uncertain educational returns in a developing economy

International Journal of Educational Development 32 (2012) 590–599

Contents lists available at SciVerse ScienceDirect

International Journal of Educational Development

journa l homepage: www.e lsev ier .com/ locate / i jedudev

Uncertain educational returns in a developing economy

Sandeep Mohapatra *, Martin K. Luckert

Department of Resource Economics and Environmental Sociology, University of Alberta, Canada

A R T I C L E I N F O

Keywords:

Education

Development

Gender

Rate of return

Distribution

Random parameters

A B S T R A C T

This paper estimates the distribution of educational returns by gender for India. While previous studies

focus on mean returns, the variance of educational returns has important implications for policy-making

and micro-level decision making with respect to education. If the variance of educational returns is large,

it can leave large sections of the population with negative returns; if the variance of educational returns

is gender specific, it can influence households’ decisions to educate girls versus boys. Our econometric

results provide evidence that India’s labor markets are characterized by significant uncertainty and that

the uncertainty is systematically larger for women.

� 2011 Elsevier Ltd. All rights reserved.

1. Introduction

The rate of return to education is a key economic parameter thathas received attention from empirical researchers for decades (seePsacharopoulos and Patrinos, 2004 for a recent review). Estimatesof educational returns are used to guage the level of educationalexpenditures undertaken by governments and individuals. Forexample, policymakers often use information on educationalreturns to assess the feasibility of student loan programs forfunding higher education (OECD, 2007; Oliveira Martins et al.,2009). Perception by parents about the return to educationinfluences their decisions regarding educating children, andeducating boys versus girls (Gertler and Glewwe, 1992; Munshiand Rosenzweig, 2006). Returns to education are also importantfrom a macroeconomic perspective since human capital accumu-lation is regarded as one of the prime movers of regional andcountry-level economic growth (Barro, 1991).

Most studies on educational returns investigate mean values. Incontrast to mean returns, the variance (henceforth uncertainty)associated with economic returns to education has received littleattention (Patrinos et al., 2006). There is, however, a long standingtheoretical literature emphasizing that when investing in humancapital, individuals and governments are not only interested inreturns, but also in the certainty of those returns (e.g. Levhari andWeiss, 1974). Understanding the uncertainty in educationalreturns is necessary for understanding both the incentives thatindividuals have to invest in education and the impacts ofgovernment policies that support increased investment in educa-

* Corresponding author at: 515 General Services Building, Department of

Resource Economics and Environmental Sociology, University of Alberta, Edmon-

ton, AB, T6G 2H1, Canada. Tel.: +1 780 492 0823.

E-mail address: [email protected] (S. Mohapatra).

0738-0593/$ – see front matter � 2011 Elsevier Ltd. All rights reserved.

doi:10.1016/j.ijedudev.2011.11.009

tion. For example, if significant portions of the distributions ofreturns to education are negative, then risk averse individuals, and/or governments averse to uncertain impacts of policies, may behesitant to invest in education. Alternatively, if investments inhuman capital decrease the variance of wages, that is, if theuncertainty in educational returns is decreasing in educationlevels, then governments will have stronger incentives to nurturehuman capital accumulation due to its insurance effect, either as asubstitute for, or in combination with social protection polices(Anderberg and Andersson, 2003).

In recent years a number of studies in labor economics, with afocus on developed economies, are increasingly paying attention tothe magnitude of uncertainty present in educational returns. Manyempirical papers confirm that educational returns exhibit signifi-cant variability across individuals (e.g. Koop and Tobias, 2004).Others show that the amount of variability present in educationalreturns may evolve over time as governments expand the totalquantity of education in an economy (Harmon et al., 2003).Uncertainty can increase, for example, if an education expansiondraws less able people into the educated pool. Implications ofuncertainty in eductional returns for wage inequality, and thepossibility that sizeable fractions of the population may facenegative returns, have also been explored (Maier et al., 2004;Lauer, 2004).

One strand of the uncertainty literature focuses on understand-ing what drives the enrollment rates of women in highereducational institutions in developed countries (e.g. Averett andBurton, 1996; Jacob, 2002; Anderson, 2002). Using theoreticalmodels of lifetime earnings, studies show that a female collegewage premium and women’s expectations of future earnings arethe major drivers of women’s college enrollment (He, 2011). Thelogic behind these models is that individuals make their collegeentry decisions by comparing the expected difference in theaverage discounted wages from two future wage distributions,

http://dx.doi.org/10.1016/j.ijedudev.2011.11.009

mailto:[email protected]

http://www.sciencedirect.com/science/journal/07380593

http://dx.doi.org/10.1016/j.ijedudev.2011.11.009

2 Models of parental investment in their children have been developed by Ben-

Porath (1967) and Heckman (1976), and have been applied in a variety of contexts

S. Mohapatra, M.K. Luckert / International Journal of Educational Development 32 (2012) 590–599 591

college and non-college. Charles and Luoh (2003) show, however,that the decision depends not only on the difference in expectedearnings, but also on differences in the anticipated variance of thetwo future distributions.

Though the literature on uncertain returns to education hasbeen focused in developed countries, the issues addressed in thisliterature are of fundamental concern to developing economies,where human capital accumulation and gender equality, bothmediated through labor markets, are key determinants ofeconomic growth (OECD, 2007; Schultz, 2002). Since uncertaintyin educational returns discourages human capital accumulation,and can also influence proportions of men and women who areeducated, it can affect the long run growth and equilibrium of aneconomy. Despite the importance of uncertain returns to educa-tion in developing economies, to our knowledge, there are noinvestigations in these settings.

The overall goal of this study is, therefore, to investigatewhether differing levels of educational returns are characterizedby uncertainty and whether the uncertainty is systematicallydifferent for men and women in a developing country context. Weconsider the case of India, one of the largest and fastest growingdeveloping economies in the world. To achieve our overall goal, weemploy three specific approaches. First, using a nationallyrepresentative dataset on over 30,000 wage workers observedduring the 2005–2006 time period, we describe how wagedistributions vary across levels of education and gender. Second,we econometrically estimate the determinants of wages for menand women and statistically test for the presence of uncertainty ateach level of education. Third, we use the estimated parameters toconstruct the distribution of annual rates of return correspondingto each level of education for each gender. As part of this thirdapproach, we also calculate the probabilities of individualsreceiving negative returns to education.

