Collective Reputation in Trade:

Evidence from the Chinese Dairy Industry ∗

Jie Bai, Ludovica Gazze, Yukun Wang

October 24, 2017

Abstract

Understanding how reputation spreads within an industry or a geo-graphic area is important for informing development and trade policy.Quality shocks about one firm’s products could affect the demand for re-lated products from the same origin country. We study this issue in thecontext of a large-scale scandal that affected the Chinese dairy industry in2008. We combine firm-product level Chinese Customs data with officialquality inspection and news data collected from internet sources. Using adifference-in-difference framework, we find a large aggregate impact of thescandal on the entire dairy industry–exports plummeted by 68% followingthe scandal, as well as a sizable spillover effect, about four fifths of the to-tal effect, on non-directly involved firms. Next, we leverage the rich microdata to separately identify the within-firm and across-firm spillovers. Ouranalysis suggests that contaminated firms saw a drop of 87.8% in exportrevenue after the scandal while the spillover on non-inspected firms selling acontaminated product is about three quarters of the direct effect. Notably,firms deemed innocent by formal inspections do not appear to be faringany better than non-inspected firms. This finding highlights the challengesof government actions in helping to signal firms’ quality and restore trust.Finally, we investigate various channels that could mediate these collectivereputational forces, including information accuracy, supply chain structure,individual reputation and third-party regulations.

∗Contact: Bai: jie bai@hks.harvard.edu; Ludovica Gazze: lgazze@uchicago.edu; Yukun

Wang: wang yukun@foxmail.com. We thank Abhijit Banerjee, Chris Blattman, Oeindrila

Dube, Rocco Machiavello, Nina Pavnick, Nancy Qian, Daniel Xu and several seminar and

conference participants for helpful comments.

1 Introduction

Quality shocks about one firm’s products can affect the demand for related products

from the same origin country, implying an important externality. Indeed, individual

firms do not fully internalize the costs on their industrial counterparts.1 When an

incident spoils the collective reputation, it can be hard for a single firm to break

away from the low trust equilibrium and new firms are “endowed” with the damaged

reputation. Such collective reputational forces are especially relevant in the context of

international trade where a longer supply chain makes it difficult to trace products to

a particular source. Moreover, these forces may be particularly relevant for developing

countries whose firms are mostly positioned at the lower end of the value-added chain

and export mainly non-branded products. Rising safety and quality concerns regarding

goods from developing countries in recent years could act as an important barrier that

hinders firms from moving up the quality ladder and penetrating the high-end markets.2

Therefore, understanding how reputation spreads within an industry or a geographic

area is important for informing development and trade policy.

In this paper, we investigate the collective reputational forces at play in the context

of a large-scale quality scandal that affected the Chinese dairy industry in 2008. Similar

to many industries in developing countries and emerging markets, the Chinese dairy

industry was dominated by a large number of small and non-established players which

exhibited rapid growth prior to the scandal. Using administrative data on quality

inspections conducted by the Chinese government following the scandal, we identify

the firms and the products at each firm that failed the inspections (contaminated firm-

product pairs) and those that passed them (which we call innocent firm-product pairs).

We merge the official inspection lists with rich firm-product level Chinese Customs

data and firm-level Manufacturing Survey data to examine both the direct effects of

the scandal on contaminated firm-product pairs, as well as the effects on non-inspected

products at contaminate firms (within-firm spillovers) and the effects on non-inspected

firms selling contaminated products (across-firm spillovers).

1The theory of collective reputation is formulated in Tirole (1996).2A list of historical and contemporary incidents on food contaminations can be found at

https://en.wikipedia.org/wiki/List of food contamination incidents#2001 to present. Some of the re-cent prominent cases include the Brazilian meat scandal in June 2017 (see the Economist arti-cle on the incidence: https://www.economist.com/news/business/21719416-chile-china-and-eu-have-banned-some-or-all-countrys-meat-meat-scandal-brazil), and the Chinese dairy scandal in 2008.

1

We begin by estimating the impact of the scandal on the export performance of the

entire Chinese dairy sector. Using a difference-in-differences framework at the industry

level, we find that the average value of dairy exports plummeted by 68% following

the scandal and had not recovered after five years. This estimate captures both the

direct impact on contaminated firm-products and the spillovers on non-contaminated

firms or products. By excluding firm-products that are identified as contaminated by

government inspections from our DD analysis, we estimate that the spillover effect of

the scandal on non-contaminated firm-products led to a decrease in exports of 57%,

about four fifths of the total effect of the scandal on the dairy sector. Our estimates

are robust to various empirical specifications that relax the classic DD assumption

and allow for unobserved interactions between time-varying factors and industry fixed

effects using interactive fixed effects (IFE) and synthetic control designs.

Next, we leverage our detailed firm-product level data on exports and inspection

outcomes to investigate how the spillover effects of the scandal are distributed within

the Chinese dairy sector. Our analysis suggests that contaminated firms saw a drop

of 87.8% in export revenue after the scandal relative to the national trend and the

firms’ average performance. These firms are also 14.7% less likely to export follow-

ing the scandal. Furthermore, we estimate that non-contaminated firms experience a

decrease in export value as a result of the scandal that is about half the size of the

effect for directly involved firms. Overall, these findings point to large within-firm and

across-firm spillovers in export performance. However, we were surprised to find that

firms deemed innocent by formal inspections do not appear to be faring any better

than non-inspected firms. In terms of domestic performance, innocent firms in fact

performed worse than non-inspected firms. These findings highlight the potentially

counterproductive role of government actions in helping firms to signal quality and

restore trust.

Finally, we discuss potential mechanisms that may underlie the spillover effects.

On the demand side, consumers might use heuristics that group firms together based

on industry and/or source location. These heuristics will be based on the consumers’

information set and will depend, for instance, on the supply chain structure. On

the supply side, a more established individual reputation could potentially mitigate

the impact of a shock to the industry collective reputation and shield firms from the

collective reputation damage. Finally, government intervention that directly affects all

2

firms from the same origin-industry might also cause spillovers. Clearly, these channels

may interact with one another. Our analysis aims to investigate the extent to which

each of these appears to play a role in the aftermath of the scandal.

First, to study the information channel, we construct measures of consumers’ knowl-

edge about the scandal across different export destinations, using Google Trends Search

indices for phrases that reflect a more or less accurate understanding of the parties di-

rectly involved. We find smaller spillover effects in export destinations where people

exhibit more accurate search behavior. Second, to examine the role of supply chain

structure, we use the transaction level trade information to identify firms’ major source

location. We find no evidence that non-inspected firms sourcing from contaminated

provinces suffer from larger spillovers than other firms, suggesting that market compe-

tition forces might countervail the reputational spillovers. Third, new firms appear to

be more vulnerable to the collective reputation shock compared to more experienced

firms, suggesting that individual reputation mitigates collective reputation forces. Fi-

nally, blanket-style third-party regulations cannot fully explain the estimated spillover

effects–we find an equally large and significant spillover effect even when excluding

countries that imposed explicit regulatory hurdles on Chinese dairy imports.

Our paper highlights the importance of understanding collective reputational forces

in the context of international trade. The most closely related work is Macchiavello

(2010) which examines the empirical relevance of reputation for firms that enter into

new export markets in the context of the Chilean wine industry. The paper points

to the importance of buyers’ beliefs about the industry-country pair and conjectures

that such beliefs may derive from buyers’ experiences with early entrants from the

same country. Our results explicitly identify this important externality, and address

the potential role and challenges of government actions in helping to certify quality

and (re)establish trust.

A growing empirical literature studies firm reputation and quality provision in mar-

kets with information frictions (Banerjee and Duflo, 2000; Jin and Leslie, 2009; Mac-

chiavello, 2010; List, 2006; Bjorkman-Nyqvist, Svensson, and Yanagizawa-Drott, 2013;

Macchiavello and Morjaria, 2015). Prior studies have also shown that information

frictions play an important role in international trade (Allen, 2014; Macchiavello and

Morjaria, 2015; Startz, 2017). We build upon this body of research by examining the

importance of “group” reputation (be it industry or country) in international trade.

3

Our results demonstrate that these group reputational forces can have important im-

plications on a country’s trade patterns, and may act as an underlying source of a

country’s comparative advantage.

Finally, the study relates to the broad literature on firm performance and quality

upgrading in development and trade.3 Previous studies have examined: (1) supply side

constraints, including credit access, lack of quality inputs, and managerial constraints

(e.g., De Mel, McKenzie, and Woodruff (2008); Harrison and Rodrıguez-Clare (2009);

Kugler and Verhoogen (2012); Banerjee (2013); Bloom, Eifert, Mahajan, McKenzie,

and Roberts (2013)), and (2) demand side factors, including access to high-income

markets (e.g., Verhoogen (2008); Park, Yang, Shi, and Jiang (2010); Manova and Zhang

(2012); Atkin, Khandelwal, and Osman (2017)). This study highlights information

problems and low collective reputation as another potential barrier.4

The remainder of the paper is organized as follows. Section 2 provides background

information on the 2008 Chinese dairy scandal and Section 3 describes the data. Section

4 presents results on the aggregate impact of the scandal on the dairy industry and the

overall spillover effect. Section 5 investigates heterogeneity in spillover effects across

different firms and products within the dairy industry. Section 6 examines mechanisms

that could mediate the observed spillover effects. Section 7 concludes.

2 Background on the 2008 Chinese Dairy Scandal

The Chinese dairy industry exhibited fast growth during the early 2000, both in terms

of domestic production and exports. Appendix A Figure 1 shows the domestic sales

revenue growth for the dairy industry from 2005 to 2009, compared to the non-dairy

food industry. The average annual growth rate is as high as 26.8% prior to 2008. In

terms of export performance, Figure 1 shows that the annual export value increased

over 3 times from 2000 to 2007. The number of exporting firms increased from 150 in

3See De Loecker and Goldberg (2014) for a comprehensive review of the empirical literature.4This paper also speaks to the literature on quality scandals and product recalls. Most of the

previous studies have either relied on lab experiments to examine consumer reactions to hypotheticalproduct scandals (Ahluwalia, Burnkrant, and Unnava, 2000; Dawar and Pillutla, 2000)), or focusedprimarily on stock market outcomes using an event-study approach (Davidson and Worrell, 1992;Marcus, Swidler, and Zivney, 1987; Zhao, Lee, Ng, and Flynn, 2009)). Furthermore, most studiesfocus on losses in own sales and stock market price (Van Heerde, Helsen, and Dekimpe, 2007), ratherthan across-firm and products spillovers, and few study this topic in the context of international trade.

4

2000 to 300 in 2007.5

Nonetheless, dairy exports still constituted a small share of the entire dairy produc-

tion in China, and accounted for only $300 million of the country’s $1.2 trillion exports

revenue in 2007. Similar to many other industries in developing countries and emerging

markets, it was an industry dominated by a large number of small and non-established

players, and demonstrated high growth potential.

Over the past decade, an increasing number of quality and safety issues regarding

food products originated from China arose. One of the most widely known incidents

is the outbreak of the contaminated baby milk formula in September 2008, hereinafter

referred to as “the scandal”. The incident led to 4 infant deaths, 51,900 children

hospitalized and 700 tons of milk powder recalled nationwide. The infant formula was

found to be illegally adulterated with the industrial chemical called “melamine”, which

was added to mimic higher protein content in the milk. Melamine is commonly used in

the manufacturing of plastics and has been linked to an increased risk of kidney stones

(Hau, Kwan, and Li, 2009).

The supplier of the product immediately linked to the outbreak, Sanlu Group, one

of the major players in the dairy industry in China, was quickly shut down. How-

ever, further investigations demonstrated that the adulteration in fact stemmed from

malpractices of some upstream milk producers, raising the suspicion that more dairy

firms downstream could have been affected.6 However, some downstream dairy firms

enforced higher quality control than others. This discovery led to three big rounds of

government inspections: the first round targeted firms producing infant formula–175

firms were inspected and 22 were found to have traces of melamine in their products.

The subsequent 2 rounds of inspections targeted producers of milk powder and liquid

milk respectively, and covered most dairy producers in China. We describe the three

rounds of inspections in greater detail in Section 3.3. These inspections uncovered con-

tamination in more dairy and dairy-related products, including yogurt, milk, cheese,

baby food with milk traces, chocolate, and cake. Product recalls were immediately

5Figure 1 shows a spike in export growth between 2006 and 2007–the total value of exports increasedby about 138%. This spike is mainly driven by new firms entering into the exporting markets.Specifically, 30 firms contributed more than 80% of the growth spike; over 50% of the spike is drivenby a single product–milk powder; finally, over 50% of the spike is driven by exports to 6 destinations,namely Thailand, Germany, Bangladesh, Taiwan, UAE and Hong Kong.

6According to subsequent investigation reports, these malpractices were in fact “open secrets” inthe industry. See https://www.wsj.com/articles/SB122567367498791713

5

issued. By the end of 2008, the Chinese government issued an official statement that

the issue was addressed and proper measures have been put in place to ensure the

safety of the dairy products on the market.7

Despite the reassuring statement, the scandal triggered widespread fears over food

safety in China. Thousands of Chinese dairy-related products were pulled from super-

market shelves across the world. Some countries strenghtened inspections for Chinese

imports, while others issued explicit import bans for products containing Chinese dairy

ingredients. For instance, the EU authorities stipulated tests for Chinese products con-

taining more than 15 percent of milk powder and announced a ban on all products for

children coming from China that contain milk; the US Food and Drug Administration

(FDA) restricted imports of all Chinese food products containing milk; India imposed

a temporary ban on import of milk and all milk-related products from China and has

extended this ban till this year. Our news search has identified 45 countries (out of

163) that imposed temporary bans on certain Chinese dairy products (see Appendix

B Table 1).

The scandal has had a long-lasting impact on the Chinese dairy industry. The

General Administration of Quality Supervision, Inspection and Quarantine (GAQSIQ)

stopped issuing national exemption status to domestic food producers8 and tightened

inspections on domestically produced food products. Dairy firms also tightened their

standards for purchased raw milk and some started to build their own upstream milk

farms or vertically integrate in order to better monitor and control quality. Despite

these actions to rectify the situation, the scandal came as a devastating blow to the

Chinese dairy industry and almost entirely wiped out the country’s dairy exports in

the ensuing years. As shown in Figure 1, Chinese dairy exports sharply collapsed after

2008 and did not recover untill the end of 2013, the last year in our data. At the

same time, Chinese dairy imports rose rapidly after the scandal (Appendix A Figure

2), suggesting that domestic consumers switched to foreign dairy products in response

to safety concerns around domestic brands.

7http://www.telegraph.co.uk/news/worldnews/3079146/China-claims-tainted-milk-scandal-is-over.html

8This was previously known as the “inspection-free” program, which gave exemption status toqualified food producers and waived various quality inspections for them.

