Year: 2018

Working Paper Number: 420

Unit: SSU

Author: Beatrice Conradie


This paper compares the total factor productivity (TFP) scores generated from two panel datasets with two common stochastic frontier methods. Likelihood ratio tests determine model specification as usual. The first dataset (sheep) reveals that while there is sometimes little practical difference between the estimates obtained with the two methods, they can vary systematically and appear to be biased upwards by the more flexible error components method. There is even stronger evidence from the wine panel that this is in fact the case, and differences increase as time passes. It also seems as if the less flexible error components model emphasises fixed factors (land, labour) over variable inputs (feed, chemicals, fuel). The two methods produced similar conclusions about the error structure. Model selection can therefore be a matter of convenience, but comparisons between studies should be handled with care.