Our study makes three contributions. First, it contributes to theempirical literature on uncertainty in educational returns, with anemphasis on its importance in developing countries. As discussedabove, only a few studies in labor economics have examineduncertainty in educational returns, and these have been indeveloped country contexts. Second, our study contributes tothe literature on the economics of gender in developing countries.Examining uncertainty in returns to education may be particularlyimportant for women in developing country settings, as theireducation has been shown to play a multifaceted role in economicdevelopment.1 Third, our study contributes to current policydiscussions regarding education reform and its link to a wide rangeof outcomes in India, including poverty, inequality and economicgrowth. In sum, by focusing on the the variance of estimatedreturns, we hope to establish more clearly the implications ofuncertain returns for the targeting of education policies.

The rest of the paper is organized as follows. In Section 2, wedevelop a conceptual framework that shows how empiricaldistributions of educational returns can affect the demand foreducation by individuals and households. In Section 3, we describethe Indian context of our study, and the data used in our analysis. InSection 4, we present our econometric model for estimatingeducational return distributions. Our results are contained in

1 Increased education of women and girls has been shown to be associated with:

decreased malnourishment (Smith and Haddad, 1999), decreased infant mortality

(Schultz, 1993; Summers, 1994), decreased fertility rates (Subbarao and Raney,

1995), decreased vulnerability to HIV/AIDS (de Walque, 2007), decreased

environmental degradation (Asian Development Bank, 1989), a greater likelihood

of resisting violence (Sen, 1999) and of participating in political meetings (UNESCO,

2000), and the adoption of more efficient farming practices (Smith and Haddad,

1999). Based on such evidence, many economists openly recommend allocating a

disproportionate share of public expenditures toward women’s education (Schultz,

2002).

Section 5, and are divided into three sections that correspond witheach of our approaches; comparisons of wage distributions,econometric estimates of uncertainty in returns to education,and a simulation of annual returns to education and the probabilityof negative returns. Section 6 concludes the paper.

2. The implications of uncertainty for gender specificinvestments in education – a conceptual framework

Charles and Luoh (2003, subsequently referred to as CL in thispaper) developed an educational investment model that is usefulfor conceptualizing the implications of our econometric findingsfor micro-level educational investment decisions. In this sectionwe describe the model and recast it to represent the case of India.We then use the results of the conceptual model to derive testablehypotheses regarding our estimates of uncertainty in educationalreturns and how they correspond to alternative incentivestructures within households to educate females and males.

A number of scholars have studied, in India and otherdeveloping country contexts, factors that influence householddecision making with regards to educating their children. Thisliterature highlights the complexity of the decisions that house-holds face (e.g. Munshi and Rosenzweig, 2006). In addition to labormarket considerations, a number of factors may influenceeducation decisions including social norms and cultural consider-ations. However, similar to a large body of past work, our focus inthis paper is on labor market considerations, while controlling forsocial considerations such as caste and religion. The conceptualmodel which follows reflects this focus.

The CL model depicts individuals who are deciding whether toattend college during a future period. In the context of our Indiancase study, consider a representative household where parentsmake the decision about whether to invest in increased educationfor their children.2 Education is valued by parents as an investmentthat is expected to yield a financial return to the household.3 Theseinvestments depend on anticipated net returns from educatingmales versus females.4

But returns to education are not certain, and can varydepending on whether a household invests in a male or a femalechild. Fig. 1, taken from the CL paper, may be used to investigatethis choice. Consider two alternative distributions, B and B0, thatrepresent, respectively, more and less certain future earnings froman investment in education. The corresponding lower futureearnings without the investment are represented by A and A0. Thelower and higher levels of education are, respectively, referred toas ‘‘no college’’ and ‘‘college’’ in the CL figure. The difference in themeans of the higher and lower education distributions represents apremium, P received from obtaining a higher education level.Under the standard approach based on average values, labormarkets are assumed to function such that households at themargin, comparing educational premiums, would be indifferentbetween A and A0 or B and B0 for a given level of costs. But if a riskaverse household is faced with the option of taking option A, or to

(Alderman and King, 1998). Though the parental investment framework more

accurately describes the situation in India than the individual investment

framework developed by CL, we follow the CL framework because it embodies

the same tradeoffs of benefits and costs while explicitly focusing on distributions

rather than on average levels of educational returns.3 Households may also have a direct consumption motive in educating children.

For example, parents may prefer educated over non educated children. We do not

include such a motive in our conceptual model because it does not change the main

point pursued here.4 Differences in the anticipated net returns to education among genders can be

due to differences in costs, benefits, or both. Empirical work using Mincer earnings

functions focuses on gross returns for a given levels of costs.

[(Fig._1)TD$FIG]

Fig. 1. A model of educational choice.

Source: Charles and Luoh (2003).

S. Mohapatra, M.K. Luckert / International Journal of Educational Development 32 (2012) 590–599592

pay costs necessary to receive options B or B0, they will be morewilling to make the investment if the option is B. Furthermore, if ahousehold invests in order to receive option B, then they are morelikely to have given up the education choice A0 rather than A. Insum, CL show that the lifetime expected utility from a particulareducation choice is strictly increasing in the average and strictlydecreasing in the cross-sectional variance of future earnings due tothe investment in education. Consequently, an individual’sexpected return from investing in education depends on theanticipated premium, as well as the difference in the anticipatedvariances of the income distributions.

The CL model may also be extended to consider educationalinvestment decisions where future payoffs differ between genders.For example in India, there is evidence to suggest that for mosteducation levels, women receive higher returns than men (e.g.Dutta, 2006; Duraisamy, 2002). In terms of Fig. 1, if we assume thatthe distributions given in the figure are for males, we could denotea separate set of distributions with a larger P for females.