6

3 Data

For this study, we assemble data from three micro-level data sources: the Chinese

Customs Database, the Chinese Manufacturing Survey, and the list of inspections

conducted by the Chinese government following the scandal. We merge together the

three datasets using detailed firm and product identifiers.

3.1 Chinese Customs Database (2000-2013)

The Chinese Customs Database provides transaction level trade flows information on

the universe of China’s exports and imports over the time period. The data is collected

and made available by the Chinese Customs Office. For the analysis in this study, we

focus on exports. For each transaction, we observe the exporting firm ID, firm name,

trade type, value and quantity of the exports, the HS eight-digit code which we use to

define products, the region or city in China where the product is exported from, the

customs office where the transaction is processed and its final destination. For each

firm, the data also provides information on its ownership type and location (the first

five digits of the firm ID). We compute unit prices for exported products by dividing

value of exports by quantity. While the data is available at daily frequency, for our

analysis we aggregate the data to firm-product-year level. Appendix A Figure 3 plots

China’s export growth by year. Exports grew rapidly during the early 2000s following

China’s entry into the World Trade Organization (WTO).

We define “dairy industry” using the HS eight-digit product information. Most of

the dairy products fall under the HS two-digit code 04, while infant dairy products fall

under 19 and milk protein products extracted from raw milk fall under 35. Appendix

B Table 2 provides the full list of the HS eight-digit codes and descriptions for dairy

products.

3.2 Chinese Manufacturing Survey (2005-2009)

The Chinese Manufacturing Survey data are compiled from annual surveys conducted

by the National Bureau of Statistics (NBS), and include all industrial firms that are

identified as being either state-owned or non-state firms with sales revenue above 5

million RMB. As described in previous studies (e.g., Brandt, Van Biesebroeck, and

7

Zhang (2012)), even though a large number of small to medium industrial firms (80%)

are excluded from the sample, they account for only a small fraction of the total

economic and export activities in China. In particular, the excluded firms employ

28.8% of the industrial workforce, but only produce 9.3% of the total output and

generate 2.5% of the export revenue. The analysis in this paper focuses on the dairy

industry within the manufacturing sector. For each firm-year observation, we observe

basic production and financial information, including firms’ 4-digit industry ID, years

of operation, firm location, total sales revenue, employment, and export revenue. We

identify the dairy and food industry according the observed industry ID.

3.3 Government Inspection Lists

The Chinese government implemented three rounds of inspections after the scandal

broke out in late 2008. The three rounds of inspections focused respectively on produc-

ers of infant formula, milk powder, and liquid milk. In the first round, the government

inspected all 109 infant formula producers in China and 22 were found to be contam-

inated. The second round inspected 154 randomly sampled milk powder producers

(together making up over 70% of the market share) out of 290 producers nationwide

and 20 were found to be contaminated. The third round targeted 466 more established

dairy brands with large market shares, and found 9 firms of 3 major brands to be

contaminated.

For each round of inspection, the Chinese government released the full list of firm

names, each firm’s products being inspected and the corresponding batch codes. We

obtained the inspection lists at the firm-product level from the official GAQSIQ web-

site. A news search through Lexis Nexis found that most of the reported cases of

contamination were covered in the Chinese official inspection lists.

We merge the firm-product level inspection lists and outcomes with the Customs

data using firm names and product information, and with the Manufacturing Survey

data using firm names only (since we do not observe product information in the Manu-

facturing Survey). Based on the firm-product level inspection information, we classify

firms in one of the following categories: contaminated, innocent, and non-inspected

firms. The first is defined as firms with at least one product found to be contaminated

during one or more rounds of the inspections. The second is defined as firms whose

8

products passed all of the tests for their inspected products. The third group includes

firms that were never inspected.

Using these definitions, in the Customs data we identify 49 contaminated firm-

product pairs in 41 contaminated firms, and 31 innocent firm-product pairs in 45

innocent firms. In the Manufacturing Survey data, we identify 65 contaminated and

369 innocent firms.

3.4 Summary Statistics

Figure 2 plots the number of exporting products and destinations for contaminated

and non-contaminated firms from 2000 to 2007 (for firms exporting in multiple years,

we calculate the median value across the 8 years). The graphs show that most firms

exported a single product and to a single destination prior to the scandal.Appendix

A Figure 4 shows similar patterns after the scanda. We discuss how these patterns

determine the variation in the data we exploit in the firm-product level analysis in

greater detail in Section 5.

Table 1 presents firm-level baseline (2000-2007) summary statistics for the dairy

industry. It reports the number of non-missing observations, mean and standard de-

viation for a set of key performance measures for exports (Panel A), and employment

and total sales revenue (Panel B). Column 7 calculates the difference in means between

the contaminated and non-inspected firms (Column 2 and 6) and Column 8 reports

the p-value of the difference. Panel A shows that on average, contaminated firms

have larger exporting revenue and are more experienced compared to innocent firms;

however, they are not systematically different from the non-inspected firms. Panel B

reinforces that contaminated firms are larger in terms both of employment and of sales

revenue than innocent firms and non-inspected firms. This is consistent with the fact

that the Chinese government targeted larger firms in the third round inspection.9

Table 2 presents product-level baseline (2000-2007) summary statistics for the same

export performance measures. Panel A shows that on average contaminated products

are larger in terms of exporting revenue and have been exported for more years compar-

ing to uncontaminated products. Contaminated products also appear to be exported

less to the member countries of the Organization for Economic Cooperation and Devel-

9Appendix A Table 1 presents the same baseline summary statistics for non-dairy food industries.

9

opment (OECD) although the difference is not significant. Panel B presents the same

summary statistics for non-dairy food products.

4 Industry Level Analysis

This section estimates the impact of the scandal on the export performance of the entire

Chinese dairy sector. Specifically, we compare changes in the value of exports at the

HS two-digit level between the dairy sector and control industries using a difference-

in-differences (DD) design as well as extentions that relax the classic DD assumptions.

Section 4.1 discusses our preferred empirical specification as well as threats to the

validity of the DD assumptions and additional tests we perform that relax these as-

sumptions. Section 4.2 estimates that the scandal decreased the value of exports for the

dairy sector by 68% over the course of five years. Sections 4.3 and 4.4 show that these

results are robust to allowing for unobserved interactions between time-varying factors

and industry fixed effects (FE) using interactive fixed effects (IFE) and synthetic con-

trol designs, as discussed more in detail in Section 4.1. Finally, Section 4.5 investigates

the extent to which the scandal affected different dairy products, as measured at the

six-digit level of the HS code, differently.

4.1 Empirical Specification

This section aims to estimate the impact of the scandal on the export performance

of the Chinese dairy sector. Equation (1) presents our baseline specification, which

compares the value of exports of dairy products, the treated industry, to the value of

exports of other two-digit level industries before and after 2008, the year of the scandal,

in a DD design.

Yjt = βdairyDairyj × Postt + γj + δt + πXjt + εjt (1)

Yjt is the natural logarithm of the value of exports for industry j in year t, Dairyj

is an indicator for the dairy industry, and Xjt include time-varying controls at the

industry level, such as the value share of the industry exported to different continents

at baseline, i.e. from the year 2000 to 2007, interacted with year indicators. We

introduce these covariates to control for differential trends in destination countries

10

that might affect different industries differently. In our preferred specification, we limit

the set of control industries to exclude non-dairy food industries. In fact, non-dairy

food industries might be affected by the scandal, too, if foreign consumers update their

priors on the quality of Chinese food products in general following the scandal. This

leaves us with 79 control industries. We cluster standard errors at the industry level

allowing for arbitrary correlation in error terms across time for a given industry, unless

otherwise noted.

The internal validity of the DD design rests on the assumptions that the treated and

control industries would be on parallel trends absent the treatment, i.e. the scandal.

In our context, we may worry that the parallel trends assumption may not hold for two

reasons: the rampant growth of the Chinese dairy industry prior to the scandal and the

global financial crisis that occurred at the same time as the scandal. First, as discussed

in Section 2 the dairy industry saw a rapid growth in exports prior to the scandal, a

growth that does not appear to be paralleled by other industries. In Section 5.4, we

show that the results from our baseline specification hold when we exclude exports by

the firms that were growing most rapidly prior to the scandal from the industry-wide

figures. Second, we may worry that the global financial crisis of 2008 affected different

industries differently. If that was the case, the parallel trends assumption would be

violated, and we could erroneously attribute to the scandal an export decline in dairy

products that might in fact be due to the crisis.

To assuage these concerns, we show that our estimates of the impact of the crisis

are robust to several specifications that relax the parallel trends assumption. First, we

perform robustness checks on our baseline specification in Equation (1) that include

a vector of industry-specific linear time trends in the vector Xjt. Second, Section 4.3

allows for higher-dimensional unobserved interactions between time-varying factors and

industry indicators using the interactive fixed effects design discussed in Gobillon and

Magnac (2016). Specifically, we use least squares minimizaton to estimate Equation

(2) where δt is an L×1 vector of time factors and γj is an L×1 vector of factor loadings

(Bai, 2009). The IFE models assumes that the interaction term δ′t ∗ γj fully describes

the unobserved heterogeneity and that the dimension of the time factors and factor

loadings, L is known. Our estimates are robust to different values of L.

Yjt = βdairyDairyj × Postt + δ′

t ∗ γj + πXjt + εjt (2)

11

In the case of a single treated unit, synthetic control methods can successfully

construct a vector of weights such that a weighted combination of control units closely

matches the time-series of the outcome variable for the treated unit in the pre-period

(Abadie and Gardeazabal, 2003; Abadie, Diamond, and Hainmueller, 2010). Therefore,

our third robustness check in Section 4.4 estimates the impact of the scandal as the

difference between the value of exports in the dairy industry and the synthetic unit

before and after the scandal, as given by Equation (3).

β =2013∑

t=2009

(Y1t −

N∑i=2

w∗i (V

∗)Yit

)−

2008∑t=2000

(Y1t −

N∑i=2

w∗i (V

∗)Yit

)(3)

We denote the dairy industry with index 1, w∗i are the optimal weights on control

units, and V ∗ minimizes the distance between the predicted pre-treatment outcomes

of the treated and synthetic control unit, with predictions based on an arbitrary set

of baseline covariates. Specifically, we use an indicator for whether an industry was

exporting in a given year and the value share of the industry exported to different

continents as the baseline covariates to predict outcomes.

Gobillon and Magnac (2016) discuss how the IFE method generalizes the synthetic

control design when the matching variables (i.e. factor loadings and exogenous co-

variates) of the treated unit do not belong to the convexified support of the matching

variables of the control units, which they call the extrapolation case. Given the unique

growth path of the Chinese dairy industry prior to the scandal, we might very well

find ourselves in the extrapolation case. Nonetheless, Section 4.4 reassuringly confirms

both our baseline and IFE estimates. Given this battery of robustness checks, we are

confident that our DD estimates capture the true effect of the scandal on the export

performance of the Chinese dairy industry, and we report these DD estimates as our

preferred ones.

4.2 Results: Difference-in-Difference

Table 3 presents estimates of Equation (1). The baseline specification in Column 1

includes only year and industry fixed effects, and estimates a decline in the value of

Chinese dairy exports following the scandal of 65.5% (= (e−1.065−1)×100%). Columns

3 and 4 build on this specification by adding time-varying controls for the value share of

12

the industry exported to different continents at baseline interacted with year indicators

and industry-specific linear trends, respectively.10 While the effect of the scandal is

estimated more imprecisely in Column 3, our preferred estimate in Column 4 is not

statistically distinguishable from the baseline estimate in Column 1. Therefore, we

conclude that the value of dairy exports plummeted by 68% following the scandal.

Finally, Column 2 expands our sample to include non-dairy food industries. The

coefficient on the interaction between the indicator for food industry and the indicator

for the post-scandal period suggests that the scandal did not affect these non-dairy

food industries in an economically or statistically significant way. Section 5 analyzes

the patterns of spillover effects of the scandal in greater details.

So far, we have estimated the impact of the scandal on the whole Chinese dairy

industry. Estimates in Table 3 capture both the direct impact on contaminated firm-

products, as well as the spillovers on non-contaminated firms or products. We can

gauge the extent to which the direct effect dominates the indirect effect by performing

our analysis on the dairy industry excluding contaminated firm-products. Table 4

shows that the indirect effect on this non-contaminated sample appears to be slightly

smaller than the total effect, although this difference is only statistically significant

in our preferred specification in Column 4 that controls for product-year fixed effects.

Specifically, we estimate a spillover effect of the scandal on non-contaminated firm-

products leading to a decrease in exports of 57%, about four fifths of the total effect

of the scandal on the dairy sector.

4.3 Results: Interactive Fixed Effects

As discussed in Section 4.1, this section investigates whether the results obtained in the

DD framework are robust to allowing for more flexible time variation at the industry

level as specified in Equation (2). We find that the DD estimates are virtually identical

to the equivalent IFE models of dimension one. For example, the baseline specification

without controls in Column 1 of Table 5 estimates that the scandal decreased the value

of Chinese dairy exports by 65.6%. Adding controls for the value share of the industry

exported to different continents at baseline interacted with year indicators in Column 5

yields an estimated decrease in the export value of 68.2%, very similar to our preferred

10Column 3 drops one industry, with two-digit level HS code equal to 77, because it has no exportsduring our baseline period.

13

estimate in Column 4 of Table 3 which controlled for industry-specific linear trend in

the DD setting. Similarly, Column 4 of Table 5 can be compared to Column 2 of Table

3 to confirm that the scandal did not significantly affect non-dairy food exports.

Finally, Columns 2 and 3 of Table 5 allow for increasingly multi-dimensional in-

teractions between time factors and industry factor loadings. These models estimate

an impact of the scandal on dairy exports that is larger in magnitude than the one

we estimate in the classic DD model. Because the dimension of these IFE models is

set somewhat arbitrarily or assumed to be known (Gobillon and Magnac, 2016), our

preferred estimate remains the more conservative one in Column 4 of Table 3, of a 68%

decrease in the value of dairy exports following the scandal.

Table 6 confirm the patterns shown in Table 4: the spillover effect of the scandal

leads to a decrease in exports of non-contaminated firm-products that is smaller than

the total effect of the scandal on the dairy sector. Specifically, using IFE we estimate

spillover effects ranging between a decrease in exports of 50 to 73%.

