CHILD LABOR AND SCHOOLING RESPONSES TO ACCESSTO MICROCREDIT IN RURAL BANGLADESH

ASADUL ISLAM and CHONGWOO CHOE∗

Microcredit has been shown to be effective in reducing poverty in many developingcountries. However, less is known about its effect on human capital formation. Inthis article, we examine the impact of access to microcredit on children’s educationand child labor using a new and large data set from rural Bangladesh. The resultsshow that household participation in a microcredit program may increase child laborand reduce school enrollment. The adverse effects are more pronounced for girls thanboys. Younger children are more adversely affected than their older siblings and thechildren of poorer and less educated households are affected most adversely. Ourfindings remain robust to different specifications and methods, and when corrected forvarious sources of selection bias. (JEL H43, I21, J13, J24, L30, O12)

I. INTRODUCTION

Microcredit programs have expanded rapidlyin recent decades in the developing world. Theyhave reached more than 20 million borrowersin Bangladesh, about 60% of the country’spoor rural households (World Bank 2006). TheUnited Nations declared 2005 as the Interna-tional Year of Microcredit, and urged multilat-eral donor agencies and developed countries tosupport the microfinance movement to achieveits Millennium Development Goal of halvingpoverty by 2015. The success and popularityof microcredit over the past decades are evi-denced by the fact that there are more than 7,000microfinance institutions (MFIs) today, servingmillions of poor people, and that microcredit

*We are thankful to Dietrich Fausten, Ronald Caldwell,John Gibson, Guang-Zhen Sun, Michael Keane, GlennHarrison, Bruce Weinberg, Mark Harris, Pushkar Maitra,Chikako Yamauchi, Gigi Foster, Frank Vella, participantsat the Western Economic Association annual conference inVancouver, Managing Selection Workshop at the Univer-sity of South Australia, the fourth Australasian DevelopmentWorkshop at the Australian National University, two anony-mous referees of this journal, and seminar participants atMonash University and Bangladesh Institute of Develop-ment Studies (BIDS) for very helpful comments and sug-gestions. The usual disclaimer applies.Islam: Lecturer (Assistant Professor), Department of Eco-

nomics, Monash University, Caulfield East, VIC 3145,Australia. Phone +61 3 9903 2783, Fax +61 3 99031128, E-mail Asadul.Islam@monash.edu

Choe: Professor, Department of Economics, MonashUniversity, Caulfield East, VIC 3145, Australia.Phone +61 3 9903 4520, Fax +61 3 9903 1128, E-mailChongwoo.Choe@monash.edu

has proved to be an important instrument inhelping “large population groups find ways inwhich to break out of poverty” (The NorwegianNobel Committee’s press release in awardingthe Nobel Peace Prize for 2006 to GrameenBank and its founder Muhammad Yunus).

If access to microcredit helps reduce poverty,then one might surmise that it could alsoimprove investment in children’s education.This is because underdeveloped credit marketscoupled with low household income (Balandand Robinson 2000; Doepke and Zilibotti 2005;Ranjan 1999) or lack of access to credit areoften considered major factors responsible forinadequate education for children in develop-ing countries (Dehejia and Gatti 2005; Edmonds2006, 2007; Jacoby and Skoufias 1997; Ranjan2001). Access to credit can have a positiveeffect on children’s education through a num-ber of channels. First, to the extent that creditmay increase the borrower’s income, the incomeeffect may positively affect the demand forchildren’s schooling (Behrman and Knowles

ABBREVIATIONSi.i.d.: Independent and Identically DistributedIV: Instrumental VariableLPM: Linear Probability ModelMFI: Microfinance InstitutionsMO: Microcredit OrganizationPKSF: Palli Karma-Sahayak Foundation2SLS: Two-Stage Least Squares

1

Economic Inquiry(ISSN 0095-2583)

doi:10.1111/j.1465-7295.2011.00400.x© 2011 Western Economic Association International

2 ECONOMIC INQUIRY

1999). Second, the vulnerability of rural house-holds to adverse exogenous shocks may forcethem to pull their children out of school in timesof need, hampering sustained school enroll-ment for their children. Loans from microcre-dit organizations (MOs) can assist consumptionsmoothing, thereby reducing the likelihood thatchildren are withdrawn from school in responseto adverse shocks. Third, preferences towardschooling may also be influenced by manda-tory adult training programs conducted by MOs.Although MOs in general do not have anydirect declared objective of improving children’seducation, they do educate members about thepotential benefits of sending children to school.For example, Grameen Bank members need tomemorize 16 decisions, one of which is, “weshall educate our children.”

On the other hand, microcredit may also haveunintended consequences on children’s educa-tion for several reasons. First, microcredit loansoften require establishment of household enter-prise that requires extra labor to work in it. Forexample, if a household uses microcredit loansto purchase livestock, it will require labor totake care of the animals, which can increase thedemand for child labor. Second, the amount ofloan is not large enough to hire external labor,which may compel the household to resort tochild labor.1 Third, the loan repayment periodis short and interest rate is high, making thehousehold myopic, which may induce parents toheavily discount the future return on their chil-dren’s education.2 To service the loan, it may benecessary to supplement household income, atleast temporarily, with the proceeds from childlabor. Therefore, the additional activities madepossible by access to microcredit and the fac-tors related to servicing the terms of microcreditloan may adversely affect children’s education.3

Children may need to be employed directly in

1. Loan size varies but is typically between US$40 andUS$150. However, members may take larger loans afterrepaying their first loan. Loans are made for any profitableand socially acceptable income generating activities such aspoultry, livestock, sericulture, fisheries, rural trading, ruraltransport, paddy husking, food processing, small shops, andrestaurants.

2. Typical interest rates on microcredit loan are above30% on a reducing-balance basis and most MOs require thathouseholds start repaying the loans 4 weeks after obtainingcredit. The effective interest rates are even higher becauseof commissions and fees charged by MOs. The frequency ofrepayments and the systems adopted to collect repaymentsalso raise the effective interest rates.

3. In an earlier version of this article (Islam and Choe2009), we provide a theoretical model showing that accessto microcredit can result in increased child labor if the credit

the newly created or expanded household enter-prises, or as caregiver for their siblings, or infarm and livestock duties, and other householdchores.