But such a difference would not imply that more females thanmales receive higher education, because costs for females may behigher too. There are three types of costs commonly associatedwith education: direct costs such as tuition fees, opportunity costssuch as forgone earnings or tasks that the child could be doing ifnot in school, and idiosyncratic costs that are randomly distributedamong individuals. Though tuition costs in India may be genderneutral, opportunity costs are not. Studies show that theopportunity cost of a girl’s time spent in education is higher thanfor males, because social norms dictate that the time required by agirl to study competes with time required to perform householdchores (Burra, 2001; Dreze and Saran, 1995). Moreover, duringtimes of household labor shortage, the burden of adjustmentfrequently falls on daughters (Jejeebhoy, 1993).5

A number of studies have confirmed that females tend to havelower schooling enrollment rates than males at various levels ofeducation (Kingdon, 2007; Nambissan, 1995; Jayachandran, 2002;Tilak, 2002). One interpretation of this evidence could be that theoverall balance between opportunity costs and education pre-miums may differ between genders, such that benefits may exceedcosts less often for females than for males. Such a conclusion would

5 Furthermore, social norms are such that females tend to move away from the

parent household after marriage while males tend to stay, thereby limiting the

returns on the educational investment in females that can be appropriated by the

parent household. All of these cost differentials may work against the household’s

incentive to educate females.

be based on examining mean values of returns and costs ofeducation. But the causes of lower educational enrolments forfemales may include considerations beyond mean values. Ratherthan characterizing distributions for each gender with differencesin mean values of premiums and costs, as discussed above, assumethat each gender faces equal premiums for a fixed level of costs, butthat uncertainty regarding returns to education is different formales and females. Assume further that these differences inuncertainty for males and females are denoted, respectively, as B

and B0 in Fig. 1. Despite equal premiums and costs betweengenders, risk averse parents will be more willing to educate malesthan females. Thus, ceteris paribus, if households have a preferencefor educating males over females at a given level of education, thiscould be caused by either a higher net return to education, and/orlower uncertainty associated with the male’s earning distributionrelative to female’s.

From the discussion above, it follows that explaining enrol-ments based solely on average returns to education can only becorrect if any of the following three conditions is true: (1) there isno uncertainty in returns to education; (2) uncertainty regardingreturns to education between genders is the same (i.e. thedistributions between genders are identical); (3) households arerisk neutral so that uncertainty does not matter. In our analysisbelow, we assume that households are not risk neutral andhypothesize that there is uncertainty in educational returns thatdiffers between genders. We therefore investigate whether thereare differences in distributions of returns to education betweengenders that could influence incentives to educate boys versusgirls.

3. Study context and data

India provides an excellent case for our analysis. After theeconomic reforms in the 1990s, economic growth in India hasaccelerated sharply, driven mostly by the services sectors. Scholarsworry, however, that the skill intensive growth is not inclusive andmay not be sustainable due to skill shortages in India’s labormarkets (Bargain et al., 2009; Banga, 2006). To deal with themultifaceted problem of closing the skill gap while making growthmore pro-poor, experts on India’s economy are emphasizing theneed for new reforms that will expand higher education (Kochharet al., 2006; Banerjee, 2006).

Special attention is being paid to expanding education amongwomen who, in many respects, remain an untapped resource inIndia’s economy. Studies show that shifts in demands of industries


within the newly liberalized and growing economy has signifi-cantly increased the wage premium for skilled or educatedworkers, especially for skilled women. Rising returns to educationhave helped women narrow their wage gap with men, but theseincreases have also been noted as being as the biggest contributorto rising wage inequality in the country (Kijima, 2006; Chamar-bagwala, 2006). These macroeconomic developments can beexpected to positively affect micro-level decisions regarding theeducation of females (e.g. see Kingdon and Theopold, 2008).However, information on the uncertainty of educational returns,and how it may be expected to affect household decisions toeducate children, is conspicuously absent for India. The currentpaper is a first attempt to addresses this gap in the literature.

We use a nationally representative dataset on labor markets inIndia, collected in 2005–2006. The data are drawn from the 62ndround of the quinquennial surveys on employment and unem-ployment conducted by the National Sample Survey Organizationof India. The survey spans over 4798 villages and 5125 urban areablocks encompassing 78,879 households and 377,377 people.Households were selected using a two stage stratified randomsampling design, whereby villages and urban area blocks areselected in the first stage and households in the second. Since thedata are not drawn from a random survey of the Indian population,multipliers representing survey probabilities for each individualare included in the survey information. We use these multipliers inthe analyses that follow.

The dataset provides detailed household and individual levelinformation on labor market engagement. Information on vari-ables such as occupational choice, type of occupation (usingindustry and occupation classifications), gender, age, educationalattainment level, region of residence, social and religious affiliationare available for all individuals in the sample. A person who couldread, write and understand a simple message in any language wasconsidered literate. For literates, the level of education recordedwas the highest level of education successfully completed. Wefocus on the wage employment portion of the labor market,6 whichprovides us with a sample of 31,541 individuals, aged 15–60 years.The sample consists of 5156 women and 26,385 men. The lownumbers for women reflects India’s notoriously low female labormarket participation.

Wage earners are defined as persons who receive a wage on aregular basis through a formal contract, though without provisionsfor the periodic renewal of the contract. Thus, our sample does notinclude casual wage laborers who work informally without formalcontracts. Information on daily wage earnings, which are based ona 7-day reference week, includes payments received in cash and inkind. Wage workers typically receive bonuses and other amenitiessuch as free lodging, medical treatment, telephones, etc. which areevaluated at retail prices and included in wage earnings.

4. Modeling approach

Empirical models of educational returns are commonly derivedfrom life cycle earnings frameworks whereby individuals facechoices among educational alternatives associated with differentprobability distributions of income (Becker, 1975). Individuals are

6 Despite the importance of the self employed and casual labor markets,

investigating returns to education among these segments of the population would

require much farther ranging analyses, which are beyond the scope of this paper.

For the self employed, earnings data is unreliable and hence not collected by the

National Sample Survey Organization. For casual labor markets, studies (e.g. Dutta,

2006) have shown that these markets are sufficiently different from regular wage

markets, so that the two types of markets should not be pooled. Accordingly, past

studies in India on returns to education do not simultaneously address all three

types of employment (e.g. Duraisamy, 2002). Therefore, our analysis, like past

studies, is not representative of the entire labor market.

assumed to invest in education up to the point where the cost ofthe investment equals the net present value of the benefits. Usingthis result, Mincer (1974) showed that the average rate of return tothe investment can be estimated, under certain restrictions, using aregression of log wage earnings on education and experiencevariables. Otherwise known as the ‘‘human capital earningsfunction’’, this empirical model is one of the most estimatedrelationships in economics. A recent evaluation of the approachand its practice for over thirty years reveals that the Mincerequation still remains a correct benchmark for wage determination(Lemieux, 2006a). Our basic econometric framework, therefore,consists of estimable human capital earnings functions for menand women.

