WPS5878 Policy Research Working Paper 5878 Average and Marginal Returns to Upper Secondary Schooling in Indonesia Pedro Carneiro Michael Lokshin Cristobal Ridao-Cano Nithin Umapathi The World Bank Development Research Group & East Asia Pacific Region Social Protection Unit November 2011 Policy Research Working Paper 5878 Abstract This paper estimates average and marginal returns to per year of schooling for those very likely to enroll in schooling in Indonesia using a non-parametric selection upper secondary schooling, or as low as −10 percent for model estimated by local instrumental variables, and those very unlikely to do so. Returns to the marginal data from the Indonesia Family Life Survey. The analysis student (14 percent) are well below those for the average finds that the return to upper secondary schooling varies student attending upper secondary schooling (27 widely across individual: it can be as high as 50 percent percent). This paper is a product of the Development Research Group; and East Asia and Pacific Region, Social Protection Unit. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank. org. The author may be contacted at numapathi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Average and Marginal Returns to Upper Secondary Schooling in Indonesia Pedro Carneiro† Michael Lokshin Cristobal Ridao-Cano Nithin Umapathi JEL Code: J31 Key words: Returns to Schooling, Marginal Return, Average Return, Marginal Treatment Effect __________________ † Pedro Carneiro is affiliated with University College London, IFS, Cemmap and Georgetown University, Nithin Umapathi is Economist with Social Protection sector in East Asia and Pacific, Michael Lokshin is Advisor, DECRG, and Cristobal Ridao-Cano is a Senior Economist with Education sector in ECA. Carneiro and Umapathi gratefully acknowledge the financial support from the Economic and Social Research Council for the ESRC Centre for Microdata Methods and Practice (grant reference RES-589-28- 0001) and the hospitality of the Development Economic Research Group of the World Bank. Carneiro gratefully acknowledges the support of ESRC-DFID (grant reference RES-167-25-0124), the European Research Council through ERC-2009-StG-240910-ROMETA and Orazio Attanasio‟s ERC-2009 Advanced Grant 249612 “Exiting Long Run Poverty: The Determinants of Asset Accumulation in Developing Countriesâ€?. These are the views of the authors and do not reflect those of the World Bank, its Executive Directors, or the countries they represent. Correspondence by email to numapathi@worldbank.org 1. Introduction The expansion of access to secondary schooling is at the center of development policy in most of the developing world. Analyzing the effects of such expansion requires knowledge of the impact of education on earnings. In contrast with the standard model, much of the recent literature on the returns to schooling emphasizes that returns vary across individuals, and are correlated with the amount of schooling an individual takes (e.g., Card, 2001, Carneiro, Heckman and Vytlacil, 2011). In terms of the traditional Mincer equation, Y  a  bS  u (where Y is log wage and S is years of schooling), b is a random coefficient potentially correlated with S. This has dramatic consequences for the way we conduct policy analysis. In this model there is no single average return that summarizes the distribution of returns to schooling in the population. For example, the individual at the margin between two levels of schooling may have very different returns from all the infra-marginal individuals. Standard instrumental variables estimates of the returns to schooling estimate the Local Average Treatment Effect (or LATE; Imbens and Angrist, 1994), which does not in general correspond to the return to the marginal person, who is most likely to be affected by the expansion of secondary schooling than anyone else in the economy. Furthermore, different policies may affect different groups of individuals. This paper studies the returns to upper secondary schooling in Indonesia in a setting where b varies across individuals and it is correlated with S (which in this paper is a dummy variable indicating whether an individual enrolls in upper secondary school or not). We find that the return to upper secondary schooling for the marginal person who is indifferent between going to secondary schooling or not is much lower than that of the average person enrolled in upper secondary schooling (14.2% vs. 26.9% per year of schooling).1 Finally, we simulate what would happen if distance to upper secondary schooling was reduced by 10% for everyone in the sample, and we estimate that the return to upper secondary schooling for those induced to attend schooling by such an incentive is 14.2%. 1 The estimated average and marginal returns to upper secondary schooling in Indonesia are 96% and 111% respectively. Average years of schooling for those who have and who have not enrolled in upper secondary schooling in Indonesia are 13.133 and 5.341, so the difference between the two is 7.79. We use this number to annualize the returns to schooling from total return. 2 When evaluating marginal expansions in access to school, the relevant quantities are the returns and costs for the marginal student, not the returns and costs for the average student. In spite of the importance of this topic, there are hardly any estimates of average and marginal returns to schooling in developing countries. Two exceptions using Chinese data are Heckman and Li (2004) and Wang, Fleisher, Li and Li (2011). Estimating the extent to which returns vary across individuals is also important if we want to understand schooling choices, and their relationship with labor market performance. Studying who is more likely to attend school, and documenting how returns differ between those who attain high and low levels of education, and across individuals more generally, is essential for learning about the main incentives and constraints to school attendance. We estimate a semi-parametric selection model of upper secondary school attendance and wages using the method of local instrumental variables (Heckman and Vytlacil, 2005). Our data comes from the Indonesia Family Life Survey. Carneiro, Heckman and Vytlacil (2011) use a similar model to estimate the returns to college in the US. Although they examine a different country in a different time period, and a different level of schooling, they also find that the returns to college vary widely across individuals in the US, and that the return to college for the marginal student is well below the return to college for the average student (see also Carneiro and Lee, 2009, 2011).2 These papers document, across very different environments, how important it is to account for heterogeneity in the returns to schooling. They also show how it is possible to take exactly the same data which is used to estimate a measure of the return to education by instrumental variables methods (IV), and extract much more information from it (allowing us to characterize the heterogeneity in returns across individuals). This can be done using fairly standard parametric methods for estimating selection models, or using a more recent non-parametric approach to the same problem. Vytlacil (2002) shows that the assumptions underlying standard IV estimates of the effect of a particular program (such as attendance of upper secondary school) are the same as the assumptions underlying a fairly standard non-parametric selection model, and 2 There exist also papers which estimate returns for average and marginal student but which account only for selection and heterogeneity given by observable variables (ignoring selection on unobservables). One example is Dearden, McGranahan and Sianesi (2004). 