SFA robustness to violated distributional assumptions: theory, simulations and empirical evidence

The well-known biases caused by taking the logarithm of a heteroskedastic dependent variable are a potential threat in stochastic frontier analyses (SFA), which are often estimated in log-linear form. This paper shows that while these biases can indeed be substantial, they do not affect the identification of best and worst performing firms, which is often more important than parameters of the frontier function. The efficiency ranks remain correctly identified except for special cases which are easily detectable. Furthermore, we document that the popular normal-truncated-normal SFA model with time-invariant efficiency shows a notable robustness to the mis-specification of the distribution of efficiency terms. The reason for this property is the model’s similarity to Gaussian pseudo-maximum likelihood estimator. Simulations are used to illustrate these properties and to show that the recovery of the correct productivity ranks is feasible even with very short panels. Empirical exercises show that using short panels is practicable in real datasets.