Uncertainty in educational returns, in a small but growingliterature, has been incorporated using alternative variancecomponent models. Two sources of uncertainty are emphasizedin the theoretical literature: (a) unobserved heterogeneity amongindividuals who are otherwise equally endowed with humancapital that create differences in wages and; (b) market risk whichmay produce unpredictable wages for individuals with the samelevel of human capital (Hartog et al., 2003). We adopt an approachthat captures both sources of variation in returns. We introduce arandom parameter for the coefficient of education in a standardMincer earnings equation (see Harmon et al., 2003; Lemieux,2006b). The random parameter on a given level of education isassociated with deeper parameters, a mean and a standarddeviation, that deliver a distribution of estimated returns at thatlevel. Examining the statistical significance of this standarddeviation allows us to formally test if returns are uncertain.

For each individual i, we specify the following wage earningsfunction:

yi ¼X

m

aimaim þX

n

bnwin þ ei (1)

where the outcome variable yi denotes wages (in logs), aim denotethe random parameters on a set of m education dummy variables(primary through college–university) and the intercept term, bm

denote nonrandom parameters on a set of n control variables(including technical degree holder, experience, caste, religion andurban residence) and e denotes a random error term whichcaptures wage uncertainty not accounted for by human capital andother factors. We estimate Eq. (1) separately for each gender. Weseek to estimate the mean and the standard deviation of each of therandom parameters. The latter is our measure of uncertainty ineducational returns. Other than the random parameter specifica-tion, Eq. (1) is a standard human capital specification that has beenused previously for studying returns to education in India (e.g.Duraisamy, 2002).

The model in Eq. (1) is estimated using simulated maximumlikelihood methods (see Green, 2008). Following this approach, therandom parameter vector is assumed to be distributed with meana and a random component e, viz., ai = a +r ei. The coefficientmatrix5 characterizes the variances and cross correlations acrossthe random parameters. All parameters of the model, including thenonrandom parameters, b, and the means and variances of therandom parameters, a, can be estimated conditional on sampledraws of ei. This set of estimates, however, represents parametervalues that are conditional on a specific draw from the distributionof the latent heterogeneity across individuals. The set ofunconditional estimates are obtained by taking the expectationof the estimated parameter values over the entire distribution of arandom parameter. The expectation is an integral defined over theconditional distribution over the range of the random parameter,and is approximated by averages of a large number of randomdraws from the distribution of ei (see Train, 2003).

Table 1Descriptive statistics of data used in the human capital earnings function.

Variables Mean Average daily wagea

All workers Male Female

164.7 171.4 132.20

Illiterate (=1) 0.17 83.38 94.59 54.39

Primary (=1) 0.11 99.76 104.12 74.45

Middle (=1) 0.19 108.15 112.44 71.35

Secondary (=1) 0.15 154.29 159.48 117.14

Higher secondary (=1) 0.10 182.11 188.31 142.72

College–university (=1) 0.28 282.23 292.22 241.69

Technical degree (=1) 0.03

Experience 21.31

Muslim (=1) 0.10

Schedule caste and

schedule tribe (SCCT = 1)

0.22

Other backward castes

(OBC = 1)

0.35

Urban location (=1) 0.58

a Mean values of wages (in Rupees).


Table 1 shows descriptive statistics of the variables used inEq. (1). Information on educational attainment is coded as fivedummy variables: primary school, middle school, secondaryschool, higher secondary school and college/university. The basecategory is below primary school and includes illiterates. The tableshows that more wage earners have a college/university level ofeducation (i.e. 28%) than any other education level.

The average daily wage was Rs. 164, with men receiving almostRs. 40 greater than women. The gender difference in average wageearnings persists at higher education levels. The table shows thatwages are an increasing function of education levels for allindividuals taken together and for men. But for women, meanwages at the middle school level actually drop relative to primaryeducation. These descriptive results are consistent with thefindings of previous studies (e.g. Dutta, 2006) who found thatwomen’s middle school education is not effective at raising theirwages.

In addition to the 6 levels of education, we also define a dummyvariable indicating technical skills possessed by an individual, asindicated by an additional graduate level technical certificate ordiploma in a specialized discipline such as business managementor engineering. The table shows that only 3% of the samplepossesses these skills. We also include a labor market experiencevariable, calculated as an individual’s age minus the number ofyears of schooling minus six.7 The table indicates that the averageindividual in our sample has 21 years of experience.

We include religious and caste classifications of workers tocontrol for the potential influence of social stratifications on wages.For religions, we use a dummy variable that indicates whether anindividual is Muslim. Table 1 indicates that 10% of the population isMuslim. Castes are social stratifications that can restrict specificgroups to specific occupations, thereby preventing their social andeconomic mobility. Though the caste system has been outlawed inIndia, evidence suggests that discrimination persists (e.g. Ito,2009). The government of India provides affirmative action againstdiscrimination in the form of preferential access to jobs andeducation for people belonging to lower caste groups. The peoplethat qualify for such affirmative action, taken as a group, arereferred to as scheduled caste or scheduled tribe (SCST). There is aheated debate regarding whether another strata of low castecategories, other backward classes (OBC), are also disadvantaged

7 For the education system in India, the number of years in school is, respectively,

5, 3, 2, 2 and 3 to complete primary, middle, secondary, higher secondary and

college education.

relative to upper castes and qualify for affirmative action. In Table1, we denote castes using three dummy variables: SCST, whichmake up 22% or our sample, OBC, which make up 35% of oursample, and upper caste (other) which make up the remainder.

Finally, we include a dummy variable to distinguish betweenurban and rural locations of the sample. The urban dummy variablein Table 1 indicates that 58% of the sample is situated in urbanareas.