3 thus the two are equivalent. Heckman and Vytlacil (2001a, 2005) explain how to estimate the model using the method of Local Instrumental Variables. We present estimates from both parametric and non-parametric models. Both show the importance of heterogeneity. The former are more precise than the latter, but the parametric model is more restrictive. This paper also proposes a methodological innovation. In the presence of multiple control variables, constructing various average parameters of interest (average returns for different groups of individuals) using the framework of Heckman and Vytlacil (2005) requires the estimation of high dimensional conditional densities, which are notoriously difficult to implement. We use instead a simulation method that avoids this high dimensional non-parametric estimation problem (in contrast, Carneiro, Heckman and Vytlacil, 2010, 2011, need to impose restrictive assumptions to reduce the dimensionality of the problem). Since schooling is endogenously chosen by individuals, we require (at least) one instrumental variable for schooling. We propose to use as the instrument the distance from the community of residence to the nearest secondary school (Card, 1995). Distance takes the value zero if there is a school in the community of residence. This variable is a strong determinant of enrolment in upper secondary school. But one could be concerned that the forces driving the location of schools and parents are correlated with wages, implying that distance is an invalid instrument. We discuss this problem in detail. We control for several family and village characteristics, namely father‟s and mother‟s education, an indicator of whether the community of residence was a village, religion, whether the location of residence is rural, province dummies, and distance from the village of residence to the nearest health post. Our assumption is that if we take two individuals with equally educated parents, with the same religion, living in a village which is located in an area that is equally rural, in the same province, and at the same distance of a health post, then distance to the nearest secondary school is uncorrelated with direct determinants of wages other than schooling. We present evidence that this assumption is likely to hold. In particular, we show that, once these variables are controlled for, there is no correlation between the distance to the nearest secondary school and whether the individual ever failed a grade in elementary school, how many times he repeated a grade in elementary school, and whether he had to work while 4 attending elementary school. In addition, we show (using a different sample) that our distance variable is uncorrelated with test scores (Math, Bahasa, Science, and Social Studies) in elementary school. These are very important dimensions of the pre-secondary school experience that are measures of early ability and early home environments, and which we would expect to be correlated with distance to the nearest secondary school if this variable was endogenously determined. Our instrumental variable estimates of the returns to schooling are higher than the returns to schooling for Indonesia found in Duflo (2000) with the qualification that the dataset, the instrumental variable, and the time period are not the same. Petterson (2010) finds similar rates of return using the same year, same data but a different sample and a different instrument variable. This paper proceeds as follows. Section 2 discusses the data. Section 3 reviews the econometric framework. Section 4 presents our empirical results. Section 5 concludes. 2. Data We use data from the third wave of the Indonesia Family Life Survey (IFLS) fielded from June through November, 20003. The IFLS is a household and community level panel survey that has been carried out in 1993, 1997 and 2000. The 2000 sample was drawn from 321 randomly selected villages, spread among 13 Indonesian provinces containing 83% of the country‟s population containing over 43,000 individuals. The sample we used consists of males aged 25-60 who are employed in the labor market and who have reported non-missing wage and schooling information. We consider only salaried workers, both in the government and in the private sector. We exclude females from the analysis because of low labor force participation, and we exclude self-employed workers because it is difficult to measure their earnings. The dependent variable in our analysis is the log of the hourly wage. Hourly wages are constructed from self-reported monthly wages and hours worked per week. The final sample contains 2608 working age males. 3 For a description of the survey see Strauss, J., K. Beegle, B. Sikoki, A. Dwiyanto, Y. Herawati and F. Witoelar. "The Third Wave of the Indonesia Family Life Survey (IFLS): Overview and Field Report", March 2004. WR-144/1-NIA/NICHD. In the appendix we list the main variables we use. 5 In our empirical model we collapse schooling into two categories: i) completed lower secondary or below, and ii) attendance of upper secondary or higher. While this division groups together several levels of schooling, it greatly simplifies the model and is standard in many studies of the returns to schooling (e.g., Willis and Rosen, 1979). The transition to upper secondary schooling is of substantial interest in the Indonesian context given its current effort to expand secondary education. We present both the return to upper secondary schooling, as well as an annualized version of this parameter which we obtain by dividing the estimated return by the difference in average years of schooling completed by those with lower secondary or less and those with upper secondary or more. Upper secondary schooling corresponds to 10 or more years of completed education.4 In order to compare our estimates with the rest of the literature (in particular, Duflo, 2000) in the appendix we also present OLS and IV estimates of returns using a continuous education variable, corresponding to years of completed schooling. We run ordinary least squares (OLS) and IV regressions of log wages on years of schooling. The control variables we use are dummies for age, indicators of the level of schooling completed by each of the parents (no education, elementary education, secondary education; we also have an indicator for unreported parental education), an indicator for whether the individual was living in a village at age 12, dummies for the province of residence, an indicator of rural residence, and distance (in Km) from the community of residence to the nearest community health post. Our instrumental variable for schooling is the distance from the office of the community head to the nearest secondary school. The distance is self-reported by the community head in the Service Availability Roster of the IFLS.5 Table 2 presents descriptive statistics on the main variables used in our analysis. It shows that individuals with upper secondary or higher levels of education have, on average, 108% higher wages than those with lower education. They also have 7.778 more years of schooling. The respondents with an upper secondary education are younger than 4 It is possible to estimate a non-parametric selection model with multiple levels of schooling but the data requirements to do it are very strong. In particular, one needs one instrumental variable for each transition. It is not feasible to pursue this with our dataset. 5 We would have liked to use instead the distance between the community of residence in childhood and the nearest school in childhood. Our hope is that current residence and current school availability are good approximations (as in Card, 1995). We show below that this measure of distance to school is a good predictor of upper secondary school attendance. 6 those without. They are also more likely to have better-educated parents, to have lived in towns or cities at age 12, and to live closer to upper secondary schools, when compared to those with less than an upper secondary education. 3. Theoretical Framework 3.1 A Semi-Parametric Selection Model We consider a standard discrete choice model of schooling, as used in Willis and Rosen (1979) or Carneiro, Heckman and Vytlacil (2011). Consider a model with two schooling levels: Y1  ï?¡1  Xï?¢1  U 1 (1) Y0  ï?¡ 0  Xï?¢ 0  U 0 S  1 if Zï?§  U s  0 (2) Y1 are log wages of individuals with upper secondary education and above, Y0 are log wages of individuals without upper secondary education, X is a vector of observable characteristics that might affect wages, and U1 and U 0 are the error terms. Z is a vector of characteristics affecting the schooling decision. Equation (2) is a reduced form model of schooling. Agents decide whether to enroll or not in upper secondary schooling based on the expected net present value of earnings with and without upper secondary schooling, and costs, which can be financial or psychic. There can also be liquidity constraints. There is heterogeneity and we expect agents with the highest returns to upper secondary schooling ( Y1  Y0 ) to be more likely to enroll in higher levels of schooling. Costs and returns to schooling can be arbitrarily correlated. For a more detailed explanation see Willis and Rosen (1979) or Carneiro, Heckman and Vytlacil (2011). It is convenient to rewrite the selection equation as: S  1 if P(Z )  V (3) P(Z )  FU S (Zï?§ ) and V  FU S (U S ) and FU S is a cumulative distribution function of Us . V is distributed uniformly by construction. This is an innocuous transformation given that US can have any density. Finally, observed wages are: 7 Y  SY1  (1  S )Y0 (4) Notice that the return to schooling is Y1  Y0  ï?¡1  ï?¡ 0  X (ï?¢1  ï?