EdSchools and Skills in Developing Countries: Education Policies and Socioeconomic Outcomesucational Policies and Socioeconomic

8/13/2019 EdSchools and Skills in Developing Countries: Education Policies and Socioeconomic Outcomesucational Policies an

1/48

American Economic Association

Schools and Skills in Developing Countries: Education Policies and Socioeconomic OutcomesAuthor(s): Paul GlewweSource: Journal of Economic Literature, Vol. 40, No. 2 (Jun., 2002), pp. 436-482Published by: American Economic AssociationStable URL: http://www.jstor.org/stable/2698384

Accessed: 13/02/2010 13:55

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at

http://dv1litvip.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless

you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you

may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at

http://www.jstor.org/action/showPublisher?publisherCode=aea.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed

page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of

content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms

of scholarship. For more information about JSTOR, please contact [email protected].

American Economic Associationis collaborating with JSTOR to digitize, preserve and extend access toJournal

of Economic Literature.

http://dv1litvip.jstor.org
http://www.jstor.org/stable/2698384?origin=JSTOR-pdfhttp://dv1litvip.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/action/showPublisher?publisherCode=aeahttp://www.jstor.org/action/showPublisher?publisherCode=aeahttp://dv1litvip.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/2698384?origin=JSTOR-pdf


2/48

Journal of Economic LiteratureVol. XL (June2002), pp. 436-482

S c h o o l s n d k i l l s n DevelopingCountries Education o l i c i e s n d

ocioeconomicutcomesPAUL GLEWWE1

1. IntroductionECONOMISTS HAVE studied economicgrowth and development since AdamSmith set out to explain the nature andcauses of the wealth of nations. In the1950s and 1960s, Gary Becker, JacobMincer, T. W. Schultz, and others turnedeconomists' attention to education andthe role it plays in a variety of economicphenomena. More recently, economistshave linked these two literatures, exam-ining the impact that education canhave, and in some countries already hashad, on economic growth (Robert Lucas1988; Robert Barro 1991; N. GregoryMankiw, David Romer, and David Weil1992).The proposition that a higher level ofeducation promotes economic growthand development suggests that govern-

ments in developing countries shouldimplement policies that raise educa-tional attainment, since growth and de-velopment are objectives of nearly alldeveloping countries. Thus manyeconomists and international organi-zations argue that investment in educa-tion is a policy priority (Becker 1995;Eric Hanushek 1995; UNDP 1990;World Bank 2001). Yet, until very re-cently, they have said little about whatgovernments in developing countriescan do to raise educational attainment.This lack of advice does not implythat schools in developing nations arealready operating effectively and effi-ciently. To the contrary, there is ampleevidence that many schools in thesecountries are not very effective, andoperate far from any conceivable effi-cient frontier (Marlaine Lockheed andAdriaanVerspoor 1991; Ralph Harbisonand Hanushek 1992; Hanushek 1995;Glewwe 1999a). It is also not the casethat governments and schools know howto improve educational outcomes butchoose not to do so because such ac-tions would not be in their interest.While there are situations where teach-ers and officials favor their interestsover those of students, it is also clear

1 University of Minnesota and the World Bank. Ithank the folTlowing people for comments, discus-sions, and/or clarification on their papers: BruceFuller, Nancy Gillespie, Eric Hanushek, Em-manuel Jimenez, Dean Jolliffe, Cigdem Kagit-cibasi, Geeta Kingdon, Michael Kremer, JuliaLane, Berk Ozler, Lant Pritchett, and Jee-PengTan. I am also grateful to John McMillan andthree anonymous referees for very detailed anduseful comments. The findings, interpretations,and conclusions expressed in this paper are en-tirely those of the author. They do not necessarilyrepresent the views of the World Bank.436


3/48

Glewwe: Schools and Skills in Developing Countries 437that ministries of education in develop-ing countries often are not sure what todo to improve their education systems(Lockheed and Verspoor 1991, p. 39).This unsatisfactory state of affairs isall the more glaring given the stagger-ing amounts of money involved; eachyear the governments of developingcountries spend about $260 billion oneducation.2Finally, this lack of knowledge onhow to operate schools most effectivelydoes not reflect lack of interest on thepart of researchers. Many studies haveaddressed these issues, but most ofthem suffer from serious shortcomings.Recently, some careful studies haveprovided more reliable findings on spe-cific policies and programs. The pur-pose of this paper is to examine thisrecent work in detail.More specifically, this paper hasthree objectives. The first is to reviewthe literature on ;the relationship be-tween school and teacher characteris-,tics, broadly defined, and the acquisi-tion of cognitive skills. The questionaddressed is: What school policies aremost cost-effective in producing studentswith particular cognitive skills, such asliteracy and numeracy? The second ob-jective is to examine the relationshipbetween schooling and labor productiv-ity, with emphasis on the relationshipbetween basic cognitive skills and laborproductivity. Knowledge of the impactof different skills on income and onother socioeconomic outcomes couldhave policy implications for school cur-ricula. For example, if literacy wereidentified as more important than, say,scientific knowledge in determining fu-

ture income, it may be desirable to re-duce classroom time devoted to sciencein order to increase the time devoted tolanguage skills. The third objective is toinvestigate the relationship betweencognitive skills and socioeconomic out-comes other than labor productivity,such as the impact of schooling onwomen's fertility and on adult and childhealth. The three main sections of thispaper cover each of these objectives inturn. A final section summarizes thefindings and provides recommendationsfor future research.Before proceeding, a few commentsare needed on the scope of the paper.First, it does not address the issue ofwhether government subsidies for edu-cation can be justified in terms of stan-dard economic theory. Other papershave argued that this is the case (see,inter alia, Daron Acemoglu 1996, andRoland Benabou 1996), and this paperneed not take a position on this issue.Second, while the paper considers the

issue of whether private schools aremore efficient than public schools, italso considers what governments can doto improve the operation of publicschools even though private schoolsmay be more efficient. The reason forthis is simple realism-many govern-ments favor public schools for a varietyof "noneconomic" reasons (examplesare perceived equity benefits and politi-cal objectives such as promoting a cur-riculum that gives students a national,as opposed to an ethnic or regional,identity) and thus policy advisors havelittle choice but to accept this con-straint and focus on ways to improvepublic schools.A final limit on the scope of this pa-per concerns the educational outcomesexamined. Schooling provides childrenwith many benefits. The most obviousare cognitive skills such as literacy,numeracy, scientific knowledge, and

2 This figure is calculated by taking the totalGNP of low- and middle-income countries in1999, which amounted to $6,311 billion, and mul-tiplying it by the (average) government expendi-tures on education as a percentage of GNP, whichwas 4.1 percent. Both of these figures are fromWorld Bank(2001).


4/48

438 Journal of Economic Literature, Vol. XL (June 2002)advanced thinking skills. Schooling canalso provide social skills and (internal-ized) values that may help children suc-ceed in the adult world. Lastly, prestigemay be attached to particular levels ofeducation, which may enable one tofind a better job or a "better" spouse. Athorough study of all these benefitscould double the length of this paper.To keep the paper to a reasonablelength, it focuses on the basic cognitiveskills that school curricula are designedto impart. However, occasional refer-ence is made to other benefits ofschooling.

2. School Characteristicsandthe Acquisition of Cognitive SkillsThis paper approaches education is-sues from an economic perspective.That is, it takes the position that amodel of "rational" behavior is neededto ensure that proper econometric andstatistical methods are used to estimate

the impact of school characteristics andpolicies on educational outcomes, andof the impact of schooling and cognitiveskills on socioeconomic outcomes. Inparticular, explicit models of human be-havior provide substantial insight intowhether assumptions underlying spe-cific econometric methods are satisfied.If a plausible model suggests that someassumptions are not satisfied, empiricalfindings based on those methods maybe invalid. The model may also suggesthow to test the econometric assump-tions, and what estimation method canbe used if they fail to hold. The sectionfirst presents such a model and exam-ines its implications for empirical analy-sis. The model is not intended to be thedefinitive model of schooling, rather itis a simple yet plausible model that illu-minates several econometric issues. Af-ter presenting the model and its impli-cations for empirical work, I examine

several recent studies of the impact ofschool and teacher characteristics onlearning.2.1 A Simple Model of SchoolingChoices

Assume that parents make decisionsfor their children and that their objec-tive is to maximize a utility functionthat has two arguments: consumption ofgoods and services and child cognitiveskills. For simplicity, assume that thereare two time periods and only one childper family.3 In period 1, a child may at-tend school, work, or both. If both, thechild first goes to school, and works af-ter schooling is completed (going toschool first is optimal in most cases; seeGlewwe 1999a, ch. 3). In period 2, thechild becomes an adult and works.When a child works in either time pe-riod, part or all of the child's earningsmay be given to his or her parents.A utility function that takes parents'consumption (C) in periods 1 and 2and child cognitive skills (A) as itsarguments is:

U=C1+6C2+GA, (1)where 6 is a discount factor for futureconsumption and 6Tindicates parentaltastes for educated children (higher val-ues imply greater utility from educatedchildren). Parents value educated chil-dren for two distinct reasons: educatingchildren can increase parents' consump-tion, and educating children directlyaffects parents' utility (through 6i).A simple production function showshow cognitive skills, A, are acquired:A= cf(Q)g(S), (2)where oc is the "learning efficiency" ofthe child, Q is school quality, and S is

3 This implies that the number of children afamily has is exogenous. Glewwe (1999b) consid-ers the possihility that the number of children is achoice variahle. That paper develops in moredetail the model presented here.