We do not control for the possible endogeneity of the educationvariables for a number of reasons. First, there is no information inthe dataset to construct commonly used instrumental variables,such as parental characteristics, to deal with the possible bias.Moreover, Harmon et al. (2003) argue, based on the work ofCarneiro et al. (2001), Card (2001), Ashenfelter et al. (1999), Manskiand Pepper (2000), that if educational returns are heterogeneousacross individuals, instrumental variable estimation does not workand can bias estimates of the average educational returns. Harmonet al. also show, based on the work of Ashenfelter et al. (1999),Manski and Pepper (2000) and Griliches (1977) that the biases dueto endogeneity and sample selection could work in oppositedirections, thereby cancelling each other and leaving the OLSestimates to be the best approximation of average returns.Moreover, Koop and Tobias (2004) show that any bias that affectsestimates of the mean return to education does not necessarilyaffect estimates of the variance, which is the main focus of ouranalysis. Finally, because no other studies in India have controlledfor endogeneity, including the benchmark study of Duraisamy(2002), our study is comparable to previous results. This point isimportant because we seek to complement previous studies thatfocus on average returns, by producing estimates of the variance ofreturns.

5. Results

We divide the discussion of our results into three sectionscorresponding to the three approaches outlined in the introduc-tion. The first section describes wage distributions conditional oneducation levels and gender. The second section discusses theeconometric estimates of the earnings functions and tests for thepresence of uncertainty. The third section presents the distribu-tions of annual rates of return to the various levels of educationundertaken by men and women, and examines the potential for anindividual to draw a negative return to education.

5.1. Descriptive analysis: conditional wage distributions

Our approach in this section is to provide a visual depiction ofwage distributions for men and women conditional on each level ofeducation. If education yields positive labor market rewards onaverage, then we would expect to see a rightward shift in wagedistributions that correspond with higher levels of education. Ifhigher education is associated with greater wage uncertainty forindividuals, then we would expect the wage distribution to bewider, with density spread more over the range of wages at highereducation levels. Moreover, gender differences in these distribu-tional patterns will allow us to gather preliminary evidence onwhether women face greater uncertainty in educational returnsthan men, and if the uncertainty increases with education levels.

Fig. 2a and b shows kernel density estimates of wages, fordifferent levels of education, for men and women, respectively. Thedensities are estimated using an Epanechnikov kernel and abandwidth of 0.25. For both genders, there is a general shift of thewage distribution to the right for each successive level ofeducation, indicating that higher education is associated withincreased expected wages. The gain in expected wages is morepronounced at higher education levels (e.g. between higher

[(Fig._2)TD$FIG]

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2.6 2.8 2.9 3.1 3.3 3.5 3.7 3.8 4.0 4.2 4.4 4.6 4.7 4.9 5.1 5.3 5.5 5.6 5.8 6.0 6.2 6.3 6.5 6.7 6.9

Den

sity

Female wage (logs)

Illiterate Primary Middle Secondary H.Secondary College-University

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

2.6 2.8 2.9 3.1 3.3 3.5 3.7 3.8 4.0 4.2 4.4 4.6 4.7 4.9 5.1 5.3 5.5 5.6 5.8 6.0 6.2 6.3 6.5 6.7 6.9

Den

sity

Male wage (logs)

Fig. 2. Wage distributions by education level and gender. India 2005–2006.


secondary and college) than at lower levels (e.g. between primaryand middle school). These findings reflect the general consensus inthe literature that investing in education is a profitable endeavor inIndia (e.g. Kingdon, 1998). In addition, the density graphs revealthat there is a general tendency for more uncertainty at higherlevels of education, indicated by wage distributions that are flatter,with density spread more over different wage levels.

There are notable gender differences in the expected wagescorresponding to higher education levels. These differences aremore evident if the distributions of both genders are shown on asingle graph, which we do in Fig. 3a–f. Specifically, the magnitudeof the increase in expected wages for an illiterate woman, whoattains primary education, is substantially larger than the gain toan illiterate man making the same transition. This result under-scores the importance of literacy for women’s welfare in India.Relative to men, however, women’s gains for a primary schoolstudent who continues on to middle school education are quitemodest. In general, the distributions of men’s wages consistentlypeak at higher wage levels than women’s, revealing gender wagegaps in India’s labor market. At most higher levels of education, thehorizontal differences between the peaks of the distributionsincrease.

There are striking differences in the nature of uncertainty facedby women and men as they become more educated. We makethree key observations in this regard. First, women face wider, andhence more uncertain, wage distributions at higher education

levels than men. Second, with higher education levels, the densityin the distribution of men’s wages loses mass in the bottom andgains mass at the top. That is, more educated men face a decreasedprobability of getting a low- and middle-wage, and an increasedprobability of getting a high wage. While this result is also evidentfor women at the college–university level, women with secondaryand higher secondary educations, in contrast, have density spreadin a bimodal pattern across low and high levels of wages. Finally,the results show that the wage gap between men and women inIndia’s labor markets is not a constant parameter but varies acrossthe wage distribution and across levels of education.

These results suggest that uncertainty is an important feature ofeducational returns in India. Further, the uncertainty appears to besystematically different between males and females. Given thesefindings, incentives for investing in education in India are unlikelyto rely solely on anticipated average returns. However, thesedescriptive findings do not control for other observed determi-nants of the wage distributions. Given the visual differences in thedistributions of wages across education levels and betweengenders, we proceed by investigating further these sources ofheterogeneity using econometric analysis.

5.2. Earnings functions and uncertainty estimates, by gender

Table 2 shows estimates of the random parameter humancapital earnings function for men and women. Most of the

[(Fig._3)TD$FIG]

0.2

.4.6

.8D

ensi

ty

3 4 5 6 7log wage

Female Male

Wage conditional on illiteracy

0.2

.4.6

.8D

ensi

ty

3 4 5 6 7log wage

Female Male

Wage conditional on primary education

0.2

.4.6

Den

sity

3 4 5 6 7log wage

Female Male

Wage conditional on middle school education

0.1

.2.3

.4.5

Den

sity

3 4 5 6 7log wage

Female Male

Wage conditional on secondary education

0.1

.2.3

.4.5

Den

sity

3 4 5 6 7log wage

Female Male

Wage conditional on higher secondary education

0.2

.4.6

Den

sity

3 4 5 6 7log wage

Female Male

Wage conditional on college education

a b

c d

e f

Fig. 3. Wage distributions for each education level by gender. India 2005–2006.

