stronger exogeneity assumptions, generate more efficient estimates than the HT estimator.20
Overall, the advantages of the Hausman-Taylor model in analyzing panel data are recognized by
many economists, who have applied the technique to a wide range of empirical topics,
particularly in human capital theory for estimating wage equations.21
We use the Hausman-Taylor model in our analysis of the wage equation. Even though the
AM and BMS estimators are demonstrated to be potentially more efficient than the HT estimator,
both estimators require stronger and more specific assumptions. Without appropriate statistical
tests to verify these assumptions, we cannot justify the applications of the AM and BMS
estimators to our data set. In contrast, the non-correlation assumption required by the HT
estimator can be conveniently tested by the Hausman specification test. Thus, the HT estimator is
preferred.
In the context of our panel data set, ai is the unobserved individual characteristic
(education quality), X2it is education level, and Z2i includes all endogenous time-invariant
explanatory variables, namely, g and all the cross-terms of g with race/gender dummy variables.
In addition, X1it includes other time-variant variables such as marital status and job tenure and Z1i
includes all the race/gender dummy variables. We use the Hausman specification test comparing
the Hausman-Taylor model with the fixed effects model to verify the assumption that all the
components of X1it and Z1i are independent of education quality.
20 The Amemiya and MaCurdy (AM) estimator requires that X1 and ai are uncorrelated every point in time. Breusch, Mison and
Schmidt (BMS) demonstrate that if one can assume that the time-varying variables in X2 are correlated with ai only through ai’s
time-invariant components, their estimator is consistent and more efficient than both the HT estimator and the AM estimator.
Cornwell and Rupert (1988) and Baltagi and Akom (1990) have conducted empirical studies and confirm that, given correct
corresponding exogeneity conditions, the AM and BMS estimators are potentially more asymptotically efficient than the HT
estimator.
21 See Contoyannis and Rice (2000), Heineck (2005), Gardebroek and Lansink (2003), and Peridy (2005).
16