5/48

Glewwe: Schools and Skills in Developing Countries 439years of schooling. The functions f and gare increasing in Q and S, respectively. Achild's learning efficiency, CC, epresentsseveral different factors, such as innate(genetically inherited) ability, child moti-vation, and parental motivation and ca-pacity to help children with their school-work. For simplicity, all these factors arecombined into oc.Parents' consumption in each timeperiod is given by:

Ci=Yi-pS+(I-S)kYc (3)C2= Y2+kYc (4)

where p is the price of schooling,4 Yi andY2 are parental income in periods 1 and2 respectively, Yc is the child's incomewhen working, and k is the fraction ofthat income given to the parents. Thelast term in (3), (1 - S)kYc, requiressome, explanation. S has been rescaled tobe the fraction of time spent in school bythe child in time period 1. The remain-ing time in the first period, 1 - S, isspent working. This is purely for nota-tional convenience; however, to keep thevocabulary simple, S is still called "yearsof schooling."Equations (3) and (4) rule out bor-rowing and saving; the only way totransfer income between periods 1 and2 is to alter investments in children'seducation. This assumption is made forsimplicity. In general, introducing bor-rowing and saving would reduce par-ents' incentive to invest in their chil-dren's education. Yet it would notcompletely eliminate this incentive be-cause almost all investments are risky,so most parents would diversify theirinvestments among several differentalternatives, including their children'seducation.

Equation (5) completes the model,relating child cognitive skills to childincome:YC=2tA, (5)where X is the productivity of cognitiveskills in the labor market.Substitution of (2) into (5), of (5) into(3) and (4), and of (2) - (4) into (1) ex-presses parents' utility as a function ofyears of schooling (S) and school quality(Q):

U=Y1-pS+6Y2 (6)+ ((1 - S + 6)kt + G)cf(Q)g(S)

Consider first the case where schoolquality is exogenous, so that S is the onlychoice variable. It is straightforward toderive the impacts of changes in themodel's various parameters on the opti-mal (utility-maximizing) value of years ofschooling (see Glewwe 1999b), all ofwhich are intuitively plausible. Optimalyears of schooling (and thus the child'scognitive skills) is an increasing functionof: the child's learning efficiency (c),school quality (Q), the relative weight (6)parents give to future consumption, andparental tastes for schooling (6). Optimalyears of schooling decreases when theprice of schooling (p) rises. Finally, opti-mal years of schooling is likely, thoughnot certain, to rise when parents expectto receive a larger proportion (k) of theirchildren's income from working andwhen the value of cognitive skills in thelabor market (X) is higher. The intuitionfor this ambiguity is that although ahigher value of cognitive skills in the la-bor market (X) raises the value of school-ing, it also makes time out of school(which increases when years of schoolingdeclines) more valuable. The sameargument applies to the proportion ofchildren's income going to parents (k).The model is easily extended to allowparents to choose school quality (Q).Assume that parents choose school

4The child's consumption while in school canbe included in p, while the child finances his orher own consumption from Yc when working.Strictlyspeaking, 'his assumes that child consump-tion while in school is exogenous, perhaps set ylocal cultural norms.


6/48

440 Journal of Economic Literature, Vol. XL (June 2002)quality, but higher quality implies ahigher price:

p =PoQ (7)where po is the "base"price of schooling.While (7) may appear to impose an arbi-trary linear functional form (why shouldthe price double if quality doubles?), thisis not the case. One should interpret Qas an index of expenditures on quality.Whether, say, doubling expendituresdoubles the impact of school quality onlearning, that is, doubles f(Q), dependson the functional form off.Replacing p with poQ in (6) yieldsan expression to be maximized withrespect to S and Q:U = YI - poQS + Y2 (8)

+ ((1 - S + 6)ki + 6)xf(Q)g(S)To simplify derivation of the impacts ofchanges in the various parameters on(optimal values of) S and Q, one moreassumption is needed on the functionalforms off and g. A convenient and plau-sible assumption is that f(Q)= Q andg(S) = SY. Different values of ,B and yyield a wide range of the shapes for bothfunctions. Both ,Band y must be positiveto ensure thatf and g are increasing in Qand S, respectively. While this assump-tion implies that the following results arenot completely general, the model is stilluseful because it demonstrates the im-plications of plausible assumptions forempirical analysis.

Using these functional form assump-tions, one can show (see Glewwe1999b) that the optimal values (denotedby asterisks) of S and Q are:5S*=(y- f)(1+6+ 6/kn)/(1+ y-D) (9)

= (ackn/po)(y- P)Y-1 (10)((1 + 6 + 6/k)/(l +1y- f))Y.

The optimal level of cognitive skills (A)is obtained by inserting (9) and (10) into(2).These optimal levels of years ofschooling (S) and school quality (Q) areintuitively plausible. Both increasewhen parents put more weight (6) onfuture consumption and when parentshave higher tastes for schooling (6).School quality (Q) increases with learn-ing efficiency (ot) but decreases as thebase price of schooling (po) rises. A lessplausible result is that years of school-ing depends neither on the base priceof schooling nor on learning efficiency.This reflects the functional forms of fand g, but is not necessarily unreason-able. Basically, when the base pricefalls or child learning efficiency rises,parents shift to higher school quality,raising their children's cognitive skillswithout changing years of schooling. Bychoosing higher quality instead of moretime in school, parents avoid a cost ofthe latter: reduced child working timein period 1; see equation (3). In devel-oping countries, grade repetition ishigh, so this can take the form ofreduced grade repetition, raising thehighest grade attained without changingyears of schooling.A second apparently counterintuitiveresult is that increases in the propensityof children to support their parents (k)and in the market return to cognitiveskills (X) decrease years of schooling.Yet these results may be reasonable;one response to such changes is tochoose higher school quality and reducetime spent in school to increase the timethe child spends working in time period1.6 Of course, other functional forms forf

5 Note that S* > 0 only if y > P3.ntuitively, if y


7/48

Glewwe: Schools and Skills in Developing Countries 441and g could lead to different impacts ofk and X on years of schooling.This simple model produces many in-tuitively plausible results. It also pro-vides some insights that go beyond sim-ple intuition. For example, when schoolquality is exogenous it is not necessarilyintuitive that parents who give greaterweight to future consumption will sendtheir children to school longer, even af-ter controlling for parental tastes forschooling. Even less obvious is the re-sult that higher returns to cognitiveskills do not necessarily increase yearsof schooling (because they raise the op-portunity cost as well as the benefit ofan additional year of school). Whenschool quality is also a choice variable,the main insights beyond simple intui-tion work through the fact that years ofschooling and school quality are alter-native inputs in the production of cog-nitive skills. This explains why the(base) price of schooling has no effecton time in school; the best response toa change in this price may be to adjustschool quality, holding years of school-ing constant (although the highestgrade attained may rise due to lessgrade repetition). While the absence ofany effect on years of schooling reflectsfunctional form assumptions, under al-most any functional forms one shouldfind that the impact of the price ofschooling on years in school diminisheswhen school quality becomes endoge-nous. Similarly, the increase in years ofschooling due to an increase in a child'slearning ability is smaller when parentshave the option of increasing schoolquality. A final insight from this modelwhen school quality is endogenous isthat the price of schooling per year ofenrollment at the chosen school, poQ, is

an endogenous variable; econometricanalyses should not treat school pricesat the school attended as exogenous.2.2 Implications of the Modelfor EconometricAnalysis

The model presented above providesa useful framework for discussing sev-eral issues concerning estimation of theimpact of school characteristics on cog-nitive skills. Most empirical studies thatattempt to estimate the cognitive skillsproduction function given in (2) assumelinear functional forms to simplifyestimation.7 Thus (2) becomes:A=,0o+JtlS+ t2a + t3Q+e (2')

where the , coefficients are unknownparameters to be estimated. The sim-plest interpretation of the residual terme is that it reflects measurement error inA, but of course it could reflect omittedvariables,or measurementerrorpertainingto cX,Q, and even S.The specification of school quality in(2') is clearly oversimplified. It is morerealistic, and more useful for policyanalysis, to decompose school qualityinto a function or index of the differentschool characteristics that promotelearning:8A= go+ .iS + J2aX+ IQI +r22(2t2 )+ -.. + tnQn + e.

