Improving Predictions of Technical Inefficiency

The traditional predictor of technical inefficiency proposed by Jondrow et al. (1982) is a conditional expectation. We study whether, and by how much, the predictor can be improved by using auxiliary information in the conditioning set. To do so, we use simulations to study two types of stochastic frontier models. The first type is a panel data model where composed errors from past and future time periods contain information about contemporaneous technical inefficiency. The second type is when the stochastic frontier model is augmented by input ratio equations in which allocative inefficiency is correlated with technical inefficiency. We consider a standard kernel-smoothing estimator and a newer estimator based on local linear random forest which helps mitigate the curse of dimensionality when the conditioning set is large. We also provide an illustrative empirical example