There is conflicting evidence on the impactof microcredit on human capital formation. Onestrand of empirical studies suggests that accessto credit can help reduce child labor and increaseschooling in developing countries. For example,Jacoby (1994) finds that unequal access to creditis an important source of inequality in school-ing investment in Peru. Dehejia and Gatti (2005)find a negative association between child laborand access to credit across various countries.Jacoby and Skoufias (1997) observe that, inIndia, the incidence of child labor increasesas access to credit becomes more difficult.4

The second strand of literature finds ambiguousresults. Wydick (1999) reports that the relationbetween access to microcredit and children’sschooling is not unambiguously positive in thecase of Guatemala. He finds that a child ismore likely to work in a household enterprisewhen the household borrowing is used for cap-ital equipment instead of working capital. Asimilar conclusion is drawn by Maldonado andGonzalez-Vega (2008), who find that house-holds demand more child labor if they cultivateland and operate labor-intensive microenter-prises. On the basis of microcredit programsin Bangladesh, Pitt and Khandker (1998) findthat girls’ schooling is positively affected whenwomen borrow from Grameen but not so whenthey borrow from other microcredit programs.Banerjee et al. (2009) find that treated microcre-dit households are not more likely to have chil-dren in school, and they do not spend more ontuition, school fees, and uniforms. Hazarika andSarangi (2008) report that, in rural Malawi, chil-dren tend to work more in households that haveaccess to microcredit. On the other hand, Karlanand Zinman (2009) find that male microen-trepreneurial borrowers in Manila use loan pro-ceeds to send their children to school.

Given the limited and conflicting evidencesummarized above, the purpose of this articleis to examine the impact of household partic-ipation in microcredit programs on both chil-dren’s schooling and child labor using a new,

cannot be used to hire external labor and the required returnson investment are high. As indicated in footnotes 1 and 2above, both conditions are likely to hold for microcreditprograms in Bangladesh that we study.

4. Similar results are reported in Beglee, Dehejia, andGatti (2005) for Tanzania and Edmonds (2006) for SouthAfrica.

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 3

large, nationally representative data set basedon a survey conducted in Bangladesh in 1998.Our empirical findings contribute to the litera-ture in three ways. First, our results show thatparticipation in microcredit programs adverselyaffects children’s schooling and exacerbates theproblem of child labor. The results overwhelm-ingly indicate that girls are more likely to beaffected adversely, although the effect on boysis ambiguous. It is also shown that younger chil-dren, who are more exposed to the program,are more likely to be put to work and lesslikely to attend school as their parents take outmicrocredit. Second, we show how the adverseeffect differs depending on the gender of par-ticipating parents. Although the adverse effectdoes not differ much whether credit is obtainedby women or men, we find some evidence forgender preferences: the adverse effect on girls’schooling tends to be smaller when credit isobtained by mother than when it is obtained byfather. Finally, we examine the possible chan-nel through which microcredit adversely affectschildren’s education. Our results show that, forchildren from participating households, the oddsof being in self-employment activities insteadof being in school are more than doubled thannonparticipating households. This suggests thatincreased child labor is in large part because ofhousehold enterprises set up with microcredit.Our empirical findings remain robust to differentspecifications and methods, and when correctedfor various sources of selection bias.

Overall, our results suggest that care needsto be taken in assessing the effectiveness ofmicrocredit programs. On one hand, successfulmicrocredit programs can alleviate poverty andcontribute to rural economy. On the other hand,they can alter parents’ incentives in a way thatadversely affects children’s schooling, whichcould exacerbate poverty in the longer term.5 In

5. One may ask if the negative effect on children’seducation is necessarily bad if children can make up forlower education by working in family enterprises, whichmay increase their future earnings. To answer the question,we need information on returns to education and returnsto working in family enterprises. Unfortunately, this isbeyond the scope of this study because we only look atthe contemporaneous effect of microcredit. A panel data setover a longer term participation in microcredit may help usunderstand the issue better. On the other hand, the long-term effect of microcredit on children’s education couldbe different from what we find in this study: the higherincome opportunities provided by microcredit over timecould lead to an income effect toward less child labor andmore education that may outweigh the substitution effect.We leave this and related questions for future study whenmore information becomes available.

addition, the adverse effect that falls unequallyon girls would reduce the effectiveness ofpolicies to promote gender equality in educationin developing countries. In summary, microcre-dit programs need to be complemented by otherpolicies to tackle the multiple goals of povertyreduction, human capital formation, and socialdevelopment.

The rest of this article is organized asfollows. Section II describes our data andpresents descriptive statistics. Section III dis-cusses issues related to our empirical methodol-ogy and Section IV reports the main empiricalfindings. Section V concludes this article.

II. THE PROGRAM, DATA, AND DESCRIPTIVESTATISTICS

A. The Program and Data

The data were collected by the BangladeshInstitute of Development Studies on behalf ofthe Palli Karma-Sahayak Foundation (PKSF;Rural Employment Support Foundation) withsupport from the World Bank.6 This surveyis the largest and the most comprehensiveof the microcredit programs that operated inBangladesh at that time period. Its geographiccoverage is spread evenly across Bangladesh,and the sub-district (thana) level comparisonsreveal that selected sub-districts are not differ-ent from the average (Zohir et al. 2001). Thedata cover 13 MOs of different sizes in terms ofoperations and membership. These MOs wereselected to constitute a nationally representativedata set for the entire microcredit program inBangladesh. The most notable MOs studied inthis article are ASA and Proshikha, the third andfourth largest MOs, respectively, in Bangladesh.All 13 MOs follow the Grameen Bank-stylelending procedure and typically give access tomicrocredit to households owning less than ahalf-acre of land.

The survey includes 13 districts covering91 villages spread over 23 sub-districts inBangladesh. A census of all households in the91 villages was conducted before the surveywas administered in early 1998. There were 80treatment villages and 11 control villages (vil-lages without microcredit program during thesurvey). The actual targeting of survey house-holds involves two stages: (a) the selection of

6. The PKSF is the apex organization for microfinance.The microlending community regards it as a regulatoryagency and it exercises authority over the MOs.

4 ECONOMIC INQUIRY

the villages where MOs operate and (b) theselection of households within the selectedvillages. In the absence of adequate controlhouseholds, some selected nonparticipants fromthe program villages who are observationallysimilar and expressed willingness to participatein microcredit program were also surveyed asthe control group. Participation in a credit pro-gram was defined in terms of current member-ship reported during the census. From the vil-lage census lists of households, 34 householdswere drawn from each program and nonprogramvillage. The selections of villages and house-hold sample are described in detail by Zohiret al. (2001).

B. Descriptive Statistics

The original survey consists of 3,026 house-holds. In this article, we consider the subset of2,034 households who have at least one childaged 7–16 at the time of the survey. Thisrepresents a total of 4,277 children of which2,658 belong to treatment households and the

remainder to the control group. Our samplecontains both male and female borrowers but theformer accounts for only 12% of all borrowers(and 133 households) representing 281 children.Among all children, 54.2% are boys.

