An efficient monte carlo study of two-step generalized least squares estimators for random-effects panel data models

Using efficient Monte Carlo methods, the performance of two-step Generalized Least Squares (GLS) estimators for the one-way error components models in small samples is analyzed. In our approach, we focus on the two-step GLS estimators provided by the programs LIMDEP, RATS and TSP, which mainly differ in the solution of negative variance components problem. Our main result is that the use of non negative first-step estimators, as RATS, produces a considerably efficiency loss. We greatly improve the efficiency of simulations using a control variate that can be implemented with no virtually computational cost.