In (2"), Q is replaced by an index of ndistinct school characteristics that affectlearning. Policymakers would like toknow the magnitude of the various t'sbecause such estimates can be combinedwith data on the costs of those same

decline, this loss in parental income is outweighedby the increased income from the child workinglonger in period 1.

Linearitycan follow from the modelresentedabove. Taking the logarithm of both sides of (2)and assuming exponential functional forms for fand g, such as fCQ) = QP and g(S) = SYyields anequation that is linear in the logarithms of thevariables.8 This linear function of the school charac-teristics can be made more realistic by addingquadratic and interaction terms. To simplify theexposition, these terms are omitted.


8/48

442 Journal of Economic Literature, Vol. XL (June 2002)school characteristics to assess the cost-effectiveness of each characteristic inpromoting learning. Indeed, this infor-mation is precisely what is needed toanswer the first of the three questionsaddressed by this paper, namely whichschool policies are most cost-effectivefor raising students' cognitive skills.A child's learning efficiency, oc, isalso multidimensional. Some factorsthat raise learning efficiency, such asparental education, are easily observed,while collecting data on others is verydifficult, if not impossible. Thus (2")can be rewritten as:A = ,o + p1S + pI(XI+ p20X2 + (2"')pm?Cm + t1Q1 +T2Q2 + ... + tnQn + U.In this equation the observed compo-nents of learning efficiency are specifiedas 0C1,OC2,tc. In contrast, the unobservedcomponents must be combined with e,which yields u, a residual term that rep-resents both random measurement errorin A and the impact of unobserved as-pects of learning efficiency (oc)on cogni-tive skill acquisition (A). In fact, u alsorepresents unobserved school qualitycharacteristics, as well as measurementerrorin S and in the ccand Q variables.Examples of difficult-to-observe learn-ing efficiency variables are the child'sinnate ability and motivation, and parents'willingness and capacity to help theirchildren with schoolwork. One can tryto measure some of these factors (suchas using an IQ test to measure innateability and using parental schooling toindicate parents' ability to assist theirchildren), but it is unlikely that one canmeasureall of them. Indeed, it is not clearthat innate ability can be measured; anytest that claims to do so (in the sense ofmeasuring a genetic endowment) almostalways reflects environmental factors(AmericanPsychologicalAssociation1995).One may be able to avoid this problem

by using data on twins (for example,Jere Behrman, Mark Rosenzweig, andPaul Taubman 1994), but such datafrom developing countries are very rare.Many aspects of school quality arealso unobserved. Most data sets haveonly a small number of school qualityvariables; many easy-to-observe schoolcharacteristics are often omitted whenthe data are collected. In addition,some aspects of school quality are in-herently difficult to measure, such asteachers' interpersonal skills and moti-vation, and the management skills ofschool principals.Suppose that (2"') is estimated usingordinary least squares (OLS). Ofcourse, the estimated parameters areunbiased only if the residual, u, is un-correlated with S and the various Q'sand oc's.Yet the model presented in theprevious subsection shows that suchcorrelation is very likely; in equation(10), higher learning efficiency (oc) in-creases school quality (Q), implyingthat u, which contains the unobservedcomponents of o, is positively corre-lated with the various Q's. Thus esti-mates of the associated parameters (t's)will be biased upward. The estimatedimpacts of observed learning efficiencyvariables are also likely to be biased,since those variables are usually corre-lated with unobserved aspects of learn-ing efficiency. Most empirical studies dolittle or nothing to avoid this problem.

If school quality were exogenous, onemight think that these estimation prob-lems could be avoided because coeffi-cients on any exogenous variables wouldbe unlikely to be biased. Yet econometrictheory shows that correlation betweenany variable and the error term is likelyto lead to biased estimates of all pa-rameters, not just the parameter of vari-ables with which the error is correlated(Russell Davidson and James McKinnon1993, pp. 211-15). In the simple model


9/48

Glewwe: Schools and Skills in Developing Countries 443given above, years of schooling is posi-tively correlated with learning abilitywhen school quality is exogenous, whichwill lead to biased estimates for theschool-quality parameters.Moreoever, school quality is likely tobe endogenous. Even in rural areas oflow-income countries, where villagesoften have only one school and are toofar apart for children to attend school ina neighboring village, parents may beable to influence school quality. First,they may directly alter the quality ofthe sole local school through theparent-teacher association (PTA) orthrough political connections. Second,they may send their children to livewith relatives (allowing them to attenda nonlocal school) or to a boardingschool. About 19 percent of secondarystudents in rural Peru live away fromtheir families (Paul Gertler and Glewwe1990), and the same holds for 27 per-cent of middle-school students inGhana (Glewwe and Hanan Jacoby1994). Third, families with higher tastesfor educated children may migrate toareas with better schools, a commonoccurrence in the United States.When parents can alter school qual-ity, overestimation is possible due topositive correlation between unob-served components of a child's learningefficiency and school quality. Endoge-nous school quality can also lead tounderestimation. Even when parentscannot alter school quality, qualitycould be correlated with the error termif governments provide better schoolsto areas with unobserved educationproblems (Mark Pitt, Rosenzweig andDonna Gibbons 1993). These unmea-sured problems would also be relegatedto u in equation (2"'), producing nega-tive correlation between the error termand the school quality variables (Q's)and thus underestimating the impact ofschool quality. On the other hand, gov-

ernments are just as likely (and somewould argue much more likely) to placebetter schools in areas that already havegood education outcomes, since bothautocratic and democratic rulers oftenderive political support from elitegroups (World Bank 2001). For empiri-cal evidence on this point, see NancyBirdsall (1988) for Brazil, and Behrmanand James Knowles (1999) for Vietnam.In theory, instrumental variablemethods can resolve this problem, butit is difficult to find plausible instru-ments. One possible instrument foryears of schooling is the price of school-ing, which should affect learning onlyby affecting years of schooling. Alterna-tively, one could estimate (2"') for asingle grade to remove variation in S.Yet both approaches have problems.First, the prices observed in the datafor the schools attended are not the po'sof equation (7) but poQ, which isendogenous if Q is endogenous. Inparticular, it will be correlated with u,invalidating its use as an instrument.Second, if some children in the relevantage range are not in school, the remain-ing children (whether in one or severalgrades) are not a random sample of thepopulation. Intuitively, communitieswith high-quality schools will keep chil-dren in school longer, leading to a stu-dent population with lower averagelearning efficiency (more "less-efficient"children stay in school). In this case uin (2"') will be negatively correlatedwith school quality, leading to underes-timation of the impact of school qualityon learning. Third, no data set includesevery component of school quality, andobserved components may be positivelycorrelated with unobserved components(because "good" schools are often goodin many ways, only some of which areobserved). Again, unobserved aspects ofschool quality are part of the residualin (2"'), causing u to be positively


10/48

444 Journal of Economic Literature, Vol. XL (June 2002)correlated with observed school-qualityvariables and causing the t parametersto be overestimated.

A final difficulty in empirical work ismeasurement error in the explanatoryvariables, both S and the various Q vari-ables. Randommeasurement rrorwillcauseunderestimation of the impact of both Sand Q on the acquisition of skills, whilenonrandommeasurementerror could leadto underestimation or overestimation.In summary, uncritical application ofsimple OLS regressionscan lead to biasedestimates of the impact of school qual-ity on learning. Some problems under-estimate the impacts, others overestimatethem, and still others could go eitherway. These difficulties are so dauntingthat some economists doubt that theycan be overcome (see Hanushek 1995).The next two subsections examine sev-eral recent studies, focusing on howthese problems have been addressed, ornot addressed, in the literature.2.3 Recent Estimates of the Impactof School Characteristicson Student Skills

How have studies of the impact ofschool characteristics on students' cog-nitive skills dealt with the problemsraised above? More generally, howmuch has been learned that govern-ments can apply to make schools moreeffective? This subsection reviews "con-ventional" studies by education special-ists and economists, where conventionalrefers to studies that attempt to esti-mate educational production functionsalong the lines of equation (2"') usingordinary (nonexperimental) variation inthe explanatory variables. The followingsubsection examines several more recent,and more innovative, papers.Most conventional studies of the im-pact of school characteristics on learningfocus on developed countries, althoughresearch on developing countries has

increased rapidly in recent years. BruceFuller and Prema Clark (1994) providea comprehensive review of the litera-ture through the mid-1990s. Earlier lit-erature reviews can be found in Har-bison and Hanushek (1992) and Fuller(1987). While these reviews are com-prehensive, they tend to take the con-clusions of the studies they review atface value. Many economists who haveexamined these studies find seriousmethodological shortcomings. For ex-ample, Hanushek (1995, pp. 231-32)claims that ". . . the standards of datacollection and analysis are so variablethat the results from this work are sub-ject to considerable uncertainty." AnneCase and Angus Deaton (1999, p. 1081)concur, stating that "the descriptions ofeconometric procedures . . . are some-times so exotic as to raise serious doubtsabout the validity of the results." My ownreading of the conventional literatureconfirms that the estimation methodsused typically ignore most of the problemsraised in the previous subsection.Given these methodological short-comings, it is not surprising that thefindings of some studies are at oddswith those of others. Fuller and Clark'ssummary conveys the uncertainty in theliterature regarding key questions.While many observers would expect re-ductions in class size to increase learn-ing, Fuller and Clark find that only 9 of26 primary-school studies and only 2 of22 secondary-school studies show a sig-nificant impact of class size on studentachievement in developing countries.Moreover, the paper reports only sig-nificant effects that are in the expecteddirection (for example, smaller classsize raises educational achievement).9Ignoring significant effects in unex-pected directions may be misleading;

9I would like to thank Bruce Fuller for explain-ing this to me.