The household-level questionnaire includesprimary and secondary activity of each child.We define “child laborer” as anyone aged 7–16who performs any economic activity (i.e., if aparent answers “employed,” “household work,”or “employed but not working”). A child isconsidered to be in school if he/she is cur-rently enrolled in school and attended schoolin the last month of the survey period. Bythis definition, 77.4% of girls aged 7–16 inthe sample were classified as being in schooland 10.4% in work. The corresponding figuresfor boys are 71.3% and 15.7%, respectively.Other children are reported to be neither work-ing nor in school, and possibly many of themare helping parents with household work. Sothere may well be under-reporting of child labor.The results by participation status are reportedin Table 1. School enrollment is lower and

TABLE 1Descriptive Statistics of Child and Household Characteristics

Variables Treatment (I) Control (II) Difference III = (I − II)

Child characteristics (7–16 years old)Child in work (%)

Boys 16.9 13.8 3.0Girls 11.0 9.3 1.7

Child in school (%)Boys 69.3 74.7 −5.5Girls 76.6 78.9 −2.3

Age of child (years) 11.497 11.494 0.003Sex of child (percentage of girls) 55.7 55.6 0.1

Household characteristicsMother age (years) 37.66 38.14 −0.48Mother schooling (years of education) 1.09 1.54 −0.45Father age (years) 45.85 46.80 −0.95Sex of household head (male = 1) 0.93 0.95 0.026Father schooling (years of education) 2.64 3.20 −0.56Household size 6.56 6.48 −0.08Number of children

0–6 years 0.81 0.79 0.026–16 years 2.79 2.66 0.13

Maximum education by any household memberMale borrower (years of education) 4.78 5.29 −0.50Female borrower (years of education) 4.17 4.57 −0.40Amount of land (decimals) 64.7 91.2 −26.6

Number of children observations 2,568 1,709 —Number of households in the sample 1,241 793 —

Notes: The last column presents the difference between columns (I) and (II). Differences that are statistically significantat less than 5% are marked bold.

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 5

FIGURE 1Proportion of Children in School

.45

.55

.65

.75

.85

Pro

port

ion

atte

ndin

g S

choo

l

7 8 9 10 11 12 13 14 15 16

Age Year

Treatment Group

.45

.55

.65

.75

.85

7 8 9 10 11 12 13 14 15 16

Age Year

Treatment GroupControl Group Control Group

Figure 1a: Proportion of Boys in School Figure 1b: Proportion of Girls in School

child labor higher among children of the treat-ment group. We find a statistically significantdifference in school enrollment and child laborbetween boys of treated and untreated house-holds, but no such difference exists for girls.However, the difference in school enrollmentbetween girls and boys is larger in the treatmentgroup.

Figure 1 plots school enrollment of childrenby age for both sex groups. Children at highschool age (12–16 years old) are less likely tobe enrolled in school because of dropouts. Atprimary school age, the proportion of childrenaged 7–8 enrolled in school is lower than theirolder counterpart (9–11 years old), indicatingthat there are a considerable number of chil-dren who start schooling at a later age. Thedifference between treatment and control groupsin school enrollment is larger for boys. Girlsaged 7–11 have a similar rate of enrollmentin both treatment and control groups, but afterage 13, girls in the control group tend to havea lower enrollment rate. On average, children

at primary school age have a higher enroll-ment rate compared with their older siblings,the latter more likely to drop out from schooland go to work. Overall, a higher proportionof children from treated households is in work(Figure 2).

Table 1 also provides other descriptive statis-tics for child and household characteristics.It shows that the average age of children is11.5 years for both groups of households. Thereis no difference between treatment and controlgroups in the gender composition of children.The treatment group has slightly more membersin the household than the control group. For eachhousehold, the average number of children underthe age of 18 is four. Nontreated households tendto be better educated, a little older but smallerin household size.

III. EMPIRICAL METHODOLOGY

In estimating the impact of microcredit onchildren’s school enrollment and child labor,we follow a standard methodology (Edmonds

6 ECONOMIC INQUIRY

FIGURE 2Proportion of Children in Work

.05

.15

.25

.35

.45

Pro

port

ion w

ork

ing

7 8 9 10 11 12 13 14 15 16Age Year

Treatment Group Control Group

Figure 2a: Proportion of Boys in Work

.05

.15

.25

.35

.45

7 8 9 10 11 12 13 14 15 16Age Year

Treatment Group Control group

Figure 2b: Proportion of Girls in Work

2006; Ravallion and Wodon 2000; Wydick1999). Let Si be a binary variable that denoteswhether child i attends school (Si = 1) or not(Si = 0).7 We estimate the impact of partici-pation in microcredit programs on children’sschooling with the following equation:

Sijkl = β0l + β1Xijkl + β2Creditjkl + β3Zk + εijkl

(1)

where the subscripts index child (i), household(j ), village (k), and district (l). X is a vec-tor of child- and household-specific covariatesand Z is a vector of village-specific covariates.β0l captures fixed effects. Credit is a continu-ous treatment variable defined by the amountof microcredit borrowed by the household. Itis equal to 0 if a household did not partici-pate in a microcredit program. Z k incorporateseither village-level fixed effects or village-level

7. A child may be reported to be both in school andin work, perhaps helping parents in household work whileattending school. In our sample, only 1% of children belongin this category, so we ignore this case from our regression.It is usual in rural areas of Bangladesh that parents arrangea modest amount of part-time work for their children whilestill keeping them at school. See, for example, Ravallion andWodon (2000).

control. The error term εijkl is assumed to beindependent and identically distributed (i.i.d.).The variables in X include the age of child, theage of child squared, dummies for birth order,the age of household head, education and sex,the level of highest education of any member inthe household, presence of father, presence ofmother in the family, mother’s schooling, andthe amount of arable land. Our results did notchange when we included a larger set of con-trols. The village-level covariates in Z include,among others, separate dummies for the pres-ence of primary school, secondary school or col-lege, religious school, health facility, post office,brick-built road, grocery market, bus stand, dis-tance to nearest sub-district, adult male wage,and price of rice. The descriptive statistics forthe full list of village-level covariates are pro-vided in Table A1. As shown in the table, thedifferences in village-level covariates are mostlyinsignificant, indicating that these differencesare not the main driving force for our results.

For the effect of microcredit on child labor,our estimating equation is as follows:

Wijkl = α0l + α1Xijkl + α2 Creditjkl(2)

+ α3Zk + υijkl

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 7

where W is a binary variable that is equal to1 if child works and 0 otherwise. All othervariables are as explained above, υ is the errorterm representing omitted characteristics and isassumed to be i.i.d.

Although the binary variable S in Equation(1) captures the current status of school agechildren, it does not measure the achievementof those who are not in school at the time ofthe survey. Nor does it tell us anything aboutthe level or quality of schooling. Therefore, weconsider the impact of microcredit participationon three other measures of children’s educa-tional achievement. They are (a) the number ofyears of school completed, (b) the “educationgap,” and (c) the “grade-for-age” variable. InBangladesh, children are expected to start schoolat around the age of six. Therefore, we can con-struct a variable “education gap” to measure theachievement in terms of grade completion for agiven age. The education gap can be defined asfollows:

Education gap = max{0, Expected education

− Actual education},where

Expected education

={

0 if age ≤ 6,age − 6 if 7 ≤ age ≤ 16.

