Upward Bias in the Estimated Returns to Education: Evidence from South Africa

摘要

Ordinary least-squares (OLS) estimates of the proportionate increase in wages due to an extra year of education in the United States (the Mincerian rate of return) are believed to be reasonably consistent. It appears that upward bias due to omitted variables is roughly offset by attenuation bias due to errors in the measurement of schooling. Orley Ashenfelter and Cecilia Rouse (1998) find a net upward bias on the order of just 10 percent of the magnitude of the OLS estimate. David Card's (2001) survey of instrumental variables-based estimates reaches a similar conclusion, as do Ashenfelter et al. (1999). This result need not obtain in all countries at all times. Some (e.g., David Lam and Robert F. Schoeni, 1993) have suggested that omitted variables bias might be larger in less developed economies, where liquidity constraints and family background are likely to be important determinants of both education and earnings. To date, however, there are few estimates of the returns to schooling in developing countries that take account of both omitted variables and measurement error; an exception is Esther Duflo (2001) who finds no net upward bias for Indonesia. In this paper I use the familiar within-family (fixed-effects) approach, as well as variants of the family-effects models described by Ashenfelter and David J. Zimmerman (1997) and Card (1999), to minimize omitted variables bias in a South African data set from 1993. I exploit the fact that about 13 percent of respondents were resurveyed (in 1998) to derive estimates of the reliability of measured schooling, which turns out to be rather low (on the order of 0.77). This suggests that the within-family fixed-effects estimates should be biased downwards to a considerable extent (Zvi Griliches, 1979). However, I also show that errors in the schooling variable are strongly correlated within the family, and that this reduces the degree of attenuation bias in the fixed-effects model. After correcting for these correlated, nonclassical measurement errors, I arrive at schooling coefficients for Africans that are less than half as large as the ordinary least-squares results. My preferred specification yields results on the order of 5 to 6 percent, whereas the initial OLS figures are 11 to 13 percent.