7. Robustness of Long Difference Regression: Including Region FixedEffect (II)

Table A.8. Robustness of Long Difference Regression: Alternative Time Interval(I)

Table A.9. Robustness of Long Difference Regression: Alternative Time Interval(II)

Model of the Growth Effect of Weather Fluctuations in a Distributed Lag Model

Human Capital as Engine of Growth: The Roleof Knowledge Transfers in Promoting Balanced

Growth within and across CountriesIsaac Ehrlich and Yun Pei∗

Unlike physical capital, human capital has both embodied and disembodieddimensions. It can be perceived not only as skill and acquired knowledgebut also as knowledge spillover effects between overlapping generations andacross different skill groups within and across countries. We illustrate theroles these characteristics play in the process of economic development,the relation between income growth and income and fertility distributions,and the relevance of human capital in determining the skill distributionof immigrants in a balanced-growth global equilibrium setting. In all threeillustrations, knowledge spillover effects play a key role. The analysis offersnew insights for understanding the decline in fertility below the populationreplacement rate in many developed countries, the evolution of income andfertility distributions across developing and developed countries, and theoften asymmetric effects that endogenous immigration flows and their skillcomposition exert on the long-term net benefits from immigration to nativesin source and destination countries.

Keywords: demographic change, economic growth, endogenous immigration,human capital, income distributionJEL codes: F22, F43, J11, J24, O15

I. Introduction

In this paper, we focus on an underlying aspect of human capital as engineof growth—the role of knowledge transfers—which has received only modestattention in the literature so far. To introduce the subject matter, we start with anattempted definition of human capital, its distinct characteristics, and its relevance

∗Isaac Ehrlich (corresponding author): Department of Finance, School of Management and Department ofEconomics, University at Buffalo; National Bureau of Economic Research; and Institute of Labor Economics(IZA). E-mail: [email protected]; Yun Pei: Department of Economics, University at Buffalo. E-mail:[email protected]. This paper is based on lectures given at the Asian Development Bank Distinguished Speaker’sProgram in Manila on 2 August 2019 and the Conference on Human Capital and Economic Development inthe People’s Republic of China on 5 October 2019, sponsored by the Center for Excellence on Human Capital,Technology Transfer, and Economic Growth and Development and the Confucius Institute at the University atBuffalo. We are indebted to Yasuyuki Sawada for very helpful comments on a previous draft and to Sungmin Parkfor dedicated assistance. The usual ADB disclaimer applies.





for the process of economic growth and development as well as the demographicchanges that appear to accompany this process.

Human capital is “an intangible asset, best thought of as a stock ofembodied and disembodied knowledge, comprising education, information, health,entrepreneurship, and productive and innovative skills, that is formed throughinvestments in schooling, job training, and health, as well as through researchand development (R&D) projects and informal knowledge transfers” (Ehrlich andMurphy 2007). It is considered a “capital” asset because of its enduring impacton the returns to knowledge. In this context, it resembles physical capital as well.Both capital assets are subject to depreciation and obsolescence. And, like physicalcapital, human capital can also be formed through investment. However, there arealso important differences between the two. We count five major distinctions:

First, human capital is embodied in people. So unlike physical capital,it is controlled by individuals with heterogeneous abilities and preferences whochoose how and where to employ it—in different occupations, markets, nonmarketactivities, or even countries in the case of international migration.

Second, being embodied in people, human capital formation cannot beseparated from population formation, by which we mean decisions about familyformation, fertility, health, and longevity.

Third, unlike physical capital, human capital has limited opportunities to bemonetized as a stock and cannot serve as a collateral, which imposes financingconstraints on investment in its formation.

Fourth, unlike physical capital, human capital has both embodied anddisembodied dimensions. Not only can it be manifested as personal knowledge orskill acquired via schooling, training, and R&D but also via knowledge spillovereffects arising from publicly accessible books and articles, interaction betweenpeople, and informal means of communication. Extant human capital can thusbe productive in creating new human capital as well as social capital by formingnetworks of people. Unlike physical capital, disembodied human capital is thus alsolikely to generate positive externalities.

Finally, because of its latter distinct property, we believe that John MauriceClark (1923) was right in proposing that human capital, or knowledge as he put it,is the only instrument of production that is not subject to diminishing returns.

We consider the last two characteristics of human capital to be key forunderstanding its critical role in determining the level and distribution of per capitaincome within and across countries at a point in time as well as income growth andincome distribution within and across countries over the long haul. We illustrateand defend this argument using three selected topics that are based on our previousand current research, as we point out in the introduction to each of the followingsections:

The process of development—the shift from a stationary state or “stagnant”economic development regime to a persistent growth regime. In section II, we show

Human Capital as Engine of Growth 227

that an exogenous shock, such as a technological advance in health maintenancethat raises sufficiently the life expectancy of children, can increase the incentive ofparents to make sufficient investments in the human capital of their children thatcan trigger a takeoff to persistent and self-sustaining growth. The underlying reasonis that the disembodied human capital level accumulated by the parents’ generationand the knowledge spillover effects they generate, running from parents to theiroffspring’s generation, are not subject to diminishing returns.

The relation between income growth and income and fertility inequality ina balanced-growth, closed-economy setting. In section III, disembodied humancapital and its spillover effects across different skill groups can explain the linkbetween per capita income growth and income distribution over the transitionfrom stagnant or lower-growth equilibrium to a higher- and persistent-growthequilibrium.

The relevance of human capital in explaining the net benefits fromimmigration in a balanced-growth, global equilibrium setting. In section IV,disembodied human capital and its spillover effects can explain the long-termconsequences of immigration on per capita income growth and income distributionacross countries.

In each of these sections we use an endogenous growth and developmentframework in which human capital is the engine or basic driver of growth,and knowledge spillover effects are essential for establishing the existence ofbalanced-growth equilibrium solutions. This framework enables us to derive thelong-term effects of shifts in external triggers and underlying exogenous factorswhich affect the equilibrium solutions via the relevant comparative dynamicsanalysis and evaluate their impact on the rate and direction of long-term incomegrowth, income distribution, and demographic changes within and across countries.It also enables us to assess their welfare implications and some of their policyramifications.

II. Human Capital and Long-Term Economic Development

A. From Sporadic to Perpetual Economic Growth

The story of “growth” in many Western countries—which can becharacterized as persistent, self-sustaining growth in per capita income—has beenin existence for about 150–200 years following the first Industrial Revolution, afterbeing more or less stagnant during the Middle Ages. The transition from stagnationto growth is one way to think about this historical pattern: What can trigger such ashift? And what is the underlying engine of growth that could explain not just theshift from a low level of development into persistent growth in per capita incomebut also the demographic transition that has accompanied such a growth pattern inmany other parts of the world in more recent decades?


The neoclassical growth model mitigated the Malthusian populationexplosion threat by showing how it could be balanced by a commensurate growthin physical capital that can produce a constant level of per capita income. But thatmodel cannot assure growth in per capita income under an exogenously given andconstant technology.

This issue has been taken up by the endogenous growth literature, whichrecognizes human capital as the “engine” or driver of growth that assures continuousproductivity growth.1 There are alternative stories to explain endogenous growththrough persistent R&D investments (such as Romer 1990), but one importantreason to consider the human capital story is that it can also explain the“demographic transition”—rising longevity and declining fertility—which is seento be an inseparable part of the endogenous growth story.

In this section, we describe a model of endogenous growth where humancapital is the only instrument of production and its knowledge spillover effects serveas both the engine of growth and the source of the demographic transition. Themodel is drawn from Ehrlich and Lui (1991).

B. Human Capital Formation Process

In discreet time the human capital production process can be described asa function of parental inputs augmented by parental transferable knowledge and atransmission technology as follows:

Ht+1 = A (Ht + H0) ht (1)

where Ht+1 denotes the human capital level of an offspring, A denotes theknowledge transmission technology, Ht + H0 represent the parent’s (generation)production capacity, consisting of raw labor (H0) and the parent’s stock of humancapital (Ht), and ht denotes the share of the production capacity allocated to buildingknowledge in the offspring.

Equation (1) formalizes the assumption that human capital is not subjectto diminishing returns since under any given transmission technology (A) andinvestment level (h) there is a linear relation between parents’ accumulatedknowledge and what is transmitted to children via parental knowledge spillovereffects. This is a necessary condition for human capital accumulation.2

Since human capital is the only capital asset in the economy having a unitaryrental cost, and labor time is normalized at 1, representing all time available for

1In this section we rely primarily on Ehrlich (1990) and Ehrlich and Lui (1991, 1998) as well as, indirectly, onthe seminal human-capital-based approach to endogenous growth by Lucas (1988) and Becker, Murphy, and Tamura(1990).

2Lucas (1988) also incorporates spillover effects in his model, coming from the average agent, but this isneither a necessary nor a sufficient condition for deriving a growth equilibrium in that model, which is based on aninfinitely lived agent.


both market and nonmarket production, the growth rate of the offspring’s potentialincome is given by

[(Ht+1 + H0)/(Ht + H0)] ≡ (1 + gt ) = Ah∗, as Ht → ∞ (2)

Persistent growth rate in potential income (production capacity), (1 + g), can thus bepositive if and only if Ah∗ > 1. If Ah∗ ≤ 1, the economy will converge to a stagnantequilibrium level of development.

The implication of this analysis is that for an economy to move from astate of “stagnation” or periodic growth to a state of endogenous growth, that is,self-sustaining, persistent growth, it is necessary for the economy to invest asufficiently high proportion of its production capacity in human capital formation.

The sufficient investment rate must exceed (1/A). Note that this requires alarger share of productive capacity to be allocated toward human capital investmentsin the least developed countries, where the technology level A is well below that inthe developed countries, which may be a major reason why many of these countriesremain in a stagnant equilibrium trap.

C. Endogenous Growth and the Demographic Transition

By this approach, whether the economy is in a stagnant or growth regimedepends on the principal objectives motivating parents (or their representativegovernment) to invest in their children as well as on external shocks that can bringabout a significant change in these incentives. The study by Ehrlich and Lui (1991),for example, recognizes two major motivating forces benefiting parents:

The first is old-age support, which can also be thought of as family insuranceor intergenerational trade.

