Rate of Return on Education and Sample Selection Bias

Abstract

The Mincerian method is extensively used to estimate rates of return on education investment, using survey data of household earnings and OLS. However, the estimated coefficients suffer from sample selection bias [Heckman, Econometrica, 1979]. The question is: Is there any difference in assessing returns on investment in education by paying attention to the sample selection bias or by ignoring it? See, for example, Psacharopoulos and Patrinos [Education Economics, 2004], Glewwe [JEL, 2002], Chase [Industrial and Labor Relations Review, 1998] and Lassibille [Economics of Education Review, 1998]. In this study, we use Household Expenditure Income Survey of Turkey (26,734 males and 7218 females) to assess sample selection bias in the estimates of investment return on education. Our data carry some bias because we only observe data on market wages when an individual chooses to enter the workforce. Dividing the workforce into wage earners and nonworkers is certainly not random; as a result the selected sample is biased. We compare two sets of estimated rates of return to education in Turkey from the Mincerian augmented earnings function for each gender: one with Fsample selection correction_ using Heckman’s two-step method and one Fwithout correction._ The independent variables are various levels of education, ranging from no education to college graduates, experience, its square, and other variables. The estimated coefficients of education in all four models are statistically significant at 1 percent. The coefficient of the inverse Mills ratio in the corrected models is also statistically significant at 1 percent, lending support for the sample selection bias. However, the differences between the estimates of the coefficients with and without corrections are not dramatic. For females, the coefficients for the seven levels of education without correction are: j0.065, 0.360, 0.682, 0.879, 1.27, 1.35, and 1.62, and with correction: j0.061, 0.404, 0.694, 0.844, 1.12, 1.38, and 1.72 respectively. For males, the coefficients for the seven education levels without correction are: 0.25, 0.314, 0.531, 0.541, 0.888, 1.06, and 1.29, and with correction: 0.28, 0.347, 0.516, 0.0.537, 0.872, 1.08, and 1.25, respectively. These coefficients for Turkey are in line with the results reported for other similar countries. The females’ rates of returns are consistently higher than those of men. The extent of the selection bias for females (4.23 percent) is higher than that of males (3.26 percent). The direction of the bias varies for some levels of education. Having a significant inverse Mills ratio with small biases may seem a paradoxical result that can be explained as follows. (1.) Since the inverse Mills ratio in an OLS relation is omitted, then the schooling coefficient bias depends on (a) the omitted variable being an important determinant of Atlantic Economic Journal (2005)33:359–360 * IAES 2005 DOI: 10.1007/s11293-005-8190-8