¢ 0 )  U1  U 0 (5) The return to schooling varies across individuals with different X‟s and different U1, U0. We require that Z is independent of ( U1 ,U 0 ) given X, and that Z is correlated with S (see Heckman and Vytlacil, 2005, for the full set of assumptions). These are the usual IV assumptions. In practice we use a stronger assumption: X, Z is independent of U1, U0, US. This stronger assumption is fairly standard in empirical applications of a selection model of the type described here. We discuss the advantages of using this stronger assumption in the empirical section (see also Carneiro, Heckman and Vytlacil, 2011). The marginal treatment effect (MTE) is the central parameter of our analysis. In the notation of our paper it can be expressed as: MTEx, v   E Y1  Y0 | X  x, V  v  (6)  ï?¡1  ï?¡ 0  x( ï?¢1  ï?¢ 0 )  E U1  U 0 | X  x, V  v  The MTE measures the returns to schooling for individuals with different levels of observables (X) and unobservables (V), and therefore it provides a simple characterization of heterogeneity in returns. Heckman and Vytlacil (2005) show how to construct parameters of interest as weighted averages of the MTE. For example: ATE ( x )   MTE ( x, v ) fV |x (v | x )dv ATT ( x )   MTE ( x, v ) fV |x (v | x, S  1)dv (7) ATU ( x )   MTE ( x, v ) fV |x (v | x, S  0)dv where ATE(x) is the average treatment effect, ATT(x) is average treatment on the treated, ATU(x) is average treatment on the untreated (conditional on X=x), and is the density of V conditional on X. A less standard parameter but equally (if not more) important is the policy relevant treatment effect (PRTE), introduced in the literature by Heckman and Vytlacil (2001b). It measures the average return to schooling for those induced to change their enrolment status in response to a specific policy. Obviously, it depends very much on the policy being considered. Consider a determinant of enrolment Z, which does not enter directly 8 in the wage equation. The policy shifts Z from Z=z to Z=z‟. The weights for the corresponding PRTE are: 3.2 Estimating the MTE Assuming that the unobservables in the wage (1) and selection (2) equations are jointly normally distributed the MTE can be estimated a standard switching regression model (see Heckman, Tobias and Vytlacil, 2001). Assume: U 0 ,U1 ,U s ~ N (0, ï?—) (8) where ï?— represents the variance and covariance matrix. Under this assumption: ï?³ U ,1 ï?³ U ,0 1 MTE( x, v)  E (Y1  Y0 | X  x, V  v)  (ï?¡1  ï?¡ 0 )  x( ï?¢1  ï?¢ 0 )  ( S  )ï?† ( P( Z )) S ï?³U S ï?³U S where ï?³ U S denotes variance of U s , ï?³ i2 variance of U i with i = 0,1, ï?³ U S 2 2 ,i covariance between U s and U i , ï?³ i2 , j the covariance between U i and U j and Φ is the c.d.f. of the standard normal. Therefore MTE can be constructed by estimating parameters ï?¡1 , ï?¡ 0 , ï?¢1 , ï?¢ 0 , ï?²1 , ï?² 2 . This model relies on strong assumptions about the distribution of the error terms in equations (1-2). To relax these restrictions, we use the method of local instrumental variables that imposes no distributional assumptions on the unobservables of the model (Heckman and Vytlacil, 2000). In particular, Heckman and Vytlacil (2000) show that: E Y | X , P  MTE( x, v)  | X  x , P v P (9) where, E (Y | X , P)  Eï?›ï?¡ 0  Xï?¢ 0  S ï?¡1  ï?¡ 0   SX ï?¢1  ï?¢ 0   U 0  S U1  U 0  | X , P ï??  ï?¡ 0  Xï?¢ 0  Pï?¡1  ï?¡ 0   PX ï?¢1  ï?¢ 0   E U1  U 0 | S  1, X , P P (10)  ï?¡ 0  Xï?¢ 0  Pï?¡1  ï?¡ 0   PX ï?¢1  ï?¢ 0   K ( P) (K(P) is a function of P, which can be estimated non-parametrically). Therefore, taking the derivative of (10) with respect to P: 9 E Y | X , P  MTE ( x, v )  |X  x ,P v  X ( ï?¢1  ï?¢0 )  K '( P) (11) P V can take values from 0 to 1. However, in practice it is only possible to estimate the MTE over the observed support of P. In our data the support of P is almost the full unit interval, so we are able to estimate the MTE close to its full support. If we had assumed that Z is independent of ( U1 ,U 0 ) given X, instead of full independence between (Z,X) and ( U1 ,U 0 ), it would be much more difficult to estimate the MTE with full support. In that case, for each value of X it is only possible to estimate the MTE over the support of P conditional on X, which in general can be much smaller than the unconditional support of P (for a detailed discussion see Carneiro, Heckman and Vytlacil, 2011). The assumption of full independence of (Z,X) and ( U1 ,U 0 ) is fairly standard and it allows us to use the full support of P. Equations (10) and (11) can be estimated using standard methods. In particular, we use the partially linear regression estimator of Robinson (1988) to estimate ( ï?¢1 , ï?¢ 0 ). Then we compute R  Y  ï?›ï?¡ 0  Xï?¢ 0  PX ï?¢1  ï?¢ 0 ï??. ( ï?¡1 , ï?¡ 0 ) cannot be identified separately from K(P). K(P) and K’(P) is estimated using a locally quadratic regression (Fan and Gijbels, 1996) of R on P. A simple test of heterogeneity in the impact by unobserved characteristics is a test of whether K’(P) is flat, or if E(Y |X, P) is nonlinear in P. If the derivative is flat then heterogeneity is not important. 3.3 Average Marginal Returns to Education Economic decisions involve comparisons of marginal benefits and marginal costs. Therefore it is especially interesting to estimate the returns to upper secondary schooling for those individuals at the margin between enrolling or not. These would be those individuals more likely to change their schooling as a response to a change in education policy. The definition of who is marginal depends on the policy being considered. This is made clear in Carneiro, Heckman and Vytlacil (2010, 2011), who consider three particularly interesting definitions of individuals at the margin: 10 P i ) P  V  ï?¥ , ii ) Zï?§  U s  ï?¥ , iii ) 1  ï?¥. U These correspond to three different marginal policy changes.6 In this paper we estimate the average marginal returns to upper secondary schooling in Indonesia according to the definition of marginal in ii) above, although we could have chosen a different one. The MTE provides a general characterization of heterogeneity in returns and from it we can construct various other parameters. Carneiro, Heckman and Vytlacil (2010) show how it is possible to write the average marginal treatment effect (or AMTE, the return for the marginal person) as a weighted average of the MTE: (12) 3.4 Estimating vs. Simulating the Weights: A New Procedure So far this section has shown how to recover the MTE from the data, and how to construct economically interesting parameters as weighted averages of the MTE. Heckman and Vytlacil (2005) and Carneiro, Heckman and Vytlacil (2010) provide formulas for the necessary weights in equations 7 and 12, conditional on X. Once these are applied it is simple to average over the relevant distribution of X, when that is what is required. In particular, these papers show that: f V | X v   1 1  FP| X v | X  f V | X v | X , S  1  E P | X  FP| X v | X  f V | X v | X , S  0   (13) E P | X  FP| X v | X   FP '| X v | X  f V | X v | X , S ( z )  0, S ( z ' )  1   ï?›F v | X   F v | X ï??dv P| X P '| X f V | X v | X , Zï?§  V  ï?¥   ï?› f P| X v | X  f U S | X F 1 US |Xv | X ï?? E ï?› f U | X Zï?§ | X ï?? S 6 The three policy changes considered are (i) a policy that increases the probability of attending college by an amount α, so that ; (ii) a policy that changes each person‟s probability of attending college by the proportion (1+ α), so that ; and (iii) a policy intervention that has an effect similar to k a shift in one of the components of Z, say Z , so that and for . 11 where f P| X  p | X  and FP| X  p | X  are respectively the p.d.f and the c.d.f. of P conditional on X, fU S | X uS | X  and FU S | X uS | X  are respectively the p.d.f and the c.d.f. of U S conditional on X, and FP '| X  p | X  is the c.d.f. of P conditional on X when Z takes value z’. In practice it is difficult to implement these formulas since they involve estimation of conditional density and distribution functions, and X is a high dimensional object in many applications (there are 28 variables in X in our empirical work). Therefore, Carneiro, Heckman and Vytlacil (2010, 2011) have aggregated X into an index, namely I  X ï?¢1  ï?¢ 0  , and proceeded by estimating conditional densities and distributions of P with respect to I. There is little theoretical basis for this aggregation which makes it quite unattractive. In this paper we use an alternative procedure, which avoids making this aggregation, and sidesteps the problem of estimating a multidimensional conditional density function. Notice that the selection equation relates S, X, Z, and or V. Using the estimates of the parameters of the selection equation, it is straightforward to simulate the following objects: Once we have these, we just need apply them to equations (7) and (12). This simulation procedure is simple, and its steps are described in detail in the appendix. 4. Empirical Results 4.1 Is Distance to School a Valid Instrument? To account for the potential endogeneity of the schooling decision we instrumented schooling with the distance to the nearest secondary school, measured in kilometers.7 In order for this to be a valid procedure distance to school needs to satisfy two assumptions: i) distance to school should affect the probability of school enrolment and ii) it should have no direct effect on adult wages. 