sfaR-package: sfaR: An R package for estimating stochastic frontier models
In sfaR: Stochastic Frontier Analysis using R

sfaR-package

R Documentation

sfaR: An R package for estimating stochastic frontier models

Description

The sfaR package provides a set of tools (maximum likelihood - ML and maximum simulated likelihood - MSL) for various specifications of stochastic frontier analysis (SFA).

Two categories of important functions are available: sfacross and lcmcross, which estimate different types of frontiers and offer nine alternative optimization algorithms (i.e. "bfgs", "bhhh", "nr", "nm", "ucminf", "mla", "sr1", "sparse" and "nlminb").

lcmcross

lcmcross estimates latent class stochastic frontier models (LCM), which accounts for technological heterogeneity by splitting the observations into a maximum number of five classes. The classification operates based on a logit functional form that can be specified using some covariates (namely, the separating variables allowing the separation of observations in several classes). Only the half normal distribution is available for the one-sided error term. Heteroscedasticity in both error terms is possible. The choice of the number of classes can be guided by several information criteria (i.e. AIC, BIC or HQIC).

sfacross

sfacross estimates the frontier for cross-sectional data and allows for ten different distributions for the one-sided error term. These distributions include the exponential, the Gamma, the generalized exponential, the half normal, the lognormal, the truncated normal, the truncated skewed Laplace, the Rayleigh, the uniform and the Weibull distributions. In the case of the Gamma, lognormal and Weibull distributions, maximum simulated likelihood (MSL) is used with the possibility of four specific distributions to construct the draws: Halton, Generalized Halton, Sobol and uniform. Heteroscedasticity in both error terms can be implemented, in addition to heterogeneity in the truncated mean parameter in the case of the truncated normal and lognormal distributions. In addition, in the case of the truncated normal distribution, the scaling property can be estimated.