A1522
Title: Likelihood specification testing in nonlinear panel data models with fixed effects
Authors: Martin Schumann - Maastricht University (Netherlands) [presenting]
Abstract: Consistent maximum likelihood estimation typically depends on the correct specification of the likelihood. If the presence of individual heterogeneity is not accounted for, estimation and inference will frequently yield misleading results. We develop an information-matrix test for likelihood-based specification testing in nonlinear panel-data models with unobserved fixed effects. We show that allowing for fixed effects leads to an incidental parameters problem, rendering the information matrix test inconsistent if the number of time periods does not grow faster than the number of individuals. The test's incidental parameters bias is shown to depend on the score and the information bias of the pseudo-likelihood on which the test statistic is based. We derive a bias-corrected version of the information matrix test and analyze its small-sample properties in a simulation study.