7 Distance to the nearest school has been used as an instrument for schooling by Card (1995), Kane and Rouse (1995), Kling (2001), Currie and Moretti (2003), Cameron and Taber (2004) and Carneiro, Heckman and Vytlacil (2011). 12 Condition ii) is controversial if families and schools do not randomly locate across locations in Indonesia. For example, Carneiro and Heckman (2002) and Cameron and Taber (2004) show that individuals living closer to universities in the US have higher levels of cognitive ability and come from better family backgrounds. In fact, it is also true Indonesia that those who have better educated parents are located closer to secondary schools. However, it is possible that school location is exogenous after we account for a very detailed set of individual and regional characteristics, namely: age (or cohort), parental education, religion, an indicator for whether the individual was living in a city or in village at age 12, an indicator for whether the individual lived in a rural area at age 12, dummies for the province of residence, and distance to the nearest health post. One way to investigate the plausibility of this assumption is to check whether distance to the nearest secondary school is correlated with pre-secondary educational outcomes of each individual. If there was non-random sorting of families and schools across locations in such a way that distance to secondary school was correlated with adult wages, it would surely appear in these variables. Table 3 examines whether distance to upper secondary school is correlated with whether an individual ever repeated a grade in elementary school, the number of repetitions in elementary school (both of which are measure of early school success), and whether the individual worked while in primary school. If our instrument is valid it should not be correlated with such early measures. Our results show no apparent correlation between distance to school and these variables. In addition, Table 4 examines comprehensive exam scores in math, science, social studies and Bahasa. The sample used in this table is not exactly the sample used in our regressions, because it is only possible to gather elementary school test scores for a very small proportion of individuals in our original sample. Therefore, in the regression showed in this table, we placed no age or gender restrictions in the sample. Again, we find no correlation between the distance to school and test scores.8 This evidence is suggestive that our empirical strategy is valid. 8 Considering a more restricted sample results in a small number of observations. Our main conclusions are unchanged, but results are fairly imprecise. 13 The first column of Table 5 shows that distance to the nearest secondary school is a strong predictor of enrolment in secondary school. We run a logit regression where the dependent variable is an indicator taking value 1 if an individual ever attended upper secondary school and the regressors include distance to the nearest secondary school and all the control variables mentioned above. The table displays marginal effects of each variable on the probability of enrolling in upper secondary education. We include distance to health post as a proxy for location characteristics and, unlike distance to school, distance to health post does not predict school enrollment. Children of highly educated parents are more likely to attend upper secondary school than children of parents with low levels of education. Catholics and Protestants are much more likely to attend secondary school than Muslims (the omitted category). Children living in small villages and in rural areas were less likely to attend upper secondary school than those living in large cities and urban areas. In the second column of table 5 we present estimates of a more flexible model where the impact of distance on secondary school attendance varies with X. In particular, we interact distance to school with age (which, for a fixed year, also captures cohort), religion, parental education, and rural residence. It is useful to estimate this richer model for two related reasons. First, it is much more flexible than the model in column 1. Second, because by allowing the impact of the instrument to vary will the variables in X we allow the effect of the instrument to vary more. As a result, the standard errors in the IV estimates and in the selection model are smaller than if we just use the model in column 1. Therefore, the basic estimates in this paper will come from this model, while estimates of the simpler model are presented in the appendix (we discuss them below). Notice also that table 5 displays p-values for tests of the null hypothesis that distance to school does not affect upper secondary school attendance. In column 1, we look to the single coefficient on distance, while in column 2 we perform a joint test on all coefficients involving distance. In both columns we reject that distance to school does not determine upper secondary school attendance. There is another important reason why condition ii) may be violated. If regions where schools are abundant are also regions where other infrastructure is also abundant, then we may be confounding the impact of school availability on wages with the impact of other 14 infrastructure on wages (see the argument in Jalan and Ravallion, 2002). This will be true unless labor is perfectly mobile, which is unlikely to be the case in Indonesia. However, we include detailed regional controls in our models which should absorb much of this variation. Therefore, our argument is that our assumption is valid conditional on all the controls we have in the model. In addition, we show that removing these detailed regional controls hardly affects our results, which indicates that this problem is unlikely to be that important in our setting. Perhaps, as Duflo (2004) argues, the response of other infrastructure (be it private or public) to school construction and to a better skilled workforce is very slow. 4.2 Standard Estimates of the Returns to Schooling In order to more easily make a comparison between our data and estimates and those in the literature we start by presenting standard OLS and IV estimates of the returns to schooling. Throughout the paper schooling takes two values: 0 for less than upper secondary, and 1 for upper secondary or above. We use the log hourly wage in 2000 as our dependent variable. The full set of controls consists of: age (or cohort), parental education, religion, an indicator for whether the individual was living in a city or in a village at age 12, an indicator for whether the individual lived in a rural area at age 12, dummies for the province of residence, and distance to the nearest health post. We present ordinary least squares (OLS) and IV results. This is shown in Table 6. The annualized OLS estimate of the return to upper secondary schooling is 9%, while the IV estimate is 12.9%. Recall from table 2 that individuals with upper secondary schooling or above have on average 13.133 years of schooling, while those with less than upper secondary have on average 5.341 years of schooling. The difference between the two groups is 7.792 years of schooling. Using this figure to annualize the returns to upper secondary education we have an OLS estimate of 9% (=70.5%/7.792) and an IV estimate of 12.9% (=100%/7.792).9 9 Appendix table A1 presents OLS and IV estimates where we use years of schooling as the main explanatory variable (as opposed to upper secondary schooling). The first column in this table shows coefficients of an OLS regression of log wages on years of schooling and several controls. The estimated return to a year of schooling is 9.6%. The second column shows the first stage of the two stage least squares estimator, i.e., a regression of years of schooling on the instrument and the control variables. It shows that distance to school is negatively related to schooling attainment. Finally, column 3 shows the IV estimate of the return to schooling, which is 15.7%. 