4.4 Results: Synthetic Control

The solid line in Figure 3 plots the natural logarithm of value of exports for the Chinese

dairy industry over our sample period. As discussed in Section 2, we observe that

Chinese dairy exports grow substantially prior to the scandal. The dashed line plots

the natural logarithm of the value of exports for the synthetic control unit, created using

industry and covariate weights specified in Appendix A Tables 2 and 3 respectively.11

Reassuringly, the synthetic control unit mimics the growth of the dairy sector quite

closely prior to 2009, the first year after the scandal. Starting in 2009, while the

synthetic control unit continues to grow through 2011, the dairy industry experiences

a drop in exports that persists through 2010, stabilizing around value levels observed

in 2004-2006. Averaging the difference between the logarithm of the value of exports

of the dairy industry and the synthetic control unit before and after the scandal, as

described in Equation (3), we obtain an estimated impact of the scandal of −71%, in

line with our DD estimate.

11These tables show that the weights selected by the data-driven algorithm in the synthetic controlmethodology appear to be somewhat disconnected from economic theory, leaving doubts as to theinterpretation of the results.

14

Abadie, Diamond, and Hainmueller (2010) prove that the bias of the synthetic

control estimator can be bounded by a function that goes to zero as the number of

pretreatment periods increases. Intuitively, a longer baseline allows for a more precise

calibration of the weights, which improves the match between the treated and control

outcomes. As our sample only provides eight years of pre-scandal data, we might

worry that the estimated impact of the scandal is confounded by residual unobserved

differences between the dairy industry and the synthetic control unit. To assuage this

concern, we perform inference by permuting the treatment to each unit in our donor

pool of control industries, following Abadie, Diamond, and Hainmueller (2010). Figure

4 plots the difference in the logarithm of the value of exports between each industry

and its synthetic control unit over time. We can contrast all the placebo differences (in

grey) with the difference for the dairy industry (in black), and we can observe that the

dairy industry sythetic control provides a good match in the pre-scandal period, i.e.

the pre-scandal difference lies comfortably within the placebo band. Moreover, there

are only five other industries that display a larger treatment effect in the post-scandal

period. Given that we have 76 placebo differences, the p-value on our estimate, i.e.

likelihood that of erroneously rejecting the hypothesis of a null effect of the scandal on

the dairy industry, is 6/76, or 0.08.12

Analogous to the analysis in Table 4, Figure 5 shows the spillover effect of the scan-

dal on the sample of non-contaminated dairy firm-products. Differently from the DD

and IFE estimates, however, the synthetic control methodology estimates an indirect

effect of −74%, with a point estimate that is larger than the total effect of −71%.13

Nonetheless, Figure 6 shows that there are four industries with a larger placebo dif-

ference than the dairy industry, implying a p-value on our estimate of 5/76, or 0.06.

Therefore we cannot reject the hypothesis that the indirect effect of the scandal on

non-contaminated firm-products is as large as the total effect on the entire dairy in-

dustry.

12The algorithm for the construction of the synthetic control unit for industries with HS codes 28,32, and 99 fails to converge.

13Industry and covariate weights used to select the synthetic control unit plotted in Figure 5 areshown in Appendix A Tables 4 and 5 respectively.

15

4.5 Heterogeneous Effects across Dairy Products

In this Section we investigate the extent to which the scandal affects different products

within the Chinese dairy industry differently. Specifically, we examine the export

performance of 25 dairy product categories as defined by their six-digit level HS code,

shown in Appendix B Table 2. To do so, we estimate a DD model at the six-digit

level, as described in Equation (4). The sample of control industries includes 5889 HS

six-digit industries excluding non-dairy food-related industries.

Yjt = βjIndustryj × Postt + γj + δt + εjt (4)

Figure 7 plots the βj coefficients estimated from Equation (4) ordered from the most

negative to the most positive, together with 95% confidence intervals estimated from

robust standard errors. While some of these coefficients are estimated imprecisely, the

figure shows that the scandal appears to have had heterogeneous effects on different

dairy products, damaging the exporting performance of some, while boosting the ex-

porting performance of others. One potential explanation is the positive countervailing

market competition effect besides the negative collective reputation effect. We discuss

this point further in Section 6.2. To provide further insights on the heterogeneous

effects of the scandal and the mechanisms through which it affects firms’ performance,

the next section studies the impact of the scandal at the firm-product level.

5 Firm-Product Level Analysis

This section investigates the impact of the scandal on export performance at the firm-

product level. We restrict the analysis to the dairy industry. We are interested in ex-

amining both the direct impact on contaminated firm-products, as well as the spillover

effects within firms and across firms in the industry. To do so, we take advantage of

the rich microdata on export activities at the firm-product level, as well as the firm-

level performance measures, and merge them with inspection outcomes as described in

16

Section 3. Our main regression specification is as follows:

Yfpt = βdirectContaminatedFirmProductfp × Postt

+ βwithin-firmContaminatedFirmf × Postt

+ βacross-firmContaminatedProductp × Postt

+ λ1InnocentFirmProductfp × Postt

+ λ2InnocentFirmf × Postt

+ γfp + δt + εfpt (5)

The dependent variable Yfpt is an outcome for firm f product p at year t, including

log (export revenue), log (export quantity), log (price), and a dummy variable for

exporting. Except for the price outcome, we first rectangularize the data at firm-

product and year level in order to capture extensive margin responses (i.e., entry and

exit into exporting). ContaminatedFirmProductfp is an indicator for firm-product

pairs directly involved in the scandal: the indicator equals 1 if a given product of a

firm was inspected and failed the quality test. ContaminatedFirmf is an indicator for

firms involved in the scandal: the indicator equals 1 if a firm was inspected and at least

one of its products failed the test. ContaminatedProductp is an indicator for products

involved in the scandal: the indicator equals 1 if at least one of the inspected firms failed

the quality test for the given product. InnocentFirmProductfp and InnocentFirmf are

defined similarly: InnocentFirmProductfp is an indicator for firm-product pairs that

were inspected and passed the quality test. InnocentFirmf is an indicator for firms

that passed the quality test for all of their inspected products.

Our preferred specification controls for firm-product (γfp) and year (δt) fixed effects,

so that we capture differential changes in performance across firm-products over time,

netting out common nationwide dairy industry time trends and different average levels

by firm-product cell. To examine the sensitivity of our estimates, we estimate two

alternative specifications exploring different identification assumptions, one controlling

for firm, product and year fixed effects, and one controlling for year fixed effects only.

Results are presented in Appendix A Tables 6 and 7. Standard errors are always two-

way clustered at the product-year and firm level allowing for arbitrary correlation in

error terms across time for a given firm and for a given product-year across firms.

17

To interpret the β and λ coefficients, note that the omitted group is non-contaminated

products from non-inspected firms. Therefore, βacross-firm identifies the impact on non-

inspected firms of selling one of the contaminated products (i.e., across-firm spillovers).

βwithin-firm identifies the overall impact on the contaminated firms (i.e., within-firm

spillovers) whereas βdirect captures the additional impact on their directly involved

products. Similarly, λ1 and λ2 capture the impact on the innocent firms and inno-

cent firm-products relative to the omitted group. Appendix A Table 8 examines the

variation in the data in each group that allows us to identify these various effects.

5.1 Results: Direct and Spillover Effects on Exports

Table 7 reports the main estimates from Equation (5). Column 1 examines the impact

on log export revenue and shows large within-firm and across-firm spillovers. Specif-

ically, the estimate for βwithin-firm is -2.1 and is significant at the 1 percent level. The

magnitude suggests that contaminated firms experienced a drop of 87.8% in export rev-

enue after the scandal relative to the national trend and the firms’ average performance.

Within contaminated firms, directly involved products are affected more–the estimated

coefficient βdirect is large (-0.637, or -46.8%) but quite inprecisely estimated, suggesting

that there may be a lot of heterogeneity among the subset of directly contaminated

firm-products. Next, in line with the industry-level analysis results in Section 4, we see

a large and negative spillover impact on firms selling one of the contaminated prod-

ucts: the estimate for βacross-firm is -1.111 (or -67%) and is significant at the 1 percent

level. Finally, the results for innocent firms and products are very mixed: while the

coefficient on InnocentFirmProduct×Post is large and positive, the overall impact on

innocent firms, relative to non-inspected firms, is negative. Both of these estimates are

statistically insignificant.

Column 2 examines log export quantity and finds similar results.14 Combining the

estimates of βacross-firm in Column 1 and 2, this decrease in quantity contributes to

97.5% of the across-firm spillovers, after taking into account entry and exit.

Column 3 examines changes in unit price. The regression is estimated on the

unbalanced panel of firm-product-year observations with positive export activity. The

estimate of βacross-firm is -0.151 (-14%) and is significant at the 5 percent level. The

14There are very few changes in unit within a HS eight-digit industry in the Customs data.

18

direct impact on contaminated firm-products is -0.391 (-32.4%) and is significant at

the 5% level, whereas the within-firm spillover βwithin-firm is positive at 0.293 (34%) with

a standard error of 0.195. The results suggest that firms dropped prices for products

that were found to be contaminated, while at the same time they increased prices in

other product lines.

Finally, Column 4 examines the impact of the scandal on the extensive margin,

and finds that contaminated firms are 14.7% less likely to export after the scandal.

The estimate for βwithin-firm is significant at 1 percent level, whereas that for βdirect is

non-distinguishable from 0. All Chinese dairy firms, regardless of whether innocent,

contaminated or not inspected, are 8.5% less likely to export contaminated products,

and the estimate of βacross-firm is significant at the 1 percent level. The results on

innocent firms and products are again inconclusive.

5.2 Interpretation: The Role of Government Inspections

Overall, the results show large within-firm and across-firm spillovers in export perfor-

mance: both the non-inspected products of contaminated firms and the contaminated

products of non-inspected firms were significantly adversely affected by the scandal.

On the other hand, the effects on innocent firms and products are highly mixed. The

inspection outcomes released by the Chinese government could be regarded as a form of

government signaling or certification. This section discusses two potential explanations

for our mixed results on innocent firms and products.

The first explanation is that the public might perceive government inspections as a

bad signal for potential quality violations. The fact that a firm went under inspection

may indicate that something was not right, even though some of the inspections were

claimed to be targeted at random (see discussions in Section 2). An alternative ex-

planation is that any firm that appears in the post-scandal news reports is associated

implicitly with the scandal stigma if people do not pay attention to the details of the

news.15

To shed light on these potential explanations, we turn to firms’ overall performance,

including both domestic sales and exports. Since the list of inspections might be more

accessible to domestic consumers (most were published in Chinese news sources), both

15Ma, Wang, and Khanna (Working Paper) discusses this “reminder effect”.

19

explanations would suggest that domestic consumers may also react negatively, if not

more, to the inspections. To test this hypothesis, we estimate the following regression

at the firm-year level using the manufacturing survey data:

Yft = b1ContaminatedFirmf × Postt + b2InnocentFirmf × Postt + γf + δt + εft (6)

The outcome variables include log total sales revenue and log total employment for firm

f in year t. We include both firm and year fixed effects and cluster standard errors at

the firm level.

Table 8 reports the results. Column 1 shows that both contaminated and innocent

firms perform worse, relative to non-inspected firms, after the scandal. Total sales

revenue dropped by 42.9% for contaminated firms and 16.7% for innocent firms, and

the estimates are significant at the 1 percent and 5 percent level respectively. The

effects on domestic sales in Column 2 are qualitatively similar but noisier, whereas the

effects on employment in Column 3 are much smaller and statistically insignificant,

suggesting that the hiring margin might be more difficult to adjust.

Overall, these findings are consistent with the two potential explanations discussed

above about the potentially counterproductive role of government signaling and certi-

fication. These results speak broadly to the importance of understanding the process

of information acquisition in thinking about collective reputational forces. We come

back to this point in Section 6 when we discuss the mechanisms.

5.3 Heterogeneous Effects on Sister Firms

Some non-inspected firms are are related to contaminated firms through their ownership

structure: we call these sister firms. For example, a popular dairy brand in China, Yili,

has a mother company and several regional subsidiaries that appear in our dta. The

mother company, Inner Mongolia Yili Industrial Group, is contaminated, while its

subsidiary Hefei Yili Dairy was never inspected. To the extent that sister firms share a

common brand name, one may expect larger reputational spillovers for these firms than

for other non-related non-inspected firms. We investigate this possibility by creating

an indicator variable for non-inspected sisters and estimating Equation (5) with this

additional interaction. In total, there are 45 such non-inspected sister firms in our

merged Customs sample.

20

The results are shown in Table 9. The coefficients for log export revenue, log

quantity and exporting dummy are all negative and the magnitudes are quite large;

however, the effects appear to vary tremendously among firms as indicated by the large

standard errors.

Table 10 examines firms’ domestic performance among sisters and non-sisters.16

Interestingly, sister firms appear to perform worse than non-sister firms in terms of do-

mestic sales. In particular, among non-inspected firms, sister firms experience a 20.3%

slower growth rate in total sales revenue compared to the non-sister firms. Similarly,

among innocent firms, sister firms performed significantly worse than non-sister firms:

a test of equality for the two coefficient estimates is rejected at the 10 percent level

(p-value=0.063). The results on domestic sales are qualitatively similar though less

precisely estimated (Column 2).

One explanation for the contrast between the domestic and export performance is

that domestic consumers have more knowledge about the individual brands and are

thus able to identify the contaminated brands better than foreign consumers. We

investigate the information channel further in Section 6.

5.4 Robustness Checks

This section examines the robustness of our main results in Section 5.1 to alternative

specifications. In our analysis, identification relies on the assumption that unobserved

firm-product-year specific shocks that affect the outcomes are uncorrelated with the

initial inspection status. In other words, absent the scandal, all firm-products would

have seen the same growth in export performance over time. In Section 3, we see

that inspected firms are on average larger than non-inspected firms. Therefore, one

concern is that firms of different sizes may lie on different growth trajectories. To

address this concern, we allow for differential time trends across firms with different

baseline characteristics, including firm’s total sales value, firm’s year of exporting and

the average exporting revenue of the HS four-digit industry. The results, reported in

Appendix A Table 9, are very similar to the results in Table 7.

A related concern is that different sub-industries (within dairy) may be on different

growth trajectories, in particular contaminated products may grow faster or slower

16Note that there are some innocent sister firms in the merged Manufacturing Survey data, andtherefore, we create additional group dummies for these firms.

21

than other products even absent the scandal, leading to biased estimates of the spillover

effects. Appendix A Table 10 allows in addition for differential time trends across sub-

industries. Reassuringly, the results are qualitatively similar to the results in Table 7.

Finally, as discussed in Section 2, a few predominant firms and export destinations

were behind the growth spike in the pre-scandal period (2006-2007). We examine to

what extent the estimated coefficients are mechanically driven by these fast-growing

parties scaling down their production and reducing exports after the scandal by drop-

ping these firms and destinations. Results are shown in Appendix A Table 11.