For example, if a child successfully stayedat school as expected, the gap is 0. If a childencountered problems such as late entry, failedgrades, or dropping out, then the gap is a posi-tive number. If a child never attended school,then the gap is the level of expected educa-tion at that age. For another measure of edu-cational achievement, we follow Patrinos andPsacharopoulos (1997) to define a grade-for-agedependent variable as follows:

Grade-for-age = 100 × (Education grade/

Expected education)

where “Education grade” is the number ofyears a child successfully completed in school.Denoting these variables by Edu, the estimatingequation is as follows:

Eduijkl = γ0l + γ1Xijkl + γ2 Creditjkl(3)

+ γ3Zk + ςijkl .

However, estimating Equations (1)–(3) di-rectly is problematic. There are a number ofdifferent potential sources of bias that need

to be addressed in examining the effect ofmicrocredit. First, program placement may notbe random. Selection for placement could beinfluenced by biases in favor of high-incomevillages—because they may have higher par-ticipation rates—or by official bias in favor ofpoorer villages. We control for village-level non-random program placement using village fixedeffects. In another specification, we use village-level control assuming that village-level pro-gram placement is a “selection-on-observables.”The survey covers a wide range of village-levelvariables. Therefore, we use a set of control vari-ables at the village level, which are includedin the vector Z and are reported in Table A1.When using village-level control, we also usedistrict-level fixed effects to remove any unob-served heterogeneity across different geographi-cal areas. Because we have 13 MOs, each froma different district, this fixed effect also cap-tures the heterogeneity across different MOs.Thus, we tackle the potential problem of non-random program placement using (a) villagefixed effects and (b) geographical and MO-levelfixed effects and village-level observed covari-ates. Our results are not affected by using eitherspecification.

Second, households self-select into the pro-gram but not all of them are able to obtainmicrocredit. Generally, only eligible poor house-holds receive microcredit, the eligibility beingtypically determined based on the amount oflandholding. However, other factors that influ-ence whether a household has access to micro-credit could also affect outcomes for children ofthat household. For example, microcredit loansoften require that family enterprises be estab-lished because they provide less opportunityfor misuse of the loan. Poor households thatoperate an enterprise are also more likely toemploy their children in that enterprise, andthus less likely to send them to school. It islikely that participants differ from nonpartici-pants in the distribution of observed character-istics, leading to a “selection-on-observables”bias. There are also problems due to “selection-on-unobservables” because of self-selection intothe program (and subsequent decision on howmuch to borrow). Thus we need to consider theendogeneity of participation in microcredit pro-grams at the household level. The endogeneityproblem implies that selection into treatment canbe affected by unobserved characteristics εijkl inEquation (1), hence a potential nonzero corre-lation between εijkl and Creditjk . Consequently,

8 ECONOMIC INQUIRY

impact estimates that use a simple probit/linearprobability model (LPM) may not reflect theprogram’s causal effect on children’s schoolenrollment or child labor.

To account for self-selection into the program,we consider a source of exogenous variation.The MOs set the eligibility criteria for participat-ing in the program. A household is eligible if itdoes not own more than a half-acre of land. Theland ownership criterion is mainly used as a tar-geting mechanism to identify the poor. Becausepoverty does not exclusively depend on landownership, however, the administrator, localloan officer, or branch manager sometimes takeinto account other socio-economic conditionsof a household. Consequently, there are someineligible households that receive microcredit.Although these households are a distinct minor-ity (70% of the treatment group in our sample iseligible), the participation in the program based

on eligibility is probabilistic because the pro-gram eligibility criterion is not strictly followed.Thus our approach in estimating the treatmenteffect is similar to the use of fuzzy regressiondiscontinuity design (Van der Klaauw 2002),which we implement using an instrumental vari-able (IV) approach.

A household does not receive microcreditif the program is not available in the village.Therefore as an instrument for the actual receiptof microcredit, we may consider the eligibilitystatus interacted with an indicator for presenceof program in a given village. Instead of usingthis instrument directly, however, we utilize anunexploited exogenous source of variation inthe treatment intensity based on a household’sexposure to the program in different villages.As shown in Figure 3, treated households indifferent villages appear to borrow differentamounts. Intensity of treatment varies widely

FIGURE 3Years of Microfinance Program in a Village and the Amount of Credit Borrowed by Households

2000

4000

6000

8000

Ave

rage

cre

dit p

er h

ouse

hold

in a

vill

age

2 6 10 14 18Number of years an MO is available in a village

Notes : Average credit per household in a village is the amount of credit borrowed (in taka) by all households divided bythe number of participating households in the program village. Number of years an MO is available in a village is the periodsince microcredit has been available in the program village for the first time.

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 9

in different villages, depending on how long amicrocredit program has been available in thevillage. In our sample, the earliest a programwas made available in a village was 1980 andthe latest a program became available in anothervillage was 1997. As shown in Figure 3, theamount of credit a household borrows largelydepends on how long the program has beenavailable in the village. Thus the instrumentfor credit variable is the interaction amongprogram availability, household eligibility, andthe number of years microfinance has beenavailable in a village. Ownership of land is aclear measure of wealth and eligibility is definedby land ownership. Because wealth or land couldhave independent effects on child labor andschooling (see, for example, Basu, Das, andDutta 2010), we control for land ownership inour regression.8 The equation for the demandfor credit then assumes the form:

Creditjkl = δ0l + δ1jk (Mk × Ej × Nk)(4)

+ δ2Xjk + ξjkl

where X now includes land/assets and otherhousehold-level covariates. The instrument isI = Mk × Ej × Nk , where Mk is a binary vari-able that equals 1 if village k has a microcreditprogram, Ej is a binary variable that equals 1if household j is eligible (i.e., owns less thanhalf-acre of land), and Nk is the number ofyears a microcredit program has been availablein village k.9

The impact estimates are obtained using thetwo-stage procedure where the second-stageregression uses the value of credit from thefirst-stage credit demand Equation (4), which isestimated by a standard Tobit model. We reportthe second stage using ordinary least squaresestimations of LPM. In an earlier version of thisarticle (Islam and Choe 2009), we also reportedthe second-stage results using probit and theresults are very similar.

A potential problem with interpreting theseresults when using credit as the treatment vari-able is that the reported amount of creditis subject to misreporting or other types ofmeasurement error because households may

8. We also tried different land and asset-related variablesincluding their polynomials. Our results are robust to suchvariations in the specification.

9. We also estimated a credit demand equation followingPitt and Khandker (1998) where interactions of eligibilityand all of the household exogenous variables are used ascontrols. The results are similar and are available from theauthors.

forget or not report the amount correctly.10 Thismeasurement error is likely to impart attenuationbias to the estimated coefficients. However, wedo not think the problem of measurement error isserious in our case because we are using instru-mented credit variable as the treatment variable.Nonetheless, we also use a binary treatment indi-cator, that is, whether or not a household iscurrently a member of a microcredit program,which is unlikely to be measured or reportedwith error. It can also serve as a robustnesscheck of our main results. It should be noted,however, that the use of binary treatment indi-cator raises another issue as dummy endogenousregressors with limited dependent variables raisesome econometric problems. Angrist (2001)advocates using simple IV estimators as an alter-native because they require weaker assumptionsand are often sufficient to answer questions ofinterest in empirical studies. We therefore esti-mate the treatment effect by using an LPM inthe second stage of the IV regression.