8 In addition to accounting for regional heterogeneity with the urban rural

differentiation, we also ran our models with regional fixed effects. Following Das

(2008) we included a set of regional dummies (north, south, east, west and north-

east). For both men and women, the results were robust and stayed qualitatively the

same. That is, the individual heterogeneity in returns to wages remained significant

even after accounting for aggregate regional level variation. Results with these

regional fixed effects are available on request.


estimated coefficients are significant with the expected signs. TheR-squares in both equations suggest an acceptable fit. Thecoefficients for the means of the random parameters associatedwith the schooling variables are positive, statistically significantand increasing in levels for both genders, conforming to thepositive education-earnings profiles estimated for India in previ-ous studies (e.g. Duraisamy, 2002; Dutta, 2006; Tilak, 2007). Theeffect of experience on earnings, for both men and women, alsoexhibits the familiar concave pattern, though the non-squaredexperience variable for males in not significant. A technical degreeis found to have a positive and significant effect on wages for bothgenders.

Religious, social and geographical factors also affect labormarket earnings. Overall, Muslims receive lower earnings thannon-Muslims. Moreover, relative to the upper caste category,belonging to SCST or OBC has a negative effect on wages, althoughthe impact of SCST is statistically significant only in the maleequation. The magnitude of the disadvantage of belonging to alower caste, as indicated by the coefficient of the caste dummies, ismuch higher for OBCs relative to SCST, providing some evidence tojustify affirmative action towards OBCs. The dummy variable for

urban residence is positive and significant, reflecting higher dailywages in urban areas.8 We estimate wages in urban areas to be 9%higher for men, and almost 20% higher for women.

Table 2 also shows the presence of significant uncertainty ineducational returns. With the exception of primary education forfemales and middle education for males, the standard deviationcoefficient for each education level for each gender is statisticallysignificant. Moreover, coefficients of these standard deviationsgenerally become larger with education levels, suggesting thathigher education levels are associated with greater uncertainty inreturns. We also see that the standard deviation coefficients aregenerally larger for women than they are for men. This resultsupports our descriptive evidence that in India’s labor marketswomen face more uncertain wage distributions compared to men.

Table 2Human capital earnings function with individual heterogeneity.

Men’s human capital earnings function Women’s human capital earnings function

Coefficient Standard error Coefficient Standard error

Random parameters

Primary: mean 0.156 (0.014)** 0.195 (0.038)**

Standard deviation 0.020 (0.010)* 0.0004 (0.032)

Middle: mean 0.346 (0.012)** 0.476 (0.035)**

Standard deviation 0.005 (0.008) 0.151 (0.027)**

Secondary: mean 0.545 (0.012)** 0.847 (0.033)**

Standard deviation 0.192 (0.008)** 0.413 (0.027)**

Higher secondary: mean 0.758 (0.013)** 1.096 (0.036)**


Graduate/university: mean 1.198 (0.012)** 1.559 (0.031)


Constant term: mean 3.354 (0.016)** 2.777 (0.044)**


Non-random parameters

Post graduate degree 0.158 (0.020)** 0.269 (0.047)**

Experience 0.062 (0.001) 0.059 (0.002)**

Experience squared �0.001 (0.002)** �0.001 (0.004)**

Muslim �0.027 (0.010)** �0.021 (0.033)

SCST �0.028 (0.009)** 0.031 (0.023)

OBCj �0.145 (0.008)** �0.187 (0.022)**

Urban 0.098 (0.008)** 0.196 (0.021)**

R-squared 0.44 0.42

* Significant at the 10% level.** Significant at the 5% level.

Table 3Annual returns to education (%).

Annual returns

(male)

Annual returns

(female)

Mean S.D. Mean S.D.

Primary 3.00 0.40 3.80 0.00

Middle 6.33 0.83 9.33 5.00

Secondary 10.00 9.75 18.50 28.00

Higher secondary 10.50 22.00 12.50 45.00

Graduate–university 14.67 19.33 15.33 30.33


5.3. Distributions of returns to education by gender and education

level

Following the format of previous human capital earningsstudies, we use the estimated coefficients in Table 2 to calculate, inTable 3, the annual rate of return for each level of education formen and women. The annual rate of return to education level k isthe difference in the estimated coefficients for level k and theprevious level, over the number of years spent in level k (seeDuraisamy, 2002).9 We use the same formulation to obtain theannual standard deviation of returns. Further, to facilitatecomparisons across genders, we graph the distributions inFig. 4a–d.10 Using the property of normal distributions we alsoshow in the same graphs the probability that an individual willreceive a negative educational return.

Consistent with the previous literature, we find that educa-tional returns, with the exception of higher secondary education,are increasing in education levels for both men and women (Table3). In contrast to the gender wage gaps in our descriptive analysiswhich show women lagging behind men, the average returns toeducation are higher for women at each level – with the biggestgender gap in returns occurring at the secondary education level

9 Footnote 7 contains the number of years in school for each level of education.10 Distributions of returns to education are shown only for the education levels

with statistically significant uncertainty coefficients (using confidence levels of 10%

or less).

(Table 3, row 3). The latter finding can be taken to suggest that,holding constant other factors such as the non-pecuniary benefitsof education, it may be beneficial to expand women’s enrollment atthe secondary level to help close the gender wage gap in India’slabor markets.

Table 3 also shows that the uncertainty in educational returns,ranges from approximately 0% at the primary level to as high as45% at the higher secondary level. The uncertainty increases witheach level of education for both men and women up to the highersecondary level, and then drops somewhat for the graduate–university level. Moreover, the standard deviations are higher formen than for women at each level of education, except theprimary level (Table 3). Overall, these uncertainty estimates ofeducational returns are much higher than that observed indeveloped countries. For example, Koop and Tobias (2004)estimated the standard deviation of educational returns in theU.S. to be around 7%.

These results provide strong support for our hypothesis, thatuncertainty in educational returns plays a role in explaining gendergaps in enrolment. Our finding, of more uncertain returns forfemales, imply that risk averse households would invest less infemales even if the gross returns and costs of education weregender neutral.

Though average returns to education, also documented inprevious studies, appear promising, our results show that asubstantial number of people – both men and women – are likelyto receive a negative return to education (Fig. 4a–d). Aftercontrolling for observable characteristics, we find that men facea negligible probability of getting a negative return to middleschool education. At higher education levels this likelihood rises.We estimate the probabilities associated with a man receiving anegative return to be 15%, 31% and 22% at secondary, highersecondary and college/university education levels, respectively.For women, although the returns are much higher than men’s onaverage, the corresponding likelihoods of negative returns arelarger and increasing almost monotonically in education levels. Weestimate the likelihoods of negative returns for women to be 3%,25%, 39% and 30% for middle school, secondary school, highersecondary school and college/university education, respectively.

