B1540
Title: Estimation for the partially linear ZIP regression model: A robust proposal
Authors: Anne Francoise Yao - Universite Clermont Auvergne/LMBP (France)
Andrea Bergesio - Universidad Nacional del Litoral - Departamento de Matematica Facultad de Ingenieria Quimica (Argentina)
Maria Jose Llop - Universidad Nacional del Litoral (Argentina) [presenting]
Abstract: In different areas of knowledge, distributions such as Poisson or negative binomial are used to model count data. However, when the data has an excess of zeros, these models may not be suitable. The zero-inflated Poisson regression (ZIP) model uses the binomial distribution to discern whether an observation comes from the zero structural process or the Poisson distribution. The partially linear ZIP regression model is estimated, which includes a non-parametric component that flexibilizes the parametric nature of the model. The EM algorithm is used, including auxiliary variables as if they were observable and specifying the likelihood function as the sum of components that can be optimized separately. Then, a three-step procedure is used that estimates linear and non-parametric components sequentially. The main drawback of the likelihood-based estimators is that the estimation can be affected when the assumed model is not completely valid. Therefore, outliers, both in the response and in the covariates, can considerably affect the estimators. In order to obtain robust estimators, robust loss functions are used and weights are included that control for the effect of covariates on the resulting estimator. The behaviour and performance of the estimators are compared in different contamination scenarios through simulation studies.