5.5 Persistent Effects Over Time

We examine how persistent the direct and spillover effects are by fully interacting

the firm-product group dummies with year dummies in Equation (5). Figure 8 plots

the regression coefficients with standard errors for two outcome variables: log value

of exports and exporting dummy. We see that both the within-firm and across-firm

spillover effects persist over five years after the scandal and display little sign of recovery.

6 Mechanisms

In this section, we investigate four potential mechanisms for the spillover effects doc-

umented in Section 4 and 5, namely information accuracy, supply chain structure,

individual reputation and the role of third-party or government regulations. In this

and other real world contexts, all of the channels may act jointly and interact with

one another. Rather than trying to disentangle and quantify the effects of the various

channels, the goal of our exercise here is to show whether a particular channel has bite.

6.1 Information Accuracy

A producer’s reputation, broadly speaking, is consumers’ belief that it provides quality.

Hence, the ways that consumers gather information and learn crucially matter for

the extent of reputational spillovers. We can imagine two scenarios: one in which

consumers perfectly understand the evolution of the scandal and are able to closely

keep track of the inspection outcomes, and the other in which consumers have trouble

identifying the contaminated firms and begin to worry about Chinese dairy products

22

in general as a result of the scandal. Collective reputational forces would be stronger

in the latter scenario compared to the former.

To construct a systematic measure of consumers’ information accuracy across dif-

ferent export destinations, we leverage Google Trends data. Google Trends is a public

web facility providing time series indices based on Google Search data, which capture

how often a search term is entered relative to the total search volume in a given ge-

ographical area. To compare relative popularity across search terms, each data point

of Google Trend indices is divided by the total searches of the geographical area and

time range it represents, and scales on a range of 0 to 100 for any given period. We

collect two types of search indices, a web search index and a news search index, for

different countries. For each index, we constructed the relative search intensity ratio

for the two keywords, “Sanlu” versus “2008 Chinese milk scandal”. Figure 9 displays

the relative search intensity for several countries. It shows that web users in countries

on the left panel–China, Hong Kong and New Zealand–search “Sanlu” much more than

the generic phrase, suggesting that consumers in these countries may be more informed

about the parties involved in the scandal. In comparison, information appears to be

much less specific in countries on the right panel–the US, Canada and UK.17

We collect the web and news search intensity ratio for all countries available on

Google Trends–there are 31 countries in total. We classify countries into two groups

according to the relative search intensity: “high” indicates a higher ratio and thus

higher information accuracy. Appendix B Table 3 describes the classification.

Table 12 reports the heterogeneous impact of the scandal on export performance

to destinations with high and low search intensity ratio, using the web search index.

Consistent with the discussion above, the across-firm spillover effect (the coefficient on

ContaminatedProduct×Post) is smaller (-0.256) for exports to destinations with high

information accuracy compared to exports to countries with low information accuracy

(-0.563). The p-value for testing equality of two coefficients is 0.0288. Results using the

news search index are presented in Appendix A Table 12 and the qualitative takeaways

are very similar.

The results show that information accuracy might play an important role in me-

diating the collective reputational forces. Moreover, the observed differential search

17Appendix A Figure 5 shows the search behavior across provinces in China. Hebei province, wherethe headquarter of Sanlu was located, has the highest search intensity for the keyword “Sanlu”.

23

behavior across countries can be partly due to heterogeneity in media contents in dif-

ferent countries, which can generate different information sets among local consumers.

For example, Appendix A Figure 6 Panel A shows a typical Chinese media report on

the scandal which usually came with a full list of contaminated firms and products.

On the other hand, Panel B shows an example from the Western media, the New York

Times, which only reported an estimated number of contaminated firms but did not

mention any specific name.

6.2 Supply Chain Structure

The production process and supply chain structure could matter for the spillover effects.

If the root cause of a quality defect is limited to an individual firm, consumers may not

be wary of other firms from the same origin-industry. One example of this is the case

of the Samsung Galaxy battery fire. On the other hand, if the quality defect stems

from issues in the upstream supply chain, then consumers might have reason to worry

about quality in other firms that source from the same upstream locations. This is the

case for the Chinese dairy scandal as discussed in Section 2.

To shed light on the role of supply chain structure, we take advantage of the Chinese

Customs Data in which firms are required to report the sourcing location for each of

their export transactions. We leverage this information and identify a firm’s major

sourcing province based on the pre-scandal export activities. We define a dummy

variable called “ContaminatedSource” if a province hosts at least one contaminated

firm. Similarly, we define a dummy variable called “ContaminatedSourceProduct” if a

province hosts at least one contaminated firm for a given contaminated product.

Table 12 reports the impact of being in a contaminated sourcing location, separately

for innocent and non-inspected firms. The results are quite mixed. In particular, being

in the same sourcing location as the contaminated firms does not necessarily lead to

lower performance.

One challenge for the power of this test is that foreign consumers might not be

aware of the sourcing location of Chinese dairy firms. Another challeng is that in many

cases, a firm’s sourcing location coincides with its physical location: for instance, Sanlu,

located in Hebei province, primarily sourced from milk farmers in Hebei. To the extent

that local dairy products are imperfect substitutes and firms in one location compete

24

in the same labor market and for the same upstream suppliers, there is a countervailing

market share reallocation effect that would benefit the non-contaminated firms. This

may explain the mixed results in Table 12 as well as the results in Appendix A Table

13 on the overall and domestic performance of firms sourcing in different areas.

6.3 Individual Reputation

Individual reputation can mitigate the impact of a shock to the industry collective

reputation. In light of a collective reputation shock, a more established individual

reputation can (partially) shield firms from the collective stigma. As many industries in

developing countries, most of the Chinese dairy exporters lack established international

brand. However, some have exported for more years than others before the scandal

broke out in 2008. We use a firm’s exporting experiences (years) as a proxy for its

accumulated reputational stock. We define “new” firms as firms that just entered into

exporting in 2008 and define “established” firms as firms that had exported for one or

more years prior to 2008. Table 13 examines the heterogeneous impact of the scandal

on new firms and established firms. In line with the discussion, the across-firm spillover

effect is larger among new exporters. A test of equality between βacross-firm in Column

1 and 2 has p-value 0.0475, and that for testing the equality of βacross-firm in Column 3

and 4 has p-value 0.0609.

6.4 Government Regulations

Another important channel for the observed spillover effects is government or third-

party regulations, which can directly affect all firms from the same origin-industry. As

we discussed in Section 2, in the context of the 2008 Chinese dairy scandal, several

countries imposed explicit import bans for Chinese dairy products. Appendix B Table

1 list these destinations by their value share of the total Chinese dairy exports. Most

of the bans were lifted in 2015.

One way to think about these trade policy changes is that foreign governments

react on behalf of domestic consumers as concerns regarding quality and safety issues

arise about products from a particular country. These blanket-style reactions are part

of the spillovers created by collective reputation. However, an alternative interpreta-

tion is that these trade-related regulatory hurdles arise from protectionist motives. In

25

other words, foreign government would take advantage of the scandal to raise protec-

tionist barriers. Empirically it is very difficult to distinguish these two stories. Table

14 presents results obtained by restricting the sample to only destinations without

explicit import bans. These estimates are very similar to the findings in our main

analysis, which suggests that while explicit government regulations might matter for

the observed spillovers, they are cannot explain the spillovers in full. Market-based

forces responding to the collective reputation shock also play an important role.

Summary: To summarize the discussion in this section, first, information accuracy

plays an important role in mediating the collective reputational foces–the spillover

effects are smaller in countries where people exhibited more targeted search behavior

and thus better information about parties involved in the scandal. Second, supply chain

structures may also matter; however, the reputational effect is confounded empirically

by the standard market competition effect. Third, individual reputation can mitigate

collective reputation shocks, and new firms may be the most vulnerable to collective

reputation damage. Finally, while country-specific and industry-wide regulations may

explain part of the observed effects, market-based forces via reputational spillovers also

play an important role.

7 Conclusion

Understanding how reputation spreads within an industry or a geographic area is im-

portant for informing development policy, because collective reputation implies an

important externality. This paper analyzes collective reputational forces in the con-

text of Chinese dairy firms’ exports following the 2008 scandal. To do so, we combine

rich administrative and survey firm data on exports, domestic production, and quality

inspection with news and search data collected from internet sources. In the first part

of this paper, we aggregate these data at the industry level and we use a difference-in-

differences framework to estimate the impact of the scandal on the export performance

of the Chinese dairy industry. Specifically, we find that the average value of dairy ex-

ports plummeted by 68% following the scandal. This estimate captures both the direct

impact on contaminated firm-products, as well as the spillovers on non-contaminated

firms or products. By excluding firm-products that are identified as contaminated by

26

government inspections from our DD analysis, we estimate a spillover effect of the

scandal on non-contaminated firm-products leading to a decrease in exports of 57%,

about four fifths of the total effect of the scandal on the dairy sector.

Our findings highlight collective reputation as an important factor that could affect

firm’s export performance and quality upgrading incentives. In the second part of

this paper, we use micro-level firm-product data to investigate how the effects of the

scandal are distributed within the Chinese dairy sector. Our analysis suggests that

contaminated firms saw a drop of 87.8% in export revenue after the scandal relative

to the national trend and the firms’ average performance. Additionally, we estimate

a spillover impact on firms selling a contaminated product that is about half of the

effect on directly contaminated firms. Finally, we were surprised to find that firms

deemed innocent by formal inspections do not appear to be faring any better than

non-inspected firms (and in fact they might be faring worse). This finding highlights

the role that government action can play in signaling firms’ quality and the challenges

governments might face in establishing the reputation needed to create credible signals.

We further hypothesize that various channels could mediate these collective rep-

utational effects. Specifically, we investigate whether the following factors appear to

play a role in spreading the effects of the scandal to firms not directly contaminated:

information accuracy, supply chain structure, individual firm reputation and the role

of third-party or government regulations. We split our sample of destination countries

based on the likely accuracy of their knowledge of the firms directly involved in the

scandal, and we find that the across-firm spillover effect is smaller for exports to desti-

nations with higher information accuracy. Similarly, we find smaller spillovers for firms

with a longer experience in exporting. In contrast, we find less supportive evidence that

the nature of the supply chain, with many firms sourcing from the same incriminated

producers, or blanket-style third-party regulation entirely drive the spillover effects we

estimate.

Our findings on the role played by information accuracy and firms’ experience con-

firm that collective reputational forces might be particularly relevant for development

policy. Specifically, collective reputational forces may call for government interven-

tion. However, the details of how government action is carried on appear to matter

for the effectiveness of the action, as shown by the almost counterproductive signal of

inspections for innocent firms that we estimate.

27

Ultimately, the external validity of the results is an empirical question as the exact

magnitudes of spillovers vary across industries and countries. For example, our findings

might underestimate the persistence of collective reputational forces if the scandal led to

changes in the production structure, e.g. fostering vertical integration, or to long-term

quality upgrading due to tightened quality certification and quality standard (Ruan

and Zhang, 2016). The general framework and approach set out here could be used to

study similar issues in other contexts. In future work, we plan to extend the analysis

to broader product categories and market settings, to explore the role played by the

salience of the scandal and the reputational stock of locations and parties involved.

28

References

Abadie, A., A. Diamond, and J. Hainmueller (2010): “Synthetic control meth-ods for comparative case studies: Estimating the effect of California’s tobacco controlprogram,” Journal of the American statistical Association, 105(490), 493–505.

Abadie, A., and J. Gardeazabal (2003): “The economic costs of conflict: A casestudy of the Basque Country,” The American Economic Review, 93(1), 113–132.

Ahluwalia, R., R. E. Burnkrant, and H. R. Unnava (2000): “Consumerresponse to negative publicity: The moderating role of commitment,” Journal ofmarketing research, 37(2), 203–214.

Allen, T. (2014): “Information frictions in trade,” Econometrica, 82(6), 2041–2083.

Atkin, D., A. K. Khandelwal, and A. Osman (2017): “Exporting and firmperformance: Evidence from a randomized trial,” Quarterly Journal of Economics.

Bai, J. (2009): “Panel data models with interactive fixed effects,” Econometrica,77(4), 1229–1279.

Banerjee, A. V. (2013): “Microcredit under the microscope: what have we learnedin the past two decades, and what do we need to know?,” Annu. Rev. Econ., 5(1),487–519.

Banerjee, A. V., and E. Duflo (2000): “Reputation Effects And The Limits OfContracting: A Study Of The Indian Software Industry,” The Quarterly Journal ofEconomics, 115(3), 989–1017.

Bjorkman-Nyqvist, M., J. Svensson, and D. Yanagizawa-Drott (2013):“The market for (fake) antimalarial medicine: Evidence from uganda,” .

Bloom, N., B. Eifert, A. Mahajan, D. McKenzie, and J. Roberts(2013): “Does Management Matter? Evidence from India*.,” Quarterly Journalof Economics, 128(1).

Brandt, L., J. Van Biesebroeck, and Y. Zhang (2012): “Creative accountingor creative destruction? Firm-level productivity growth in Chinese manufacturing,”Journal of development economics, 97(2), 339–351.

Davidson, W. N., and D. L. Worrell (1992): “Research notes and communica-tions: The effect of product recall announcements on shareholder wealth,” Strategicmanagement journal, 13(6), 467–473.

Dawar, N., and M. M. Pillutla (2000): “Impact of product-harm crises onbrand equity: The moderating role of consumer expectations,” Journal of marketingresearch, 37(2), 215–226.

29

De Loecker, J., and P. K. Goldberg (2014): “Firm performance in a globalmarket,” Annu. Rev. Econ., 6(1), 201–227.

De Mel, S., D. McKenzie, and C. Woodruff (2008): “Returns to capitalin microenterprises: evidence from a field experiment,” The Quarterly Journal ofEconomics, 123(4), 1329–1372.

Gobillon, L., and T. Magnac (2016): “Regional policy evaluation: Interactivefixed effects and synthetic controls,” Review of Economics and Statistics, 98(3),535–551.

Harrison, A., and A. Rodrıguez-Clare (2009): “Trade, foreign investment, andindustrial policy for developing countries,” Handbook of Development Economics, 5,4039–4214.

Hau, A. K.-c., T. H. Kwan, and P. K.-t. Li (2009): “Melamine toxicity and thekidney,” Journal of the American Society of Nephrology, 20(2), 245–250.

Jin, G. Z., and P. Leslie (2009): “Reputational incentives for restaurant hygiene,”American Economic Journal: Microeconomics, 1(1), 237–267.

Kugler, M., and E. Verhoogen (2012): “Prices, plant size, and product quality,”The Review of Economic Studies, 79(1), 307–339.

List, J. A. (2006): “The Behavioralist Meets the Market: Measuring Social Prefer-ences and Reputation Effects in Actual Transactions,” Journal of Political Economy,114(1), 1–37.