[(Fig._4)TD$FIG]

Probability of NegativeReturns:Male=1.20E-14Female=0.03

0.1

.2.3

.4.5

-10 0 10 20 30

Returns to Middle School Education

Probability of NegativeReturns:Male = 0.15Female=0.25

0.0

1.0

2.0

3.0

4

Den

sity

Den

sity

-100 -50 0 50 100 150

Returns to Secondary SchoolEducation

Male Female


0.0

05.0

1.0

15.0

2

Den

sity

-100 -50 0 50 100 150

Returns to Graduate/University Education

Male Female


0.0

05.0

1.0

15.0

2

Den

sity

-200 -100 0 100 200

Returns to Higher Secondary School Education

a

b

c

d

Fig. 4. Distributions of annual returns to education at middle school (a), secondary levels (b), higher secondary school (c), and graduate/university levels (d) for men and

women.


6. Conclusions

In this paper we examine whether educational returns in India’slabor markets are characterized by uncertainty and whether theuncertainty is systematically different for men and women. In ourdescriptive analysis, we find evidence that wages are subject tosubstantial uncertainty within each education level. Our multivar-iate analysis, which allows us to control for observed sources ofvariation in wages, shows that educational returns vary widelyacross individuals. We find the magnitude of the uncertaintycoefficients, measured by the standard deviation of the wage-education slope parameter estimated across individuals, to bestatistically significant for all levels except for women’s primaryeducation.

Our results indicate further that uncertainty in educationalreturns increases with education levels. For instance we estimatethat the uncertainty in men’s returns increases from 19% at thesecondary level to 25% and 33% at the higher secondery andcollege/university levels, respectively. Given that the mean returnsto secondary, higher secondary and college/university educationfor men are 54%, 75% and 119%, respectively, the statisticallysignificant uncertainty coefficients are in fact quite quite high. Weestimate that almost 31% of men at the higher secondary level and22% of men at college level are likely to receive a negative return,despite the high mean return.

Our results also indicate that the uncertainty in educationalreturns is higher for women than for men at every level, postprimary education. The uncertainty in women’s returns increasesfrom 41% at the secondary level to 49% and 42% at the highersecondery and college/university levels, respectively. However, themean returns to education, consistent with previous studies ofIndia, are higher for women relative to men.

Our results challenge the sufficiency of using mean returns toeducation as a measure of labor market incentives for individualsor households in India to invest in higher education. Analysesregarding the demand for education and the decision to educatemales versus females in India have thus far been based on thisparameter. For example, ‘‘Save the Children’’ (2005) state that‘‘Looking ahead to secondary school is an incentive for females toattend and perform well in primary school, and reassures familiesthat their investments will pay off’’. Our finding of pervasiveuncertainty in educational returns suggests an additional butoverlooked factor in the education decision-making process. Thestrikingly gendered nature of the uncertainty also raises concernswith regard to the implications for women’s education in a countryalready lagging behind in many gender equality scales.

References

Alderman, H., King, E.M., 1998. Gender differences in parental investment ineducation. Structural Change and Economic Dynamics 9 (4), 453–468.

Anderson, P.M., 2002. Where the Boys No Longer Are: Recent Trends in U.S. CollegePatterns. Mimeo, Department of Economics, Dartmouth College.

Anderberg, D., Andersson, F., 2003. Investments in human capital, wage uncertainty,and tax policy. Journal of Public Economics 87, 1521–1537.

Asian Development Bank, 1989. Asian Development Outlook. Asian DevelopmentBank, Manila.

Ashenfelter, O., Harmon, C., Oosterbeek, H., 1999. A review of estimates of theschooling/earnings relationship, with tests for publication bias. Labour-Eco-nomics 6, 453–470.

Averett, S.L., Burton, M.L., 1996. College attendance and the college wage premium:differences by gender. Economics of Education Review 15 (1), 37–49.

Ben-Porath, Y., 1967. The production of human capital and the life cycle of earnings.Journal of Political Economy 75, 352–365.

Banerjee, A.V., 2006. The paradox of Indian growth: a comment on Kochhar et al.Journal of Monetary Economics 53 (July (5)), 1021–1026 (Elsevier).

Banga, R., 2006. Critical issues in India’s services-led growth. Indian Council forResearch on International Economic Relations, New Delhi Working Papers 171.


Indian Council for Research on International Economic Relations, New Delhi,India.

Bargain, O., Bhaumik, S.K., Chakrabarty, M., Zhao, Z., 2009. Earnings differencesbetween Chinese and Indian wage earners, 1987–2004. Review of Income andWealth 55, 562–587.

Barro, R.J., 1991. Economic growth in a cross section of countries. The QuarterlyJournal of Economics 106 (2), 407–443.

Becker, G.S., 1975. Human Capital, 2nd ed. University of Chicago Press.Burra, N., 2001. Cultural stereotypes and household behaviour: girl child labour in

India. Economic and Political Weekly 36 (5), 481–488.Card, D., 2001. Estimating the return to schooling: progress on some persistent

econometric problems. Econometrica 69, 1127–1160.Carneiro, P., Heckman, J.J., Vytlacil, E., 2001. Estimating the Return to Education

when it Varies Among Individuals. Mimeo, Department of Economics, Univer-sity of Chicago.

Chamarbagwala, R., 2006. Economic liberalization and wage inequality in India.World Development 34 (12), 1997–2015.

Charles, K.K., Luoh, M., 2003. Gender differences in completed schooling. Review ofEconomics and Statistics 83 (3), 559–577.

Das, M.B., 2008. Minority Status and Labor Market Outcomes: Does India HaveMinority Enclaves? Policy Research Working Paper 4653. World Bank.

de Walque, D., 2007. How does the impact of an HIV/AIDS information campaignvary with educational attainment? Evidence from rural Uganda. Journal ofDevelopment Economics 84 (2), 686–714.

Duraisamy, P., 2002. Changes in returns to education in India, 1983–1994: bygender. Age Cohort and Location, Economics of Education Review 21, 609–622.

Dutta, P.V., 2006. Returns to education: new evidence for India, 1983–1999.Education Economics 14 (4), 431–451.