The second is altruism or companionship: parents enjoy vicariously both thenumber of children they have as well as the quality of their children—the humancapital or potential income they help produce in their offspring.

These motives increase the demand for both the number and quality ofchildren. Whether parents can receive any of these benefits, however, depends onthe children’s probability of survival or life expectancy.

Another important feature of the human-capital-based endogenous growthparadigm is that it can explain the “demographic transition,” which invariablyaccompanies a transition from stagnation or a low level of “development” toself-sustaining “growth.” This sets the model apart from Lucas (1988), where theagent is infinitely lived, and Romer (1990), where human capital is exogenous, sothere is never any demographic transition.

D. Model Setup

The decision maker is a representative agent—a young parent in an extendedfamily setup where young parents bear and raise children and provide old-age


support to their elderly parents. Specifically, agents live through three periods: fromchildhood to adulthood, with survival probability π1; and from adulthood to oldage, with probability π2. The unit costs to young parents of raising children (nt)and investing in their human capital are v and ht , respectively. The extended familyis assumed to be self-insured against the prospects of nonsurvival of children toadulthood and nonsurvival of adults to old age. Specifically old parents receivecompensations from their surviving adult children (π1nt) at a rate ωt+1 that isproportional to the children’s attained human capital (Ht+1), so an old parentreceives old-age material benefit in the amount π1ntωt+1Ht+1. Each adult child, inturn, pays a premium to the family insurance pool that is adjusted by the old parent’sprobability of survival, π2ωtHt . Parents are also altruistic toward their children andderive psychic benefits or “companionship” from the expected number of childrensurviving to adulthood and the human capital stock they helped build in each child.3

Under the extended family insurance setup, the representative adult’sconsumption flow at adulthood (period t) is thus given by

C1t = (Ht + H0) (1 − vnt − htnt ) − π2ωtHt (3)

The expected consumption flow parents derive at old age, funded by the earnings ofthe expected number of surviving offspring, is given by

C2t+1 = π1ntωt+1Ht+1 (4)

and the expected altruistic benefits (W ) the old parents derive from their survivingoffspring is given by

Wt+1 = B(π1nt )β (Ht+1)α (5)

where B is an altruism or “companionship” parameter, as defined in Ehrlich andLui (1991). The expected lifetime utility function, which the parent maximizes bychoosing optimal values of fertility and investment in human capital is thus

U ∗t = [1/ (1 − σ )]

[(C1

t

)1−σ − 1]

+ δπ2 [1/ (1 − σ )]{[(

C2t+1

)1−σ − 1]

+ [(Wt+1)1−σ − 1

]}(6)

where σ < 1 is the inverse value of the elasticity of substitution in consumptionand 0 < δ < 1 is a generational discount factor. The representative parent is thenmaximizing this expected lifetime utility function with respect to fertility andinvestment in human capital and thereby the consumption flows at young and oldage, subject to the budget constraints specified in equations (3) and (4) and the

3Although the probabilities of survival π1 and π2 are in principle endogenously determined by parentalinvestments in their own health and their children’s health (see Ehrlich and Kim 2005 and Ehrlich and Yin 2013), inthis analysis, they are treated as functions of the exogenously given state of the arts in medical science.


conditions assuring the concavity of equation (6). The concavity of equation (6) issatisfied subject to the following parametric restrictions:

1/ (1 − σ ) > β > α and 1 ≥ α (7)

To assure that the choice variables have interior solutions, it is necessary thatthe elasticity of substitution in consumption σ < 1 is large enough to exceed theelasticities of the marginal psychic returns to the quantity of children (β) and thatthe latter would exceed the elasticity of the marginal psychic return to the “quality”of children (α). This is because the assumed egalitarian treatment of all children byaltruistic parents implies that the opportunity cost of having another child accountsfor the desire to provide that child the same educational opportunities as those givento her siblings.

E. Endogenous Growth and the Demographic Transition

By integrating these elements of an endogenous growth model and solvingthe dynamic maximization problem, it can be shown that an upward shock inthe life expectancy of children (π1) due to technological breakthroughs in healthscience, such as sanitation or pasteurization of milk, can trigger a takeoff fromstagnant to growth equilibrium by raising h∗ sufficiently, so that the growth rate ofhuman capital (1 + g) exceeds 1 (i.e., Ah > 1). Put differently, a growth equilibriumis possible if and only if investment in human capital as a fraction of parentalproduction capacity or potential income exceeds a critical value h∗ > 1/A. Themodel suggests that this outcome can be facilitated by a technological shock thatsufficiently raises life expectancy, since a higher level of the latter raises one’spotential lifetime earnings and thus the expected return on investment in humancapital.

Initially both investment in children’s human capital and fertility go upbecause of the wealth effect generated by the shock. But as incomes go up, the netaltruistic benefits of having more children start declining because the opportunitycosts of having a larger number of children rise relative to the option of having fewerbut more educated children. Put differently, a substitution effect from “quantity”toward “quality” of children emerges, which can reduce fertility quite sharply. In theEhrlich–Lui (1991) model, this substitution effect is further reinforced by the familysecurity system whereby children provide old-age support for their old parents. Thepaper shows that this old-age support motive can be served more efficiently (i.e., ata lower cost to parents) by having fewer children but investing more in their humancapital.

The following charts simulate these results. In Figure 1, a sufficiently largeupward shock in π1 (the probability of survival from childhood to adulthood) canraise the level and slope of the stagnant equilibrium path of human capital formation


Figure 1. Takeoff to Growth

Notes: The figure shows possible dynamics after an exogenous increase in longevity: a sufficient increase in longevitycan shift the human capital phase line above the 45-degree line. Ht is parent’s human capital, and Ht+1 is offspring’shuman capital. For parameter values see footnote 4.Source: Figure reproduced from Ehrlich and Lui (1991).

(I) from its stagnant equilibrium regime (point A on the 45-degree line) into a pathof perpetual growth equilibrium regime (II). Figure 2 illustrates how this shift toperpetual growth affects the underlying time paths of fertility and investment inhuman capital.4 While investment in human capital rises sharply in absolute value,fertility declines continuously to the point where even a “corner solution” mightarise. Technically, this corner solution occurs only if the elasticity of marginalpreference for “quality,” as reflected in the level of human capital attained by anoffspring, is less than 1 (i.e., if α < 1).5 This constrained parameter implies thatover time, the marginal psychic net benefits from having children turn negative,which justifies a corner solution. Figure 2 illustrates this solution by restricting thevalue of fertility (n) to reach an arbitrarily small positive level (so countries donot disappear). A more realistic restriction is α = 1, in which case optimal fertility

4The parameters used to simulate Figures 1 and 2 are: A = 5.5; v = 0.13; α = 0; β = 0.75; σ = 0.5; π1 =0.6; π2 = 0.6; ω = 0.45; H0 = 0.5. The stationary equilibrium at point A produces optimal values of H ∗ = 0.292and n∗ = 2.793. A technological innovation shifting the value of π1 from 0.6 to 1 lifts the dynamic path I to II whereH ∗ grows without bound while n∗ in Figure 2 falls to a corner solution set arbitrarily at n∗ = 1.

5In Figure 2, the path of declining fertility per family ends arbitrarily at a rate of 1. An alternative explanationfor the decline in fertility below the replacement level is provided in Ehrlich and Kim (2007a) where fertility ismodeled as a decision variable that is determined jointly with a family formation decision. Fertility trends are shownto be the results of declining trends in the share of married relative to single households as well as in fertility withinthese families.


Figure 2. Demographic Transition

Notes: The figures show the simulated paths of human capital and fertility in response to an upward shock in the lifeexpectancy of children that lifts human capital formation from a stagnant to an endogenous growth regime, as shownin Figure 1. Ht+1 is offspring’s human capital, and nt is fertility. For parameter values see footnote 4.Source: Figures reproduced from Ehrlich and Lui (1991).

is always an interior solution and can reach a level below population-replacementlevel.

These results were considered to be harsh results in 1991 when total fertilityrates (TFR) in the most developed economies were above “replacement levels” (aTFR above 2.1 per family). But today, we have 80 countries out of 200 comprisingthe oldest countries in the world where TFRs are below 2. See Figures 3, 4, and 5.

What is notable about these figures is that in both highly developed Westerncountries—the United States (US) and the United Kingdom (UK)—where takeoffto growth started in the early 19th century, and in fast-developing Asian countries,where takeoff to growth started in the early 1960s, the pattern of the evolutionof fertility—an initial rise followed by a continuing and persistent decline—tendto be mirror images. In both groups, fertility rates rise around the same timeperiod in which the respective economies start taking off from a relatively stagnantequilibrium regime toward a regime of self-sustaining growth. And similar tosome south European countries, the developing countries in East Asia and thePacific exhibit fertility levels that are below those in the most developed Westerncountries (the US and the UK), as shown in Figures 3 and 4. One reason for thisresult is that the old-age support motive has traditionally played a stronger role inAsian culture where the family security system has been an important channel forproviding old-age support before the more recent establishment of a public socialsecurity system. This is consistent with the implications of the family securitymodel addressed in this section which accentuates the substitution of quantity forquality of children (see Figure 5). Another factor is that the more developed Western


Figure 3. Total Fertility Rates in the United Kingdom and the United States, 1800–2018

Source: Princeton European Fertility Project. http://www.gapminder.com (accessed February 26, 2020).

Figure 4. Total Fertility Rates of Developing Countries in East Asia and the Pacific,1960–2017

Source: FRED Economic Data. https://fred.stlouisfed.org/series/SPDYNTFRTINEAP (accessed July 4, 2019).


Figure 5. Total Fertility Rates of the Five Fast Developing Economies in East Asia and thePacific, 1962–2017

Source: World Development Indicators. https://data.worldbank.org/indicator/sp.dyn.tfrt.in (accessed September 8,2019).

countries attract a higher share of immigrants coming from sending countries withhigh fertility rates, which offsets to some extent the declining fertility rates in theimmigration-receiving countries, whereas in the Asian countries, immigration rateshave remained relatively low. This point will be further developed in section IV ofthis paper.