Ma, J., Z. Wang, and T. Khanna (Working Paper): “Milk Safety in China: AField Experiment,” .

Macchiavello, R. (2010): “Development uncorked: Reputation acquisition in thenew market for Chilean wines in the UK,” .

Macchiavello, R., and A. Morjaria (2015): “The value of relationships: evidencefrom a supply shock to Kenyan rose exports,” The American Economic Review,105(9), 2911–2945.

Manova, K., and Z. Zhang (2012): “Export prices across firms and destinations,”The Quarterly Journal of Economics, p. qjr051.

Marcus, R. D., S. Swidler, and T. L. Zivney (1987): “An explanation ofwhy shareholders’ losses are so large after drug recalls,” Managerial and DecisionEconomics, 8(4), 295–300.

Park, A., D. Yang, X. Shi, and Y. Jiang (2010): “Exporting and firm perfor-mance: Chinese exporters and the Asian financial crisis,” The Review of Economicsand Statistics, 92(4), 822–842.

30

Ruan, J., and X. Zhang (2016): “Low-Quality Crisis and Quality Improvement: TheCase of Industrial Clusters in Zhejiang Province,” Industrial Districts in History andthe Developing World, pp. 169–189.

Startz, M. (2017): “The value of face-to-face: Search and contracting problems inNigerian trade,” Discussion paper.

Van Heerde, H., K. Helsen, and M. G. Dekimpe (2007): “The impact of aproduct-harm crisis on marketing effectiveness,” Marketing Science, 26(2), 230–245.

Verhoogen, E. A. (2008): “Trade, quality upgrading, and wage inequality in theMexican manufacturing sector,” The Quarterly Journal of Economics, 123(2), 489–530.

Zhao, X., Y. Lee, S. Ng, and B. B. Flynn (2009): “The impact of product recallannouncements on stock market reaction: a study of Chinese listed companies,”US-China Business Cooperation in the 21st Century: Opportunities and Challengesfor Enterpreneurs, p. 28.

31

Figure 1: Event Study Graph: China’s Dairy Exports

Note: This figure plots Chinese dairy annual exporting value and quantity from 2000 to 2013.

32

Figure 2: Baseline Exporting Products and Destinations (2000-2007)

01

23

45

67

89

10N

umbe

r of a

ffect

ed fo

od fi

rms

01

23

45

67

89

10N

umbe

r of a

ffect

ed d

airy

firm

s

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17Median for Number of Products

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2000 - 2007Number of products (median) for affected firms

050

0010

000

1500

0N

umbe

r of n

onaf

fect

ed fo

od fi

rms

020

040

060

080

0N

umbe

r of n

onaf

fect

ed d

airy

firm

s

0 1 2 3 4 5 6 7 8 9 10Median for Number of Products

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2000 - 2007Number of products (median) for nonaffected firms

01

23

45

Num

ber o

f affe

cted

food

firm

s

01

23

45

Num

ber o

f affe

cted

dai

ry fi

rms

0 2 4 6 8 10 12 14 16 18Median for Number of Destinations

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2000 - 2007Number of destinations (median) for affected firms

050

0010

000

1500

020

000

Num

ber o

f non

affe

cted

food

firm

s

020

040

060

080

0N

umbe

r of n

onaf

fect

ed d

airy

firm

s

1 2 3 4 5Median for Number of Destinations

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2000 - 2007Number of destinations (median) for nonaffected firms

Note: This figure plots number of firms with different median number of products and number of destinations acrossthe whole pre-scandal period (2000-2007). We categorize the firms by affected(contaminated) firms on the left columnand nonaffected (innocent and noninspected) firms on the right column. Dairy firms are defined as firms which haveever exported dairy products during 2000 to 2013 while non-dairy food firms are defined as firms which have exportedfood products but have never exported dairy products during 2000 to 2013.

33

Figure 3: Synthetic Control Analysis: All Dairy Exports

17

.51

81

8.5

19

19

.52

0lv

alu

e

2000 2005 2010 2015year

treated unit synthetic control unit

Note: This figure plots the natural logarithm of the value of exports for the dairy industry (solid line) and the syntheticcontrol unit (dashed line). The vertical dotted line indicates the year 2009, the first year after the scandal.

34

Figure 4: Synthetic Control Analysis: All Dairy Exports, Placebo

−4

−2

02

4D

iffe

ren

ce

be

twe

en

Tre

ate

d a

nd

Co

ntr

ol U

nits

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

year

Note: This figure plots the difference in the log value of exports between each industry and its respective syntheticcontrol unit. The black line indicates the dairy industry. The vertical red line indicates the year 2009, the first yearafter the scandal.

35

Figure 5: Synthetic Control Analysis: Dairy Exports of Non-Contaminated Firm-Products

18

18

.51

91

9.5

20

lva

lue

2000 2005 2010 2015year

treated unit synthetic control unit

Note: This figure plots the natural logarithm of the value of exports for the dairy industry excluding the value ofcontaminated firm-products (solid line) and the synthetic control unit (dashed line). The vertical dotted line indicatesthe year 2009, the first year after the scandal.

36

Figure 6: Synthetic Control Analysis: Dairy Exports of Non-Contaminated Firm-Products, Placebo

−4

−2

02

4D

iffe

ren

ce

be

twe

en

Tre

ate

d a

nd

Co

ntr

ol U

nits

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

year

Note: This figure plots the difference in the log value of exports between each industry and its respective syntheticcontrol unit. The black line indicates the dairy industry excluding the value of contaminated firm-products. The verticalred line indicates the year 2009, the first year after the scandal.

37

Figure 7: Heterogeneous Effects across Products within Dairy

−2

0−

10

01

02

0

40610

40130

40210

40221

40110

40291

40299

40520

40120

350110

40490

40310

40229

190190

40390

40140

190110

40590

350220

40410

40150

40510

40690

40620

40630

HS6 Code

HS6−level Effect 95% CI

Note: This figure plots the effect of the scandal on each six-digit level dairy product (bars) and their respective 95%confidence intervals constructed from robust standard errors.

38

Figure 8: Persistent Effects Over Time

-6-4

-20

24

Dire

ct e

ffect

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CFirmsProductsXYear

-6-4

-20

24

Dire

ct e

ffect

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CFirmsProductsXYear

-50

510

With

in-fi

rm s

pillo

vers

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CFirmsXYear-5

05

10W

ithin

-firm

spi

llove

rs

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CFirmsXYear

-2-1

01

23

Acro

ss-p

rodu

cts

spillo

vers

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CProductsXYear

-2-1

01

23

Acro

ss-p

rodu

cts

spillo

vers

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

CProductsXYear

-10

-50

5Sp

illove

rs to

inno

cent

firm

-pro

duct

line

s

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

IFirmsProductsXYear

-10

-50

5Sp

illove

rs to

inno

cent

firm

-pro

duct

line

s

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

IFirmsProductsXYear

-50

510

Acro

ss-fi

rms

spillo

vers

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

IFirmsXYear

-50

510

Acro

ss-fi

rms

spillo

vers

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Event

Year

IFirmsXYear

Note: This figure plots the regression coefficients of the following five group dummies

interacted with year dummies: ContaminatedFirmProduct, ContaminatedFirm,

ContaminatedProduct, InnocentFirmProduct, and InnocentFirm. The outcome variable for

the left column is log value of exports and the outcome variable for the right column is

exporting dummy. All regressions control for firm-product and year fixed effects. The

dotted lines plot the standard error bars.

39

Figure 9: News Search Behavior Across Countries from Google Trends

Note: The figure plots google trend news search indices for China, Hong Kong and New

Zealand on the left column, and US, Canada and UK on the right column. The red line

plots the search index for the keyword “Sanlu” and the blue line plots the search index for

the keyword “2008 Chinese milk scandal”.

40

Table 1: Baseline Summary Statistics: Dairy Industry

Contaminated Firms Innocent Firms Non-inspected Firms

Number Mean Number Mean Number Mean Difference p-value(1) (2) (3) (4) (5) (6) (7) (8)

Panel A. Customs Database

Avg. yearly export revenue 15 3.299 23 .496 960 1.846 1.453 .349(in million dollars) . (6.152) . (.906) . (6.617) (1.551) .Years of exporting 15 4.8 23 3.217 960 4.222 .578 .421

. (2.859) . (2.315) . (2.43) (.718) .% exports to OECD countries 15 .49 23 .389 960 .603 -.113 .299(conditioning on exporting) . (.431) . (.422) . (.388) (.108) .

Panel B. Manufacturing Census

Employment 38 954.711 284 279.86 1407 177.244 777.467 .002. (1573.413) . (451.327) . (294.354) (252.157) .

Log (employment) 38 5.732 284 5.043 1393 4.551 1.181 0. (1.449) . (.94) . (1.053) (.234) .

Sales revenue (in million RMB) 36 982.197 264 112.279 880 73.237 908.96 .008. (2073.951) . (235.044) . (165.395) (341.242) .

Log (sales revenue) 36 5.048 263 3.707 867 3.136 1.912 0. (2.006) . (1.275) . (1.449) (.334) .

Note: The sample include only dairy products. Column 1, 3 and 5 show the number of firms fall into each category. Column 7 is the difference betweencontaminated firms (Column 2) and non-inspected firms (Column 6), obtained through a simple regression of the outcome variable on a contaminated groupdummy. Column 8 is the p-value of the difference. Standard deviations are in parentheses for Column 2, 4 and 6. For Column 7, robust standard errors arein parentheses.

41

Table 2: Baseline Summary Statistics: Products

Contaminated Uncontaminated

Number Mean Number Mean Difference p-value(1) (2) (3) (4) (5) (6)

Panel A. Dairy Products

Avg. yearly export revenue 11 6.358 12 3.989 2.369 .542(in million dollars) . (11.122) . (6.409) (3.825) .Number of exporting years 11 7.636 12 6.417 1.22 .086

. (1.206) . (1.975) (.677) .% to OECD countries 11 .552 12 .648 -.095 .437

. (.246) . (.328) (.12) .

Panel B. Non-dairy Food Products

Avg. yearly export revenue 16 38.528 944 21.227 17.301 .168(in million dollars) . (51.01) . (67.015) (12.552) .Number of exporting years 16 7.875 944 6.088 1.787 0

. (.5) . (2.507) (.146) .% to OECD countries 16 .798 944 .758 .04 .204

. (.125) . (.264) (.031) .

Note: Column 1 and Column 3 show the number of products (HS eight-digit) fall into each category. Column 5 is thedifference between contaminated products (column 2) and uncontaminated products (column 4), obtained through asimple regression of the outcome variable on a contaminated group dummy. Column 6 is the p-value of the difference.Standard deviations are in parentheses for column 2 and 4. For column 5, robust standard errors are in parentheses.

42

Table 3: Difference-in-Differences Analysis at HS two-digit Industry: All Dairy Exports

Dep Var: Log (Value) (1) (2) (3) (4)DairyXPost -1.065*** -0.892*** -0.687*** -1.140***

(0.080) (0.114) (0.238) (0.086)FoodXPost -0.174

(0.139)Observations 1120 1386 1107 1120

YearXValue Share to different continents NO NO YES NOProduct×Year Fixed Effects NO NO NO YESWhether Drop Food YES NO YES YES

Note: This table shows the regression results for equation 1. The sample contains all exporters except for non-dairy food industries in the Chinese Customs Data for Column 1, 3 and 5. We rectangularize the data to createa balanced panel at industry (HS two-digit) and year level. The dependent variable is the logarithm of theannual exporting value for each industry. The baseline specification in Column 1 includes only year and industryfixed effects. Column 3 and 4 build on this specification by adding time-varying controls for the value shareof the industry exported to different continents at baseline interacted with year indicators and industry-specificlinear trends, respectively.a Column 2 expands our sample to include non-dairy food industries. Standard errorsclustered at the product (HS2) level. Value share is the share of value exported to different continents during2000 to 2007. *** implies significance at 0.01 level, ** 0.5, * 0.1.

aColumn 3 drops one industry, with two-digit level HS code equal to 77, because it has no exports during our baseline period.

43

Table 4: Difference-in-Differences Analysis at HS two-digit Industry: Dairy Exports of Non-Contaminated Firm-Products

Dep Var: Log (Value) (1) (2) (3) (4)DairyXPost -0.915*** -0.741*** -0.554** -0.848***

(0.080) (0.114) (0.233) (0.086)FoodXPost -0.174

(0.139)Observations 1120 1386 1107 1120

YearXValue Share to different continents NO NO YES NOProduct×Year Fixed Effects NO NO NO YESWhether Drop Food YES NO YES YES

Note: This table shows the regression results for equation 1. We first drop the directly contaminated dairyfirm-products. The remaining sample contains all exporters except for non-dairy food industries in the ChineseCustoms Data for Column 1, 3 and 5. We rectangularize the data to create a balanced panel at industry(HS two-digit) and year level. The dependent variable is the logarithm of the annual exporting value for eachindustry. The baseline specification in Column 1 includes only year and industry fixed effects. Column 3 and4 build on this specification by adding time-varying controls for the value share of the industry exported todifferent continents at baseline interacted with year indicators and industry-specific linear trends, respectively.a

Column 2 expands our sample to include non-dairy food industries. Standard errors clustered at the product(HS2) level. Value share is the share of value exported to different continents during 2000 to 2007. *** impliessignificance at 0.01 level, ** 0.5, * 0.1.

aColumn 3 drops one industry, with two-digit level HS code equal to 77, because it has no exports during our baseline period.

44

Table 5: Interactive Fixed Effects Analysis at HS two-digit Industry: All Dairy Exports

Dep Var: Log (Value) (1) (2) (3) (4) (5)DairyXPost -1.067*** -1.505*** -1.356*** -0.947*** -1.146***

(0.067) (0.064) (0.078) (0.090) (0.297)FoodXPost -0.122

(0.119)Observations 1120 1120 1120 1386 1106

Dimension of Factor Model 1 2 3 1 1YearXValue Share to different continents NO NO NO NO YESWhether Dropped Food YES YES YES NO YES

Note: The sample contains all exporters except for non-dairy food industries in the Chinese Customs Data for Column 1, 2, 3and 5. We rectangularize the data to create a balanced panel at industry (HS two-digit) and year level. The dependent variableis the logarithm of the annual exporting value for each industry. Column 1 shows the results for the baseline specification withoutcontrols. Column 5 further adds controls for the value share of the industry exported to different continents at baseline interactedwith year indicators. Column 4 expands our sample to include non-dairy food industries. Columns 2 and 3 allow for increasinglymulti-dimensional in- teractions between time factors and industry factor loadings. All regressions include a factor model inproducts (HS two-digit) and time of dimension shown in each column. Standard errors clustered at the product (HS two-digit)level. Value share is the share of value exported to different continents during 2000 to 2007. *** implies significance at 0.01 level,** 0.5, * 0.1.