We also control selection bias using an alter-native econometric strategy. We consider cor-rections for endogeneity using reduced-formresiduals that lead to a control function methodof accounting for both selection and endogene-ity.11 This is also important if the effect of creditvaries across households. In that case, IV/two-stage least squares (2SLS) may not estimatethe average treatment effect of credit. Thereare, however, different approaches to estimatingcontrol functions, and not all these proceduresproduce consistent estimates of the treatmenteffect. We adopt the procedure suggested byVella (1993), which identifies correctly the treat-ment effect parameter in our context.12 We firstobtain generalized residuals using either Tobit

10. See, for example, Karlan and Zinman (2008) forproblems with self-reported credit data.

11. The control function approach estimates the averagetreatment effect by controlling directly for the correlationbetween the error term and the outcome of equations withthe treatment variable. It treats the selection bias problemas an omitted variable problem and augments the outcomeequation by a term to control for this omission. Thetraditional example is the Heckman sample selection modelthat augments the outcome equation by an estimate of theMills ratio.

12. Garen (1984) suggests a linear control functionestimator to correct for endogeneity. However, Garen’sapproach is appropriate when the dependent variable in thefirst stage can take a value over a continuous range and itshould be uncensored. Similarly, the two-stage conditionalmaximum likelihood approach of Rivers and Vuong (1988)is not applicable as the approach also requires that the creditvariable be continuous (see Ravallion and Wodon 2000;Vella 1993).

10 ECONOMIC INQUIRY

TABLE 2First-Stage Results

Participation Credit Demand

Probit Probit Tobit TobitInstrument (1) (2) (3) (4)

M × E 0.5059 (11.59)** — 1,017.2 (7.27)** —M × E × N — 0.0437 (10.07)** — 103.22 (7.61)**

Notes: Regressions include controls and fixed effects. Absolute value of t-statistics is presented in parentheses.Coefficients with ** are significant at the 1% level.

(for credit as the treatment variable) or probit(for binary treatment indicator) for the reduced-form first-stage equation, and then use the esti-mated residuals as an additional regressor in thesecond stage.13

A. First Stage and the Instrument Validity

Before we move on to estimation results,we offer further discussions on our identifica-tion strategy and the validity of our instrument.Identification requires that land ownership beexogenous conditional on program participation.The validity of the land-based eligibility crite-rion as an instrument for microcredit participa-tion in Bangladesh is also defended at lengthby Pitt and Khandker (1998). Because credit isextended mainly for self-employment activities,households having more land are exogenouslyruled out. However, some participating house-holds own more than half an acre of land. Thesehouseholds are either currently not activelyengaged in agriculture or their land is not fertilefor cultivation, or they participate in microcre-dit programs because of mistargeting as per-fect monitoring is not possible. The eligibilityrule is set to simply identify the poverty sta-tus of the household. Because land price andquality also vary between different regions, ahousehold having more than half an acre ofland may be considered to be poor in someregions, as subjectively judged by the loan offi-cer or branch manager. On the other hand, richerhouseholds typically obtain credit from formalmarkets, or through other means. Also thereare social norms that bar them from becomingmembers of an MO. Moreover, rich people inrural areas hesitate to become members of MFI,because they consider MFI as an organizationfor the poor. Thus the use of eligibility criterion

13. This model is identified even without the exclusionrestrictions because of the nonlinearity of the residuals.

as an instrument for treatment in microfinance iswell justified. Finally, we stress that our identi-fication strategy does not depend exclusively onthe eligibility rule because we also exploit thevariation in credit demand among householdsin different villages based on the availabilityof program in different villages. Thus, it is theinteractions of these three variables (M, E, andN ) that need to be exogenous.

In the following, we discuss the validity ofour instrument. We test whether the IVs arecorrelated with the endogenous regressor andorthogonal to the error process. By definition,the instrument should be correlated with theendogenous regressor (“credit” variable), andit should not be correlated with outcomes ofinterest (child labor or schooling) through anychannel other than via its influence on endoge-nous regressor. The instrument should also beorthogonal to any other omitted characteris-tics reflected in the error term in Equation (1).The first-stage regression results are reported inTable 2 using both probit (for participation)14

and Tobit (for credit demand) models based onEquation (4). The Tobit results show that bothinstruments M × E and M × E × N are highlystatistically significant with t-statistics aboveseven in both cases. The coefficient estimateis positive and also economically significant,implying that our instrument is significantlyrelated to the demand for credit. We estimatethe participation decision equation by regress-ing a binary indicator for participation on an

14. For the first-stage regression, we use the nonlinearprobit model instead of the LPM. We prefer probit becausein the case of IV estimation, the untestable assumption is theexclusion restriction. However, this assumption is not strictlyrequired when a nonlinear binary response regression is usedinstead of the LPM (see Ravallion 2008). We have also triedlinear models in both stages of the IV estimation but ourconclusions remain intact. The first-stage results using theLPM for binary participation measure show that t-statisticsare always greater than 10.

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 11

TABLE 3Reduced-Form Results—Effects of Eligibility

on Schooling/Work

Child in School Child in Work(1) (2)

Eligibility −0.036 (0.042) −0.019 (0.026)Observations 4,277 4,277R2 0.245 0.246

Notes: Regressions also include instrumented creditvariable, child and household characteristics, and villagefixed effects. Standard errors presented in parentheses arecorrected for clustering at the village level.

indicator of interaction between eligibility andprogram village dummies (plus all controls). Theresults are stronger with t-statistics above 10.

Because we have a single instrument for thecredit variable, we cannot test the exogene-ity of the instrument as in an over-identifiedmodel. The remaining concern is whether theinstrument satisfies the exclusion restriction,that is, whether eligibility affects child laboror school enrollment only through participationin the credit program or the amount of creditborrowed conditional on the amount of land,demographic, and other socio-economic char-acteristics. Although the exclusion restriction isnot directly testable, we address this concern ina number of ways. First, we estimate a reduced-form regression to examine the effect of loaneligibility on school enrollment/child labor. Theresults, reported in Table 3, indicate that there isno effect of eligibility or program placement onschool enrollment and child labor. We also esti-mate an equation in which credit is instrumentedbut instrument eligibility enters the second-stageregression directly (and naturally in the first-stage regression). By definition of IV, the instru-ment should be uncorrelated with the outcomesof interest through any channels other than theireffects via the endogenous regressors. There-fore, once the credit is instrumented, eligibilityitself should have no effect on schooling or childlabor when both instrumented credit and eligibil-ity are entered as controls for child labor/schoolenrollment equation. The results in Table 3 donot indicate any significant effect of eligibilityin any of the specifications.