15 These estimates are higher than (but of comparable magnitude to) those in Duflo (2001), although we use more recent data. Petterson (2010) finds a return of 14% using the same data from the same year but a different sub-sample and instrument. As in most of the literature, our IV estimates of the return to education are larger than OLS estimates. Card (2001) suggests that such a finding indicates that returns to schooling are heterogeneous and the marginal individual induced to enroll in school by the change in the instrument has a higher return than the average individual. Carneiro, Heckman and Vytlacil (2011) show that in the case of college attendance in the US, IV estimates can be above OLS estimates even if the marginal individual has a lower return than the average. Another reason why IV can exceed OLS is measurement error in schooling. Although schooling is relatively well measured in the US (Card, 1999) and other developed countries, that is not necessarily the case in Indonesia. In appendix table A2 we also present IV estimates of returns for models where we do not interact the instrument with the variables in X. The point estimate is smaller than the one in Table A1, and the standard error is larger. Nevertheless, the main pattern remains: the IV estimate is much higher than the OLS estimate. In a model with heterogeneous returns, it is not surprising that the instrumental variable is sensitive to the choice of instrument. For the remaining of the paper, we present a parallel set of results in the appendix in which we do not interact the instrument with X in the selection equation.10 Finally, in appendix table A3 we present results were we omit all regional dummies from the model. Our IV estimate is very similar to the ones in tables A1 and A2. This indicates that regional variation in infrastructure, which is correlated with the availability of schooling, is unlikely to be driving our results. OLS and IV estimates hide considerable heterogeneity in returns and, as emphasized in Heckman and Vytlacil (2005), Heckman, Urzua and Vytlacil (2006), and Carneiro, Heckman and Vytlacil (2011), it is not clear what economic question is answered by the IV estimate. In order to further investigate this issue we use the framework of section 3. 10 We do this for two reasons. First, to show that the main patterns in our results are not driven by choosing the specific way the instrument enters the model. Second, because the first stage F-statistic is higher in the case where we use a single IV (F=11.34) than when we use multiple IVs (F=3.62) consisting of distance interacted with different components of X. We will see throughout the paper that using the expanded set of instruments allows us to get similar results and lower standard errors than we use a single (but apparently stronger) instrument. 16 We estimate parametric (assuming joint normality of (U1, U0, US)) and semi-parametric versions of the model (which does not put assumptions on the joint distribution of (U1, U0, US)). 4.3 Average and Marginal Treatment Effect Estimates We start with the semi-parametric model. We construct P as a predicted probability of ever attending upper secondary school from a logit regression of upper secondary school attendance on the X and Z variables of section 3. Table 5, discussed above, reports the coefficients of the logit model. All average derivatives presented in the table have the expected sign. It is only possible to identify the MTE over the support of P. Therefore, we need to examine the density of P for individuals who attend upper secondary school or above and those who do not. This is done in both panels of Figure 1, which shows the distributions of the predicted propensity score (P) for these two groups. The supports for these distributions overlap almost everywhere, although the support at the tails is thin for low values of P among those with upper secondary school or above. We construct the MTE as described in Section 2. In order to estimate K(P) we run a local quadratic regression of R on P, using a Gaussian kernel and a bandwidth of 0.2. The implied MTE(x,v) is computed by calculating the slope on the linear term of the local quadratic regression.11 Figure 2 displays the estimated MTE (which we evaluate at the mean values of the components of X). The MTE is monotonically decreasing for all values of V. Returns are very high for individuals with low values of V (individuals who are more likely to enroll in upper secondary school or facing high costs). The figure demonstrates substantial heterogeneity in the return to schooling, which ranges from 34% for individuals with V around 0.1 to 13% for those with V close to 0.5, and becomes negative for those with values of V close to 1. The fact that returns are the lowest for individuals who are least likely to go to school is consistent with a simple economic model where agents sort themselves based on their comparative advantage. Unfortunately the standard errors on our estimated MTE are quite wide (standard errors are estimated using the bootstrap). However, it is still possible to reject that the 11 The coefficients on X in the outcome equations are presented in table A4 in the appendix. 17 MTE is flat. Table 7 tests whether adjacent segments of the MTE are equal. Take, for example, the first line of the table. In the first column we show the average value the MTE takes when X is fixed at its mean and V takes values between 0.1 and 0.2, while the second column corresponds to values of V between 0 and 0.1. The third column shows the t-statistic for a test of whether the numbers in the first two columns are equal. We reject equality in almost all lines of the table at the 5% significance level. Therefore, we are able to reject that the MTE is flat, even with the large standard error bands shown in figure 2. Standard errors improve when we estimate the MTE assuming joint normality of (U1, U0, US) as shown in Figure 3. The shape of the MTE is declining as in Figure 2, although the normality assumption does not allow the MTE to have a flat section as in Figure 2 so the MTE is declining everywhere, again taking negative values for very high values of V. Table 8 presents average returns to upper secondary schooling for different groups of individuals. The return to upper secondary school for a random person is 12.3%. The return for the individuals who were enrolled in upper secondary schooling is considerably higher, at 26.9%. If individuals who did not go to upper secondary school would have gone there, they could expect the returns of 1.7%. In the parametric case the average parameters are estimated with the assumption of full support. Estimates of the return to the marginal student (AMTE) are robust to the lack of full support. The return to the marginal student is 14.2%, well below the return to the average student in upper secondary school (26.9%). Finally, the last line of Table 8 reports the average return for those induced to attend upper secondary school by a particular policy shift: a 10% reduction in distance to an upper secondary school. This is the parameter needed to understand the impacts of such an education expansion. By coincidence, it is remarkably similar to the MPRTE. In the appendix we show that results are similar but more imprecise when we do not interact Z and X in the selection equation. This is reassuring, and shows the usefulness for the precision of our estimates of accounting for a more flexible model.12 12 See tables A2 , A5 and A6, and figures A1, A2 and A3. 18 5. Conclusion Indonesia has an impressive record of educational expansion since the 1970s. The enrollment rates are nearly universal for elementary schooling and are around 75% for secondary education. There is an ongoing effort to extend universal education attainment to the secondary level. And although enrollment in secondary education continues to rise we find striking inequality in returns to education. The individuals who are most likely to be attracted by educational expansions at the upper secondary level (marginal) have lower average returns than those already attending upper secondary schooling. In this paper we document a large degree of heterogeneity in the returns to upper secondary schooling in Indonesia. We estimate the return to upper secondary education to be 12 percentage points higher (per year of schooling) for the average than for the marginal student. Therefore, efforts aimed at educational expansion will attract students with much lower levels of returns, although the returns are still fairly high for the marginal person, and therefore further expansions are probably justified. But our estimates also show that it is probably not optimal to bring everybody into upper secondary education. Is it possible to reduce such a high degree of inequality in the returns to schooling? There is a growing body of literature that argues that human capital outcomes later in life are largely determined early in life (e.g., Carneiro and Heckman, 2003). It is therefore important for the design of schooling policy to determine whether the inequality in secondary schooling outcomes can be remedied at earlier stages, for example during elementary education or even earlier. In an impressive drive to increase the quantity of education, there should be a renewed emphasis on quality of education that ensures a more relevant learning environment for the disadvantaged children that reduces the inequity in lifetime outcomes. This can be achieved by raising the return for the marginal student, and therefore the increased equity need not cost in terms of efficiency. 