Robust inference for generalized linear models

By starting from a natural class of robust estimators for generalized linea r models based on the notion of qua-si-likelihood, we define robust devianc es that can be used for stepwise model selection as in the classical framew ork. Wc derive the asymptotic distribution of tests based on robust devianc es, and we investigate the stability of their asymptotic level under contam ination. The binomial and Poisson models are treated in detail. Two applica tions to real data and a sensitivity analysis show that the inference obtai ned by means of the new techniques is more reliable than that obtained by c lassical estimation and testing procedures.