Abstract
This work focuses on Markov and semi-Markov switching regression models, i.e. finite mixtures of regression models with (semi-)Markovian dependencies. These statistical models enable to analyse data structured as a succession of stationary phases that are asynchronous between individuals, influenced by time-varying covariates and which present inter-individual heterogeneity. The proposed inference algorithm for (semi-)Markov switching generalized linear models is a gradient EM algorithm. For (semi-)Markov switching linear mixed models, we propose MCEM-like algorithms whose E-step decomposes into two conditional restoration steps: one for the random effects given the state sequences (and the observed data) and one for the state sequences given the random effects (and the observed data). Various conditional restoration steps are presented. We study two types of random effects: individual-wise random effects and environmental random effects. The relevance of these models is
illustrated by the analysis of forest tree growth influenced by climatic covariates. These models allow us to identify and characterize the three main growth components (ontogenetic component, environmental component and individual component). We show that the weight of each component varies according to species
and silvicultural interventions.