Estimation of the Truncated Bivariate Normal Stochastic Frontier Model by EM Algorithm

The truncated bivariate normal stochastic frontier model (TBN-SFM), which is globally unidentifiable on its natural parameter space, has been recently shown to be locally near-identifiable even under the identifying parametric restrictions. Maximum Likelihood estimation of such a model is problematic as its likelihood function is 'ill-shaped' around the near-identifiable parameter. In this paper we show that the EM algorithm provides a viable method for estimating the parameters of a locally near-identifiable statistical model that can be recast in the "missing data" framework. We next develop the EM algorithm for the TBN-SFM and apply it to estimate a truncated bivariate normal cost frontier using live data. We also carry out Monte Carlo simulation experiments to study the small sample performance of the EM estimates of locally near-identifiable TBN-SFM