F. From Positive to Higher Growth—the United States’ Ascendancyas Economic Superpower

Another example of a shock that started a new takeoff from a relativelylow growth equilibrium to a higher and persistent long-term growth regime is theMorrill Act of 1862 in the US that established for the first time—to our knowledge,anywhere in the world—a public higher education system. The Act led to theformation of the land grant university system, which established most of the majorleading public universities in the US (for example, the University of California,Berkeley; University of California, Los Angeles; Pennsylvania State University; andMichigan State University). Unlike the takeoff from stagnant to growth equilibriumwe discussed in the previous subsection, which was simulated from a theoretical


model where the trigger was assumed to be an upward shock in the probability ofsurvival from childhood to adulthood, the trigger implicit in the Morrill Act wasthe significant reduction in the financing costs of college education. The land grantsmade enrollment in higher education possible for all students, especially those fromlow- and middle-income families, who could not previously afford the prohibitivetuition costs of higher education in the few exclusive older private colleges inthe US that were accessible primarily to students from wealthy families. A recentpaper by Ehrlich, Cook, and Yin (2018) has developed and applied empirically aLucas-type endogenous growth model of institutional change indicating that thisAct may have been a major factor that promoted the US growth rate in the latterpart of the 19th century and enabled the US to overtake the UK as an economicsuperpower in the 20th century. The UK lagged 50 years behind the US in launchinga comparable system of public universities.

III. Income Growth and Income Distribution6

A. Linking Income Growth and Income and Fertility Distribution

The homogeneous-agent, human-capital-based endogenous growth modelgenerates important and testable implications about intergenerational knowledgetransfer, the transitional development phase, and the demographic transition,leading to a balanced growth in per capita income. But the model is silentabout how the distribution of income evolves over this transitional phase. Thisissue has been of major interest in more traditional development literature. Inparticular, Kuznets (1955, 1963) proposed his “inverted-U curve” hypothesis tocharacterize the behavior of income inequality in the transition from a lower to ahigher development stage. The ensuing literature, however, has offered conflictingevidence and inference about Kuznets’ hypothesis and the direction of causalitybetween income growth and income inequality.

Any model dealing with the income distribution issue must start witha recognition of the sources of income diversity and why they vary over thedevelopment process. While there is a vast literature on this topic, a major voidin this literature is that it has largely overlooked the endogeneity of not just thelevel and distribution of income but also the level and distribution of demographicvariables over the transition from a stagnant or low development level to anendogenous growth equilibrium regime. This is especially important when theGini coefficient is used as an inequality measure, since it is a function of bothfamily-income inequality and the distribution of families across income brackets.

6This section draws largely from Ehrlich and Kim (2007b). See also Kuznets (1955, 1963); Tamura (1991);Zhong (1998); and De la Croix and Doepke (2003).


The latter depends on the distribution of the relative family group sizes and hencethe fertility rates across different family groups.

Following Ehrlich and Kim (2007b), we take a holistic approach to deal withthis latter issue by attempting to explain income growth and income distribution asendogenous outcomes of a paradigm that allows for heterogeneity across individualsor family groups and accounts for what separates these groups as well as whatlinks them together. Human capital remains the engine of growth and the soleinstrument of production in this model, while labor time is normalized at 1, as seenin section II. However, we assume here the existence of different “skill groups”due to differences in their inherited ability and family background. We recognizethree sources of inherent heterogeneity: (i) differences in learning or productionabilities (Ai); (ii) differences in income-yielding “endowments” (H i

0) stemmingfrom inherited social status or wealth; and (iii) differences in education-financingcosts (θ i). The latter can be affected by government educational policies that enablethe introduction of social mobility in the model. We abstract from differences inpreferences or external production technologies, since these need not be related toidiosyncratic personal differences.

The inherited sources of family differences are what separates the agentsin the model. What links them, however, are knowledge spillover effects acrossdifferent skill groups. The role of spillover effects from disembodied human capitalthus becomes a central focus of this section.

B. Basic Elements of the Model

To recognize inherent heterogeneities in the population, we assume foranalytical simplicity that the population is composed of two family types withvarying skills or earning capacities (i = 1, 2) due to positive assortative mating.Family group 1 is identified as the higher (initial) skill or leading group, whilegroup 2 is the (initial) lower skill or follower group. Specifically, the two skill groupsdiffer in ability (Ai), family endowments (H i

0), and unit cost of investment in humancapital as a fraction of potential income (θ i), but they share the same personal andaltruistic preferences. In addition, we also restrict the unit cost of rearing and raisingoffspring as a fraction of potential income to be uniformly distributed among allagents to enable a stable solution to the model, so v i = v . As seen in the modelpresented in the previous section, human capital is the only productive capitalasset in the economy. In this heterogeneous-agents model, however, we abstractfrom any uncertainty of survival from childhood to adulthood and assume that theagents’ lifespan includes just two effective periods—childhood and adulthood—soall agents live through these two periods. We further abstract from any need forold-age support, although the introduction of savings in the model would not affectany of its qualitative implications. For analytical simplicity, we also assume thataltruism is the only motive for having children.


The central assumption in the model is that while the human capitalproduction function of the top skill group (i = 1) remains the same as that specifiedin equation (1), the corresponding production function for the lower skill group(i = 2) is augmented by the social interaction term (S2

t )γ , reflecting the hierarchicalknowledge spillover effects (S2

t ) coming from the more knowledgeable skill group 1:

H 2t+1 = A2h2

t

(H 2

0 + H 2t

)1−γ [(H 1

0 + H 1t

) (N1

t /Nit

)]γ

≡ A2h2t

(H 2

0 + H 2t

) (E2

t P2t

)γ ≡ A2h2t

(H 2

0 + H 2t

) (S2

t

)γ(8)

In equation (8), hit measures investment rate per child as a fraction of the parent’s

earnings capacity; E2t ≡ (H 1

0 + H 1t )/(H 2

0 + H 2t ) measures relative production

capacity and thus potential “family-income inequality”; P2t ≡ N1

t /N2t denotes the

relative income-group size in the population; Nit stands for the number of workers

in the skill group i; S2t ≡ E2

t P2t is a measure of the knowledge spillover effect; and

γ < 1 indicates the latter’s intensity.The production function (8) expresses the role of two types of spillover

effects or social interactions in production. The first is within families—humancapital formation crucially depends on parental inputs or transfer of knowledgefrom parents to children. The other is across families or skill groups indicatingthat knowledge transmission or learning is also a social process. The rationalebehind S2

t ≡ E2t P2

t serving as a spillover of knowledge from high- to low-skillgroups is that its first term, E2

t ≡ (H 10 + H 1

t )/(H 20 + H 2

t ), which measures therelative earnings capacity of the two, also indicates the relative gap in knowledgebetween the two groups and thus the potential gain to skill group 2 from its socialinteraction with group 1. The second term P2

t ≡ N1t /N2

t is a proxy for the oddsof personal interaction between individual members of these skill groups, which isproportional to their respective population sizes. An alternative interpretation is thatP2

t captures the interaction intensity, as it represents the ratio of “teachers” (leaders)to “students” (followers), assuming that agents of type 1 are the effective source ofknowledge transfer.

Note that E2t is a measure of “family-income inequality,” while S2

t ≡ E2t P2

t

is a measure of “income-group” or “income-class” inequality, since P2t ≡ N1

t /N2t

represents the share of the total population (N) concentrated at the top relativeto the bottom income brackets. A third measure of inequality used in the modelis the Gini coefficient. In a population comprised of just two income groups, theGini coefficient becomes Gt ≡ (S2

t P2t )/[(1 + S2

t )(1 + P2t )], which is monotonically

rising with S2t and falling with P2

t .

C. Objective Function and Optimization

The representative family heads (i = 1, 2) in this model maximize theirlifetime utility comprised of the utility from their own lifetime consumption as well


as the altruistic benefits they are assumed to derive from their offspring’s potentiallifetime income (production capacity):

U(Ci

t ,W it+1

) = [1/ (1 − σ )][(

Cit

)1−σ − 1]

+ δ [1/ (1 − σ )][(

W it+1

)1−σ − 1]

(9)

with respect to ht and nt , where

Cit = (

H i0 + H i

t

) (1 − v ini

t − θ ihitn

it

)(10)

is the parent’s own lifetime consumption benefits at adulthood (period 1), and

W it+1 ≡ B(ni

t )β(H i

0 + H it+1

)α, with β > α and α = 1 (11)

is the parent’s altruistic benefits from having children. Note that equation (11)is similar to the altruism function (5) in section II, with two exceptions. First,the psychic benefits from the offspring’s well-being are specified as a functionof the latter’s potential income rather than just their human capital attainments,since this specification can be shown to allow for the existence of equilibriumunder both stagnant and growth regimes. Second, the restriction α = 1 is made inorder to guarantee interior solutions in fertility, which are necessary to preserveheterogeneity in the model. As in equation (3), v i and hi are fractions of the parenti’s production capacity that are spent on raising and investing in the human capitalof each child, and θ i is the unit cost of financing educational investment per child.Savings can be added as a separate choice variable so that earnings may stand forincome.

Solving the first-order conditions we can obtain optimal interior solutions forfertility, ni, and human capital investments, hi, over two possible stable equilibriumregimes and a transition path connecting the two. Stability is guaranteed by the forceof the knowledge spillover effects term, S2

t , and the parameter restrictions assumedin this model:

1/ (1 − σ ) > β > α = 1 (12)

D. Equilibrium Regimes7

Recall that human capital is the sole productive asset in this model, and timeat work is normalized at 1 to reflect potential income in both market and nonmarketactivities. In this case, equilibrium depends on the way human capital evolves overthe generations. Equations (10) and (11) represent a recursive model, since theleading group 1 arrives at all of its choices independently as a function of its ownparameters, while the following group 2’s fertility choices are affected by those offamily group 1 through the social interaction term (S2

t )γ . From the human capital

7For a more detailed technical analysis of the equilibrium conditions, see Ehrlich and Kim (2007b).


formation function and the optimal investment condition for group 1, we derive anexplicit, linear law of motion of human capital for this group:

H 1t+1 = [

v1(A1/θ1

)/ (β − 1)

]H 1

t + [v1

(A1/θ1

)/ (β − 1)

]H 1

0 (13)

Since the economy is dictated by family group 1, the equation of motion (13)indicates the existence of two equilibrium regimes, depending on the magnitudesof the model’s basic parameters: If the slope dH 1

t+1/dH 1t = v1(A1/θ1)/(β1) = A1h1

t

exceeds 1, H 1t grows exponentially without bound and the economy is in a persistent

growth equilibrium regime. If the slope is below 1, H 1t becomes constant and a

stagnant equilibrium regime ensues. The transitional development phase connectingthe regimes’ steady states is supported by the same parameter set that sustains thegrowth regime.

