A Cautionary Note on the Assessment of Goodness-of-Fit in Models Estimated by PPML

Abstract

We argue that, in general, measures of goodness-of-fit based on the value of the likelihood function should not be used when estimation is performed by pseudo maximum likelihood. We illustrate this by showing that, when the dependent variable are not counts, some well-known measures of goodness-of-fit for Poisson regression depend on the scale of the data, and therefore are uninformative.

Keywords: Information criteria, Poisson regression, Pseudo R-squared