Returns to Education: New Evidence for India, 1983–1999

Abstract

This paper estimates the returns to education for adult male workers in regular and casual wage employment using Indian national survey data at three points in time spanning almost two decades. Both standard and augmented Mincerian wage equations are estimated using a set of human capital measures and other controls after addressing the issue of potential selection bias. This paper finds that the returns to education are significantly different for the two types of workers—while casual workers face at best flat returns to education, the returns to education for regular workers are positive and U‐shaped with respect to education levels. There is also some evidence of a widening wage gap between regular workers with graduate and primary education that could possibly be a consequence of trade liberalization and other reforms pursued during the 1990s.

Key Words:

• Rate of return
• human capital
• India

Acknowledgements

This research was carried out as part of my D.Phil. thesis at the University of Sussex and was supported by grants from the BFWG Charitable Foundation, the Department of Economics at the University of Sussex, the International Foundation of University Women, Universities UK, the Wingate Foundation and the Royal Economic Society and support from the Poverty Research Unit at Sussex. I am very grateful to my supervisors, Dr Barry Reilly and Professor Alan Winters, for their guidance and suggestions. Comments on earlier drafts from two anonymous referees, participants at the PRUS seminar (2003) and from Patricia Justino, Julie Litchfield, Andy Newell, Yoko Niimi and Howard White were also much appreciated.

Notes

1. In this paper we focus on the wage determination process for adult male workers; we extend this analysis to female workers in our other work.

2. Although there has been some criticism of this approach in the literature (see Bourguignon et al., 2004), the Lee correction was chosen over alternative parametric and semi‐parametric approaches (see Vella, 1998) because of its simplicity, computational convenience and transparent interpretation of the selection effect. The Bourguignon et al. alternative allows for the estimation of one outcome equation only, whereas in this paper there are two outcome equations—wage regression models for regular and casual workers. It should be noted that the parameter estimates of the wage equations using power series approximations for the selection term following the semi‐parametric approach advocated by Newey (1999) were very similar to those obtained using the Lee correction.

3. The employment survey for 1987–88 could not be used as over 76% of observations on rural wages for persons participating in wage employment are missing (see also Duraisamy, 2002).

4. Results are available on request.

5. As the choice of identifying variables is necessarily ad hoc the MNL model was estimated for different specifications of identifying variables. The parameter estimates in the wage equations are not sensitive to the choice of the identifying variables and the coefficient on the correction term itself was not materially different across specifications. On balance, these instruments were also not found to strongly influence wages in most specifications in most years.

6. The statistics are 19 087.91, 21 281.50 and 22 068.91 for the three years, respectively. At the same time, the results of the Small–Hsiao test for the independence of irrelevance alternatives assumption are inconclusive, implying correlations in the extreme value errors. The Hausman–McFadden test could not be conducted as the matrix of the differences in the variance–covariance matrices of the restricted and full models was not positive definite.

7. For the standard Mincerian specification, the statistics are 244.66 between 1983 and 1993, 152.28 between 1993 and 1999, and 794.73 between 1983 and 1999 for regular workers, and 47.30, 22.78 and 119.85 for the three years for casual workers. The corresponding figures for the augmented specification are 95.78, 24.54 and 193.21 for regular workers and 22.36, 20.52 and 66.42 for casual workers.

8. The omitted category for the education dummy variables is those who are illiterate or have less than two years of formal education.

9. The 1983 survey has additional information on the number of years of schooling and somewhat confirms this correspondence—on average, individuals that had completed primary, middle, secondary and graduate education had done so in 5.18, 8.17, 10.79 and 14.16 years, respectively.

10. The standard specification yields returns to primary education for casual workers that are comparable with those for regular workers in the 1990s as well as a positive, but very small, return to middle schooling (and to secondary schooling in 1983). Note, however, that the human capital variables as captured by age and education explained only about 5% of the variation in log real hourly wages for casual workers in the standard Mincerian specification (regression not reported).

11. Card (1999) concludes in his survey of the empirical literature that the OLS estimates of returns to education are not widely different from instrumental variables (IV) estimates obtained after correcting for ability bias and measurement error.

12. The education effects are marginally higher than those reported by Duraisamy as the current paper is unable to include a technical education dummy in the analysis—all values for this variable equalled one in the 1983 survey.

13. The estimates from the standard Mincerian model used in this paper are closer to Duraisamy’s estimates than the augmented model (see Table 4).

14. Although using the constructed real hourly wage variable (see Appendix 1) introduces measurement error in the dependent variable, this would be captured in the error term in the wage regression models. On the other hand, using the reported real weekly wage as the dependent variable without controlling for the hours worked or including this as a control might lead to biased parameter estimates due to omitted variable bias or because of correlation of measurement error with the error term.

15. A broadly similar trend for returns to regular workers is apparent from the standard Mincerian model—with the returns to primary and secondary schooling declining between 1983 and 1993 while the returns to graduate education rose during both subperiods. For casual workers, on the other hand, there is no significant change in the returns to primary education over time, although the returns to secondary and graduate education decline between 1983 and 1993. A Wald test of shifts in the education coefficients from the wage equations between 1983 and either of the later two years (although not for the 1990s) is decisively rejected for regular workers. For casual workers, on the other hand, the null hypothesis of no change during any pair of years cannot be rejected.