IV. EMPIRICAL FINDINGS

This section reports our empirical findingswhere the estimated value of credit from the

first-stage regression is used as the regressorin the second-stage estimation. We estimate theimpact of credit extended to women and menseparately.15 This is to see how the gender ofparticipants in the microcredit program affectsschooling and work decision for their children.We estimate the results separately for boys andgirls by credit given to both women and men.

Table 4 reports the estimates of the second-stage regression. All coefficients are estimatedusing LPM.16 The results in columns (1) and(2) indicate that participation in microcredit pro-grams significantly increases the probability ofchild labor for girls. For boys, there is someindication that microcredit reduces the prob-ability of child labor especially when creditis obtained by women. Overall, the impact ofmicrocredit on child labor is positive and sig-nificant. The qualitative results are independentof whether credit is obtained by men or women.For example, microcredit increases the probabil-ity of child labor for girls by 13.7% using bothvillage fixed effects and village controls. Theprobability increases to 14%–16% when womenare borrowers. For boys, women’s credit has aninsignificant but detrimental effect. Table 4 alsoshows that girls are affected more adversely, andboys more favorably, when credit is obtained bymen than by women, although these estimatesare not statistically significant. A Hausman-liketest does not support the difference in treatmenteffect between men and women borrowers. Theoverall finding is that participation in microcre-dit programs increases the likelihood of childlabor for girls while the impact on boys is lessclear.

Columns (3) and (4) in Table 4 report theeffect of microcredit on school enrollment. Theresults overwhelmingly indicate that access tomicrocredit negatively affects children’s schoolenrollment. This is true across all regressionmodels and regardless of whether we use vil-lage fixed effects or village controls. The neg-ative effect is especially pronounced for girlsalthough, for boys, it is statistically insignifi-cant in many cases. For example, microcreditdecreases the probability of school enrollment

15. Although credit is given to both women and men indifferent villages, credit groups are never mixed by gender.Households do not have choice over which gender is toparticipate as MOs select one or the other gender, but notboth.

16. The results do not change when using probit in thesecond stage, which are reported in an earlier version of thisarticle (Islam and Choe 2009).

12 ECONOMIC INQUIRY

TABLE 4Impact Estimates on Child Work and Schooling

Child in Work Child in School

Village Control Village Fixed Effects Village Control Village Fixed Effects(1) (2) (3) (4)

Women and men’s creditAll 0.0801 (0.0441)† 0.119 (0.041)** −0.1588 (0.0688)** −0.219 (0.065)**Boys 0.0034 (0.0583) 0.075 (0.070) −0.0756 (0.0825) −0.159 (0.079)*Girls 0.1367 (0.0594)** 0.137 (0.054)* −0.2261 (0.0905)** −0.246 (0.093)**

Women’s creditAll 0.087 (0.0456)† 0.134 (0.066)* −0.1717 (0.0694)** −0.213 (0.093)*Boys 0.0129 (0.0590) 0.067 (0.109) −0.0965 (0.0846) −0.102 (0.123)Girls 0.1426 (0.0610)** 0.159 (0.089)† −0.2296 (0.0921)** −0.287 (0.123)*

Men’s creditAll 0.0774 (0.0746) 0.133 (0.040)** −0.1423 (0.1108) −0.246 (0.067)**Boys −0.0239 (0.0912) 0.089 (0.065) −0.0194 (0.1372) −0.195 (0.079)*Girls 0.1507 (0.1094) 0.147 (0.055)** −0.2635 (0.1496)† −0.256 (0.099)*

Notes: All the results are the marginal effects of instrumented credit variable using IV regressions. The regressions alsoinclude child and household characteristics. Standard errors presented in the parentheses are corrected for clustering at thevillage level using the formulas by Liang and Zeger (1986). The coefficients and the standard errors are multiplied by theaverage credit borrowed by the respective group of households.

Coefficients with † are significant at the 10% level, those with ** at the 5% level, and those with * at the 1% level.

TABLE 5Impact Estimates Based on Binary Participation Measure

Child in School Child in Work

Village Control Village Fixed Effect Village Control Village Fixed Effect

Women’s creditBoys −0.1690 (0.1374) −0.223 (0.179) 0.0193 (0.0956) 0.132 (0.146)Girls −0.4782 (0.1359)* −0.471 (0.176)** 0.2569 (0.0946)* 0.281 (0.119)*

Men’s creditBoys −0.1190 (0.1744) −0.248 (0.128)† 0.0097 (0.1175) 0.087 (0.091)Girls −0.5828 (0.1782)* −0.408 (0.134)** 0.3019 (0.1357)** 0.194 (0.084)*

Notes: All the results are the marginal effects of instrumented binary treatment indicator variable using IV regressions. Theregressions also include child- and household-specific covariates. Standard errors presented in the parentheses are correctedfor clustering at the village level using the formulas by Liang and Zeger (1986).

Coefficients with † are significant at the 10% level, those with ** at the 5% level, and those with * at the 1% level.

for girls by 22.6% using village controls and24.6% using village fixed effects. The negativeeffect on boys’ school enrollment is larger whenwomen are borrowers although it is statisticallyinsignificant in Model (3). One might surmisethat this could be an indication of gender pref-erence by parents. However, Hausman-type testsdo not reject the equality of the coefficientsbetween the sexes of the borrower. Overall,the (negative) effects of credit on schooling arelarger than the effects on child labor, suggestingthat households might be under-reporting childlabor or some children are away from schooland doing nothing.

Table 5 shows the treatment-on-treated effectusing a binary participation indicator as thetreatment variable. The estimated effect using2SLS is identical to the indirect least squaresestimate obtained from taking the ratio of thereduced-form coefficients, because we are esti-mating a just identified equation. The results arequalitatively similar to those in Table 4. Girls’education continues to be affected adverselyby parental participation in microcredit pro-grams whether credit is obtained by men orwomen. Using village fixed effects, for example,we find that women’s microcredit borrowingincreases the probability of girls’ child labor

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 13

TABLE 6Impact Estimates on Children’s School Achievement

Boys Girls

GradeCompletion

EducationGap Grade-For-Age

GradeCompletion

EducationGap Grade-For-Age

Female borrowerTreatment

effect−0.196 (0.617) −0.092 (0.611) −21.64 (12.912)† −2.95 (0.738)* 2.752 (0.696)* −48.393 (16.089)*

Controlfunction

0.158 (0.624) 0.143 (0.617) 22.29 (13.283)† 2.963 (0.745)* −2.755 (0.705)* 49.405 (16.256)*

Male borrowerTreatment

effect−0.423 (0.846) 0.043 (0.817) −23.727 (17.603) −3.77 (0.961)* 3.39 (0.923)* −72.49 (21.299)*

Controlfunction

0.327 (0.787) 0.047 (0.765) 25.084 (16.855) 3.537 (0.958)* −3.199 (0.921)* 69.1 (21.373)*

Notes: All the results are estimated using the control function method. The regressions include child- and household-specific covariates, village-level covariates, and district fixed effects. The coefficients and the standard errors of treatmenteffects are multiplied by the average credit borrowed by male and female borrowers. Standard errors presented in theparentheses are corrected for clustering at the village level using the formulas by Liang and Zeger (1986).

