Effect of misspecification of random effects distribution on the performance of parameters estimation methods in binary logistic mixed models.

Abstract

We empirically compared a Bayesian estimation method (Integrated Nested Laplace Approximation, INLA) to three classical estimation methods (Penalized Quasi-Likelihood, PQL; Hierarchical Likelihood Method, HLM and Adaptive Gauss-Hermite Quadrature, AGHQ) under six random effect distributions in binary logistic mixed models. Results revealed that AGHQ and HLM had best performance for all distributions considered in the case of fixed effects. For the random effects, classical methods showed best performance for the symmetric distributions (normal, uniform and mixture-normal). AGHQ, HLM and INLA outperform PQL for normal and uniform distributions whatever the sample considered.