11/48

Glewwe: Schools and Skills in Developing Countries 445the literature summary by Harbisonand Hanushek (1992) included thirtystudies that examined the impact ofteacher-pupil ratios and found that ofthe sixteen with statistically significanteffects, eight were positive and eightwere negative These problems castdoubt on whether any conclusions canbe drawn with confidence from theconventional literature. This pessimisticinterpretation includes meta-analysesalong the lines suggested by MichaelKremer (1995), since that approach isonly as plausible as the studies onwhich it is based.Can more careful conventional esti-mates produce useful results? The restof this subsection addresses this ques-tion. Before doing so, an importantpoint needs to be made regarding stud-ies that have attempted to draw infer-ences about school quality based onwage data, such as David Card and AlanKrueger's (1992) study of U.S. schoolsand Behrman and Birdsall's (1983)study of Brazil. The point is that verylittle can be inferred from such studiesregarding what makes one school betterthan another, because such studies typi-cally have only one indicator of schoolquality, such as spending per pupil orthe average education level of teachers.Clearly, any single indicator of schoolquality is likely to be correlated withmany other school-quality variables, sosuch studies cannot determine whichschool variables improve children'slearning. To make further progress,data are needed on schools, teachers,and students' cognitive skills.Four studies completed in the earlyto mid-1990s attempted to estimateeducational production functions usingdata specifically collected for that pur-pose: Harbison and Hanushek's (1992)book on Brazil; Glewwe and Jacoby's(1994) study of Ghana; the analysis ofJamaican data by Glewwe et al. (1995);

and Geeta Kingdon's (1996a) paper onIndia. These are probably the best"conventional" studies, so it is worth-while to see how they address, or do notaddress, the problems raised in section2.2 and, more generally, how useful theirresults are for making education policydecisions in developing countries.10Harbison and Hanushek examinedthe performance of primary-school chil-dren in rural areas of northeast Brazilin reading (Portuguese) and mathemat-ics. Tests were administered in 1981,1983, and 1985. The school charac-teristics examined were a facilitiesindex (of about ten building charac-teristics), a writing materials index(chalk, notebooks, pencils, etc.), theavailability of textbooks, and a dummyvariable indicating graded classrooms(as opposed to multigrade classrooms).Both the facilities and the writing mate-rials indices had significantly positiveimpacts in most specifications for bothreading and math. The textbook vari-able was significantly positive for threeof five specifications in math and two offive in reading. Graded classrooms wasnever significantly positive; in somecases it was significantly negative. Thestudy also examined teacher charac-teristics. Neither the pupil-teacher rationor teacher experience had consistenteffects in either subject, but teachersalaries had significantly positive im-pacts in both subjects. Teacher educa-tion almost always had insignificant im-pacts for reading, but usually had asignificantly positive impact for math.Finally, the impact of teacher trainingprograms was mostly insignificant.11 To

1 The study of Pakistan by Harold Aldermanet al. (1996a) is not discussed here because itdoes not estimate the impact of school or teachercharacteristics on student achievement.11Another explanatory variable in the Brazilstudy is teachers' test scores, but the table anddiscussion in the text (p. 114) contradict theresults given in the appendix tables.


12/48

446 Journal of Economic Literature, Vol. XL (June 2002)give an idea of the size of one ofthe significant impacts, consider theteacher-salary variable. In the 1983level specification (second-grade stu-dents), doubling teachers' salariesraised reading test scores by 0.14 stan-dard deviations and math scores by 0.15standard deviations. These effects arenot particularly large compared to thoseof the three other studies, as will beseen below.In Glewwe and Jacoby's study ofGhana, achievement tests were given in1988-89 in reading (English) andmathematics in middle schools (grades7-10). Many school and teacher vari-ables were examined. Most estimatedeffects were small and not statisticallysignificant. The only statistically signifi-cant teacher variable was teaching expe-rience, but its effect was only indirect;it raised children's grade attainment,which then increased both reading andmath test scores. The estimated impactof repairing leaking classrooms, whichpresumably reduced school closings dueto rain, was much larger; the overall(direct plus indirect) impact was anincrease of 2.0 standard deviations inreading scores and 2.2 in math scores.Blackboards also had large estimatedimpacts (direct plus indirect), raisingreading scores by 1.9 standard devia-tions and math scores by 1.8. Adding alibrary led to smaller increases, 0.3standard deviations for reading and 1.2for math scores.The Jamaica study used data col-lected in 1990 on the performance ofprimary-school students in reading(English) and mathematics. Over fortyschool and teacher characteristics wereexamined, including pedagogical pro-cesses and management structure. Mostvariables had statistically insignificanteffects. The school variables with sig-nificantly positive impacts were admini-stration of eye examinations (reading

only), teacher training within the pastthree years (math), routine academictesting of students (reading and math),and the use of textbooks in class (read-ing). Class time devoted to written as-signments had a significantly negativeimpact in both subjects. The sizes ofthese estimated impacts (in standarddeviations of the test score variable)were lower than those for Ghana. Thelargest impact is a change from neverusing textbooks in instruction to usingthem almost every lesson, which raisesreading scores by 1.6 standard devia-tions. The smallest is from teachertraining; a school in which all teacherswere trained is estimated to have mathscores 0.7 standard deviations higherthan an otherwise identical school withuntrained teachers.Kingdon's study of India is based ondata collected in 1991. Tests in reading(Hindi and English) and mathematicswere given to students in "class 8"(grade 8). She examined five teachervariables (years of general education,years of teacher training, marks re-ceived on official teacher exams, yearsof teaching experience, and salary) andthree school variables (class size, an in-dex of seventeen physical characteris-tics, and hours per week of academicinstruction). The teacher variables withsignificant effects were teacher exammarks, which had significantly positiveimpacts on both math and readingscores, and teachers' years of education,which had a significantly positive im-pact on reading scores. Two of thethree school variables, the physicalcharacteristics index and time in aca-demic instruction, had significantlypositive effects on both reading andmath scores. Class size has no signifi-cant impact on math, and a significantlypositive impact on reading. The impactof the teacher's exam marks is not ro-bust to attempts to control for selection


13/48

Glewwe: Schools and Skills in Developing Countries 447into schools (an issue discussed furtherbelow). These impacts are not particu-larly large. An additional year ofteacher's education raises readingscores by 0.13 standard deviations. Go-ing from zero to all seventeen physicalfacilities (which would be quite costlysince this includes toilets, computers,and musical instruments) increasesmath scores by 0.7 standard deviationsand reading scores by 1.0 standard de-viations. Adding another hour per weekof instructional time raises math andreading scores by only 0.04 and 0.02standard deviations, respectively.How much confidence can be placedin the results of these studies? Of theissues raised in the previous subsection,consider first the problem that unob-served components of a child's learningability, such as a child's innate abilityand motivation and parents' willingnessto help their children with their school-work. This leads to upwardly biased es-timates of the impact of school-qualityvariables. The Ghana and India studiesused data from an "intelligence" test,the Raven's Coloured Progressive Ma-trices test, to control for innate ability.The Ghana study concedes that this testmeasures not only innate ability (how-ever defined) but also reflects environ-mental influences, including time inschool. It used a simple "family fixedeffects" procedure to extract what isprobably a cleaner estimate of innateability from the Raven's test, but thismethod relies on several rather simplis-tic assumptions. The India study usedthe Raven's test score directly, withoutany refinement, and the Brazil and Ja-maica studies had no variables to con-trol for child innate ability. Only one ofthe four studies, the one on India, at-tempted to control for child motivationas a factor that is distinct from innateability. (Another possible exception isthe value-added estimates in the Brazil

study, which are discussed below.) Re-garding parents' motivation and abilityto help their children, none of thesestudies goes beyond the common prac-tice of using mother's and father's yearsof education. On a more positive note,all of these studies use standard selec-tivity correction methods (primarily toaccount for choices among different typesof schools); this may reduce bias causedby a variety of unobserved variables,including innate ability.Another potential problem is bias dueto omitted school- and teacher-qualityvariables. If unobserved school andteacher variables are positively corre-lated with observed school and teachervariables, the estimated impacts on theobserved variables will be biased up-ward. At first glance, all four studiesseem to minimize this problem by in-cluding large numbers of school andteacher variables. The Brazil study useddata on at least twenty school andteacher characteristics (the exact num-ber is unclear because many were ag-gregated into indices). The originalGhana study used eighteen school vari-ables (see Glewwe and Jacoby 1992),and the Jamaica study had 42, includingvariables on pedagogical techniques and"school organization, climate and con-trol." Finally, the India study used dataon about 24 variables, although seven-teen were aggregated into a single in-dex. Yet some variables, such as teachermotivation, are inherently difficult tomeasure and thus are not used in any ofthese studies, so the large number ofschool variables used does not necessar-ily avoid bias due to omitted school andteacher characteristics. Moreover, in allfour studies, most school and teachervariables were not significantly differ-ent from zero, which reflects both lowsample sizes (163 students in Ghana,355 in Jamaica, and about 250 in Brazilfor the authors' preferred value-added