19 References Bjorklund, A. and R. Moffitt (1987) , “The Estimation of Wage Gains and Welfare Gains in Self-Selection Models,â€? Review of Economics and Statistics , 69:42-49. Cameron, S. and C. Taber (2004), “Estimation of Educational Borrowing Constraints Using Returns to Schoolingâ€?, Journal of Political Economy , part 1, 112(1): 132- 82. Card, D. (1995), “Using Geographic Variation in College Proximity to Estimate the Return to Schooling “, Aspects of Labour Economics: Essays in Honour of John Vanderkamp , edited by Louis Christofides, E. Kenneth Grant and Robert Swindinsky. University of Toronto Press. Card, D. (1999), “The Causal Effect of Education on Earnings,â€? Orley Ashenfelter and David Card, (editors), Vol. 3A, Handbook of Labor Economics, Amsterdam: North-Holland. Card, D. (2001), “Estimating the Return to Schooling: Progress on Some Persistent Econometric Problems,â€? Econometrica , 69(5): 1127-60. Card, D. and T. Lemieux (2001), “Can Falling Supply Explain the Rising Return to College For Younger Men? A Cohort Based Analysis,â€? Quarterly Journal of Economics 116: 705-46. Carneiro, P. and J. Heckman (2002), “The Evidence on Credit Constraints in Post- secondary Schooling,â€? Economic Journal 112(482): 705-34. Carneiro, P. and J. Heckman (2003), “Human Capital Policy,â€? in Inequality in America: What Role for Human Capital Policy, J. Heckman, A. Kruger and B. Friedman eds, MIT Press. Carneiro, P., J. Heckman and E. Vytlacil (2010), “Evaluating Marginal Policy Changes and the Average Effect of Treatment for Individuals at the Marginâ€?, Econometrica. Carneiro, P., J. Heckman and E. Vytlacil (2011), “Estimating Marginal Returns to Education “, forthcoming in American Economic Review. Carneiro, P. and S. Lee (2009), “Estimating Distributions of Potential Outcomes using Local Instrumental Variables with an Application to Changes in College Enrolment and Wage Inequalityâ€?, Journal of Econometrics. 20 Carneiro, P. and S. Lee (2011), “Trends in Quality Adjusted Skill Premia in the US: 1960 to 2000â€?, forthcoming in American Economic Review Currie, J. and E. Moretti (2003), Mother's Education and the Intergenerational Transimission of Human Capital: Evidence from College Openings, Quartely Journal of Economics , 118:4. Dearden, L., L. McGranahan, B. Sianesi (2004), “Returns to Education for the „Marginal Learner‟: Evidence from the BCS70â€?, CEEDP 45, Center for the Economics of Education, London School of Economics and Political Science. Duflo, E. (2001), “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experimentâ€?, American Economic Review, 91(4), 795-813. Duflo, E. (2004), “The medium run effects of educational expansion: evidence from a large school construction program in Indonesiaâ€?, Journal of Development Economics, 74, 163-197. Fan, J. and I. Gijbels (1996), Local Polynomial Modelling and its Applications, New York, Chapman and Hall. Griliches, Z. (1977), “Estimating the Return to Schooling: Some Persistent Econometric Problemsâ€?, Econometrica. Petterson, G. "Do supply-side education programs targeted at under-served areas work? The impact of increased school supply on education and wages of the poor and women in Indonesia." PhD dissertation (Draft), Department of Economics, University of Sussex Heckman, J. and X. Li (2004), “Selection Bias, Comparative Advantage, and Heterogeneous Returns to Education: Evidence from China in 2000â€?, Pacific Economic Review. Heckman, J., S. Urzua and E. Vytlacil (2006), “Understanding What Instrumental Variables Really Estimate in a Model with Essential Heterogeneityâ€?, Review of Economics and Statistics. Heckman, J. and E. Vytlacil (1999), Local Instrumental Variable and Latent Variable Models for Identifying and Bounding Treatment Effects, Proceedings of the National Academy of Sciences, 96, 4730-4734. 21 Heckman, J. and E. Vytlacil (2001a), “Local Instrumental Variables,â€? in C. Hsiao, K. Morimune, and J. Powells, (eds.), Nonlinear Statistical Modeling: Proceedings of the Thirteenth International Symposium in Economic Theory and Econometrics: Essays in Honorof Takeshi Amemiya , (Cambridge: Cambridge University Press, 2000), 1-46. Heckman, J. and E. Vytlacil (2001b), “Policy Relevant Treatment Effectsâ€?, American Economic Review Papers and Proceedings. Heckman, J. and E. Vytlacil (2005), “Structural Equations, Treatment, Effects and Econometric Policy Evaluation,â€? Econometrica, 73(3):669-738. Imbens, G. and J. Angrist (1994), “Identification and Estimation of Local Average Treatment Effects,' Econometrica, 62(2):467-475. Jalan, J. and M. Ravallion (2002), “Geographic poverty traps? A micro model of consumption growth in rural Chinaâ€?, Journal of Applied Econometrics, 17, 329- 346. Kane, T. and C. Rouse (1995), “Labor-Market Returns to Two- and Four-Year Collegeâ€?, American Economic Review, 85(3):600-614. Kling, J. (2001), “Interpreting Instrumental Variables Estimates of the Returns to Schoolingâ€?, Journal of Business and Economic Statistics, 19(3), 358-364. Robinson, P. (1988), “Root-N-Consistent Semiparametric Regressionâ€?, Econometrica, 56(4), 931-954. Vytlacil, E. (2002), “Independence, Monotonicity, and Latent Index Models: An Equivalence Resultâ€?, Econometrica, 70(1), 331-341. Wang, X., B. Fleisher, H. Li and S. Li, “Access to Higher Education and Inequality: the Chinese Experimentâ€?, working paper, Ohio State University. Willis, R. and S. Rosen (1979), “Education and Self-Selection,â€?Journal of Political Economy, 87(5):Pt2:S7-36. 22 Table 1: Definitions of variables used in the empirical analysis Variable Definition Y Log hourly earnings for salaried males S=1 Ever enrolled in upper secondary school; zero otherwise X Age, age squared, respondent‟s religion – protestant, catholic and other, mother‟s and father‟s education – elementary, secondary or higher, distance to the nearest health post in km from the community, rural residence, province of residence – West Sumatra, South Sumatra, Lampung, Jakarta, Central Java, Yogyakarta, East Java, Bali, West Nussa Tengara, South Kalimanthan, South Sulawesi Z Distance in km to nearest secondary school from community heads office, interactions of distance with age, age squared, religion, parental education and rural residence 23 Table 2: Sample statistics for the treatment groups Upper secondary or higher Less than upper secondary N = 1085 N = 1523 Log hourly wages 8.198 7.481 Years of education 13.133 5.341 Distance to school in km 1.053 1.564 Distance to health post in km 0.889 1.079 Age 37.058 38.675 Religion Protestant 0.050 0.022 Catholic 0.029 0.009 Other 0.062 0.043 Muslim 0.860 0.927 Father uneducated 0.130 0.383 …elementary 0.503 0.507 ...secondary and higher 0.330 0.061 ...missing 0.020 0.037 Mother uneducated 0.201 0.425 …elementary 0.484 0.406 ...secondary and higher 0.204 0.022 ...missing 0.098 0.133 Rural household 0.240 0.476 North Sumatra 0.057 0.063 West Sumatra 0.047 0.058 South Sumatra 0.048 0.032 Lampung 0.016 0.027 Jakarta 0.181 0.095 Central Java 0.085 0.163 Yogyakarta 0.092 0.054 East Java 0.121 0.180 Bali 0.056 0.038 West Nussa Tengara 0.050 0.048 South Kalimanthan 0.040 0.020 South Sulawesi 0.035 0.035 Source: Data from IFLS3. Sample restricted to males aged 25-60 employed in salaried jobs in government and private sectors. Hourly wages constructed based on self-reported monthly wages. 24 Table 3: Regression of elementary education experience on distance to school Number of Failed grade Worked repeats Dist to nearest secondary school in km 0.007 0.011 -0.001 (0.007) (0.008) (0.005) Individual and family controls Yes Location fixed effect Yes Number of observations 2,248 2,244 2,250 R2 0.041 0.043 0.043 Note: Sample restricted to males with the repeated grade information non-missing. The individual and family controls include age, age squared, religion, fathers and mother‟s schooling levels completed, distance to local health outpost, rural and province dummies. Standard errors are robust to clustering at the community level. Standard Errors are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. Table 4: Regression of comprehensive exam test scores from elementary school on distance to school Math Bahasa Science Social Studies Distance to nearest secondary school 0.001 -0.002 -0.004 -0.005 (0.005) (0.005) (0.005) (0.005) Individual and family controls Yes Location fixed effect Yes Number of observations 1,652 1,668 1,621 1,605 R2 0.134 0.187 0.124 0.115 Note: Sample includes everyone with non-missing test scores. Test scores recorded from score cards. The individual and family controls include age, age squared, religion, fathers and mother‟s schooling levels completed, distance to local health outpost, rural and province dummies. Standard errors are robust