Dreze, J., Saran, M., 1995. Primary education and economic development in Chinaand India: overview and two case studies. In: Basu, K., Pattanaik, P., Suzumura,K. (Eds.), Choice, Welfare, and Development: Essays in Honour of Amartya Sen.Clarendon Press, Oxford.

Gertler, P., Glewwe, P., 1992. The willingness to pay for education for daughters incontrast to sons: evidence from rural Peru. World Bank Economic Review 6 (1),171–188.

Green, W.H., 2008. Econometric Analysis, 6th ed. Prentice-Hall, Upper Saddle River,NJ.

Griliches, Z., 1977. Estimating returns to schooling: some econometric problems.Econometrica 45, 1–22.

Harmon, C., Hogan, V., Walker, I., 2003. Dispersion in the economic return toschooling. Labour Economics 10 (2), 205–214 (Elsevier).

Hartog, J., Plug, E., Serrano, L., Vieira, J., 2003. Risk compensation in wages, areplication. Empirical Economics 28, 639–647.

He, H., 2011. Why have girls gone to college? A quantitative examination of thefemale college enrollment rate in the United States: 1955–1980. Annals ofEconomics and Finance 12 (2011), 43–66.

Heckman, J., 1976. A life cycle model of earnings, learning and consumption. Journalof Political Economy 84, 511–544.

Ito, T., 2009. Caste discrimination and transaction costs in the labor market:evidence from rural North India. Journal of Development Economics 88 (2),292–300.

Jacob, B.A., 2002. Where the boys aren’t: non-cognitive skills, returns to school andthe gender gap in higher education. Economics of Education Review 21 (6), 589–598.

Jayachandran, U., 2002. Socio-economic determinants of school attendance in India.Delhi School of Economics Working Paper, No. 103.

Jejeebhoy, S.J., 1993. Family size, outcomes for children, and gender disparities.Economic and Political Weekly 28 (35), 1811–1821.

Kingdon, G., 1998. Does the labour market explain lower female schooling in India?Journal of Development Studies 35 (1), 39–65.

Kingdon, G.G., 2007. The progress of school education in India. Oxford Review ofEconomic Policy 23 (2), 168–195.

Kingdon, G., Theopold, N., 2008. Do returns to education matter to schoolingparticipation. Education Economics 16 (4), 329–350.

Kijima, Y., 2006. Why did wage inequality increase? Evidence from urban India1983–1999. Journal of Development Economics 81 (1), 97–117.

Kochhar, K., Kumar, U., Raghuram, R., Subramanian, A., Tokatlidis, I., 2006. India’spattern of development: what happened, what follows? Journal of MonetaryEconomics 53 (July (5)), 981–1019 (Elsevier).

Koop, G.M., Tobias, J., 2004. Learning about heterogeneity in returns to schooling.Journal of Applied Econometrics 19 (7), 827–849 ISSN: 0883-7252.

Lauer, C., 2004. Education, gender and earnings in France and Germany: level anddispersion effects. ZEW Discussion Paper 04-54, Mannheim.

Lemieux, T., 2006a. The Mincer equation thirty years after schooling, experience,and earnings. In: Grossbard-Shechtman, S. (Ed.), Jacob Mincer, A Pioneer ofModern Labor Economics. Springer-Verlag.

Lemieux, T., 2006b. Post-secondary education and increasing wage inequality.American Economic Review 1–23.

Levhari, D., Weiss, Y., 1974. The effect of risk on the investment in human capital.American Economic Review 64, 950–963.

Maier, M., Pfeiffer, F., Pohlmeier, W., 2004. Returns to education and individualheterogeneity. ZEW Discussion Paper 04-32, Mannheim.

Manski, C., Pepper, J., 2000. Monotone instrumental variables: with an applicationto the returns to schooling. Econometrica 68, 997–1013.

Mincer, J., 1974. Schooling, Experience and Earnings. Columbia University Press,New York.

Munshi, K., Rosenzweig, M., 2006. Traditional institutions meet the modern world:caste, gender and schooling choice in a globalizing economy. American Eco-nomic Review 96 (4), 1225–1252.

Nambissan, G., 1995. The social context of schooling girl children in India. Per-spectives in Education 11. (3–4).

OECD, 2007. OECD Economic Survey: India. OECD, Paris.Oliveira Martins, J., Boarini, R., Strauss, H., de la Maisonneuve, C., 2009. The policy

determinants of investment in tertiary education. OECD Journal: EconomicStudies 1, 5 OECD Publishing.

Patrinos, A.H., Ridao-Cano, C., Sakellariou, C.N., 2006. Estimating the returns toeducation: accounting for heterogeneity in ability. World Bank Policy ResearchWorking Paper 4040, World Bank. Human Development Network. EducationTeam.

Psacharopoulos, G., Patrinos, H., 2004. Returns to investment in education: a globalupdate. Education Economics 12 (2), 111–134.

Schultz, T.P., 1993. Investment in the schooling and health of women and men.Journal of Human Resources 28 (4), 694–734.

Schultz, P.T., 2002. Why governments should invest more to educate girls. WorldDevelopment 30 (2), 207–225.

Sen, A., 1999. Development as Freedom. Alfred A. Knopf, New York.Smith, L.C., Haddad, L., 1999. Explaining Child Malnutrition in Developing Coun-

tries: A Cross-Country Analysis. Discussion Paper 60, International Food PolicyResearch Institute, Food Consumption and Nutrition Division, Washington, DC.

Summers, L.H., 1994. Investing in All the People: Educating Women in DevelopingCountries. EDI Seminar Paper 45. World Bank, Washington, DC.

Subbarao, K., Raney, L., 1995. Social gains from female education: a cross-nationalstudy. Economic Development and Cultural Change 44 (1), 105–128.

Tilak, J., 2002. Determinants of Household Expenditure on Education in Rural India.Working Paper Series 88. NCAER.

Tilak, J., 2007. Post-elementary education, poverty and development in India.International Journal of Educational Development 27 (4), 435–445.

Train, K., 2003. Discrete Choice Methods with Simulation. Cambridge UniversityPress, Cambridge, UK.

UNESCO (United Nations Educational, Scientific, and Cultural Organization), 2000.Women and Girls: Education, Not Discrimination. UNESCO.

uncertain educational returns in a developing economy

Documents