Bayesian Model Averaging for Latent Class Models in Capture-Recapture

Model selection appears to be crucial in capture-recapture problems as it is common that different models with an equally good level of adaptation to the observed data lead to rather different estimates of the undercounts. We consider log–linear Latent Class Models as our capture-recapture model and propose Bayesian model averaging to overcome the difficulties of model selection within this class. We show that, by focusing on graphical decomposable models, we can design a simple Gibbs–based MCMC to sample over the space of eligible models.