However, Cawley et al. (1996, 1999) have not verified that the strict exogeneity condition
is satisfied in their random effects regression model.15 Random effects estimates suit the purpose
of their analysis better, which examines the estimated coefficient of intelligence on earnings.
Cognitive ability, as measured by g, is time-invariant in the data set and drops out in the fixed
effects model; but the random effects model allows for the possibility of obtaining an estimate
for the coefficient of intelligence. However, as indicated by Wooldridge (2005), if the strict
exogeneity assumption is violated, the random effects model does not yield consistent estimates
and the fixed effects model must be used instead.
Regressions 2A to 2D are designed to test for the strict exogeneity condition. We run
separate regressions under both the random and fixed effects specifications. 2A and 2C are the
random effects regressions; 2B and 2D are the fixed effects ones.16 Instead of separating the
sample into six sub-samples, we include dummy variables for race and gender, as well as the
cross-terms of g with the race/gender dummy variables in the panel regressions. Regressions 2A
and 2B use annual observations from 1979 to 1994, while 2C and 2D employ biennial data from
1994 to 2002.17 We have also included several current family and social background variables,
namely, marital status for each year, number of children for each year, residence type in each
year, and local unemployment rate for each year, in the set of control variables.
Regression 2A, which is a random effects model using annual data from 1979 to 1994, is
the closest replication of the econometric analysis of the Cawley et al. (1996) -type specification
(though we do not run fully-interacted regressions and these are thus not identical specifications
15 Two estimation methods are used widely in panel data analysis, namely, the fixed effects model and the random effects model.
The fixed effects model allows correlation between the unobservable characteristic and the independent variables. The random
effects model requires satisfaction of the strict exogeneity assumption that the unobservable characteristic is not correlated with
any independent variable to yield consistent estimates.
16 We have also run regressions with Eicker-White standard errors for both the fixed and random effects specifications; the
results are the same as those in regressions 2A to 2D.
17 This separation of survey years results from the fact that the NLSY79 data is yearly until 1994 and biennially onwards.
11