14/48

448 Journal of Economic Literature, Vol. XL (June 2002)regressions)12 and high correlationamong many of these variables.A third problem is sample selection.In many developing countries somechildren never attend school, graderepetition is quite common, and a sub-stantial fraction of children drop out ofschool after only a few years. Estima-tion problems can also arise due to thechoices parents make regarding theschools their children attend and ac-tions parents may take to change thoseschools. Each of these studies at-tempted to address at least some ofthese problems. The Brazil study is theleast satisfactory because of the as-sumptions used to achieve identifica-tion of the sample selection terms. It isnot clear why the variables in the selec-tion equation for on-time promotionthat are omitted from the achievementregressions (such as mother's educa-tion, number of students in the school,and type of school) can be excludedfrom the latter regressions. The authorsconcede that their selection correctionprocedure "does rely heavily on the as-sumption that the probit errors are nor-mally distributed" (footnote 103). TheIndia study has similar problems. It ap-peals to the Brazil and Ghana studiesfor evidence that selection of students(in terms of "survival" to higher grades)does not matter. It does address selec-tion into public and private schools butdoes not explain how the selection termis identified. The efforts to deal with se-lection bias are better in the Ghana andJamaica studies. Both clearly explainthe identification strategy (the identify-ing variables are characteristics of theschool not chosen), and the Ghanastudy accounts for sample selection ef-fects due to delayed enrollment and

dropping out (using a similar identifica-tion strategy). In both cases controllingfor sample selection has little impact onthe results, which is consistent with theBrazil study (but not the India study).While this "regularity" may be goodnews, because it implies that bias dueto school selection is probably small, re-sults from more countries are neededbefore drawing general conclusions.The fact that years of schooling, orgrade in school, could be endogenous isa fourth problem. The India study ap-pears to avoid it because all students inthe sample are in the same grade, butthere is still a sample selection issueregarding which children reach thatgrade. The Brazil study mentions it butdoes nothing more. The Ghana studytreats it as a sample selection problemcaused by delayed enrollment; lowgrade repetition in Ghana implies thatnothing further need be done. In Ja-maica, delayed enrollment is not com-mon, and grade repetition is moderate(a typical child repeats once during sixyears in primary school), so that studyignores this issue.A fifth potential problem is measure-ment error in the regressors. None ofthe four studies addresses or mentionsit. A plausible case can be made thatmost such errors are random, which im-plies underestimation of true effects.This may explain why in each studymost of the teacher and school variableswere insignificant. While it is not clearhow serious a problem this is, futurestudies must address it, although how todo so will depend on the details ofthose studies.A final issue is the specification ofthe dependent variable. All four studiesused test scores in level form. A notablealternative, used only in the Brazilstudy, is the "value-added approach,"which is motivated in part by fixed-effects estimation that has long been

12Although the sample size in the India studywas larger, with 902 students, they are concen-trated in thirty,schools, which limits variation inschool characteristics.


15/48

Glewwe: Schools and Skills in Developing Countries 449used in analysis of panel data. The basicidea is quite simple. Suppose, for exam-ple, that one has test scores for a sam-ple of children for two consecutiveyears, say, grades five and six. Assumeas well that one has current data on theschools those children attend, but nodata on the schools attended by thosechildren when they were in grades onethrough four. In addition, one has nodata on the innate ability of those chil-dren nor on a host of other unobservedcharacteristics of children and schools.Such data can be used to estimatevalue-added specifications, of whichthere are two variants.The first approach uses the change inthe test score over the two points intime as the dependent variable, withcurrent child, household, and schoolcharacteristics as the explanatory vari-ables. The second uses the more recenttest score as the dependent variable andincludes the prior test score as an addi-tional regressor. The prior test score isalmost certainly measured with error,so the second variant requires one ormore variables that can serve as instru-ments. Under certain assumptions, thevalue-added approach can reduce biasin estimates of the impact of schoolcharacteristics on student achievement.In particular, if the first test measuresthe impact of all child, household, andschool variables that precede it in time,there will be no omitted variable biasdue to lack of data (child, household,or school) that predate the first test. Inaddition, if innate ability (or child moti-vation) is a fixed effect in a level regres-sion, differencing test scores at two dif-ferent periods of time should differenceout this, and any other, fixed effect.The usefulness of the value-added ap-proach, however, is open to challenge. Ifone is examining student performancein primary schools, and school charac-teristics change slowly over time, the

first advantage is minimal. Morever, theinformation contained in, say, a fifth-grade test score may have a highersignal-to-noise ratio than the informa-tion in the difference of the fifth- andfourth-grade scores. Only a comparisonof a level specification with a value-added specification will clarify this.More importantly, innate ability maynot be a fixed effect. A more plausiblespecification is to interact innate abilitywith school quality, in which case itcannot be differenced out. Finally, allthe other problems raised in section 2.2still apply to the value-added approach.Thus, while value-added specificationsare worth exploring (if the requisitedata exist), findings based on themmust be treated with caution.This review of conventional studiesleads to several conclusions. Many stud-ies suffer from multiple estimationproblems and show only limited aware-ness of them. Recent studies have madesome progress, but many problemsremain. In particular, they use moresophisticated econometric methods, orat least show a clear awareness of manypotential estimation problems, but havenot overcome all of these problems.Third, in my opinion there are two re-lated problems that are difficult toresolve in conventional studies that at-tempt to estimate the impact of schoolcharacteristics on student achievement:omitted school characteristics and un-observed characteristics of children andtheir households. Regarding the firstproblem, although the Brazil, Ghana,Jamaica, and India studies includedlarge numbers of school characteristicvariables in their regressions, there maybe very important but hard-to-observecharacteristics, such as teacher motiva-tion, that are highly correlated with thevariables that are observed, which willlead to biased estimates. Some results seemrather counterintuitive; for example,


16/48

450 Journal of Economic Literature, Vol. XL (June 2002)the most important single school char-acteristic in the Ghana study wasleaking roofs. Perhaps the underlyingrelationship is that more motivatedteachers, principals, and parents weremore likely to keep the building in goodrepair. The inability to observe certainchild and household characteristics,such as the child's innate ability andparental tastes for education, alsoleaves lingering doubts.On a more positive note, if a largenumber of good conventional studiesshow that a specific school charac-teristic increases learning, there is agood chance that these studies are de-tecting a strong causal relationship, andpolicies could be based on such findings(the alternative being choosing policieswithout any evidence whatsoever). Yetthere are only a small number of rigor-ous conventional studies. Fortunately,in the past few years several new ap-proaches have been used to understandhow school characteristics affect stu-dent achievement. These are discussedin the following subsection.2.4 New Approaches to Estimatingthe Impact of Policieson Education Outcomes

In recent years, both education re-searchers and economists have triednew methods to avoid the problemsraised in section 2.2. These can be di-vided into two types. The first retainsthe goal of estimating an education pro-duction function, or at least a reducedform version of it. The second abandonsaltogether attempts to identify specificschool characteristics that make someschools better than others; instead itasks whether certain policies-such asvouchers, decentralized administrationof public schools, or promotion ofprivate schools-can raise students'cognitive skills.Education production functions such

as equation (2) in section 2.2 containmost of the information that a ministryof education wants to know.13 Thesefunctions are technological relation-ships that show how much studentslearn when placed in certain types ofschools with certain types of teachers(conditional on student and householdcharacteristics). Education planners canuse this information to assess the im-pact of each school and teacher charac-teristic on learning. Combined with costdata on these characteristics, they can"design" schools to maximize learningper dollar spent.Suppose that it were impossible to es-timate an education production func-tion using conventional econometricmethods, due to the problems raisedabove. An equally useful, though prob-ably more expensive, approach is toconduct a series of randomized trials,one per school characteristic, to evalu-ate the impact of changes in schooland teacher characteristics on learning.Randomized trials are very commonin medicine but very rare in the fieldof education.14 Labor economists haveconducted randomized trials to in-vestigate the impact of welfare reform,

13One type of informationnot provided by edu-cation production functions is behavioral re-sponses of households to education policies. Asexplained below, such information can be veryuseful.14Three possible reasons why education re-searchers rarely use randomized trials are: a) mosteducation policies are implemented at the class-roomor school evel,whichgreatly aises he costsof randomized rials-in contrast,most medicaltrials are randomizedat the individualevel; b)medicalresearchershave more experiencewithrandomizedrialsbecausethey often implementthem usinganimals, ince animal tudiesare muchmore relevant for understanding umanhealththan for understandingducation ssues; and c)findingson humanhealth in one countryusuallyapply o humans enerallydue to commonphysio1-ogy, but results in educationare typicallymuchmore specificto the local culture andschoolsys-tem. Theremaybe other reasons,but I will notpursue hem.


