Finite-Sample Diagnostics for Random-Effects Misspecification in Poisson Generalized Linear Mixed Models

Poisson mixed-effects models are essential for analyzing repeated count data, relying on latent random effects to account for unobserved heterogeneity and longitudinal dependence. However, the validity of likelihood-based inference in these models is highly sensitive to the specification of both the fixed-effects structure and the distributional assumptions of the random effects. While diagnostics based on the information matrix equality (IME) provide a theoretical framework for detecting misspecification, their high dimensionality and reliance on second-order derivatives often result in numerical instability and poor finite-sample performance in nonlinear settings. Here we introduce the Contrast of Information by Volume (CIV) test, a low-dimensional information-based diagnostic test for Poisson generalized linear mixed models (GLMMs). By integrating the scalar CIV statistics with novel graphical diagnostics, our approach facilitates the interpretation of specification errors in the random-effects structure. We derive the asymptotic behaviour of the CIV statistics under local misspecification and evaluate their properties through Monte Carlo simulations. To ensure robust inference in moderate samples, a parametric bootstrap procedure is employed for size calibration. Simulation results demonstrate that the CIV diagnostics maintain accurate Type I error control and achieve competitive power against common misspecification, including heteroskedasticity, correlation, and heavy-tailed random-effect distributions. Compared to traditional IME diagnostics, estimator-comparison tests, and GMM-based procedures, the CIV approach offers a superior balance between finite-sample stability and detection power. Finally, an empirical application illustrates the utility of the CIV framework in diagnosing latent misspecification and guiding the selection of random-effects covariance structures in applied research.