Regression 3B is the same specification as 3A except using the 1994 to 2002 data. The
change in the coefficient of g is negligible.22 In contrast, the magnitude of the coefficients of the
human capital variables, including education, job tenure, and work experience, all decrease
significantly. Most changes in education level after 1994 are due to the respondents returning to
school for graduate degrees that yield a lower marginal pecuniary return than college degrees.
Therefore, the marginal return to education drops in the 1994 to 2002 period. A similar argument
can be made concerning job tenure and work experience because both variables have
diminishing marginal returns with respect to income. Job tenure and work experience are both
positively correlated with the wage rate, but the rate of return actually decreases as tenure and
experience accumulate. A graph of wage rate versus job tenure or work experience should be
concave, with the slope of the curve, i.e., the marginal wage return, positive but decreasing as
tenure/experience increases. After 1994, all the respondents of the NLSY79 are around 40 to 50
years old, and in the later stage of their careers. Thus, they are on the flatter part of the wage-
tenure/wage-experience curve, and the marginal return to tenure/experience is expected to be
lower than in the 1979 to 1994 time period when they just started working.
The Hausman specification test is employed to compare regressions 3A and 3B to
regressions 2B and 2D in the previous section, which are the fixed effects panel regressions on
exactly the same set of control variables, in order to verify the legitimacy of the application of
the Hausman-Taylor model. The key assumption being tested is that the time-variant control
variables, including marital status, number of children, residential region, job tenure, and labor
market experience, are not correlated with the unobservable individual characteristics so that
they can be used as valid instrumental variables in the model. For both time periods, the null
22 A Chi square test shows that we cannot reject the null hypothesis that the coefficients of g in the two time period are the same:
Chi square = 0.00, p>Chi square = 0.986.
18