45

Table 6: Interactive Fixed Effects Analysis at HS two-digit Industry: Dairy Exports of Non-Contaminated Firm-Products

Dep Var: Log (Value) (1) (2) (3) (4) (5)DairyXPost -0.872*** -1.296*** -1.304*** -0.748*** -0.699**

(0.076) (0.062) (0.062) (0.089) (0.319)FoodXPost -0.122

(0.119)Observations 1120 1120 1120 1386 1106

Dimension of Factor Model 1 2 3 1 1YearXValue Share to different continents NO NO NO NO YESWhether Dropped Food YES YES YES NO YES

Note: We first drop the directly contaminated dairy firm-products. The remaining sample contains all exporters except fornon-dairy food industries in the Chinese Customs Data for Column 1, 2, 3 and 5. We rectangularize the data to create abalanced panel at industry (HS two-digit) and year level. The dependent variable is the logarithm of the annual exportingvalue for each industry. Column 1 shows the results for the baseline specification without controls. Column 5 further addscontrols for the value share of the industry exported to different continents at baseline interacted with year indicators. Column4 expands our sample to include non-dairy food industries. Columns 2 and 3 allow for increasingly multi-dimensional in-teractions between time factors and industry factor loadings. All regressions include a factor model in products (HS two-digit)and time of dimension shown in each column. Standard errors clustered at the product (HS two-digit) level. Value share is theshare of value exported to different continents during 2000 to 2007. *** implies significance at 0.01 level, ** 0.5, * 0.1.

46

Table 7: Firm-Product Level Analysis: Impact of the Scandal on Export Performance

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

(1) (2) (3) (4)CFirm-ProductXPost -0.637 -0.427 -0.393** -0.032

(1.362) (1.347) (0.173) (0.082)CFirmXPost -2.097*** -2.141*** 0.293 -0.147***

(0.507) (0.534) (0.195) (0.034)CProductXPost -1.111*** -1.063*** -0.147** -0.085***

(0.285) (0.262) (0.072) (0.021)IFirm-ProductXPost 1.124 1.040 -0.302* 0.077

(1.114) (1.062) (0.173) (0.071)IFirmXPost -0.932 -0.866 0.178 -0.072

(0.811) (0.755) (0.149) (0.058)Observations 26152 26152 2498 26152

Firm-product FE YES YES YES YESYear FE YES YES YES YES

Note: This table shows the regression results for the effects of the scandal on exports. The samplecontains all dairy exporters in the Chinese Customs Data. We rectangularize the data to create abalanced panel at firm-product (HS eight-digit) and year level for outcomes in Column 1, 2 and 4.We do so by filling in a small value for missing year observations before taking log. The interactionterms are the products of the post-scandal dummy (2009-2013) with the following five group indica-tors: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProduct, InnocentFirmProduct,and InnocentFirm. The omitted group is non-contaminated products from non-inspected firms. Allregressions control for firm-product and year fixed effects. Standard errors are two-way clustered atthe firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

47

Table 8: Firm-Product Level Analysis: Impact of the Scandal on Overall Performance

Sample: Dairy firms (Manufacturing Survey 2005-2009)

Log (total sales revenue) Log (domestic sales revenue) Log (employment)(1) (2) (3)

CFirmsXPost -0.357** -0.275* -0.138(0.158) (0.151) (0.101)

IFirmsXPost -0.155*** -0.084 -0.019(0.060) (0.062) (0.040)

Observations 6372 4813 9898

Firm FE YES YES YESYear FE YES YES YES

Note: This table shows the regression results for the effects of the scandal on sales and employment. The samplecontains all dairy firms in the Manufacturing Survey data. The interaction terms are the post-scandal dummy (2009-2013) with the following two group indicators: ContaminatedFirm and InnocentFirm. The omitted group is non-inspected firms. All regressions control for firm and year fixed effects. Standard errors are clustered at the firm level.*** implies significance at 0.01 level, ** 0.5, * 0.1.

48

Table 9: Export Performance: Heterogeneous Impact on Sister Firms

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

(1) (2) (3) (4)CFirm-ProductXPost -0.637 -0.428 -0.391** -0.032

(1.362) (1.347) (0.172) (0.082)CFirmXPost -2.100*** -2.144*** 0.295 -0.147***

(0.508) (0.534) (0.195) (0.034)CProductXPost -1.110*** -1.061*** -0.151** -0.085***

(0.285) (0.262) (0.071) (0.022)InnocentSisterFirmsXPost 0.000 0.000 0.000 0.000

(0.000) (0.000) (.) (.)InnocentSisterFirmsProductsXPost 0.000 0.000 0.000 0.000

(0.000) (0.000) (.) (.)InnocnetNonsisterFirmsXPost -0.935 -0.869 0.182 -0.073

(0.811) (0.755) (0.149) (0.058)InnocnetNonsisterFirmsProductsXPost 1.124 1.039 -0.303* 0.077

(1.114) (1.062) (0.173) (0.071)NoninspectedSisterFirmsXPost -0.493 -0.547 0.074 -0.047

(0.606) (0.520) (0.103) (0.043)NoninspectedSisterFirmsProductsXPost 0.000 0.000 0.000 0.000

(.) (.) (.) (.)Observations 26152 26152 2498 26152

Firm-product FE YES YES YES YESYear FE YES YES YES YES

Note: This table shows the regression results for the effects of the scandal on exports. The sample contains all dairy exportersin the Chinese Customs Data. We rectangularize the data to create a balanced panel at firm-product (HS eight-digit) and yearlevel for outcomes in Column 1, 2 and 4. We do so by filling in a small value for missing year observations before taking log. Wefurther separate the sister firms of the contaminated firms from the innocent firms and non-inspected firms comparing to Table7. It turns out that no innocent sister firms kept exporting in post-scandal period thus the exporting value for firm-productsfall into this category is 0. The interaction terms are the products of the post-scandal dummy (2009-2013) with the followingsix group indicators: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProduct, InnocentNonsisterFirmProduct,InnocentNonsisterFirm and NoninspectedSisterFirm. The omitted group is non-contaminated products from non-sister andnon-inspected firms. All regressions control for firm-product and year fixed effects. Standard errors are two-way clustered atthe firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

49

Table 10: Domestic Performace: Heterogeneous Impact on Sister Firms

Sample: Dairy firms (Manufacturing Survey 2005-2009)

Log (total sales revenue) Log (domestic sales revenue) Log (employment)(1) (2) (3)

CFirmsXPost -0.393** -0.292* -0.153(0.159) (0.152) (0.101)

InnocentSisterFirmXPost -0.523*** -0.413 -0.004(0.188) (0.261) (0.094)

InnocentNonsisterFirmXPost -0.172*** -0.083 -0.036(0.063) (0.064) (0.041)

NoninspectedSisterFirmXPost -0.227** -0.120 -0.108(0.104) (0.106) (0.072)

Observations 6372 4813 9898

Firm FE YES YES YESYear FE YES YES YES

Note: This table shows the regression results for the effects of the scandal on sales and employment. The sample contains all dairy firms inthe Manufacturing Survey data. We further separate the sister firms of the contaminated firms from the innocent firms and non-inspectedfirms comparing to Table 8. The interaction terms are the post-scandal dummy (2009-2013) with the following four group indicators:ContaminatedFirm, InnocenSistertFirm, InnocenNonsistertFirm and NoninspectedSisterFirm. The omitted group is non-sister and non-inspected firms. All regressions control for firm and year fixed effects. Standard errors are clustered at the firm level. *** implies significanceat 0.01 level, ** 0.5, * 0.1.

50

Table 11: Information Accuracy: Heterogeneous Impact Based On Google Web Search Index

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

High Low Rest High Low Rest High Low Rest High Low Rest(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

CFirm-ProductXPost -1.403 0.195 -0.113 -1.188 0.253 -0.074 -0.539** 0.000 0.000 -0.081 0.030 -0.002(1.635) (1.083) (0.463) (1.598) (1.045) (0.439) (0.219) (.) (.) (0.101) (0.064) (0.033)

CFirmXPost -0.614 -2.300*** -1.026** -0.757 -2.218*** -0.974** 0.519** 0.000 -0.031 -0.047 -0.168*** -0.075**(0.772) (0.489) (0.408) (0.793) (0.487) (0.399) (0.242) (.) (0.039) (0.051) (0.037) (0.030)

CProductXPost -0.256** -0.563*** -0.564*** -0.261** -0.514*** -0.533*** -0.036 -0.124 0.054 -0.019* -0.044*** -0.044***(0.127) (0.172) (0.152) (0.118) (0.158) (0.142) (0.079) (0.080) (0.070) (0.010) (0.013) (0.011)

IFirm-ProductXPost 0.689 0.628 0.638 0.664 0.571 0.601 -0.009 -0.465*** -0.431 0.041 0.035 0.056(0.565) (0.846) (0.673) (0.563) (0.814) (0.635) (0.055) (0.082) (0.261) (0.038) (0.059) (0.045)

IFirmXPost -0.417 -0.344 -0.809 -0.388 -0.320 -0.741 0.030 0.382*** 0.391 -0.034 -0.024 -0.062(0.364) (0.424) (0.719) (0.367) (0.386) (0.673) (0.066) (0.067) (0.253) (0.033) (0.031) (0.049)

Observations 26152 26152 26152 26152 26152 26152 1121 1183 644 26152 26152 26152

Firm-product FE YES YES YES YES YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YES YES YES YES YES

Note: This table shows the regression results for the information accuracy heterogeneous effects of the scandal on exports. The sample contains all dairy exporters in the Chinese Customs Data. Werectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level for outcomes in Column 1, 2, 3, 4, 5, 6, 10, 11 and 12. We do so by filling in a small value for missingyear observations before taking log. We categorize countries by high and low information accuracy about the scandal, using google web search intensity ratio. Thus the outcomes for columns with title”high” are for countries with high information accuracy, outcomes for column with title ”low” are for countries with low information accuracy. The title ”rest” is for countries without google web searchindex. The interaction terms are the products of the post-scandal dummy (2009-2013) with the following five group indicators: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProduct,InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products from non-sister non-inspected firms. All regressions control for firm-product and year fixed effects. Standarderrors are two-way clustered at the firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

51

Table 12: Supply Chain Structure: Heterogeneous Impact By Sourcing/Business Location

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

Innocent Noninspected Innocent Noninspected Innocent Noninspected Innocent Noninspected(1) (2) (3) (4) (5) (6) (7) (8)

CProductXPost -2.770* -1.004*** -2.710* -0.960*** -0.069 -0.119 -0.210** -0.083***(1.422) (0.264) (1.350) (0.244) (0.110) (0.086) (0.098) (0.021)

CSource-ProductXPost 0.328 -0.300 0.333 -0.292 0.098 -0.031 0.059 -0.005(1.368) (0.446) (1.296) (0.412) (0.145) (0.073) (0.088) (0.032)

CSourceXPost -2.211 1.019*** -2.012 0.944*** 0.011 0.006 -0.187 0.074***(2.672) (0.321) (2.477) (0.298) (0.077) (0.053) (0.171) (0.024)

Observations 770 24598 770 24598 124 2201 770 24598

Firm-product FE YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YES

Note: This table shows the regression results for the sourcing/business location heterogeneous effects of the scandal on exports. The sample contains innocent andnon-inspected dairy exporters in the Chinese Customs Data. We rectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level foroutcomes in Column 1, 2, 3, 4, 7 and 8. We do so by filling in a small value for missing year observations before taking log. Column 1, 3, 5 and 7 use the subsample ofinnocent firms, Column 2, 4, 6 and 8 use the subsample of non-inspected firms. The interaction terms are the products of the post-scandal dummy (2009-2013) with thefollowing three group indicators: ContaminatedProduct, ContaminatedLocationProduct and ContaminatedLocation. All regressions control for firm-product and yearfixed effects. Standard errors clustered at the product-year and location (province) level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

52

Table 13: Individual Reputation: Heterogeneous Impact By Age of Exporting

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Exporting

Established New Established New(1) (2) (3) (4)

CFirm-ProductXPost -0.801 -1.385*** -0.042 -0.094***(1.634) (0.196) (0.096) (0.011)

CFirmXPost -1.646*** -1.972*** -0.107*** -0.160***(0.538) (0.729) (0.035) (0.056)

CProductXPost -0.958*** -1.665*** -0.075*** -0.124***(0.268) (0.403) (0.020) (0.029)

IFirm-ProductXPost -0.903 3.254* -0.053 0.202*(0.926) (1.958) (0.055) (0.118)

IFirmXPost -0.815 -1.737 -0.062 -0.138(0.823) (1.381) (0.059) (0.093)

Observations 19768 6384 19768 6384

Firm-product FE YES YES YES YESYear FE YES YES YES YES

Note: This table shows the regression results for the individual reputation heterogeneouseffects of the scandal on exports. The sample contains all dairy exporters in the ChineseCustoms Data. We rectangularize the data to create a balanced panel at firm-product (HSeight-digit) and year level. We do so by filling in a small value for missing year observationsbefore taking log. Column 1 and 3 use the subsample of established firms, which are firms thathave exported more than 1 year before 2008. Column 2 and 4 use the subsample of new firms,which are firms that haven’t exported before 2008. The interaction terms are the productsof the post-scandal dummy (2009-2013) with the following five group indicators: Contam-inatedFirmProduct, ContaminatedFirm, ContaminatedProduct, InnocentFirmProduct, andInnocentFirm. The omitted group is non-contaminated products from non-inspected firms.All regressions control for firm-product and year fixed effects. Standard errors clustered atthe firm-product and year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

53

Table 14: The Role of Third-Party Regulations: Explicit Import Bans

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

w/o Bans All w/o Bans All w/o Bans All w/o Bans All(1) (2) (3) (4) (5) (6) (7) (8)

firmXproductsXpost -1.037 -0.637 -0.767 -0.427 -0.425** -0.393** -0.053 -0.032(1.916) (1.362) (1.866) (1.347) (0.170) (0.173) (0.115) (0.082)

firmXpost -1.733* -2.097*** -1.834** -2.141*** 0.343* 0.293 -0.124** -0.147***(0.924) (0.507) (0.917) (0.534) (0.195) (0.195) (0.057) (0.034)

productsXpost -1.294*** -1.111*** -1.231*** -1.063*** -0.125 -0.147** -0.097*** -0.085***(0.277) (0.285) (0.255) (0.262) (0.078) (0.072) (0.021) (0.021)

ifirmXproductsXpost 1.620 1.124 1.521 1.040 -0.165 -0.302* 0.105 0.077(1.250) (1.114) (1.183) (1.062) (0.158) (0.173) (0.078) (0.071)

ifirmXpost -0.801 -0.932 -0.740 -0.866 0.164 0.178 -0.061 -0.072(0.904) (0.811) (0.844) (0.755) (0.157) (0.149) (0.064) (0.058)

Observations 18452 26152 18452 26152 1707 2498 18452 26152

Firm-product FE YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YES

Note: This table shows the regression results of the scandal on exports for countries without imposing bans on Chinese dairy products. Thesample of Column 1, 3, 5 and 7 contains all dairy exporters in the Chinese Customs Data, dropping firm-products exported to countries with banson Chinese diary products. Column 2, 4, 6 and 8 come from Table 7 for comparison. We rectangularize the data to create a balanced panel atfirm-product (HS eight-digit) and year level. We do so by filling in a small value for missing year observations before taking log. The interaction termsare the products of the post-scandal dummy (2009-2013) with the following five group indicators: ContaminatedFirmProduct, ContaminatedFirm,ContaminatedProduct, InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products from non-inspected firms. Allregressions control for firm-product and year fixed effects. Standard errors clustered at the firm-product and year level. *** implies significance at0.01 level, ** 0.5, * 0.1.

