This dissertation provides frameworks to select production function estimators in both the state-contingent and the general monotonic and concave cases. It first presents a Birth-Death Markov Chain Monte Carlo (BDMCMC) Bayesian algorithm to endogenously estimate the number of previously unobserved states of nature for a state-contingent frontier. Secondly, it contains a Reversible Jump Markov Chain Monte Carlo (RJMCMC) algorithm to determine a parsimonious piecewise linear description of a multiplicative monotonic and concave production frontier. The RJMCMC based algorithm is the first computationally efficient one-stage estimator of production frontiers with potentially heteroscedastic inefficiency distribution and environmental variables. Thirdly, it provides general framework, based on machine learning concepts, repeated learning-testing and parametric bootstrapping techniques, to select the best monotonic and concave functional estimator for a production function from a pool of functional estimators. This framework is the first to test potentially nonlinear production function estimators on actual datasets, rather than extrapolation of Monte Carlo simulation results.