17/48

Glewwe: Schools and Skills in Developing Countries 451guaranteed minimum incomes, and jobtraining on labor force outcomes (seeJames Heckman, Robert Lalonde, and Jef-frey Smith1999; Charles ManskiandIrwinGarfinkel 1992; and the special issue ofthe Journal of Labor Economics 1993).Results from several randomized tri-als on different school or teacher char-acteristics cannot be assembled into aneducation production function, becausesuch trials provide only reduced formestimates of the impacts of those char-acteristics. Yet this is not a problem forpolicymakers; indeed, a limitation ofknowing only the "true"education pro-duction function is that it does notincorporate households' behavioral re-sponses. For example, suppose schoolquality increases in some way. One pos-sible response of parents to higher qual-ity is to reduce the time they spendhelping their children with schoolwork.Such a behavioral response is not mea-suredin an education productionfunction,but would be measured in a randomizedtrial of that quality improvement (as-suming that the randomized trial en-counters no serious problems, an issuediscussed further below).The next paragraphs review twomethods that, in principle, providereduced form estimates of educationproduction functions: randomized trials,and natural experiments.151. Randomized Trials. The basic ideaof randomized evaluations of any kind isto compare two groups of observationsthat have no systematic differencesother than that one group received the

"treatment" and the other did not. Thesimplest method is to sample a popula-tion of interest and randomly divide thesample into "treatment" and "control"groups. If this can be done without fur-ther complications-a big "if'- differ-ences in the variables of interest acrossthe two groups are unbiased estimates ofthe (reducedform)effect of the treatment.In theory, randomized trials avoid allthe problems discussed in section 2.2.Random assignment of observationsinto treatment and control groups im-plies that both observed and unob-served characteristics of those observa-tions are uncorrelated with treatmentstatus. In econometric terms, the out-come of interest is the dependant vari-able and the only regressor is treatmentstatus. That regressor is uncorrelatedwith everything in the error term be-cause treatment status is uncorrelatedwith virtually everything. Another prob-lem that randomized studies shouldresolve is measurement error; in anywell-managed study treatment statusshould be measured without error.In practice, randomized trials canhave serious problems. First, child,household, and school characteristicsmay change in response to the treat-ment. For example, if treatment schoolsare provided with abundant school sup-plies, parental efforts to improve thoseschools (such as fund-raising activities)may decline. Even so, the only implica-tion of this is that the impact of thetreatment is a reduced form effect,rather than a structural parameter. Asexplained above, the former is oftenmore useful for making policy choices.Yet structural estimates may also be ofinterest. Even if the reduced form ef-fect on student learning is zero, a policymay still raise the welfare levels of par-ents and others. In the above example,parents' welfare rises due to less timespent on fund-raising.

15A third approach is matched comparisons,which have been used to analyze U.S. job trainingprograms (Heckman et al. 1997, 1998). Yet thesemethods offer only a modest extension of the con-ventional approach of controlling for observedschool, teacher, and child characteristics becausethey do not avoid the problem that observed andunobserved characteristics may be correlated.Moreover, I know of no studies on education indeveloping countries that use this method.


18/48

452 Journal of Economic Literature, Vol. XL (June 2002)Another set of problems is sampleselection issues, which are difficult toavoid. Parents of students in the control

schools (or schools not included in theevaluation) may try to enroll their chil-dren in the treatment schools. This mayaffect the results by increasing classsize (if class size affects the outcome ofinterest). This is not part of a reducedform effect because a nationwide adop-tion of the policy would not have thiseffect. In addition, children who trans-fer into the treatment schools may notbe a random sample of the general stu-dent population. A related problem isthat marginal students in the treatmentschools are less likely to drop out (ifthe intervention raises student achieve-ment), which will underestimate the im-pact of the policy on learning if com-parisons are made based on all studentscurrently enrolled in school. As dis-cussed below, there are ways to reducethese problems, but they may not alwayswork.The first randomized trials of educa-tion policies in developing countrieswere done in the early 1980s byStephen Heyneman, Dean Jamison, andtheir collaborators. Jamison et al.(1981) conducted a randomized trial inNicaragua in which 48 first-grade class-rooms received radio mathematics in-struction, twenty received mathematicsworkbooks, and twenty served as con-trols. After one year, students in theclassrooms that received radio instruc-tion scored more than one standard de-viation higher on mathematics teststhan students in the control group, andstudents in the classrooms that receivedmathematics workbooks scored about athird of a standard deviation higher. Bothdifferences were highly statisticallysignificant.In the second study, Heyneman,Jamison, and Montenegro (1984) stud-ied the first two grades of 104 primary

schools in the Philippines. The schoolswere divided into three groups: 26 re-ceived mathematics, science, and Fili-pino textbooks at a ratio of one forevery pupil; 26 received the same text-books at a ratio of one for every twopupils; and 52 served as controls. Be-cause textbooks were distributed to allschools in the 1977-78 school year, thecontrol schools were evaluated in termsof student test scores in the previousschool year (1976-77). Students in thetwo groups that received textbooks per-formed similarly, even though one hadtwice as many textbooks as the other;their test scores were about 0.4 stan-dard deviations higher than those in thecontrol schools (averaged over twogrades and three subjects). These dif-ferences were also highly statisticallysignificant.The randomized experiments in Nica-ragua and the Philippines were well de-signed and executed. Yet a potentiallyserious problem of both studies is sam-ple selection and attrition. It is possiblethat enrollment increased in the treat-ment schools, which may have affectedindicators of student performance. Thedirection of bias depends on the charac-teristics of the students attracted tothose schools. If they were relativelyweak students who otherwise would nothave been in school, the bias is down-ward, but if they were strong studentsfrom other schools the bias is upward. Asimilar result on downward bias holdsfor attrition; if the intervention causedrelatively weak students to stay inschool longer, the estimated effect ofthe program is biased downward. An-other potential problem in the Philip-pines study is that students' test scoresin the control group were collected oneyear earlier than those of the studentsin the treatment groups. It is possiblethat other differences in those twoyears could lead to biased results.


19/48

Glewwe: Schools and Skills in Developing Countries 453Did the results of these two studieslead to changes in education policies?The strong impact of radio instruction

in Nicaragua may explain the expansionof educational radio to other LatinAmerican countries (and a few coun-tries in Africa and Asia) in the 1980s(John Newman, Laura Rawlings, andGertler 1994). Ironically, Nicaraguaabandoned educational radio after theSandinista government came to powerin 1979; that government favored hugeliteracy campaigns, and its disputeswith the U.S. government ended theUSAID funding that had financed itseducational radio program. It is lessclear whether the textbook results inNicaragua and the Philippines led topolicy changes; since most educationofficials would view this result as un-surprising, it may have had little effect.Unfortunately, no more randomizedstudies in education were done until themid-1990s. The following paragraphsreview recent studies done in Turkey,the Philippines, and Kenya.A recent study by Turkish educa-tional psychologists is the only random-ized study in a developing countrynot initiated by economists. CigdemKagitcibasi, Diane Sunar, and SevdaBekman (2001) examined the impactof a mother-education program onpreschool-aged children. They consid-ered three preschool settings: "educa-tional centers," which attempted toteach children specific skills; "custodialcenters," which had no specific educa-tional objectives; and children cared forat home. Within each group, about halfof the children were three years old andhalf were five years old. For each of thesesix age/preschool categories, motherswere randomly assigned to receive (ornot receive) intensive "mother train-ing." Ten mothers assigned to betrained "declined to participate" andwere placed in the "no training" cate-

gory. While this attrition is small, it maylead to overestimation of the impact ofthe program if these ten mothers hadlower-than-average tastes for their chil-dren's education. After four years, 25 (9percent) of the original 280 mothershad dropped out of the program, leav-ing 255 children in the sample, 64 ineducational centers, 105 in custodialcenters, and 86 cared for at home. Nodifferences were found in an IQ test ad-ministered at the start of the programbetween the 25 children who droppedout and the 255 that remained. Aftertwo years of mother training, a varietyof sociological, psychological, andachievement tests were administered.There were significant differences be-tween the treatment and control groupsfor some outcomes but not for others.The study found no significant programimpact in terms of mathematics and(Turkish) reading ability, although thepoint estimates were positive, but didfind a statistically significant positive im-pact on IQ scores and on "general ability"(spatial, numeric, and verbal reasoning).This is puzzling because students withhigher "ability" should be better atlearning academic subjects. The magni-tude of the estimated impacts is unclearbecause the study does not report thestandard deviations of the test scores.The Turkish study, while innovative,is open to several criticisms. The poten-tial for bias caused by the ten motherswho declined training could have beenavoided by using an instrumental vari-ables estimation procedure, where ac-tual treatment is instrumented by theoriginal random assignment. This wouldmeasure the impact of the program onmothers who were trained, the "effectof the treatment on the treated." Re-taining the ten mothers and regressingthe outcome(s) of interest on the origi-nal random assignment would measure theeffect of being offered the treatment.