to clustering at the community level. Absolute t-statistics are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 25 Table 5: Upper school decision model – Average Marginal Derivatives Coef/(se) Average Derivative/(se) Dist to sec school in km -0.123*** -0.0300** (0.040) (0.0127) Age 0.077* 0.0130 (0.044) (0.0090) Age Squared -0.096* -0.0162 (0.055) (0.0111) Protestant 0.730*** 0.1382*** (0.264) (0.0484) Catholic 1.211*** 0.2123** (0.395) (0.0890) Other religions 0.245 0.0552 (0.363) (0.0878) Fathers education elementary 0.766*** 0.1342*** (0.127) (0.0217) Father higher education 1.835*** 0.3769*** (0.178) (0.0320) Mother education elementary 0.443*** 0.0852*** (0.123) (0.0230) Mother higher education 1.851*** 0.3730*** (0.237) (0.0418) Rural -0.593*** -0.1143*** (0.110) (0.0276) Distance to health post in km -0.017 0.0000 (0.040) (0.0083) Location fixed effect Yes Test for joint significance of 9.42/0.0021 22.26/0.051 instruments: Chi-square/p-value Note: This table reports the coefficients and average marginal derivatives from a logit regression of upper school attendance (a dummy variable that is equal to 1 if an individual has ever attended upper school and equal to 0 if has never attended upper secondary school but graduated from lower secondary school. Type of location is controlled for using province dummy variables. Dummy variable for missing parental education is included in the regressions but not reported in the table. The first column presents coefficients of logit where only distance to school is used an IV. In the second column average derivatives are presented and instruments include distance to secondary school and interactions with all the Xs. Reference categories are Muslim, not educated. Standard errors are robust to clustering at the community level. Standard Errors are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 26 Table 6: Annualized OLS and IV estimates of the return to upper secondary schooling OLS 2SLS IV estimate *** Upper secondary (annualized) 0.090 0.129*** (0.005) (0.048) Age 0.052*** 0.048** (0.019) (0.020) Age Squared -0.042* -0.037 (0.023) (0.025) Protestant 0.182** 0.142 (0.084) (0.104) Catholic 0.059 0.001 (0.189) (0.202) Other religions 0.109 0.097 (0.126) (0.125) Fathers education elementary 0.135*** 0.091 (0.048) (0.070) Fathers education secondary or higher 0.215*** 0.101 (0.067) (0.153) Mother‟s education elementary -0.052 -0.080 (0.048) (0.060) Mother‟s education secondary or higher -0.031 -0.128 (0.078) (0.136) Rural household 0.111** 0.152** (0.045) (0.068) Distance to health post in km -0.023 -0.020 (0.018) (0.017) Location controls YES YES Number of observations 2,608 2,608 Test for joint significance of instruments: F-stat/p-value 2.22/0.00 R2 0.210 0.190 Note: This table reports the coefficients for OLS and 2SLS IV for regression of log of hourly wages on upper school attendance (a dummy variable that is equal to 1 if an individual has ever attended upper school and equal to 0 if has never attended upper secondary school but graduated from lower secondary school. controlling for parental education, religion and location. First stage results report average marginal derivative. Excluded instruments are distance to secondary school and interactions with parental education, religion and age. Type of location is controlled using province dummies. Dummy variable for missing parental education is included in the regressions but not reported in the table. Reference categories are Muslim, not educated.Standard errors are robust to clustering at the community level. Standard Errors are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 27 Table 7: Testing for heterogeneity in returns: comparing adjacent sections of the semi-parametric MTE Ranges of US for (0,0.1) (0.1,0.2) (0.2,0.3) (0.3,0.4) (0.4,0.5) (0.5,0.6) (0.6,0.7) (0.7,0.8) (0.8,0.9) LATEj Ranges of US for (0.1,0.2) (0.2,0.3) (0.3,0.4) (0.4,0.5) (0.5,0.6) (0.6,0.7) (0.7,0.8) (0.8,0.9) (0.9,1) LATEj+1 Difference in -0.078 -0.039 -0.013 -0.012 0.00 0.005 -0.014 -0.024 -0.04 LATEs p-value 0.00 0.00 0.00 0.00 0.597 0.759 0.005 0.00 0.00 Note: In order to compute the numbers in this table we construct groups of values of Us and average the MTE within these groups, where and are the lowest and highest values of Us defined for interval j. Then we compare the average MTE across adjacent groups and test whether the difference is equal to zero using the bootstrap with 250 replications. Table 8: Estimates of Average Returns to Upper Secondary Schooling with 95% confidence interval Parameter Non parametric Estimate Normal selection model ATT 0.269*** 0.201*** (.069, 0.47) (0.05,0.35) ATE 0.123* 0.066 (-0.019, 0.266) (-0.029,0.163) ATU 0.017 -0.029 (-0.236, 0.27) (-0.175,0.116) MPRTE 0.142*** (.038, 0.246) PRTE 0.142*** (.038, 0.247) Note: This table presents estimates of various returns to upper secondary school attendance for the semi- parametric and normal selection models: average treatment on the treated (ATT), average treatment effect (ATE), treatment on the untreated (ATU), and marginal policy relevant treatment effect (MPRTE). Returns to upper school are annualized to show returns for each additional year. Standard errors bootstrapped using 250 repetitions. 95% confidence interval in parentheses. Absolute t-statistics are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 28 Figure 1: Propensity score (P) support for each schooling group S = 0 and S = 1 .1 .08 .06 Propensity score by treatment status Fraction .04 .02 0 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 less than upper secondary upper secondary Note: P is estimated probability of going to upper secondary school. It is estimated from a logit regression of upper school attendance on Xs, distance to school, interactions of X and distance to school (See Table 5). 29 Figure 2: Marginal treatment effect with 90% Confidence Interval – Semi-parametric regression estimates 0.9 0.7 0.5 0.3 MTE 0.1 -0.1 0.00 0.10 0.20 0.30 0.40 0.51 0.61 0.71 0.81 0.91 -0.3 -0.5 lower bound MTE upper bound Note: To estimate the E(Y1-Y0|X, Us) function we used a partial linear regression of log wages on X and K(P) ,with a bandwidth of 0.2. X includes age, age squared, religion, parental education, rural and province dummy variables. 90% confidence interval constructed using 250 boostrap repetitions. Values of V on the x-axis. Figure 3: MTE with 90% Confidence Interval – Parametric normal selection model estimates 0.80 0.60 0.40 0.20 MTE 0.00 0.01 0.11 0.21 0.31 0.41 0.51 0.61 0.71 0.81 0.91 -0.20 -0.40 -0.60 MTE Upper Lower Note: Parametric MTE is estimated using the standard switching regression model. 30 Appendix Simulation-based approach for estimating average treatment effects in equations 7 and 12. Step 1: Estimate MTE(x, v) as described in section 3. Step 2: For each individual in the sample construct the corresponding P(Z) and take n draws from (recall that we assumed that V was independent of X and Z). Since there are 2608 individuals in the sample this creates a simulated dataset of size 2608*n (we use n=1000). Evaluate the MTE(x,v) for each value of X and each value of simulated V. Step 3: In this simulated dataset both X and V are observed for all 2608*n observations. In addition, we have estimates of MTE(x,v) for each of them. Therefore it is trivial to construct the following quantities: by respectively averaging the MTE for everyone in the simulated sample, for those who have P>V, and for those with P≤V. Step 4: There is one parameter that remains to be estimated: the AMTE. The version of the AMTE we use in this paper defines marginal individuals as those for whom: Carneiro, Heckman and Vytlacil (2010) show that this is equivalent to estimating the average return to schooling for those induced to enroll in upper secondary schooling when one of the components of Z, say the intercept, changes my a marginal amount. This is exactly what we do in our simulations: we change the intercept of the selection equation marginally and we see which members of our simulated dataset change their schooling decision. Finally, we average the MTE for this group. 