Coefficients with † are significant at the 10% level and those with * at the 1% level.

by 28% and decreases the probability of theirschool enrollment by 47%. The magnitude ofthe impact estimates is similar in case the bor-rower is a man. The corresponding coefficientestimates for child labor for boys are not statis-tically significant and have mixed signs. Overall,binary participation measures generate consider-ably larger coefficient estimates for girls. How-ever, these results are only indicative as theydo not take into account the variation of treat-ment intensity, and assume the treatment effectto be the same for all children in the treatmentgroup.

The estimates using educational achievementmeasures are reported in Table 6.17 The negativecoefficients for grade completion and grade-for-age, and the positive coefficient for educationgap all imply that participation in microcre-dit programs adversely affects children’s educa-tional achievement. Once more, girls are moreadversely affected than boys: coefficient esti-mates for girls are larger than for boys andstatistically significant at the 1% level. Women’sparticipation in microcredit reduces girls’ edu-cation by about 3 years of schooling, whereasthe corresponding decrease for boys’ educationis about 0.2 years of schooling. Men’s partic-ipation has a larger negative impact on girls’

17. The results reported in this section use the controlfunction approach discussed in the previous section. Theresults using the IV/2SLS approach are similar to thosereported here and are available upon request from theauthors.

grade completion—a reduction of 3.8 years inschooling compared with the girls from the con-trol group. The effects on boys’ school achieve-ment are not statistically significant in general.The results from all three measures of educa-tional achievement are similar, which are notparticularly surprising because the three mea-sures are likely to be highly correlated.

In further tests reported in an earlier versionof this article (Islam and Choe 2009), we alsofind that younger children are more adverselyaffected than their older siblings and the childrenof poorer and less educated households areaffected more adversely. These results indicatethat microcredit to the poorest of the poorhouseholds neither alleviates the problem ofchild labor nor improves children’s schoolingas poorer households among the clients aremore adversely affected than their less poorcounterparts.

A possible explanation for the adverse effectof microcredit on children’s schooling is thatmicrocredit increases demand for labor in house-hold enterprises set up with microcredit, whichmay cause children’s time to be diverted awayfrom school into household enterprises. Weexamine this issue below. We classify a child’scurrent status into five different categories basedon the detailed occupational information col-lected during the survey. They are (a) self-employment activity (in household enterprise),(b) agricultural activity, (c) day labor, (d)service-related activity, and (e) student (enrolled

14 ECONOMIC INQUIRY

TABLE 7Multinomial Logit Model for Children’s

Work/School Status

Child Status Coefficient Marginal Effect

Self-employment activity 2.03 (0.71)* 0.266 (0.095)*Agriculture −6.48 (1.73)* 0.000 (0.0000)*Day laborer 1.33 (1.07) 0.0000 (0.001)Service-related activity 2.42 (4.50) 0.0000 (0.0000)Enrolled in school — −0.266 (0.096)*

Notes: The regressions include child- and household-specific covariates, village-level covariates, and districtfixed effects. Standard errors presented in the parenthesesare corrected for clustering at the village level using theformulas by Liang and Zeger (1986).

Coefficients with * are significant at the 1% level.

in school).18 We run a multinomial logit modelwhere the parameter of interest is the coefficientcorresponding to the instrumented credit vari-able obtained from Equation (4). Table 7 reportsthe odds ratios and corresponding marginaleffects of the treatment variable. The resultsshow that, for children from treated households,the odds of being in self-employment activitiesinstead of being in school are more than dou-bled. The corresponding marginal effect indi-cates that children from treated households are26.6% more likely to work in self-employmentactivities than those from nontreated households.The odds ratio is higher and negative for agri-cultural activity. However, the correspondingmarginal effect is economically insignificant.The rest of the coefficient estimates is not sta-tistically significant. Finally, the marginal effectfor the student status implies that children fromtreated households have a 26.6% lower chanceof being enrolled in school than those from non-treated households. Overall, these results sup-port the explanation that children from treatedhouseholds are more likely to work in householdenterprises set up with microcredit.

V. SUMMARY AND CONCLUSION

This article has studied the impact of accessto microcredit on children’s education and childlabor using a new and large data set consist-ing of treated and nontreated microcredit house-holds from Bangladesh. The results overwhelm-ingly indicate that household participation in

18. Only a very few households (less than 1%) reportedthat their children are only doing household work. Therefore,we do not include them in our regression.

microcredit programs has adverse effects onchildren’s schooling, which are especially pro-nounced for girls. It appears that children takenout from school are more likely to work inhousehold enterprises that are set up with micro-credit than in other types of work. Overall,our results suggest that care needs to be takenin assessing the effectiveness of microcreditprograms. Although microcredit programs canalleviate poverty and contribute to rural econ-omy in the short term, they can also result inunintended consequences of adversely affect-ing children’s schooling, which could exacerbatepoverty in the longer term. An additional con-cern relates to the gender-asymmetric impactof access to microcredit. Government policiesaimed at rectifying gender imbalance in edu-cation may turn out to be less effective in thepresence of many active microcredit programs.

A number of policies can be adopted to miti-gate the adverse effect on child labor and school-ing so that microcredit can benefit both currentand future generations (Wydick 1999). At thelevel of MOs, the gestation period betweenactual loan disbursement and the start of repay-ment can be extended. This allows many house-holds to invest in suitable investment projectswhere they may find a greater balance betweenemploying children at household enterprises andsending them to school. Reduced interest ratesand longer repayment periods can also helphouseholds to become less myopic. In addi-tion, increases in the size of credit allowingemployment of external labor can take the bur-den off from households to resort to child labor.The latter suggests that MOs may eventuallyneed to look further and consider financing ruralenterprises at the village level rather than atthe household level. These measures that aredirected at MOs alone are by no means suffi-cient in reducing child labor and improving childschooling. They need to be complemented bypolicies that directly target children’s education,of which the essence is to increase the returnon education perceived by parents. In summary,microcredit programs need to be complementedby other policies to tackle the multiple goalsof poverty reduction, human capital formation,and social development. It is, however, to benoted that we estimate here only the contem-poraneous effect of microcredit. It is entirelypossible (and plausible) that the long-run resultscould be opposite. That is, the higher incomeopportunities provided by microcredit over timemay lead to an income effect toward less child

ISLAM & CHOE: CHILD LABOR, SCHOOLING, AND MICROCREDIT 15

labor and more education that may outweigh thesubstitution effect. Therefore, our results shouldbe interpreted with caution, and perhaps a panel

data set over a longer term participation inmicrocredit can address the issue concerningshort- versus long-run effects.

