out that the non-correlation assumption can be verified with the Hausman specification test that
compares the Hausman-Taylor model with the fixed effects model.
Another necessary condition for consistent estimation in the Hausman-Taylor model is
that the columns of X1it provide sufficient instruments for the columns of Z2i, there must be at
least as many exogenous time-varying variables as there are endogenous time-invariant variables.
If there are more variables in Z2i than in X1it, the model is under-identified and it does not
generate consistent estimates of β and γ.
Hausman and Taylor (1981) apply their model to a wage equation to estimate the returns
to schooling, treating cognitive ability as an unobservable individual characteristic. They
compare OLS, random effects, fixed effects, and Hausman-Taylor models, recognizing that the
potential correlation between individual specific latent ability and schooling, which is time-
invariant in their data set, causes the OLS and the random effects model to generate biased
coefficients. Then, the authors employ the Hausman specification test to verify that in the
Hausman-Taylor model, X1 and Z1 are not correlated with ai. Therefore, they conclude that the
Hausman-Taylor model is the most appropriate among these models for estimating the wage
equation and returns to schooling because it provides consistent estimated coefficients for time-
invariant independent variables.
Various economists have responded favorably to the Hausman and Taylor panel data
estimation technique. Baltagi, Bresson and Pirotte (2002) verify that the Hausman-Taylor (HT)
estimator is consistent under the non-correlation assumption and should be preferred over the
random effects model. Based on the HT estimator, Amemiya and MaCurdy (1986), and Breusch,
Mison and Schmidt (1989), have respectively proposed new estimation techniques that, given
15