1. Stagnant Equilibrium Steady State

A stagnant equilibrium can be shown to exist when levels of the parameterset {Ai/θ i, v i} are sufficiently low, so the optimal value of investment in humancapital of family group 1 (h1) is below the critical level enabling a takeoff to growth,while fertility rates are relatively high when compared to their values at the growthequilibrium steady state, as shown in section II. Except for population size (Ni), allcontrol and state variables, including the equilibrium income and fertility inequalitymeasures, are then constant over time.

Figure 6 illustrates the conditions that must be satisfied in this equilibriumregime. A stable stagnant equilibrium exists if the equation of motion—theevolution paths of human capital of each family group H i

t+1 as a function ofH i

t —intersects the 45-degree line from above at a slope ait (s) ≡ dH i

t+1/dH it < 1.

This solution seems to reflect by and large the long historical period known asthe Middle Ages. Note that the stagnant equilibrium regime is consistent withoccasional fluctuations in per capita income and distribution as a result of changes inexternal conditions. However, small parameter shifts are not sufficient to generatepersistent growth in per capita income or the long-term distributions of incomeor fertility. This may explain the absence of continuous growth and distributionalchanges over the Middle Ages.

2. Transitional Development Phase

An upward shock in {Ai/θ i, v i} may trigger a takeoff into a growthequilibrium regime by raising the slopes of the evolution paths of human capital foreach family above 1, which is also the condition for attaining a growth equilibriumregime. While the parameter values that may trigger a takeoff guarantee the


Figure 6. Stagnant and Growth Equilibrium

Notes: Ht is parent’s human capital, and Ht+1 is offspring’s human capital.Source: Authors’ illustration.

existence of a growth equilibrium steady state, the effects on the pattern of thedynamic paths of income and fertility distributions will depend on the type of shockthat triggers the takeoff: whether the shock is uniform and occurs synchronouslyacross family groups, whether it benefits one group more than the other, or whetherit impacts one of the family groups with a lag. A skill-biased technological changethat generates an upward increase in the technology of knowledge transfer (Ai)and favorably affects the higher skill group can produce an inverted-U pattern ofthe evolution of income inequality over the transition phase, while governmentsubsidization of education that affects favorably the lower skill group can producea U-shaped pattern of the evolution of income over that phase. However, all takeofftriggers affect the fertility distribution in the same direction (see next subsection).

3. Growth Equilibrium Steady State

Human capital formation in this case grows continuously and without boundas does per capita income, while fertility and fertility inequality are falling. Thenecessary condition for attaining a growth equilibrium steady state is that theoptimal investment in human capital must be sufficiently high such that hi

t >

1/A. The necessary conditions for attaining a balanced-growth equilibrium are


opposite to those for obtaining a stagnant equilibrium steady state: (i) the pathof the evolution of human capital must have a slope higher than 1, i.e., ai

t (g) ≡dH i

t+1/dH it > 1, as is the case in Figure 6; and (ii) the fertility rate of a specific

skill group (ni) cannot be increasing with group size Ni. Further increases in thegrowth rate would again require an upward shock in one of the basic parameters ofthe model (Ai/θ i and v i). The impact on the distribution of income and fertilitygenerally depends on the parameter effecting the change as well as the way itimpacts the two family groups.

E. Some Testable Implications

The model offers several propositions and insights concerning observedempirical evidence about the pattern of long-run income growth and distributionas well as accompanying demographic trends.

1. Convergence in Marginal Growth Rates of Income and Fertility

The analysis implies that both fertility rates and marginal rates of change ofhuman capital must converge over time across families in any stable steady state.Formally, the analysis implies:

n1t = n2

t as t → ∞; and

a1t = (

dH 1t+1/H 1

t

) = A1h1t = a2

t ≡ (dH 2

t+1/H 2t

) = A2h2t

(S2

t

)γ(14)

The rationale follows partly from the definition of a stable steady state, in whichthe distribution of both income and fertility attains a stable level, but also fromthe role of knowledge spillover effects, which function like the glue or magnetthat holds equal the rates of change of population groups and their human capitalgrowth rate, thus eliminating the prospect of a blowup in the distribution of both.Figure 7 provides a graphical proof of equation (14). If there is an exogenous shockin the economy that lowers the marginal rate of growth of human capital for therepresentative agent in group 2, the income inequality ratio E2

t will rise in favorof group 2, which will then intensify the knowledge spillover effect S2

t flowingfrom group 1 to 2. The process would continue until the marginal changes in thegrowth rates of human capital of the two groups equalize again. The same appliesto the stability of the ratio of the populations of the two groups (N1

t /N2t ) in the

steady state, which can be maintained only if n1t = n2

t . Empirical evidence indicatesthat such convergence is typically observed following an external shock, such asthe digital revolution in the US, that may temporarily raise income inequalitysignificantly, despite the tendency of inequality to stay constant over long periods of


Figure 7. Equilibrium Solutions for Marginal Growth Rates (ai) and Spillover Effects (S2)

Notes: a1t ≡ (dH 1

t+1/dH 1t ) = A1h1

t ; a2t ≡ (dH 2

t+1/dH 2t ) = A2h2

t (S2t )γ ; S2

t ≡ [(H 10 + H 1

t )/(H 20 + H 2

t )][(N1t /N2

t )].Source: Authors’ illustration.

time. The secular decline in average fertility has also tended to reduce substantiallythe coefficient of variation of fertility over time in the US.

2. Income Inequality in Less-Developed Countries

In a stable stagnant-equilibrium steady state, family-income inequality,denoted by Ei(s), as well as the families’ relative human capital attainments equaltheir relative inherited endowments. This result can be shown to follow fromequation (14), the solution for the optimal values of investment in human capitalwhereby hi∗

t = [v i/θ i(1)], and the need to restrict the values of the unit cost offertility as a proportion of income to be the same for both skill groups (v1 = v2).This gives the following interesting result:

E2 (s) = H 1t /H 2

t = H 10 /H 2

0 (15)

This result implies that status differences are the key factor determining family-income inequality in economies that are stagnant over long periods, an inferencethat seems compatible with historical evidence, such as the pre-IndustrialRevolution period in Europe. Equation (15) is also dynamically stable: if aparameter shock lowers a2, raising H 1

t /H 2t above H 1

0 /H 20 , then E2 and S2 would


also rise initially. This would raise h2 and, in turn, increase a2t and lower H 1

t /H 2t

until the initial equilibrium is restored.

3. The Impact of Investment Efficiencies on Income-Group Distribution

By combining equations (14) and (15) and using the optimal values ofinvestments in human capital and fertility (derived through the first-order optimalityconditions) we can show that the value of the income-group inequality measurein the stagnant equilibrium steady state is just a function of the relative ratios(A1/θ1)/(A2/θ2), which we term the relative “investment efficiencies” of investingin the human capital of skill groups 1 and 2. Specifically, the group-incomedistribution is found to be a function of the relative investment efficiencies of thetwo groups:

S2 (s) ≡ E2 (s) P2 (s) = [(A1/θ1

)/(A2/θ2

)]1/γ ≡ (

e1/e2)1/γ

(16)

The same result is obtained for the value of the income-group inequality at thegrowth equilibrium steady state, which converges on the same value this measureobtained in the stagnant growth equilibrium steady state, provided the relativevalues of the parameters Ai and θ i associated with skill groups i = 1, 2 remainunchanged. This implies the strong result that if relative investment efficienciesof the two groups (e1/e2) remain the same at the stagnant and growth equilibriasteady states, the income-group inequality measure would also become equal atthese steady states.

F. Main Takeaways

The distribution of personal income is highly linked to the distribution ofhuman capital formation measures, not just at a point in time and over the lifecycleas a vast literature in labor economics has documented, but also under dynamic,long-term conditions that allow for the processes of development and sustainablegrowth as well as their accompanying demographic changes. Our analysis suggeststhat both processes, and thus the pattern of distribution of human capitalattainments and family income, are driven not just by the heterogeneous conditionsseparating the different population groups—what we term in this analysis as skillgroups—but also by the effects of knowledge spillover effects generated by thedisembodied component of human capital that binds them.

No general propositions can be established regarding the shape of thedistribution of income inequality over the transition from a stagnant to a growthequilibrium. The shape of the dynamic path depends on the type of shock thattriggers the transition and the way it impacts different family groups. While askill-biased technological shock can generate an inverted-U shape, government


educational policies favoring lower income groups can produce a U-shaped path.However, in all cases, the fertility distribution is expected to become attenuatedover the transition to a steady state.

Moreover, the dynamic evolution of the level and distribution of familyincome over the development process, and the distributional measures we use toquantify them, cannot be fully understood without recognizing their link to thecorresponding trends in the distribution of fertility rates, which are critical forunderstanding trends in income-group inequality or the Gini coefficient over time.The underlying trends in the level and distribution of fertility affect measures ofincome distribution differently, depending on how sensitive they are to changes inthe population shares of different skill groups in the economy.

IV. Equilibrium Migration and Human Capital and Physical Capital Formation

A. Is Human Capital Relevant for Immigration?

The relevance of human capital to immigration harkens back to our pointin the introductory section that human capital cannot be separated from people.People invest in this asset in order to guarantee a larger earning capacity for boththemselves and their offspring. But they have the choice to do that in the place wherethey were born, or they can seek a higher return on their investments by moving toeven distant locations where they have opportunities for a better life.

Thus, migration, whether within or across countries, is itself a kindof investment in human capital that has potential returns but also significantcosts—financial and emotional—to individuals who choose to migrate.

There is one implication that immediately arises from this feature ofimmigration: those who freely choose to migrate are expected to obtain a positiverate of return on migrating. But this may not always be the consequence fornatives of the affected destination (receiving) or source (sending) countries.Thus, understanding the “equilibrium” properties of immigration is an economicchallenge which also has important policy implications. This is the topic of ourcurrent evolving immigration study.8

B. International Migration—Common Features and Some Stylized Facts

International migration has been a persistent phenomenon since the dawnof civilization. Net migration is typically from source countries with high fertilityand low human capital formation to destination countries with lower fertility andpersistently higher human capital and per capita income levels. While immigration

8This section builds on our working paper (Ehrlich and Pei 2019) that is a sequel to an earlier study by Ehrlichand Kim (2015).