20/48

454 Journal of Economic Literature, Vol. XL (June 2002)Other problems are harder to solve.First, the small sample of 255 childrenmay explain why most of the results,which were in the expected direction,were insignificant. Second, there is noinformation on the costs of the interven-tion, which hampers cost-effectivenesscomparisons with other studies. Theprogram description suggests very highcosts. Third, the mother training mayhave been implemented by highlymotivated and highly trained individ-uals; implementing the program on alarger scale may draw less-educated andless-motivated trainers, reducing theprogram's effectiveness.A second recent randomized study,Jee-Peng Tan, Julia Lane, and GerardLassibille (1999), was also done in thePhilippines. It examined four educationpolicies: school feeding; multilevellearning materials (pedagogical materi-als for teachers); and combinations ofeach with "parent-teacher partnerships"(structured meetings between parentsand school officials). Thirty schoolswere randomly assigned to five groups:five schools each for the four policy in-terventions and ten control schools. Theauthors examined dropout rates andstudent test scores after one year. Theyfound almost no effects on droppingout; only the provision of multilevel ma-terials had a significant impact (andonly at the 10 percent level), reducingthe dropout rate by about five percent-age points. In contrast, most of the poli-cies had significant impacts on testscores, though statistical significancevaried with the estimation procedureused. Simple estimates that ignore se-lection bias due to differential dropoutrates produced large impacts (as highas 0.87 standard deviations), althoughmost were statistically insignificant.Correction for selection bias yieldedsignificant effects more often, but withlittle effect on the point estimates.

School feeding combined with parent-teacher partnerships most often pro-duced sizeable and statistically signifi-cant impacts, ranging from 0.28 to 0.44standard deviations for math, Filipino,and English test scores. Multilevel ma-terials with parent-teacher partnershipsalso had significant impacts, from 0.23to 1.05 standard deviations for Filipinoand English (but not math). Schoolfeeding alone had statistically significantimpacts on English (and for math in oneof three specifications), while multi-level materials alone had small impactsthat were rarely statistically significant.The authors conclude that combiningmultilevel learning with parent-teacherpartnerships seems to be the most cost-effective policy, partly because of theirregression results and partly becauseschool feeding programs are expensive.Yet they recognize the imprecise andtentative nature of their results. Themethods used to control for sample se-lectivity raise some doubts; for example,one of the identifying variables in theselection correction term is distance to thenearest school, but this could directlyaffect learning by causing children to beabsent or tardy more often. Overall, theimprecision of these results and theirsensitivity to estimation methods suggestthat they be interpreted with caution.The most recent set of randomizedstudies on education are those beingconducted in Kenya by Michael Kremer,Glewwe, and other collaborators. Sixrandomized trials have been conductedin rural Kenyan primary schools: a stan-dard package of inputs (textbooks,school uniforms, and construction mate-rials); textbooks only; block grants; flipcharts; a package of teacher incentives;and treatment of intestinal parasites.16

16 A standard package of preschool assistance isalso being evaluated. Preliminary evidence indi-cates no effect of that package, but the finalresults are not yet available.


21/48

Glewwe: Schools and Skills in Developing Countries 455Results are currently available for fourof these studies.The first study in Kenya, Kremer etal. (1997), examined the standard assis-tance package of a Dutch nongovern-mental organization (NGO). Fourteenschools participated, of which half wererandomly chosen to receive assistance.There were no statistically significantimpacts of the package on student testscores (English, mathematics, science,Kiswahili, geography/history/civics, andart/craft/music), and the point estimateswere small (less than 0.1 standard de-viations). On the other hand, the pro-gram did reduce dropout rates. Thisstudy faced two serious problems. First,the sample size was small (in terms ofthe number of schools), which led toimprecise estimates. Second, the pro-gram increased enrollment in thetreatment schools by an average of 35percent, while in comparison schoolsenrollment declined by 10 percent. Ifhigher class size lowers student achieve-ment (and one study, discussed below,supports this hypothesis), the estimatedimpact of the program is biased down-ward. The authors attempt to correct forthis problem, but they have difficultyisolating the impact of the package fromthe impact of higher class size.The second Kenya study, Glewwe,Kremer, and Sylvie Moulin (2001), ex-amines provision of textbooks. Ruralprimary schools in Kenya rarely providetextbooks; parents are expected to buythem, but few do. In 1995, one hundredrural primary schools were randomly di-vided into four groups of 25 schools. In1996, textbooks were provided to chil-dren in grades 3-8 in the first group of25 schools. After four years, there isvery little evidence of a sizeable impactof textbooks on the average test scoresof students. Point estimates are usually0.1 standard deviations or less, and inalmost all cases impacts of 0.3 or higher

can be ruled out. However, the authorsdo find evidence that textbooks bene-fited the better students. These overallresults are at odds with the first ran-domized studies in Nicaragua and thePhilippines. Two possible reasons forthe lower impact in Kenya are: (a) Theteachers were not trained in the use oftextbooks (extensive training was pro-vided in the Philippines, but only mini-mal training was given in Nicaragua);and (b) The textbooks were too difficultfor the average student in rural Kenya.The authors show that the typical me-dian child in grades 3-5 could not readthe textbooks provided (the official text-books recommended by the Ministry ofEducation), although this was not thecase for grades 6-8. Unlike the firstKenya study, provision of textbooksdid not increase enrollment in the 25treatment schools.The third Kenyan intervention, exam-ined in Glewwe et al. 2002, focused onflip charts: large poster-sized chartswith instructional material that can bemounted on walls or placed on easels.This intervention covered 178 primaryschools, half of which were randomlyselected to receive flip charts coveringscience, math, geography, and health.Despite a large sample size and twoyears of follow-up data, the estimatedimpact of flip charts on students' testscores is essentially zero and completelyinsignificant. In contrast, several con-ventional OLS estimates, which maysuffer from many of the problems de-scribed in section 2.2, show impacts aslarge as 0.2 standard deviations, 5-10times larger than the estimates basedon randomized trials.The most recent intervention inKenya examines student health. Intesti-nal parasites (roundworm, whipworm,hookworm, and schistosomiasis) are en-demic in rural areas of Kenya and manyother developing countries. Medical


22/48

456 Journal of Economic Literature, Vol. XL (June 2002)research shows that high "loads" of in-testinal worms lower scores on IQ tests,but almost no research has been done ontheir long-term impact on academic tests.Fortunately, treatment with albenda-zole every six months eliminates round-worm, whipworm, and hookworm, andannual doses of praziquantel cure schis-tosomiasis. A sample of 75 schools wasdivided into three groups of 25 schools.The first group was treated in 1998, thesecond in 1999, and the third was a con-trol group treated in 2001. Analysis oftwo years of data by Edward Migueland Kremer (2000) indicates that provi-sion of albendazole and praziquantel in-creased student participation (fewer ab-sences and reduced dropout rates) buthad no significant effect on test scores.In fact, the program slightly reducedtest scores (by -0.04 standard devia-tions after one year and -0.07 after twoyears; these are averages over English,math, and science), but these impactswere statistically insignificant.

