Stochastic frontier models with time-varying conditional variances

In this paper we introduce stochastic frontier models in which either the inefficiency or the noise component or both the components follow Generalized AutoRegressive Conditional Heteroskedasticity GARCH(1,1) process. Bayesian estimation of the technology parameters are proposed using a half-normal (exponential) distribution for the inefficiency component, and a normal distribution for the noise component. We show, in simulations, that predictions of inefficiency ignoring the GARCH(1,1) process are not aligned with their true values. We use real panel data on electricity distribution and show how to estimate our proposed model, and predict inefficiency. Moreover, we examine the effect of ignoring GARCH specification on economic measures like input elasticities, technical change and returns to scale. We also provide test for GARCH vs no GARCH models using Bayes factors. Finally, we examine variants of the GARCH family such as the ARCH and EGARCH models and we compared GARCH models of different orders.