Figure 8. Immigrant Share of the United States Population, 1850–2000

Sources: The United States Census Bureau “Historical Census Statistics on the Foreign-Born Population of the UnitedStates: 1850–2000” and Pew Research Center tabulation of 2010 and 2016 American Community Survey (IPUMS).

has been growing dramatically in recent decades—by 50% worldwide from 2000 to2017—trends may shift, as the long-term share of immigrants in the US populationindicates (see Figure 8). Immigration may also come in waves—there were 30million asylum seekers in 2016 alone.

There are also significant variations in the volume of immigration withinand across different regions of the world. Asia and the Pacific is the largest regionwhere immigrants live, due to internal migration to destinations such as the RussianFederation and Australia. Asia as a whole, however, is a net migration region withIndia and the People’s Republic of China being the largest source countries.9 Onewould expect a useful model of immigration to account for both the persistenceof immigration and its variability across time and space. This is the focus of ouranalysis in this section.

C. Our Approach to Explain and Assess These Immigration Features

There is a vast literature on immigration, most of which focuses onits short-term effects in the context of a static model or the neoclassical

9See United Nations. 2017. “Trends and Drivers of International Migration in Asia and the Pacific.” https://www.unescap.org/sites/default/files/GCMPREP_1E.pdf.


growth model. The common feature of both approaches is the treatment ofimmigration as an exogenous variable. This approach is evident, with someexceptions, even in the recent authoritative report by a National Academy ofSciences panel on the economic and fiscal consequences of immigration (see NAS2017).10

The basic innovation of the analysis in this section is treating both migrationand economic growth as endogenous variables—coming from individual and familychoices that are subject to external triggers—and accounting for the economicconsequences of such triggers, especially in the long term. We model immigration inan open-economy setting that explores its consequences on income growth, incomedistribution within and across countries, and the net benefits from immigration tothe native populations in both the source (S) and destination (D) countries— knownas the “immigration surplus” (IS).

The following supply and demand analysis illustrates how to figure out theIS for destination and source countries in the short term under static conditions.

1. Why Is Immigration Surplus Expected to Be Positivein Country D in the Short Term?

In the short term, D’s economy has a fixed quantity of technology andphysical capital, so its labor demand curve is downward sloping. Any rise in laborsupply due to migration lowers labor wages. Consequently, labor income falls.

But total output rises: more labor yields more output, so aggregate incomerises. There are therefore conflicting consequences for natives. The interestingimplication of the conventional literature on immigration is that the net effect onnatives as a whole is positive—this is summarized by the well-known result that theshort-term immigration surplus is positive.

The reason is that while labor income falls as a result of diminishing returnsto labor, the returns to physical capital rise. The higher return to physical capitaloutweighs the fall in labor income because of the larger output generated byimmigrants who increase labor supply and thus aggregate output in the economy.Immigrants are willing to accept the lower wages they induce at the destinationcountry, since these wages remain higher than those in their home countries.

The net effect is what the literature calls the immigration surplus (see e.g.,Borjas 1995). This is the area E in Figure 9. Immigration thus creates a positive ISin the receiving country D. By the converse logic, however, the immigration surplusmust be negative in the sending country S in the short term, before any adjustmentsin physical or human capital stocks take place.

10For fair disclosure, one of the authors of this paper Isaac Ehrlich was a member of this panel.


Figure 9. The Short-Term Immigration Surplus

Notes: The recent National Academy of Sciences Report on the economic and fiscal consequences of immigration(see NAS 2017) has estimated the immigration surplus in the United States to be 0.3% of its gross domestic product,which amounts to $54.2 billion.Source: Authors’ illustration.

2. Limitations of the Static, Short-Term Immigration Surplus Measure

The short-term IS has some limitations that need to be highlighted. First, itconsiders only net income benefits to natives, not to the immigrants. This is the casein both D and S.

Second, as Figure 9 shows, the immigration surplus in country D can bepositive if and only if labor wages fall. Immigration yields a net gain to capitalowners but a loss to workers who do not own capital. Put differently, IS ignoresdistributional effects.


Third, a major limitation of the immigration surplus highlighted by ouranalysis is that it treats immigration as an exogenous variable and ignores whathas triggered it as well as its dynamic, long-term consequences. A major result ofour study is that the long-term consequences of immigration can be different fromits short-term effects due to its influence on the average level of human capital andits spillover effects across destination and source countries.

We also compute the immigration surplus after extending the concept toaccount for interaction effects—a two-way knowledge transfer between natives andimmigrants—as a result of complementarities in knowledge formation (“diversityeffects”), which enhance human capital formation. This concept is a naturalextension of knowledge transfer as an enhancer of human capital formation andeconomic growth.

D. Elements of the Model

Our story involves two countries—destination (D) and source (S)—whichare both in a growth equilibrium regime. As a point of reference, we allow forfree international mobility of labor. We allow for two periods of life: childhoodand parenthood. Each country (i = d, s) also has two segmented goods-producingsectors ( j = 1, 2)—high tech and low tech—which exclusively employ high-skilled and low-skilled workers ( j = 1, 2), respectively. In country D, theseworkers include immigrants as well. Immigrants’ children, however, are treated asnatives.

D is more developed technologically than S: both technologies of knowledgetransfer (A) and technologies of goods production (Г) are higher in D relativeto S (Adj > As j and Гdj > Гs j, j = 1, 2), while, by definition, the correspondingtechnologies are higher for high-skilled relative to lower-skilled workers and thehigh-tech sector relative to the low-tech sector (Ai1 > Ai2 and Гi1 > Гi2, i = d, s),respectively. At the same time, we need to assume that the unit costs of fertilityas a fraction of potential income must be higher in D relative to S, but theyare the same across the two skill groups within each country (vd > v s for all j).These are necessary modeling conditions that must be imposed to obtain interior,balanced-growth equilibrium solutions that allow for the possibility of continuousmigration across countries while also preserving the populations of different skillgroups within countries.

The global economy thus comprises six distinct population groups andrepresentative decision-makers: high- and low-skilled natives in two countries,including high- and low-skilled immigrants. Parents are altruistic. They want abetter life for their children so they make consumption, bequest, fertility, and humancapital investment decisions as well as migration decisions. Children thus benefitfrom both parental human capital investment and any bequests they receive inadulthood. Optimal immigration is determined at the point where the net utility


benefits to those who choose to migrate from S to D are equally productive on themargin as those who choose to stay in S—an arbitrage condition.

1. Human Capital Spillover Effects and Equilibrium Migration

To derive a balanced-growth equilibrium in D and S, we must allow forspillover effects of human capital, not just between parents and offspring or acrossskill groups within each country (including the immigrant groups in D) but alsoacross countries. International spillover effects may include financial remittances.But we focus especially on the role of knowledge transfers via spillover effects.In our benchmark model they are one-sided and hierarchical: they flow from thehigh-skilled sector to the low-skilled sector within countries and from each sectorin D to the corresponding sector in S, due to the superior technologies of producingboth human capital and industrial products in D relative to S. The magnitudes ofthese spillover effects are proportional to the weighted average of human capitalin D relative to S, where the weights are the relative population sizes of therelevant skill groups. The larger the size of each skill group in D—includingimmigrants—the bigger the spillover effects on the corresponding groups in S.

The one-sided spillovers are also gravitational: they pull all skill groupswithin and across countries together into equilibrium growth paths that grow atthe same rate over time. This assures the existence of a balanced-growth globalequilibrium that avoids “corner solutions.”

2. Formal Setup in the Benchmark Model

Formally, our specification of the human capital formation process andassociated spillover effects are given by the following six production functions ofsix skill groups of workers in the global economy: two skill groups of natives in twocountries (i = d, s; j = 1, 2) and two skill groups of immigrants (i = m; j = 1, 2).The corresponding production functions are

H d1t+1 = Ad1hd1

t H d1t for skilled natives in D, (17a)

H d2t+1 = Ad2hd2

t H d2t

(dd2

t

)γ 1

for unskilled natives in D, (17b)

H m1t+1 = Ad1hm1

t H s1t for skilled immigrants in D, (17c)

H m2t+1 = Ad2hm2

t H s2t

(dd2

t

)γ 1

for unskilled immigrants in D, (17d)

H s1t+1 = As1hs1

t H s1t

(ds1

t

)γ 2

for skilled natives in S, (17e)

H s2t+1 = As2hs2

t H s2t

(ds2

t

)γ 2(ss2

t

)γ 1

for unskilled natives in S, (17f)


where H i jt is the attained level of human capital of the six skill groups, hi j

t is theshare of earning capacity that parents in these groups invest in each child, and Ai j

are the endowed knowledge transfer technologies which augment the productivityof the corresponding investments in human capital hi j

t . As indicated earlier, theknowledge-production-and-transfer technology is higher in D relative to S, Adj >

As j, while it is also higher for high-skilled relative to lower-skilled workers withineach country, Ai1 > Ai2.

The spillover effects i j in equations (17b) and (17d) to (17f) areproportional to the relative human capital attainments of the interacting groups,weighted by their corresponding population shares—the equivalent of theknowledge spillover effect terms in equation (8).11 The spillover effects flow fromhigher to the lower skill groups within D and S and from skill j in D to the respectiveskill j in S. These effects are subject to diminishing returns: γ j < 1.

