Bootstrapping a Conditional Moments Test for Normality after Tobit Estimation

Abstract

Categorical and limited dependent variable models are routinely estimated via maximum likelihood. It is well-known that the ML estimates of the parameters are inconsistent if the distribution or the skedastic component is misspecified. When conditional moment tests were first developed by Newey (1985) and Tauchen (1985), they appeared to offer a wide range of easy-to-compute specification tests for categorical and limited dependent variable models estimated by maximum likelihood. However, subsequent studies found that using the asymptotic critical values produced severe size distortions. This paper presents simulation evidence that the standard conditional moment test for normality after tobit estimation has essentially no size distortion and reasonable power when the critical values are obtained via a parametric bootstrap.