Stochastic production function estimation: small sample properties of ML versus FGLS
Just-Pope production functions have been traditionally estimated by feasible generalized least squares (FGLS). This paper investigates the small-sample properties of FGLS and maximum likelihood (ML) estimators in heteroscedastic error models. Monte Carlo experiment results show that in small samples, even when the error distribution departs significantly from normality, the ML estimator is more efficient and suffers from less bias than FGLS. Importantly, FGLS was found to seriously understate the risk effects of inputs and provide biased marginal product estimates. These results are explained by showing that the FGLS criteria being optimized at the multiple stages are not logically consistent.