3. Optimization and Arbitrage Condition

Each representative parent in country i = {d, s} and skill group j = {1, 2} isassumed to maximize the utility of own consumption and the psychic returns fromaltruistic sentiments. Note that the altruism function in this model is expanded toaccount for the utility to parents from the bequest of physical capital they leavefor their children, which becomes the physical capital endowment inherited by theoffspring’s generation. Formally, the optimization problem for each of the heads ofthe six population groups comprising the world’s population is maximizing

U i j = [1/ (1 − σ )](Ci j

t

)1−σ

+ δ [1/ (1 − σ )](W i j

t+1

)1−σ

(18)

where the altruism benefits are given by

W i jt+1 = B

(ni j

t

)β(w

i jt+1 H i j

t+1

)η(ri j

t+1 Ki jt+1

)1−η

(19)

subject to the budget constraint on native parents’ lifetime consumption (Ci jt ) and

bequest (Ki jt+1):

Ci jt + ni j

t Ki jt+1 = ri j

t Ki jt +

(1 − v i ni j

t − θ i j hi jt ni j

t

)w

i jt H i j

t (20)

and the production functions of human capital (equations 17a–17f) and high- andlow-tech goods:

Qi jt = �i j

(Ki j

t

)α(H i j

t Li jt

)1−α

(21)

11More specifically, the relevant spillover effects are: dd2t = (Nd1

t H d1t + Ms1

t H s1t )/(Nd2

t H d2t + Ms2

t H s2t );

ss2t = (Ns1

t H s1t − Ms1

t H s1t )/(Ns2

t H s2t − Ms2

t H s2t ); ds1

t = (Nd1t H d1

t + Ms1t H s1

t )/(Ns1t H s1

t − Ms1t H s1

t ); and ds2t =

(Nd2t H d2

t + Ms2t H s2

t )/(Ns2t H s2

t − Ms2t H s2

t ).


where ni jt and hi j

t are the number of children and investment in human capital perchild, respectively, and v i and θ i j are the unit costs of raising and financing a child’seducational investments, respectively. In the goods production function in equation(21), �i j denotes the productivity parameter or production technology augmentinggoods production, and H i j

t Li jt stands for the effective labor input in production,

which in country D includes immigrants as well. While we assume that �i j ishigher in D relative to S (see section IV.D), the ratio �i1/�i2 must be equal acrosshigh- and low-tech sectors in each country.

For migrants, the optimization problem is the same, except that the budgetconstraint is given by

Cm jt +nm j

t Km jt+1 = (

1−τ k j)

rdjt Ks j

t +(

1−v s nm jt −θ s j hm j

t nm jt −τ h j

)w

djt H s j

t (22)

where τ h j and τ k j stand for migration costs in terms of forgone wage and nonwageincome (opportunity and transaction costs on assets), respectively. The arbitragecondition for the representative migrant underlying the decision to migrate from Sto D is given by(Cs j

t

)1−σ

+ δ(W s j

t+1

)1−σ

=(Cm j

t

)1−σ

+ δ(W m j

t+1

)1−σ

(23)

As we showed in section II, the necessary condition for attaining a balanced-growthequilibrium is that the rate of investment in human capital as a fraction of productioncapacity of the leading skill group in country D must exceed a critical level, givenby12

dH i jt+1/dH i j

t = Ad1(hd1

t

)∗ = (vdη)(Ad1/θd1

)/ (β − 1) = (1 + g∗) > 1 (24)

This model is too complex to be solved analytically via any closed-form solutions.We therefore resort to simulation analysis to solve for the key control and statevariables of the model. What facilitates the existence of general equilibrium interiorsolutions, which include an equilibrium flow and stock of migrants, are (i) theknowledge spillover effects running from the top skill group in the high-tech sectorof country D (which includes high-skilled migrants as well) to both its counterpartin country S and the lower skill group in the low-tech sector in country D (whichincludes low-skilled migrants as well) and (ii) the similar knowledge spillovereffects running from the top skill group in the high-tech sector in country S to thelower skill group in the low-tech sector in the same country. The global equilibrium,stabilized by hierarchical knowledge spillover effects, is illustrated in Figure 10.

12Additional constraints must be imposed on the ratios of human capital investment efficiencies and goodsproduction efficiencies to be equal across all skill groups within and across countries, i.e., (A1/θ1 )/(A2/θ2 ) and�i1/�i2 must be equal across countries (i = d, s) in order for the required balanced-growth equilibrium to exist.


Figure 10. Equilibrium Migration with Hierarchical and Gravitational Spillover Effects:A Global View

HC = human capital, HCPF = human capital production function, PF = production function.Source: Authors’ illustration.

4. Extended Model: Recognizing Spillover Effects from Diversity

Our baseline model is based on the premise that knowledge spillover effectsare largely hierarchical, running from the population group with the highest levelof acquired knowledge to groups with lesser attained knowledge, which cantherefore benefit from knowledge spillover effects. However, knowledge transferscan also be two sided, or interactive, as well as hierarchical. The interactionbetween immigrants and natives can produce more human capital because ofcomplementarities in knowledge production, or two-way spillover effects, betweenthe diverse interacting groups, essentially because diversity in the process ofknowledge formation is itself a source of efficiency in knowledge creation.

A formal way to express the benefits from diversity in knowledge formationis by modifying the knowledge transfer technological parameter Ai j in equation (1)to account for the two-way interaction variable d j as follows:

Adj(d j

) = Adj(1 + d j

)ρ(25)

where d j = Ms j/(Ndj + Ms j ) represents the share of the immigrants’ stock incountry D’s population.

While diverse workers may raise communication costs due to differences inlanguage and culture, the interaction between natives and immigrants raises theproductivity of knowledge generation, especially across workers of the same skill


level who acquired their knowledge and skill in independent environments (theclassical example being the Manhattan project). We consider these diversity effectsin our extended model when estimating the relevant immigration surplus impliedby that model.13

E. Numerical Solutions of the Benchmark Model and the Impactof Parameter Shifts

Table 1 presents numerical solutions of the global equilibrium systemillustrated in Figure 10, generated by calibrated simulations using data associatedwith the average groups of source and destination countries in Asia and the Pacificfrom 1962 to 2017. The group of destination countries and economies includeAustralia; Japan; the Republic of Korea; Singapore; and Hong Kong, China; whilethe group of source countries include the countries of origins of immigrants in thesedestination countries and economies.

The top row records the steady-state solutions for the control and statevariables of the benchmark model: fertility, investment in human capital, the growthrate of human and physical capital, country- and skill-specific knowledge spillovereffects, and the share of high-skilled migrants in the total stock of migrants residingin the destination locations. The latter variable exerts a critical influence on thelong-term net benefits of immigration for the destination and source locations. Thelower rows of Table 1 indicate the way the equilibrium values of the same controland state variables change as a consequence of exogenous shocks in some of thebasic parameters of the model. These shocks can also be thought of as representingtwo main immigration-inducing factors:

(i) Pull factors. For example, a skill-biased technological shock (SBTS)—for example, the digital/internet revolution starting in the late 1960s—which weinterpret as shifting the technology of knowledge generation and the technologicalprowess of high-skilled workers upward in both D (Ad1) and S (As1). This shockis shown to increase the rate of growth of human and physical capital and thusincome per capita as well in both D and S, as illustrated in the second row ofTable 1.

Another pull factor illustrated in the third row of Table 1 is a sharp drop infertility level in the destination due to exogenous factors, such as a structural changein the economy that induces a larger participation of women in the labor force andthus raises the opportunity costs of bearing and raising children, vd . This lattereffect also generates an upward change in the growth rate of human and physical

13Some evidence on favorable productivity effects of diversity is provided in Alesina, Harnoss, and Rapoport(2016).


Tabl

e1.

Com

para

tive

Dyn

amic

sun

der

the

Ben

chm

ark

Mod

el

nd

1=

ns1

=n

m1

=h

d1

=h

s1=

hm

1=

Hij t+

1/H

ij t=

�d

d2

=�

ds1

=M

s1/

Par

amet

erSh

ift

nd

2n

s2n

m2

hd

2h

s2h

m2

Kij t+

1/K

ij t�

ss2

�d

s2(M

s1+

Ms2

)

1.B

ench

mar

k0.

9561

1.55

301.

2045

0.29

690.

1858

0.18

582.

9693

5.65

695.

6386

0.5

2.A

d1

=12

,A

s1=

9.6

0.96

301.

5657

1.21

550.

2969

0.18

580.

1858

3.56

328.

9234

5.63

860.

5182

3.v

d=

0.17

050.

8645

1.55

961.

1939

0.32

660.

1858

0.18

583.

2663

5.65

697.

1558

0.5

4.�

s1=

�s2

=8.

404

0.95

911.

5530

1.20

590.

2969

0.18

580.

1858

2.96

935.

6569

5.63

860.

5

Ms1

/N

d1

=M

s1/N

s1=

Kd

1/H

d1

=K

s1/H

s1=

Hd

1/H

s1=

Kd

1/K

s1=

wd

1/w

s1=

rd1/rs1

=H

d1/H

d2

=K

d1/K

d2

=P

aram

eter

Shif

tM

s2/N

d2

Ms2

/N

s2K

d2/H

d2

Ks2

/H

s2H

d2/H

s2K

d2/K

s2w

d2/w

s2rd

2/rs2

Hs2

/H

s2K

s1/K

s2

1.B

ench

mar

k0.

1517

0.26

670.

3230

0.12

832.

2004

5.53

891.

5753

0.67

415.

6569

5.65

692.

Ad

1=

12,A

s1=

9.6

0.17

800.

2467

0.24

720.

0980

2.88

637.

2831

1.57

910.

6708

8.92

348.

9234

3.v

d=

0.17

050.

2878

0.22

540.

3256

0.11

146.

7921

19.8

442

1.66

220.

6055

5.65

695.

6569

4.�

s1=

�s2

=8.

404

0.08

880.

3135

0.32

090.

1278

1.00

722.

5280

1.57

080.

6828

5.65

695.

6569

Ben

chm

ark

para

met

ers:

σ=

0.9,

δ=

0.7,

B=

1,β

=1.

3,α

=1/

3,η

=2/

3,γ

1=

γ2

=0.

4,�

d1

=�

d2

=10

,�s1

=�

s2=

8.42

4,A

d1

=10

,Ad

2=

5,A

s1=

8,A

s2=

4,θ

d1

=θ

d2

=θ

s1=

θs2

=1.

16,v

d=

0.15

5,v

s=

0.09

7,τ

h1=

τh2

=0.

052,

τk1

=τ

k2=

0.5.

Sou

rce:

Aut

hors

’si

mul

atio

nre

sult

s.


capital, largely because it induces a decline in fertility but also a rise in parentalinvestments in human capital per child (a quantity–quality substitution effect) andthus in per capita income in both D and S.

