Diagnóstico de influência e validação em modelos não lineares com efeitos mistos

Abstract

Nonlinear mixed-effects (NLME) models are a widely used statistical tool for modeling data with hierarchical structure and nonlinear relationships among variables. Despite their flexibility, the adequacy and robustness of NLME models depend on careful evaluations of the influence of individual observations and the predictive performance of the fitted models. This thesis develops and evaluates methods aimed at influence diagnostics and validation of NLME models. Initially, results are presented for local influence diagnostics in nonlinear mixed-effects models under the Scale Mixture of Skew-Normal (SMSN) family of distributions, where several perturbation schemes are applied to analyze model sensitivity using two pharmacokinetic datasets. Subsequently, a systematic analysis of different validation methods for normally distributed NLME models is developed, comparing error and variability measures across resampling schemes. Finally, new diagnostic approaches are proposed for NLME models under the normal distribution. The results contribute to improving predictive assessment and influence analysis in nonlinear mixed-effects models.