Tricks with metrics: combining statistics for improved inference in regression analysis

In maximum likelihood methods, the three classical tests statistics are often unreliable for inference in small samples, even under the correct model specification. In this article, I discuss how likelihood ratio and Wald test statistics can be combined to obtain highly improved parameter inference in regression models with small samples. I consider modifications obtained from both the Barndorff-Nielsen and the Lugannani and Rice likelihood approximations, and I show how they can produce highly accurate parameter inference in a general (possibly non linear) regression model with possibly non spherical disturbances. I discuss the underlying theory and provide Monte Carlo simulations demonstrating the superior accuracy of the proposed procedures over the first-order classical likelihood methods (i.e., the signed log-likelihood ratio test and the Wald test). An empirical application to a regression model of mobile money ("M-pesa") adoption in Kenya is provided as an illustration of the usefulness of these methods in practice.