This more recent experience withrandomized studies of education in de-veloping countries provides several use-ful lessons. First, sample sizes shouldbe quite large, at least fifty to one hun-dred schools, to avoid imprecise esti-mates. Second, problems of differentialselection into the initial sample (firstKenya study) and attrition (Philippines)across the two types of school are realpossibilities; sound estimation methodsthat address these problems must beplanned before data collection becausethey may require additional baselinedata. Third, school outcomes should befollowed for more than one year to seewhether program impacts increase orfade over time. Fourth, a large amountof school data should be collected tocheck for other possible biases. Anexample of this is in the paper ontextbooks in Kenya; it examinedwhether biases could be caused by re-

duced school fundraising, reduction inthe purchase of textbooks by parents,and a greater tendency to promotestudents to the next grade, and foundthat none of these potential problemsappears to overturn the result thattextbooks had little or no impact.2. Natural Experiments. Althoughwell-executed randomized studies canavoid many econometric problems, theycan be very expensive to implement. Anappealing (though rare) alternative is tofind "natural"variation in a school char-acteristic that is uncorrelated with vir-tually anything else that determineschild learning. Two recently publishedstudies demonstrate what can and can-not be learned from such "natural ex-periments."17 The first, by Case andDeaton (1999), examined educationaloutcomes in South Africa. The dataused were collected in 1993, when gov-ernment funding for schools was highlycentralized, and blacks (people of Afri-can descent) had virtually no politicalrepresentation of any kind. The authorsargue that blacks did not control thefunds provided to their children'sschools, and that tight migration con-trols limited their ability to migrate toareas with better schools. They showthat pupil-teacher ratios varied widelyacross black schools, and argue that thisvariation, combined with migration bar-riers and black South Africans' lack ofcontrol over their schools, generated akind of natural experiment.The South Africa study examineswhether increased school resourceslead to better educational outcomes.Most economists and other observers

17 See Rosenzweig and Wolpin (2000) for a thor-ough discussion of "natural" natural experiments,i.e., natural experiments whose parameters of in-terest are identified by date of birth, twin births,gender of newborn child or siblings, and weather.The issues raised in that paper also apply to "lessnatural" experiments, and many are discussedbelow.


23/48

Glewwe: Schools and Skills in Developing Countries 457would probably expect the answer to be"yes," but Case and Deaton argue thatsome economists have claimed other-wise. (For example, Hanushek 1995said that "providing more inputs . . . isfrequently ineffective," yet this seemsto allow room for inputs to be effectivein some cases.18) They present severalregressions that show the impact ofschool resources (primarily measuredby student-teacher ratios) on years ofcompleted schooling, enrollment, andtest scores, and find evidence thatgreater school resources increase allthree outcomes. Decreasing the student-teacher ratio from forty to twenty(the approximate means in black andwhite schools, respectively) increasesgrade attainment by 1.5 to 2.5 years andraises students' reading test scores(conditional on years of school atten-dance) by the same amount as does twoadditional years of schooling (in con-trast, there was no significant impact onmath scores).

While the South Africa study hassome data problems (e.g., the childrentested were not a random sample ofhousehold members, and data from theMinistry of Education are not highlycorrelated-an R2 coefficient of 0.15-with the authors' community data),most readers would agree that, in prin-ciple, resources matter. One criticism isthat even if blacks could not influenceclass size in their children's schools,certainly someone, presumably somegovernment officials, made decisionsthat influenced class sizes in SouthAfrica's black schools. If these decisionswere influenced by education outcomesin those schools, they could yield biasedestimates of the impact of class size(and, more generally, school resources)

on those outcomes. This is a well-known problem of endogenous pro-gram placement (see, inter alia, MarkRosenzweig and Kenneth Wolpin 1986).A further limitation is that this paperdoes not tell us how educational re-sources should be used; ministers ofeducation in developing countrieswould like to know how to spend anyadditional resources they may receive.The other recent study based on anatural experiment is that of JoshuaAngrist and Victor Lavy (1999a), whoexamine the impact of class size in Is-rael.19 The natural experiment is astrictly enforced rule that limits classsizes to forty or fewer students (a ruleproposed by Moses Maimonides, atwelfth-century Talmudic scholar). Thelimits on class size determined by thisrule vary in a highly nonlinear way withtotal enrollment in a given grade, pro-viding an unusually credible instrumen-tal variable to get around the problemthat class size may be correlated withunobserved determinates of studentlearning. The authors use data from theearly 1990s on a national test for Israelithird, fourth, and fifth graders. Most ofthe data are at the classroom level, sothe analysis is at that level. The data arelimited to Jewish public school stu-dents; private schools (mostly Jewish re-ligious schools) are excluded due totheir different curriculum, and Arabpublic schools (Arabs and Jews attendseparate public schools) are excludeddue to lack of data on "percentage dis-advantaged"in Arab schools.20 For each

18 Hanushek is more pessimistic on the impactof increased inputs in the United States and otherdeveloped countries; see, for example, Hanushek(1996).

19 Another paper by the authors (Angrist andLavy 1999b) examines computer-assisted instruc-tion in Israeli schools. The identification strategyin that paper is less appealing because there are nolarge discontinuity points in the function generat-ing the instrumental variable. Moreover, most ofthe results based on that strategy are statisticallyinsignificant.20 In fact, Arab schools could have been ana-lyzed because the percent disadvantaged variable


24/48

458 Journal of Economic Literature, Vol. XL (June 2002)grade the sample is approximately 2000classrooms from about 1000 schools.The only explanatory variables usedby Angrist and Lavy are class size, thepercent of disadvantaged students inthe school (averaged over all grades),and total enrollment for the grade. Inmost contexts this paucity of schoolvariables would lead to omitted variablebias. Yet all one needs to obtain consis-tent estimates is an instrumental vari-able that predicts class size and is un-correlated with the error term in thetest score regression. The application ofMaimonides' rule is promising becauseit generates an oddly shaped relation-ship between class size and total schoolenrollment. In grades with an enroll-ment of forty or less, class size willequal total enrollment. When total en-rollment hits 41 the class must be splitinto two, so that class size is half of totalenrollment for grades with 41 to 80 stu-dents. When total enrollment hits 81 athird teacher must be hired, so thatclass size is one-third of enrollment forgrades with total enrollment from 81 to120. This "zigzag" relationship betweentotal enrollment and class size gener-ated by Maimonides' rule allows theauthors to create an instrument forclass size that is not highly correlatedwith total enrollment, so they can in-clude total enrollment and its square asadditional regressors.Before examining the results, twocomments are in order. First, as in ran-domized trials, the estimated impact ofclass size is not a production functionparameter but a reduced form effect.When class size shifts abruptly due toapplication of Maimonides' rule otherclassroom characteristics may also

change, such as teaching methods ortime spent on various activities. Yetfrom a policy perspective this informa-tion is very useful, as explained above.Second, even this estimation strategymay have problems. Some parents mayknow how Maimonides' rule is applied,and those with high tastes for child edu-cation may transfer their children out ofschools in which that rule leads to highclass sizes. This can cause correlationbetween unobserved parental tastes forchild education and the instrumentalvariable used to predict class size. Theauthors claim that this bias should benegligible (for example, Israeli parentswould have to move to transfer theirchild into another school, or at leastswitch the child from a secular to a reli-gious school), but there is no rigorousway to test for this problem.Angrist and Lavy find a significantlynegative impact of class size on thereading and mathematics scores of fifthgraders. The estimated effects of a onestandard-deviation decrease in class size(reduction of 6.5 pupils) are increasesin reading scores of 0.2 to 0.5 standarddeviations and in math scores of 0.1 to0.3 standard deviations (the range re-flects differences in the sample and inthe other covariates). The effects onfourth graders are less precisely esti-mated; sometimes they are significantlynegative for reading scores, but formath scores the effects are all insignifi-cant. For third graders all estimated im-pacts are insignificant; the authors sug-gest that this may reflect difficulty inmeasuring a presumably cumulative ef-fect at lower grades. They also point outthat testing conditions for the thirdgraders were different from those forfourth and fifth graders.These two studies of natural experi-ments in education in developing coun-tries demonstrate both promise and pit-falls. In the South Africa study little was

is not needed; it should be uncorrelated with theinstrument constructed using Maimonides' rule.The only reason to include that variable is to try toincrease the precision of the estimates.


25/48

Glewwe: Schools and Skills in Developing Countries 459learned that school officials could use.While the Israeli study is probably thebest study of the impact of class size onstudent performance in a developingcountry, it also highlights how much isleft to learn. First, Israel is in manyways closer to a developed country thanto a developing country. Second, thefinding that class size matters is alreadyassumed to be true by most officials inministries of education, so it is unlikelythat policies will change in response tothis research. Third, this method prob-ably cannot be applied to other coun-tries because Maimonides' rule is usedonly in Israel. On a more positive note,both studies highlight what can belearned from a natural experiment andraise the intriguing possibility that morenatural experiments are waiting to bediscovered in developing countries. Avery recent example is Esther Duflo's(2001) study of Indonesia; it is notreviewed here because it does notexamine cognitive skills.3. Studies on Private Schools andDecentralization of Public Schools. Theimplicit assumption thus far is thatgovernments will use the estimatesobtained to improve public schools.Another strand of recent research goesbeyond this policy framework and in-stead considers decentralized manage-ment of public schools and privateprovision of educat

Documents

EdSchools and Skills in Developing Countries: Education Policies and Socioeconomic Outcomesucational Policies and Socioeconomic