(ii) Push factors. The factor we illustrate in the fourth row of Table 1is an adverse productivity shock due to war, famine, or political instability inthe source country, captured by a fall in the technological parameter affectinggoods production in both sectors of the source country—a downward shift in theparameters �s j. This change initially raises aggregate income in D and lowers it inS, but it causes no growth effects in either D or S.

F. Long-Term Effects of Immigration Triggers on Immigration Surplusunder the Benchmark Model

In the static context, the effects of any immigration triggers on the short-termimmigration surplus are always positive in D and negative in S, as predicted by theconventional or neoclassical models of immigration. In the context of our model,however, where both growth and immigration as well as their basic determinants areendogenous variables, the static analysis cannot predict the long-term evolution ofthe immigration surplus following a change in immigration generated by a specificimmigration trigger. To do that, we need to assess the level and direction of changesin the immigration surplus by isolating the “pure” immigration effects generatedby the rise in migration flows of high-skilled and low-skilled workers from S toD from the total effect of the trigger on the economy, as summarized in Table 1.More specifically, we assess the percentage difference between the unrestricted totaleffects of any immigration trigger and the latter’s hypothetical effects if immigrationflows were restricted to remain unchanged at their initial equilibrium levels. Table 2summarizes the dynamic effects of these “pure” immigration effects in the contextof our benchmark model.

Generally, the long-term effects of immigration changes on the immigrationsurplus depend on the type of external shock—the pull or push factors that havetriggered them. Under an SBTS, which is a pull factor, the immigration surplusrises in D and falls in S. This is essentially because the SBTS induces a rise inthe skill composition of migrants—the share of high-skilled migrants in the totalflow and stock of skilled workers in the population. Some evidence supportingthe validity of this result is illustrated in Figure 11. The figure shows that in fourof the largest destination countries—the UK, France, Australia, and Canada—theweighted averages of the skill composition of the migrant population have evenexceeded those of the total domestic populations in the corresponding countriesover the period 1975–2010, a period in which the digital revolution may havetriggered an SBTS in the destination countries. More importantly, this inducedimmigration effect raises the average human and physical capital stocks in theeconomy and thus the level of per capita income in D, which is the key determinant


Table 2. Benchmark Model: Immigration Surplus Changes Due to Induced Migrationby Triggers

Destination D Source S

3rd Generation 5th Generation 10th Generation 10th GenerationTrigger after a Shock after a Shock after a Shock after a Shock

1. Skill-biased technological shock in D and S

Human capital 0.04 0.25 1.16 −14.18Physical capital 0.06 0.25 1.10 −13.98Full income 0.05 0.26 1.14 −14.11Utility 0.007 0.04 0.23 −2.30

2. Adverse fertility shock in D

Human capital 0 0 0 67.96Physical capital −0.20 −0.50 −1.29 54.48Full income −0.16 −0.35 −0.77 64.03Utility −0.004 −0.01 −0.03 10.76

3. Adverse productivity shock in S

Human capital 0 0 0 40.39Physical capital −0.08 −0.26 −0.42 37.99Full income −0.10 −0.16 −0.24 39.61Utility −0.001 −0.005 −0.009 5.01

Notes: Values are percentage changes in welfare measures. Positive values mean a loss from restrictingmigration, hence a gain from unrestricted migration.Source: Authors’ simulation results.

of the level and direction of the change in the immigration surplus. The rise inthe skill composition of emigration from S exerts the opposite effects in S, whichexperiences a “brain drain.” Consequently, the immigration surplus in S falls butrises in D.

Under a downward fertility shock in D or an adverse productivity shock in S,the immigration surplus falls in D and rises in S. This is because both shocks raisethe volume of immigration flows from S to D, without any increase in their skillcomposition. Since immigrants come from source countries where the average levelof human and physical capital is lower by our model, the corresponding immigrationwaves into D thus lower the average per capita human and physical capital levelsin D, causing the immigration surplus to decline in D. But the higher population ofskilled and unskilled workers in D raise the intensity of the knowledge spillovereffects running from D to S, which raises the average human capital in S. Theimmigration surplus in S thus increases.

In general, under the benchmark model, immigration triggers generallyproduce asymmetric changes in the immigration surplus in destination and sourcecountries D and S. The underlying reason is that individual migrants do notinternalize the externalities generated by the spillover effects of their decision tomigrate on the native populations in both their home and destination countries.


Figure 11. Weighted Averages of the Skill Composition of the Migrant Population (WBand IAB) versus Total Domestic Population in Major Destination Countries (BL):

Population Age 25+ with 13+ Years of Education

BL = Barro and Lee, IAB = Institute for Labor Research, WB = World Bank.Sources: Authors’ illustration using WB data from Schiff and Sjoblom (2011); IAB data from Brücker, Capuano, andMarfouk (2013); and BL data from Barro and Lee (2013).

G. Long-Term Effects of Immigration Triggers on Immigration Surplusunder the Extended Model

Allowing for spillover effects from diversity, which generatecomplementarities in knowledge production between migrants and natives, altersto some extent the level and direction of the induced immigration effects of the pulland push factors considered in the previous section.

Any external shock that increases immigration from S to D is expected toraise the growth rate of human capital formation in D and thus in per capita incomein D. The knowledge spillover effects running from the increased levels of skillgroups in D assure that the growth rate of per capita income will rise in S as well, asthe growth rates of all skill groups ultimately converge to that of the top skill groupin D. However, the net effects on the immigration surplus also depend on the impactof the trigger on the skill composition of workers in the two countries, as indicatedby Table 3.


Table 3. Extended Model: Immigration Surplus Changes Due to Induced Migrationby Triggers

Destination D Source S

3rd Generation 5th Generation 10th Generation 10th GenerationTrigger after a Shock after a Shock after a Shock after a Shock

1. Skill-biased technological shock in D and S

Human capital 0.01 0.21 1.35 −13.99Physical capital 0.05 0.22 1.17 −14.43Full income 0.03 0.22 1.28 −14.12Utility 0.004 0.04 0.25 −2.40

2. Adverse fertility shock in D

Human capital 0.36 1.20 5.15 71.48Physical capital −0.12 0.06 2.51 58.20Full income 0.10 0.63 3.95 67.66Utility 0.04 0.11 0.50 11.79

3. Adverse productivity shock in S

Human capital 0.40 −0.08 −2.65 39.33Physical capital 0.70 0.13 −2.21 37.27Full income 0.40 −0.09 −2.61 38.65Utility 0.05 −0.01 −0.27 4.84

Notes: Values are percentage changes in welfare measures. Positive values mean a loss from restrictingmigration, hence a gain from unrestricted migration.Source: Authors’ simulation results.

Under an SBTS, the asymmetry in the consequences of the SBTS triggerbetween D and S remains: while the gain in the immigration surplus becomes largerin D—it rises by 1.28% in D after 10 periods in the extended model relative to1.14% in the benchmark model—there is still a loss in the immigration surplus inS due to a reduction in the population share of high-skilled natives and thus in theaverage corresponding levels of human and physical capital and per capita income.Indeed, the loss level remains similar to that under the benchmark model: –14.12%after 10 periods relative to –14.11% in the benchmark model.

Under a downward fertility shock in D, the immigration surplus now turnspositive rather than negative as it was in the benchmark model, while it continues tobe positive in country S. This is a win-win outcome due to the assumed interactivediversity effects of immigration in the receiving country. This may provide someexplanation to the receptiveness of a number of European countries to immigrationsurges in recent decades.

Under an adverse productivity shock in S, the immigration surplus turnspositive in D for three consecutive periods, but it then slides down in a negativedirection. The initial gain in the immigration surplus is due to the interactivediversity effects, while the later loss is due to the lower average human capital of


migrants relative to those of the natives in D, which outweighs the diversity benefits.The immigration surplus continues to be positive in the sending country S.

H. Policy Implications

1. Asymmetries in the Net Gain from Immigration across Destinationand Source Countries

Our simulations, especially under the benchmark model, reveal that changesin the immigration surplus in D and S tend to be asymmetrical in response to bothpull and push factors, not just in the short run but in the long run as well. This callsfor an explanation, which we take up in the next subsection.

For example, an SBTS raises immigration surplus (IS) in D and lowers it inS, while the adverse fertility shock in D or productivity shock in S lowers IS in Dand raises it in S. At the same time, we find that in most cases, immigration lowersincome inequality across D and S because of the spillover effects running fromskilled groups in D to their counterparts in S, which lower the percentage incomegap between the two countries.

2. Externalities and the Political Stability of Immigration

The essential reason for the asymmetry in the immigration surplus generatedby both pull and push forces may be related to the fact that free migration mayinevitably impose negative externalities on natives, since individual decisions tomigrate do not consider any adverse consequences to natives in either D or S. Thus,while there exists a balanced-growth equilibrium under unrestricted immigration, itmay be unstable politically. Note that in this analysis we consider only the economicconsequences of migration. We ignore the fiscal burdens of immigration in D, whichmay worsen the IS for natives in D even in the short term.

This explains why most independent countries resort to regulatoryintervention in the form of restrictive immigration policies or to bargainingsolutions, for example, the US aid to Central America when immigration waveswere triggered by adverse productivity shocks in these sending countries.

A superior solution, however, would be more investments in human capital inthe sending countries—especially in higher education—which puts a check on theincentive to emigrate while allowing higher education to produce higher growthin sending countries. But this solution works if the economic policies promotecompetitive labor markets to provide an adequate rate of return to investment inhigher education. This may have been one of the reasons the Morrill Act had beensuccessful in raising the rate of productivity growth in the US due to the dominantshare of the private sector in the US in the 19th century.


There is, of course, an altruistic solution as well: a special allowance, forexample, for asylum seekers and refugees.

V. Concluding Remarks—The Role of Knowledge Spillover Effects

We began this paper by highlighting the role of disembodied human capitaland its spillover effects in enabling the attainment of a balanced-growth equilibrium.Each of the studies we covered in this paper relies on at least one of the followingchannels that play a critical role in enabling a balanced-growth equilibrium:

(i) knowledge spillover effects across generations in models of homogeneousagents or a representative agent;

(ii) knowledge spillover effects across different skill groups in models ofheterogeneous agents in a closed economy, which enable a balanced-growthequilibrium involving both income growth and income distribution within aneconomy; and

(iii) knowledge spillover effects across heterogeneous skill groups within andacross countries.

