On the estimation of total factor productivity: A novel Bayesian non-parametric approach

This paper provides an alternative general empirical method for the estimation of Total Factor Productivity (TFP). We use a decomposition which allows non-parametric estimation and at the same time addresses the issue of endogeneity of inputs. In this way, we also deal with the unavailability of input prices which is common in the TFP literature. We apply the new techniques to U.S four-digit manufacturing data using a novel Bayesian nonparametric model based on local likelihood. We use Markov Chain Monte Carlo (MCMC) techniques organized around the method of Girolami and Calderhead (2011). We compare and contrast the estimates from the proposed new method with standard parametric methods such as the translog, the Generalized Leontief and the Normalized Quadratic and we also propose novel diagnostic tests for correct specification and validity of instruments. We show that parametric methods lead to biased estimation of TFP growth. Our empirical findings reveal that the new model passes successfully a battery of robustness checks including diagnostic testing and tests for weak identification as well as weak instruments. Finally policy implications relating to the nature of TFP growth are also provided.