54

APPENDICES

• Appendix A: Additional Tables and Figures

• Appendix B: Data Appendix and Codebooks

55

Appendix A: Additional Tables and Figures

Appendix A Figure 1: Domestic Dairy Growth: Sales Revenue

6000

0080

0000

1000

000

1200

000

1400

000

Tota

l sal

es re

venu

e of

non

-dai

ry fo

od in

dust

ry

1000

0012

0000

1400

0016

0000

1800

0020

0000

Tota

l sal

es re

venu

e of

dai

ry in

dust

ry

2005 2006 2007 2008 2009Year

Dairy industry Non-dairy food industry

Note: This figure plots domestic sales revenue for dairy industry, with comparison to non-dairy food industry.

56

Appendix A Figure 2: Chinese Imports Growth

020

0040

0060

0080

00D

airy

and

milk

pow

der i

mpo

rts (i

n m

illion

USD

)

050

0000

1000

000

1500

000

2000

000

2500

000

Valu

e of

impo

rts (i

n m

illion

USD

)

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Year

Total imports Dairy imports Milk powder imports

Note: This figure plots all Chinese imports value, dairy products exports value and milk powder imports value from 2000 to2013.

57

Appendix A Figure 3: Chinese Exports Growth

010

0000

020

0000

030

0000

0Va

lue

of e

xpor

ts (i

n m

illion

USD

)

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013Year

Total exports Exports to major dairy destinations

Note: This figure plots all Chinese exports value and exports value to major dairy products destinations from 2000 to 2013.

58

Appendix A Figure 4: Post-scandal Exporting Products and Destinations (2009-2013)

01

23

45

67

8N

umbe

r of a

ffect

ed fo

od fi

rms

01

23

45

67

8N

umbe

r of a

ffect

ed d

airy

firm

s

0 1 2 3 4 5 6 7 8 9Median for Number of Products

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2009 - 2013Number of products (median) for affected firms

050

0010

000

1500

0N

umbe

r of n

onaf

fect

ed fo

od fi

rms

010

020

030

040

050

0N

umbe

r of n

onaf

fect

ed d

airy

firm

s

0 1 2 3 4 5 6 7 8Median for Number of Products

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2009 - 2013Number of products (median) for nonaffected firms

01

23

45

Num

ber o

f affe

cted

food

firm

s

01

23

45

Num

ber o

f affe

cted

dai

ry fi

rms

0 2 4 6 8 10 12 14 16 18 20 22 24Median for Number of Destinations

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2009 - 2013Number of destinations (median) for affected firms

050

0010

000

1500

0N

umbe

r of n

onaf

fect

ed fo

od fi

rms

010

020

030

040

050

0N

umbe

r of n

onaf

fect

ed d

airy

firm

s

0 1 2 3 4 5 6 7 8Median for Number of Destinations

Dairy Firms Non-dairy Food FirmsNotes: Dairy firms are not included in food firms

2009 - 2013Number of destinations (median) for nonaffected firms

Note: This figure plots number of firms with different median number of products and number of destinations across the wholepost-scandal period (2009-2013). We categorize the firms by affected(contaminated) firms on the left column and nonaffected(innocent and noninspected) firms on the right column. Dairy firms are defined as firms which have ever exported dairy productsduring 2000 to 2013 while non-dairy food firms are defined as firms which have exported food products but have never exporteddairy products during 2000 to 2013.

59

Appendix A Figure 5: Google Trend Search Across Regions in China

Note: This figure plots Google trends web search indices across regions (provinces) in China.

60

Appendix A Figure 6: Typical News Report Examples: Domestic and Foreign

Note: The figure in the left column is a typical screenshot of a Chinese news reporting the results of the first round of inspectionon infant formula. The figure on the right is a typical English news reporting the same news. Chinese news usually gives fulllist of contaminated firms and foreign new usually doesn’t.

61

Appendix A Table 1: Baseline Summary Statistics: Non-Dairy Food Industry

Contaminated Firms Innocent Firms Non-inspected Firms

Number Mean Number Mean Number Mean Difference p-value(1) (2) (3) (4) (5) (6) (7) (8)

Panel A. Customs Database

Avg. yearly export revenue 26 2.908 19 .581 37264 .554 2.354 .014(in million dollars) . (4.98) . (1.014) . (3.685) (.958) .Years of exporting 26 5.231 19 4.316 37264 2.751 2.48 0

. (2.438) . (2.335) . (2.053) (.469) .% exports to OECD countries 26 .682 19 .567 37264 .688 -.005 .929(conditioning on exporting) . (.32) . (.402) . (.403) (.062) .

Panel B. Manufacturing Census

Employment 23 920.346 30 402.022 24239 174.298 746.048 .025. (1629.059) . (821.764) . (435.535) (332.241) .

Log (employment) 23 5.943 29 5.31 23849 4.399 1.544 0. (1.254) . (1.058) . (1.152) (.256) .

Sales revenue (in million RMB) 19 268.453 19 242.377 13164 61.137 207.316 .039. (449.599) . (542.848) . (255.915) (100.426) .

Log (sales revenue) 19 4.611 19 3.959 12850 2.913 1.698 0. (1.509) . (1.698) . (1.453) (.337) .

Note: The sample include only non-dairy food products. Column 1, 3 and 5 show the number of firms fall into each category. Column 7 is the differencebetween contaminated firms (Column 2) and non-inspected firms (Column 6), obtained through a simple regression of the outcome variable on a contaminatedgroup dummy. Column 8 is the p-value of the difference. Standard deviations are in parentheses for Column 2, 4 and 6. For Column 7, robust standard errorsare in parentheses.

62

Appendix A Table 2: Industry Weights for Synthetic Control Analysis: All Dairy Exports

HS2 Code Weight

1 0.20031 0.28247 0.36598 0.12399 0.030

Note: The table showsthe weights for indus-tries for synthetic con-trol analysis. HS two-digit codes representsdifferent industries, HS01 represents live ani-mals. HS 31 representsfertilizers industry. 47represents fibrous cellu-losic material and re-covered paper or paper-board. HS 98 com-prises special classifica-tion provisions, and HS99 contains temporarymodifications pursuantto a parties’ national di-rective or legislation.

Appendix A Table 3: Covariate Weights for Synthetic Control Analysis: All Dairy Exports

Covariate Weight (X1,000)

Log Value 0.4083810Positive Exports 0.0007236Value Share to Asia 0.0003530Value Share to Europe 0.0002267Value Share to Africa 999.5845556Value Share to Oceania 0.0000001Value Share to North America 0.0056985Value Share to Latin America 0.0000571

Note: The table shows the weights for covariates for syntheticcontrol analysis.

63

Appendix A Table 4: Industry Weights for Synthetic Control Analysis: Dairy Exports ofNon-Contaminated Firm-Products

HS2 Code Weight

23 0.15931 0.09247 0.01860 0.20293 0.50299 0.026

Note: The tableshows the weights forindustries for syntheticcontrol analysis.HS two-digit codesrepresents differentindustries, HS 23 rep-resents food industriesproducing residuesand wastes thereof orprepared animal fod-der. HS 31 representsfertilizers industry. HS47 represents fibrouscellulosic materialand recovered paperor paperboard. HS60 represents fabricsindustry. HS 93 is forarms and ammunitionindustry and HS 99is for special importreporting provisions.

Appendix A Table 5: Covariate Weights for Synthetic Control Analysis: Dairy Exports ofNon-Contaminated Firm-Products

Covariate Weight (X1,0000)

Log Value 5126.6235Positive Exports 56.7377Value Share to Asia 3.1233Value Share to Europe 133.4432Value Share to Africa 4617.8919Value Share to Oceania 25.3966Value Share to North America 0.0001Value Share to Latin America 36.7834

Note: The table shows the weights for covariates for syntheticcontrol analysis.

64

Appendix A Table 6: Firm-Products Analysis: Alternative Specification Controlling forFirm, Product and Year Fixed Effects

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting(1) (2) (3) (4)

CFirm-ProductXPost 1.238 1.320 -0.163 0.077(1.258) (1.229) (0.162) (0.088)

CFirmXPost -2.866*** -2.857*** 0.182 -0.192***(0.450) (0.461) (0.149) (0.030)

CProductXPost -1.126*** -1.077*** -0.179** -0.086***(0.283) (0.260) (0.076) (0.021)

IFirm-ProductXPost 0.932 0.901 0.019 0.059(0.980) (0.938) (0.203) (0.064)

IFirmXPost -0.822 -0.785 0.020 -0.062(0.777) (0.723) (0.152) (0.055)

Observations 26152 26152 2892 26152

Firm FE YES YES YES YESProduct FE YES YES YES YESYear FE YES YES YES YES

Note: This table shows the regression results for the effects of the scandal on exports using an differentspecification from Table 7. The sample contains all dairy exporters in the Chinese Customs Data. Werectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level foroutcomes in Column 1, 2 and 4. We do so by filling in a small value for missing year observations beforetaking log. The interaction terms are the products of the post-scandal dummy (2009-2013) with thefollowing five group indicators: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProd-uct, InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products fromnon-inspected firms. All regressions control for firm, product and year fixed effects. Standard errorsare two-way clustered at the firm and product-year level. *** implies significance at 0.01 level, ** 0.5,* 0.1.

65

Appendix A Table 7: Firm-Products Analysis: Alternative Specification Controlling YearFixed Effects

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting(1) (2) (3) (4)

CFirm-ProductXPost -0.637 -0.427 -0.990*** -0.032(1.271) (1.273) (0.276) (0.076)

CFirmXPost -2.097*** -2.141*** 0.257 -0.147***(0.445) (0.485) (0.232) (0.029)

CProductXPost -1.111*** -1.063*** 0.359* -0.085***(0.290) (0.274) (0.204) (0.022)

IFirm-ProductXPost 1.124 1.040 -0.111 0.077(1.110) (1.061) (0.390) (0.071)

IFirmXPost -0.932 -0.866 -0.070 -0.072(0.814) (0.759) (0.235) (0.058)

Observations 26152 26152 3681 26152Year FE YES YES YES YES

Note: This table shows the regression results for the effects of the scandal on exports using an differentspecification from Table 7. The sample contains all dairy exporters in the Chinese Customs Data. Werectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level foroutcomes in Column 1, 2 and 4. We do so by filling in a small value for missing year observationsbefore taking log. The interaction terms are the products of the post-scandal dummy (2009-2013) withthe following five group indicators: ContaminatedFirmProduct, ContaminatedFirm, Contaminated-Product, InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated productsfrom non-inspected firms. All regressions control for year fixed effects. Standard errors are two-wayclustered at the firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

66

Appendix A Table 8: Variation in the Data Across Firms, Products and Years

Firm-Product(HS8)-Year Counts (All Food Industry)

2000-2007 2009-2013Contaminated Products Uncontaminated Products Contaminated Products Uncontaminated ProductsDairy Non-dairy Dairy Non-dairy Dairy Non-dairy Dairy Non-dairy

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A. All Destinations

Contaminated firm 77 196 41 269 25 132 12 113Uncontaminated firm 922 17503 1056 481814 353 9516 850 277038

Panel B. Dropping Destinations with Bans

Contaminated firm 69 182 34 225 25 123 12 101Uncontaminated firm 758 13757 610 322554 284 7484 529 191602

Note: This table shows number of observations fall into different cells.

67

Appendix A Table 9: Robustness Check: Baseline Characteristics × Post

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

(1) (2) (3) (4)CFirm-ProductXPost -0.632 -0.422 -0.407** -0.031

(1.414) (1.390) (0.174) (0.083)CFirmXPost -1.308* -1.413** 0.267 -0.088*

(0.726) (0.706) (0.209) (0.050)CProductXPost -0.549** -0.558** -0.195** -0.052***

(0.253) (0.232) (0.083) (0.020)IFirm-ProductXPost 0.972 0.895 -0.269 0.063

(1.017) (0.972) (0.182) (0.062)IFirmXPost -1.437** -1.340** 0.173 -0.114**

(0.680) (0.631) (0.157) (0.047)Observations 26152 26152 2498 26152

Firm-product FE YES YES YES YESYear FE YES YES YES YESBaselineChar×Post YES YES YES YES

Note: This table shows the regression results for the effects of the scandal on exports using an differentspecification from Table 7. The sample contains all dairy exporters in the Chinese Customs Data. Werectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level foroutcomes in Column 1, 2 and 4. We do so by filling in a small value for missing year observations beforetaking log. The interaction terms are the products of the post-scandal dummy (2009-2013) with thefollowing five group indicators: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProd-uct, InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products fromnon-inspected firms. All regressions control for firm-product fixed effects, year fixed effects and somebaseline characteristics of the outcomes. Baseline characteristics include firm’s total sales value, firm’syear of exporting and the average exporting revenue of the HS4 industry. Standard errors are two-wayclustered at the firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

68

Appendix A Table 10: Robustness Check: Time Trends

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

(1) (2) (3) (4) (5) (6) (7) (8)CFirm-ProductXPost -0.686 -0.707 -0.468 -0.492 -0.461** -0.368 -0.037 -0.038

(1.396) (1.394) (1.373) (1.371) (0.183) (0.224) (0.082) (0.082)CFirmXPost -1.362* -1.367* -1.465** -1.464** 0.296 0.218 -0.091* -0.092*

(0.716) (0.713) (0.695) (0.693) (0.198) (0.232) (0.050) (0.049)CProductXPost -0.388 -0.330 -0.419* -0.356 -0.121 -0.249 -0.035** -0.032*

(0.235) (0.250) (0.216) (0.227) (0.126) (0.179) (0.018) (0.019)IFirm-ProductXPost 0.937 0.920 0.865 0.845 -0.255 -0.265 0.059 0.058

(1.022) (1.025) (0.976) (0.979) (0.172) (0.193) (0.062) (0.062)IFirmXPost -1.420** -1.406** -1.327** -1.309** 0.188 0.158 -0.111** -0.111**

(0.682) (0.683) (0.632) (0.633) (0.137) (0.166) (0.046) (0.046)Observations 26152 26152 26152 26152 2498 2498 26152 26152

Firm-product FE YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YESBaselineChar×Post YES YES YES YES YES YES YES YES

HS2digit×Year YES NO YES NO YES NO YES NOH2digit×Post NO YES NO YES NO YES NO YES

Note: This table shows the regression results for the effects of the scandal on exports using an different specification from Table 7. Thesample contains all dairy exporters in the Chinese Customs Data. We rectangularize the data to create a balanced panel at firm-product(HS eight-digit) and year level for outcomes in Column 1, 2 and 4. We do so by filling in a small value for missing year observationsbefore taking log. The interaction terms are the products of the post-scandal dummy (2009-2013) with the following five group indicators:ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProduct, InnocentFirmProduct, and InnocentFirm. The omitted groupis non-contaminated products from non-inspected firms. All regressions control for firm-product fixed effects, year fixed effects andsome baseline characteristics of the outcomes. Baseline characteristics include firm’s total sales value, firm’s year of exporting and theaverage exporting revenue of the HS4 industry. In this specification, we further allow in addition for differential time trends acrosssub-industries by adding interactions between dummies for industries and year dummies, or interactions between dummies for industriesand post-scandal dummy. Standard errors are two-way clustered at the firm and product-year level. *** implies significance at 0.01level, ** 0.5, * 0.1.