APPENDIX

TABLE A1Descriptive Statistics of Village-Level Controls

VariableControl

Village (I)Program

Village (II)Difference

III = (II–I)

Education facilitiesPrimary school 90.91 86.25 −4.66Secondary school 27.27 31.25 3.98Maktab/Madrasa (religious school) 81.82 90.00 8.18

Health facilitiesUnion health center 10 17.5 7.5Allopathic doctor 50 42.5 −7.5Homeopathic doctor 40 38.75 −1.25

Transport, communication, and infrastructureElectricity connection 17 26 9.0Presence of pucca road 10.6 34.8 24.2Distance to nearest thana (km) 11.91 7.14 −4.77Presence of grocery market 18.2 22.5 4.3Presence or absence of frequent haat (big market) 27.3 32.5 5.2Presence of bus stand 9.1 15 5.9Presence of post office 18.2 20 1.8Presence of telephone office 9.1 6.3 −2.8Presence of Union Parishad (local Government) office 18.2 13.8 −4.4

Irrigation equipmentNumber of low lift pumps 0.27 0.44 0.16Number of shallow tube wells 11.82 12.13 0.31Number of hand tube wells for drinking water 68 78.04 10.04

Credit-related variablesPercentage of crop received by land owner in sharecropping 49.55 47.53 −2.02Number of people who provide advances against crops 2.73 3.85 1.12Number of small credit/savings groups 0.91 0.76 −0.15

Notes: The first two columns show the mean of each variable for the control and treatment villages. The third column isthe difference between the two, with differences that are statistically significant at less than 5% marked bold.

REFERENCES

Angrist, J. “Estimation of Limited Dependent Variable Mod-els with Dummy Endogenous Regressors: SimpleStrategies for Empirical Practice.” Journal of Business& Economic Statistics, 19(1), 2001, 2–16.

Baland, J.-M., and J. A. Robinson. “Is Child Labor Ineffi-cient?” Journal of Political Economy, 108(4), 2000,663–79.

Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan. “TheMiracle of Microfinance? Evidence from a Random-ized Evaluation.” Working Paper, Department of Eco-nomics, MIT, 2009.

Basu, K., S. Das, and B. Dutta. “Child Labor and HouseholdWealth: Theory and Empirical Evidence of an Inverted-U.” Journal of Development Economics, 91(1), 2010,8–14.

Behrman, J., and J. Knowles. “Household Income and ChildSchooling in Vietnam.” World Bank Economic Review,13(2), 1999, 211–56.

Dehejia, R., and R. Gatti. “Child Labour: The Role ofIncome Variability and Access to Credit across Coun-tries.” Economic Development and Cultural Change,53(4), 2005, 913–32.

Doepke, M., and F. Zilibotti. “The Macroeconomics ofChild Labor Regulation.” American Economic Review,95(5), 2005, 1492–524.

Edmonds, E. “Child Labour and Schooling Responses toAnticipated Income in South Africa.” Journal of Devel-opment Economics, 81(2), 2006, 386–414.

. “Child Labour,” in Handbook of Development Eco-nomics, Vol. 4, edited by T. P. Schultz and J. Strauss.Amsterdam, The Netherlands: North-Holland ElsevierScience, 2007, 3607–709.

Garen, J. “The Returns to Schooling: A Selectivity BiasApproach with a Continuous Choice Variable.” Econo-metrica, 52(5), 1984, 1199–218.

Hazarika, G., and S. Sarangi. “Household Access to Micro-credit and Child Labour in Rural Malawi.” WorldDevelopment, 36(5), 2008, 843–59.

16 ECONOMIC INQUIRY

Islam, A., and C. Choe. “Child Labour and School-ing Responses to Access to Microcredit in RuralBangladesh.” MPRA Paper 16842, University Libraryof Munich, Germany, 2009.

Jacoby, H. “Borrowing Constraints and Progress throughSchool: Evidence from Peru.” Review of Economicsand Statistics, 76(1), 1994, 151–60.

Jacoby, H., and E. Skoufias. “Risk, Financial Markets, andHuman Capital in a Developing Country.” Review ofEconomic Studies, 64(3), 1997, 311–35.

Karlan, D., and J. Zinman. “Lying About Borrowing.” Jour-nal of the European Economic Association, 6(2–3),2008, 510–21.

. “Expanding Microenterprise Credit Access: UsingRandomized Supply Decisions to Estimate the Impactsin Manila.” Review of Financial Studies, 23(1), 2009,433–64.

Liang, K., and S. Zeger. “Longitudinal Data Analysis UsingGeneralized Linear Models.” Biometrika, 73(1), 1986,13–22.

Maldonado, J., and C. Gonzalez-Vega. “Impact of Microfi-nance on Schooling: Evidence from Poor Rural House-holds in Bolivia.” World Development, 36(11), 2008,2440–55.

Patrinos, H., and G. Psacharopoulos. “Family Size, School-ing and Child Labour in Peru—An Empirical Anal-ysis.” Journal of Population Economics, 10(4), 1997,387–405.

Pitt, M., and S. Khandker. “The Impact of Group-BasedCredit Programs on Poor Households in Bangladesh:Does the Gender of Participation Matter?” Journal ofPolitical Economy, 106(5), 1998, 958–96.

Ranjan, P. “An Economic Analysis of Child Labour.” Eco-nomics Letters, 64(1), 1999, 99–105.

. “Credit Constraints and the Phenomenon of ChildLabour.” Journal of Development Economics, 64(1),2001, 81–102.

Ravallion, M. “Evaluating Anti-Poverty Programs,” inHandbook of Development Economics, Vol. 4, editedby P. Schultz and J. Strauss. Amsterdam, The Nether-lands: North-Holland, 2008, 3787–846.

Ravallion, M., and Q. Wodon. “Does Child Labour DisplaceSchooling? Evidence on Behavioural Responses toan Enrolment Subsidy.” Economic Journal, 110(462),2000, C158–75.

Rivers, D., and Q. Vuong. “Limited Information Estimatorsand Exogeneity Tests for Simultaneous Probit Mod-els.” Journal of Econometrics, 39(3), 1988, 347–66.

Van der Klaauw, W. “Estimating the Effect of FinancialAid Offers on College Enrolment: A Regression-Discontinuity Approach.” International EconomicReview, 43(4), 2002, 1249–87.

Vella, F. “A Simple Estimator for Simultaneous Modelswith Censored Endogenous Regressors.” InternationalEconomic Review, 34(2), 1993, 441–57.

World Bank. Microfinance in South Asia: Towards a Finan-cial Inclusion for the Poor. Washington, DC: TheWorld Bank, 2006.

Wydick, B. “The Effect of Microenterprise Lending onChild Schooling in Guatemala.” Economic Develop-ment and Cultural Change, 47(4), 1999, 853–69.

Zohir, S., S. Mahmud, B. Sen, and M. Asaduzzaman. Mon-itoring and Evaluation of Microfinance Institutions.Dhaka: Bangladesh Institute of Development Stud-ies, 2001. Accessed April 26, 2011. http://www.pksf-bd.org.