These effects could be hierarchical, flowing from the group with greaterknowledge to the one with lesser knowledge, or it could be two sided because ofcomplementary “diversity effects.” Equilibrium immigration flows under balancedgrowth regimes are enabled due to the spillover effects operating through all of theabovementioned channels.

Spillover effects by definition impart external effects on others, which inmany cases may also involve externalities, since not all the benefits from spillovereffects can be internalized. The extent of externality effects may be minor whenspillover effects take place between parents and offspring, but it may be moreserious in the case of immigration for two reasons:

First, cross-country externalities, like intellectual property, are more difficultto internalize. Their presence generally leads to underinvestment in humancapital and explains why destination countries may be especially concerned aboutprotecting technological property rights as part of trade agreements. A case in pointmay be the current trade dispute between the US and the People’s Republic ofChina.

Second, the immigration decision by itself has an inherent externalityproblem because individual decisions to migrate are motivated by the net benefit tothe individual but they do not account for their long-term consequences for nativesin both destination and source countries. As our paper points out, they generally


result in opposite net benefits from immigration across countries in the long term.The asymmetry leads to politically unstable equilibria and explains why, in mosteconomies, immigration is subject to some restrictive policies, especially in thecase of unskilled migration.

Recognizing these externalities opens the door for policy interventions aimedat maximizing the benefits from human capital as engine of national and globaleconomic development and balanced growth.

References

Alesina, Alberto, Johann Harnoss, and Hillel Rapoport. 2016. “Birthplace Diversity andEconomic Prosperity.” Journal of Economic Growth 21 (2): 101–38.

Barro, Robert J., and Jong-Wha Lee. 2013. “A New Data Set of Educational Attainment in theWorld, 1950–2010.” Journal of Development Economics 104 (C): 184–98.

Becker, Gary S., Kevin M. Murphy, and Robert Tamura. 1990. “Human Capital, Fertility, andEconomic Growth.” Journal of Political Economy 98 (5, Part 2): S12–S37.

Borjas, George J. 1995. “The Economic Benefits from Immigration.” Journal of EconomicPerspectives 9 (2): 3–22.

Brücker, Herbert, Stella Capuano, and Abdeslam Marfouk. 2013. “Education, Genderand International Migration: Insights from a Panel-Dataset 1980–2010.” Institute forEmployment Research. http://www.iab.de/en/daten/iab-brain-drain-data.aspx.

Clark, John Maurice. 1923. “Overhead Costs in the Modern Industry.” Journal of PoliticalEconomy 31 (1): 47–64.

De la Croix, David, and Matthias Doepke. 2003. “Inequality and Growth: Why DifferentialFertility Matters.” American Economic Review 93 (4): 1091–113.

Ehrlich, Isaac. 1990. “The Problem of Development: Introduction.” Journal of Political Economy98 (5): S1–S11.

Ehrlich, Isaac, Adam Cook, and Yong Yin. 2018. “What Accounts for the US Ascendancy toEconomic Superpower by the Early Twentieth Century? The Morrill Act-Human CapitalHypothesis.” Journal of Human Capital 12 (2): 233–81.

Ehrlich, Isaac, and Jinyoung Kim. 2005. “Endogenous Fertility, Mortality and Economic Growth:Can a Malthusian Framework Account for the Conflicting Historical Trends in Population?”Journal of Asian Economics 16 (5): 789–806.

______. 2007a. “Social Security and Demographic Trends: Theory and Evidence from theInternational Experience.” Review of Economic Dynamics 10 (1): 55–77.

______. 2007b. “The Evolution of Income and Fertility Inequalities over the Course of EconomicDevelopment: A Human Capital Perspective.” Journal of Human Capital 1 (1): 137–74.

______. 2015. “Immigration, Human Capital Formation and Endogenous Growth.” Journal ofHuman Capital 9 (4): 518–63.

Ehrlich, Isaac, and Francis T. Lui. 1991. “Intergenerational Trade, Longevity, and EconomicGrowth.” Journal of Political Economy 99 (5): 1029–59.

______. 1998. “Social Security, the Family, and Economic Growth.” Economic Inquiry 36 (3):390–409.

Ehrlich, Isaac, and Kevin M. Murphy. 2007. “Why Does Human Capital Need a Journal?” Journalof Human Capital 1 (1): 1–7.


Ehrlich, Isaac, and Yun Pei. 2019. “Endogenous Immigration, Human and Physical CapitalFormation, and the Immigration Surplus.” Working paper.

Ehrlich, Isaac, and Yong Yin. 2013. “Equilibrium Health Spending and Population Aging in aModel of Endogenous Growth: Will the GDP Share of Health Spending Keep Rising?”Journal of Human Capital 7 (4): 411–47.

Kuznets, Simon. 1955. “Economic Growth and Income Inequality.” American Economic Review45 (1): 1–28.

______. 1963. “Quantitative Aspect of the Economic Growth of Nations: VIII. Distribution ofIncome by Size.” Economic Development and Cultural Change 11 (2, Part 2): 1–80.

Lucas, Robert Jr. 1988. “On the Mechanics of Economic Development.” Journal of MonetaryEconomics 22 (1): 3–42.

National Academies of Sciences, Engineering, and Medicine (NAS). 2017. The Economic andFiscal Consequences of Immigration. Washington, DC: The National Academies Press.

Romer, Paul M. 1990. “Endogenous Technological Change.” Journal of Political Economy 98 (5,Part 2): S71–S102.

Schiff, Maurice, and Mirja C. Sjoblom. 2011. “World-Panel Data on International Migration1975–2000.” Washington, DC: World Bank.

Tamura, Robert. 1991. “Income Convergence in an Endogenous Growth Model.” Journal ofPolitical Economy 99 (3): 522–40.

United Nations. 2017. “Trends and Drivers of International Migration in Asia and the Pacific.”https://www.unescap.org/sites/default/files/GCMPREP_1E.pdf.

Zhong, Jianguo. 1998. “Income Growth and Income Inequality: Theory and Evidence.” PhDdissertation, State University of New York at Buffalo.

Asian Development ReviewThe Asian Development Review is a journal of the Asian Development Bank and the Asian Development Bank Institute. ADR publishes research on economic and development issues relevant to the Asia and Pacific region.

ADR is Open Access. All content is freely available online from the MIT Press. Print copies of each issue are produced in limited quantities for subscribers and are made available for sale as single copies.

To see current pricing for subscriptions and single issues, please refer to the inside back cover of this issue or visit mitpressjournals.org/adr. You may also contact our Journals Customer Service group to place orders by phone or fax.

Journals Customer ServiceMIT Press JournalsOne Rogers StreetCambridge, MA 02142-1209Tel: 617-253-2889US/Canada: 1-800-207-8354Fax: 617-577-1545

Single copies of issues 29:2 or older can be purchased directly from ADB.

Publishing TeamAsian Development BankFax: +63 2 636 [email protected]

mitpressjournals.org/adr

ADEV3401-Ad.indd 207 2/16/17 10:20 PM

The Asian Development Review (ISSN 0116-1105, E-ISSN 1996-7241) is published twice a year (March, September) by the MIT Press, One Rogers Street, Cambridge MA 02142-1209, USA. A free electronic full-text version of the Asian Development Review is available from the MIT Press.

Vol. 37 No. 2 © 2020, Asian Development Bank and Asian Development Bank Institute Publication Stock No. JRN200232-2

The Asian Development Review is available under the Creative Commons Attribution 3.0 (CC BY 3.0) license: https://creativecommons.org/licenses/by/3.0. By using the content of this publication, you agree to be bound by the terms of said license. This CC license does not apply to non-ADB and ADBI copy right materials in this publication. If the material is attributed to another source, please contact the copyright owner or publisher of that source for permission to reproduce it. ADB, ADBI, and MIT Press cannot be held liable for any claims that arise as a result of your use of the material. Please address inquiries to the Permissions Department of MIT Press Journals at [email protected].

The MIT Press colophon is registered in the U.S. Patent and Trademark Offi ce.

Subscription InformationSubscription and address changes should be addressed to MIT Press, Journals Customer Service, One Rogers Street, Cambridge MA 02142-1209; phone 617-253-2889, US/Canada 800-207-8354; fax 617-577-1545; e-mail: [email protected].

Subscriptions begin with the fi rst issue of the volume. Subscription rates are: International/DMCs* $22.00, International/non-DMCs* $40.00, United States $40.00. Canadians add 5% GST. Outside the US and Canada add $18.00 for postage and handling. Single Issues: International/DMCs* $12.00, International/non-DMCs* $20.00, United States $20.00. Canadians add 5% GST. Outside the US and Canada add $9.00 per issue for postage and handling.

*Purchases of the journal from within the Developing Member Countries (DMCs) of the Asian DevelopmentBank (ADB) can be made at a reduced rate.

Claims for missing issues will be honored free of charge if made within three months of the publication date of the issue. Claims may be submitted to: [email protected]. Prices are subject to change without notice. Send address changes to Asian Development Review, MIT Press Journals, One Rogers Street, Cambridge, MA 02142-1209.

Advertising and Mailing List RentalsTo advertise within the pages of Asian Development Review, or on the ADR homepage of the MIT Press Journals website, or to rent the subscriber mailing list, contact the Marketing Department, MIT Press Journals, One Rogers Street, Cambridge, MA 02142-1209; phone: 617-253-2866; fax: 617-253-1709; e-mail: [email protected].


Asian Development Review

The Asian Development Review (ADR) is a professional journal that publishes research on development issues relevant to the countries of the Asia and Pacific region, specifically members of the Asian Development Bank. The ADR publishes high-quality empirical papers, survey articles, historical analyses, and policy-oriented work. The ADR bridges theoretical work and rigorous empirical studies that advance the understanding of Asia’s development; and it is open to discussions of alternative perspectives on all aspects of development, including globalization, inequality, structural transformation, and poverty. The ADR aims to disseminate the best research, with contributions from scholars in all fields, with the objective of providing the knowledge necessary for designing, implementing, and sustaining effective development policies. Its intended audience comprises a worldwide readership of economists and other social scientists.


Asian D

evelopm

ent Review

Volume 37 N

umb

er 22020


asian development review...of the philippines); and martin alexander cruz (university of the...

Documents