69

Appendix A Table 11: Robustness Check: Dropping Firms and Destinations for the Spike

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)CFirm-ProductsXpost 0.988 0.363 1.009 1.264 0.474 1.207 -0.508* 0.000 0.000 0.070 0.021 0.087

(1.298) (0.689) (1.022) (1.273) (0.671) (1.015) (0.270) (.) (.) (0.095) (0.051) (0.080)CFirmXPost -2.486*** -2.014*** -1.910** -2.605*** -1.985*** -1.965** 0.351 0.597 1.099*** -0.180*** -0.143*** -0.143**

(0.606) (0.549) (0.871) (0.642) (0.545) (0.887) (0.278) (0.439) (0.372) (0.041) (0.042) (0.073)CProductsXPost -1.092*** -1.416*** -1.452*** -1.045*** -1.306*** -1.343*** -0.146* -0.221** -0.189* -0.085*** -0.119*** -0.123***

(0.273) (0.310) (0.294) (0.251) (0.285) (0.272) (0.077) (0.104) (0.104) (0.021) (0.024) (0.023)IFirm-ProductsXPost 0.963 -0.581 -0.819 0.880 -0.617 -0.860 -0.248 -0.718*** -0.914*** 0.067 -0.040 -0.062

(1.270) (0.822) (0.846) (1.210) (0.789) (0.828) (0.181) (0.143) (0.132) (0.083) (0.054) (0.057)IFirmXPost -0.474 -0.177 0.154 -0.434 -0.118 0.209 0.111 0.521*** 0.714*** -0.051 -0.010 0.006

(0.828) (0.604) (0.551) (0.782) (0.551) (0.504) (0.160) (0.124) (0.116) (0.062) (0.043) (0.040)Observations 25074 13552 12782 25074 13552 12782 2265 1248 1111 25074 13552 12782

Firm-product FE YES YES YES YES YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YES YES YES YES YES

Note: This table shows the main results dropping the driving firms and destinations (contributed to over 1% of the growth in 2007) of spike in dairy product exports growth between 2006 and 2007. Thesample of Column 1, 4, 7 and 10 contains all dairy exporters in the Chinese Customs Data, dropping driving firms. Column 2, 5, 8 and 11 drop driving exporting destinations. Column 3, 6, 9 and 12 drop bothdriving firms and destinations. We rectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level. We do so by filling in a small value for missing year observations beforetaking log. The interaction terms are the products of the post-scandal dummy (2009-2013) with the following five group indicators: ContaminatedFirmProduct, ContaminatedFirm, ContaminatedProduct,InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products from non-inspected firms. All regressions control for firm-product and year fixed effects. Standard errors clusteredat the firm-product and year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

70

Appendix A Table 12: Information Accuracy: Heterogeneous Impact Based On Google News Search Index

Sample: Dairy exporters (Chinese Customs Data 2000-2013)

Log (Value) Log (Quantity) Log (Price) Exporting

High Low Rest High Low Rest High Low Rest High Low Rest(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

CFirm-ProductXPost -1.351 0.179 -0.113 -1.176 0.269 -0.074 -0.303** 0.000 0.000 -0.077 0.025 -0.002(1.637) (1.076) (0.463) (1.596) (1.045) (0.439) (0.118) (.) (.) (0.101) (0.062) (0.033)

CFirmXPost -0.496 -2.286*** -1.026** -0.612 -2.236*** -0.974** 0.284*** 1.376*** -0.031 -0.040 -0.163*** -0.075**(0.770) (0.489) (0.408) (0.789) (0.494) (0.399) (0.074) (0.072) (0.039) (0.051) (0.036) (0.030)

CProductXPost -0.178* -0.634*** -0.564*** -0.187* -0.583*** -0.533*** -0.059 -0.167 0.054 -0.013 -0.049*** -0.044***(0.107) (0.189) (0.152) (0.098) (0.175) (0.142) (0.084) (0.102) (0.070) (0.008) (0.015) (0.011)

IFirm-ProductXPost 0.663 0.651 0.638 0.639 0.594 0.601 -0.028 -0.478*** -0.431 0.039 0.036 0.056(0.567) (0.844) (0.673) (0.564) (0.812) (0.635) (0.061) (0.083) (0.261) (0.038) (0.059) (0.045)

IFirmXPost -0.383 -0.370 -0.809 -0.358 -0.343 -0.741 0.018 0.439*** 0.391 -0.033 -0.025 -0.062(0.365) (0.420) (0.719) (0.368) (0.382) (0.673) (0.064) (0.100) (0.253) (0.033) (0.031) (0.049)

Observations 26152 26152 26152 26152 26152 26152 884 1385 644 26152 26152 26152

Firm-product FE YES YES YES YES YES YES YES YES YES YES YES YESYear FE YES YES YES YES YES YES YES YES YES YES YES YES

Note: This table shows the regression results for the information accuracy heterogeneous effects of the scandal on exports. The sample contains all dairy exporters in the Chinese Customs Data.We rectangularize the data to create a balanced panel at firm-product (HS eight-digit) and year level for outcomes in Column 1, 2, 3, 4, 5, 6, 10, 11 and 12. We do so by filling in a small value formissing year observations before taking log. We categorize countries by high and low information accuracy about the scandal, using google news search intensity ratio. Thus the outcomes for columnswith title ”high” are for countries with high information accuracy, outcomes for column with title ”low” are for countries with low information accuracy. The title ”rest” is for countries withoutgoogle news search index. The interaction terms are the products of the post-scandal dummy (2009-2013) with the following five group indicators: ContaminatedFirmProduct, ContaminatedFirm,ContaminatedProduct, InnocentFirmProduct, and InnocentFirm. The omitted group is non-contaminated products from non-sister non-inspected firms. All regressions control for firm-product andyear fixed effects. Standard errors are two-way clustered at the firm and product-year level. *** implies significance at 0.01 level, ** 0.5, * 0.1.

71

Appendix A Table 13: Impact of Sourcing/Business Location on Domestic Performance

Sample: Dairy firms (Manufacturing Survey 2005-2009)

Log (total sales revenue) Log (domestic sales revenue) Log (employment)Innocent Noninspected Innocent Noninspected Innocent Noninspected

(1) (2) (3) (4) (5) (6)FirmLocXPost -0.084 0.044 -0.129 0.106 0.002 0.102**

(0.091) (0.089) (0.090) (0.094) (0.066) (0.052)Observations 1754 4404 1334 3293 2525 7029

Firm FE YES YES YES YES YES YESYear FE YES YES YES YES YES YES

Note: This table shows the regression results for the impact of sourcing/business location on sales and employment. The samplecontains innocent and non-inspected dairy firms in the Manufacturing Survey data. Column 1, 3 and 5 use the subsample ofinnocent firms, Column 2, 4 and 6 use the subsample of non-inspected firms. The interaction terms are the post-scandal dummy(2009-2013) with the group indicator: ContaminatedFirmLocation (Province). The omitted group is firms from uncontaminatedbusiness locations. All regressions control for firm and year fixed effects. Standard errors are clustered at the firm level. ***implies significance at 0.01 level, ** 0.5, * 0.1.

72

Appendix A Table 14: Individual Reputation: Heterogeneous Impact By Age of Operation

Sample: Dairy firms (Manufacturing Survey 2005-2009)

Log (total sales revenue) Log (domestic sales revenue) Log (employment)Established New Established New Established New

(1) (2) (3) (4) (5) (6)CFirmsXPost -0.287* -0.908*** -0.225 -0.496 -0.145 0.154*

(0.166) (0.231) (0.156) (0.600) (0.105) (0.087)IFirmsXPost -0.098 -0.374** -0.038 -0.239 0.001 -0.233

(0.063) (0.184) (0.066) (0.160) (0.041) (0.163)Observations 5172 1200 4248 565 7806 2092

Firm FE YES YES YES YES YES YESYear FE YES YES YES YES YES YES

Note: This table shows the regression results for the impact of heterogeneous effect of individual reputation on salesand employment. The sample contains all dairy firms in the Manufacturing Survey data. Column 1, 3 and 5 use thesubsample of established firms, which are firms that have operated more than 1 year before 2008. Column 2, 4 and 6 usethe subsample of new firms, which are firms that haven’t operated before 2008. The interaction terms are the post-scandaldummy (2009-2013) with the following two group indicators: ContaminatedFirm and InnocentFirm. The omitted group isnon-inspected firms. All regressions control for firm and year fixed effects. Standard errors are clustered at the firm level.*** implies significance at 0.01 level, ** 0.5, * 0.1.

73

Appendix B: Data Appendix and Codebooks

74

Appendix B Table 1: Destinations that Imposed Bans on Chinese Dairy Exports

Country Value Share of dairy Value share of milk powder Year lifted

Poland .051 .05 2015Netherlands .033 .039 2015Slovakia .024 .011 2015Finland .011 0 2015Belgium .01 0 2015Singapore .008 .014 2009India .006 .001 2017Spain .006 .011 2015Ivory Coast .004 .01Romania .003 .008 2015Latvia .003 .004 2015Kyrghyzstan .003 .006Luxembourg .002 .001 2015Ghana .001 .002Vietnam .001 0England .001 .001 2015Columbia .001 .001America 0 .001Tanzania 0 .001 2010Gabon 0 .001Bangladesh 0 0Slovenia 0 0 2015Bulgaria 0 0 2015Indonesia 0 0Albania 0 0Croatia 0 0 2015Hungary 0 0 2015Malaysia 0 0

Note: This table shows the list of country names, value share of dairy products exported to thedestination, value share of milk powder products exported to the destination and the year of liftingthe ban for the countries imposed bans on Chinese dairy/food products due to the scandal.

75

Appendix B Table 2: Dairy 8-digit HS Codebook

8-digit HS Code Product Category Product Description

04011000 fresh milk products Milk and cream, not concentrated nor containing added sugar or other sweetening matter, of a fat content, by weight, not exceeding 1%04012000 fresh milk products Milk and cream, not concentrated nor containing added sugar or other sweetening matter, of a fat content, by weight, exceeding 1% but not exceeding 6%04013000 fresh milk products Milk and cream, not concentrated nor containing added sugar or other sweetening matter, of a fat content, by weight, exceeding 6%04014000 fresh milk products Milk and cream, not concentrated nor containing added sugar or other sweetening matter, of a fat content, by weight, exceeding 6% but not exceeding 10%04015000 fresh milk products Milk and cream, not concentrated nor containing added sugar or other sweetening matter, of a fat content, by weight, exceeding 10%04021000 milk powder Milk and cream in solid forms of ≤ 1.5% fat04022100 milk powder Milk and cream in solid forms of > 1.5% fat, unsweetened04022900 milk powder Milk and cream in solid forms of > 1.5% fat, sweetened04029100 condensed milk products Concentrated milk and cream, unsweetened ( excl. in solid form )04029900 condensed milk products Sweetened milk and cream ( excl. in solid form )04031000 cultured milk products Yogurt04039000 cultured milk products Buttermilk, curdled milk and cream, etc ( excl. yogurt )04041000 whey products Whey and modified whey04049000 whey products Other products consisting of natural milk constituents04051000 milk fat products Butter04052000 milk fat products Dairy spreads04059000 milk fat products Other fats and oils derived from milk04061000 cheese products Fresh cheese, incl. whey cheese and curd04062000 cheese products Grated or powdered cheese, of all kinds04063000 cheese products Processed cheese, not grated or powdered04064000 cheese products Blue-veined cheese and other cheese containing veins produced by penicillium requefort04069000 cheese products Other cheese19011000 infant formula Preparations for infant use, for retail sale19019000 malted milk products Other food preparations of malt extract, flour; dairy products(Cocoa content:¡40% of powder, starch or malt extract, or Cocoa contents:¡5% of dairy products)35011000 milk protein produts Casein35022000 milk protein produts Milk albumin, incl. concentrates of two or more whey proteins

Note: Specifically, the HS code for infant formula is 1901100010, however, since the HS10 information is not avaliable in Customs data, here we use HS8 19011000 to represent infant formula.

76

Appendix B Table 3: Destinations with Google trends info

Country Ratio of Google Web Search index Ratio of Google News Search index

China High HighHong Kong High HighMacao High HighJapan High HighNew Zealand High HighNetherlands High HighSweden High HighSouth Africa High LowSouth Korea High LowPakistan Low LowMalaysia Low LowSwitzerland Low LowBurma Low LowIndia Low LowAustria Low LowUK Low LowFrance Low LowIndonesia Low LowBelgium Low LowSpain Low LowCanada Low LowItaly Low LowUAE Low LowUS Low LowGerman Low LowAustralia Low LowPhilippines Low LowTaiwan Low LowSingapore Low LowThailand Low LowVetnam Low Low

Note: This table shows the list of countries which we have the Google trends indices and the category oftheir ratios of Google trends indices. Column 2 shows the category for ratio of Google web search index,Column 3 shows the category for ratio of Google news search index. For each index, we constructed a relativesearch intensity ratio for the two keywords, “Sanlu” versus “2008 Chinese milk scandal” during 09/01/2008 and10/31/2008, which is the outbreak period of the scandal.

77