31 Table A1: OLS and IV estimates of the return to a year of schooling OLS First stage IV estimate Average Coef se Marginal se Coef se Derivative Years of education 0.096*** 0.005 0.157*** 0.037 Age 0.058*** 0.017 0.027 0.078 0.055*** 0.018 Age Squared -0.047** 0.022 -0.062 0.098 -0.042* 0.022 Muslim Protestant 0.084 0.082 2.033 0.381 -0.037 0.118 Catholic 0.003 0.152 2.196 0.856 -0.117 0.149 Other religions 0.055 0.121 0.987 0.754 0.002 0.128 Father uneducated … elementary 0.062 0.048 1.759 0.228 -0.049 0.080 … secondary or higher 0.135** 0.067 3.627 0.312 -0.083 0.144 Mother uneducated … elementa -0.086* 0.046 1.000 0.216 -0.147** 0.063 … secondary or higher -0.119 0.078 3.173 0.344 -0.316** 0.145 Rural household 0.149*** 0.044 -1.146 0.301 0.234*** 0.073 Distance to health post in km -0.020 0.015 0.037 0.084 -0.015 0.013 Location controls Yes Dist to nearest sec school -0.298*** 0.102 Number of observations 2,608 2,608 Test for joint significance of 3.62/0.000 instruments: F-Stat/p-value R2 0.260 0.204 Note: This table reports the coefficients for OLS and 2SLS IV for regression of log of hourly wages on years of schooling controlling for parental education, religion and location. First stage results report average marginal derivative. Excluded instruments are distance to secondary school and interactions with parental education, religion, age and distance to health center. Type of location is controlled using province dummies. Dummy variable for missing parental education is included in the regressions but not reported in the table. Standard errors are robust to clustering at the community level. Standard Errors are in parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 32 Table A2: IV estimates of the return to a year of schooling without distance and X interactions IV estimate First stage coef se coef se Years of education 0.144*** 0.053 *** Age 0.056 0.017 0.036 0.077 Age Squared -0.043* 0.022 -0.072 0.096 Muslim Protestant -0.011 0.141 2.050*** 0.380 Catholic -0.091 0.164 2.229** 0.906 Other religions 0.014 0.128 0.839 0.778 Father uneducated … elementary -0.025 0.102 1.800*** 0.231 … secondary or higher -0.036 0.198 3.525*** 0.316 … education missing -0.034 0.109 0.353 0.444 Mother uneducated … elementary -0.134* 0.073 0.973*** 0.215 … secondary or higher -0.274 0.185 3.180*** 0.331 … education missing -0.183*** 0.063 0.367 0.301 Rural household 0.215** 0.091 -1.144*** 0.302 Distance to health post in km -0.016 0.013 0.007 0.082 W Java N Sumatra 0.114 0.088 -0.615 0.500 W Sumatra 0.282** 0.112 -0.704 0.476 S Sumatra 0.137 0.125 0.667 0.476 Lampung -0.044 0.108 0.149 0.477 Jakarta -0.077 0.078 0.752* 0.421 * C Java 0.051 0.091 -0.937 0.498 Yogyakarta -0.303*** 0.100 1.128** 0.570 E Java -0.007 0.066 -0.300 0.411 Bali -0.197 0.159 1.027 0.946 W Nusa Tenggara -0.176 0.107 0.715 0.839 *** *** S Kalimantan 0.298 0.114 1.726 0.540 S Sulawesi 0.032 0.097 0.226 0.702 Dist to nearest sec school -0.244*** 0.072 Number of observations 2,608 Test for joing significance of instruments: 11.34/0.00 F-stat/p-value R2 0.206 Note: This table reports the coefficients for 2SLS IV for regression of log of hourly wages years of schooling, controlling for parental education, religion and location. Excluded instruments are distance to secondary school. Type of location is controlled using province dummies. Dummy variable for missing parental education is included in the regressions but not reported in the table. Reference categories are Muslim, not educated. Standard errors are robust to clustering at the community level. Standard Errors are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 33 Table A3: IV estimates of the return to a year of schooling without regional dummies IV estimate coef se *** Years of education 0.135 0.034 Age 0.059*** 0.018 Age Squared -0.046** 0.022 Muslim Protestant -0.032 0.100 Catholic -0.153 0.154 Other religions -0.109 0.091 Father uneducated … elementary -0.006 0.077 … secondary or higher -0.004 0.141 … education missing -0.002 0.107 Mother uneducated … elementary -0.074 0.057 … secondary or higher -0.190 0.131 … education missing -0.156*** 0.060 Rural household 0.227*** 0.072 Distance to health post in km -0.008 0.014 Number of observations 2,608 Test for joing significance of instruments: F-stat/p-value 4.08/0.00 R2 0.22 Note: This table reports the coefficients for 2SLS IV for regression of log of hourly wages years of schooling, controlling for parental education, religion and location. Excluded instruments are distance to secondary school. Type of location is controlled using province dummies. Dummy variable for missing parental education is included in the regressions but not reported in the table. Reference categories are Muslim, not educated. Standard errors are robust to clustering at the community level. Standard Errors are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 34 Table A4: Outcome equation: Partial linear regression estimates Coeffients Standard Errors Age 0.070* 0.042 Age Squared -0.076 0.051 Protestant -0.022 0.368 Catholic -0.816 0.634 Other religions 0.786* 0.406 Father with elementary education 0.042 0.192 … secondary or higher 0.103 0.675 … education missing 0.425 0.292 Mother with elementary education -0.144 0.156 … secondary or higher -1.570* 0.938 … education missing -0.173 0.170 Rural household 0.288* 0.161 Distance to health post in km -0.016 0.030 N Sumatra 0.333 0.214 W Sumatra 0.177 0.218 S Sumatra 0.233 0.309 Lampung 0.253 0.294 Jakarta -0.248 0.233 C Java 0.071 0.153 Yogyakarta -0.127 0.301 E Java -0.071 0.149 Bali -1.022** 0.478 W Nusa Tenggara -0.267 0.325 S Kalimantan 0.013 0.451 S Sulawesi -0.434 0.274 N Sumatra -0.550 0.465 S Sumatra -0.134 0.595 C Java -0.197 0.415 Yogyakarta -0.127 0.602 E Java 0.326 0.357 Bali 1.660* 0.898 W Nusa Tenggara 0.192 0.711 S Kalimantan 0.367 0.860 W Sumatra*P 0.465 0.535 Lampung*P -0.993 0.839 Jakarta*P 0.394 0.452 S Sulawesi*P 0.979 0.598 Age*P -0.069 0.097 Age Squared*P 0.124 0.121 Protestant*P 0.130 0.639 Catholic*P 1.171 0.931 Other religions*P -1.261* 0.703 Father with elementary*P 0.053 0.605 Father with secondary/higher*P 0.002 1.280 Father education missing*P -1.322 0.942 Mother with elementary*P 0.187 0.393 Mother with secondary/higher *P 1.977 1.433 Mother education missing*P 0.109 0.458 Rural *P -0.275 0.362 Distance to health post*P 0.037 0.082 Number of observations 2,608 R2 0.080 *** note: ** * p<0.01, p<0.05, p<0.1 The table presents the coefficients on X and P*X from the Robinson‟s (1988) double residual semi-parametric regression estimator. The logit estimated pscore (P) enters the equation nonlinearly according to a non-binding function and estimated using a gaussian kernel regression with bandwidth equal to 0.2. 35 Table A5: Testing for equality of LATEs over different Intervals of MTE Ranges of US for (0,0.1) (0.1. 0.2) (0.2,0.3) (0.3,0.4) (0.4,0.5) (0.5,0.6) (0.6,0.7) (0.7,0.8) (0.8,0.9) LATEj Ranges of US for (0.1. 0.2) (0.2,0.3) (0.3,0.4) (0.4,0.5) (0.5,0.6) (0.6,0.7) (0.7,0.8) (0.8,0.9) (0.9,1) LATEj+1 Difference in -0.078 -0.04 -0.014 -0.012 -0.010 -0.011 -0.012 -0.014 -0.014 LATEs p-value 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Note: In order to compute the numbers in this table we construct groups of values of Us and average the MTE within these groups, where and are the lowest and highest values of Us defined for interval j. Then we compare the average MTE across adjacent groups and test whether the difference is equal to zero using the bootstrap with 250 replications. Table A6: Estimates of Average Returns to Upper Secondary Schooling with 95% confidence interval Parameter Non parametric Estimate Normal selection model ATT 0.217 0.198** (-.1, 0.525) (-0.041,0.438) ATE 0.13 0.065 (-0.06, 0.32) (-0.099, 0.231) ATU 0.07 -0.028 (-0.227, 0.365) (-0.217, 0.160) Note: This table presents estimates of various returns to upper secondary school attendance for the semi- parametric and normal selection models: average treatment on the treated (ATT), average treatment effect (ATE), treatment on the untreated (ATU), and marginal policy relevant treatment effect (MPRTE). Returns to upper school are annualized to show returns for each additional year. Standard errors bootstrapped using 250 repetitions. 95% confidence interval in parentheses. Absolute t-statistics are in the parentheses with significance at *** p<0.001, ** p< 0.05, * p<0.1 indicated. 36 Figure A1: Propensity score (P) support for each schooling group S = 0 and S = 1 Propensity score by treatment status .1 .08 .06 Fraction .04 .02 0 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 less than upper secondary upper secondary Note: P is estimated probability of going to upper secondary school. It is estimated from a logit regression of upper school attendance on Xs, distance to school, interactions of X and distance to school (See Table 5). Figure A2: Marginal treatment effect with 90% Confidence Interval – Semi-parametric regression estimates (without distance and Xs interactions) 1 0.8 0.6 0.4 0.2 0 0 0.10 0.20 0.30 0.40 0.51 0.61 0.71 0.81 0.91 -0.2 -0.4 -0.6 lower bound MTE upper bound Note: To estimate the E(Y1-Y0|X, Us) function we used a partial linear regression of log wages on X and K(P) ,with a bandwidth of 0.2. X includes age, age squared, religion, parental education, rural and province dummy variables. 90% confidence interval constructed using 250 boostrap repetitions. Values of V on the x-axis. 37 Figure A3: MTE with 90% Confidence Interval – Parametric normal selection model estimates 0.8 0.6 0.4 0.2 0 0.01 0.11 0.21 0.31 0.41 0.51 0.61 0.71 0.81 0.91 -0.2 -0.4 -0.6 MTE upper lower 38