Efficiency comparison of random effects two stage least squares estimators

For two stage least squares estimation (2SLS) with panel data, Baltagi and Li (1992) have shown that Baltagi's (1981) error-component 2SLS and Balestra and Varadharajan-Krishnakumar's (1987) generalized 2SLS are equivalent in terms of asymptotic variance under the random effects error-component assumption. In the present paper this asymptotic equivalence is extended to models with heteroskedastic and serial correlated errors. However, it is shown that the equivalence does not hold if fixed effects are present in the reduced-form equations. The theoretical claims are verified by simulations and are examined for Cornwell and Trumbull's (1994) model and data.