Title: A robust alternative to the Poisson hurdle model
Authors: Conceicao Amado - Universidade de Lisboa (Portugal) [presenting]
Manuela Souto de Miranda - University of Aveiro (Portugal)
Abstract: Hurdle Poisson models are mixed models that can deal with an excess of zeroes by considering two separate components, namely, a binary process and a truncated discrete distribution. They are particularly adequate for modelling counting processes when the occurrence of zero observations does not depend on the main generating process of strictly positive counting. It is often the case when counting low-probability incidents, as they appear in some health econometric statistics, or extreme-value events per unit of time. Maximum likelihood is used to fit the Poisson hurdle model under exact and strict stochastic assumptions. However, the performance of these estimators degrades when these assumptions are not verified. The aim is to compare robust estimators, particularly considering minimum distance estimators for the parameters of the hurdle model when the positive component is modelled by a truncated Poisson. We investigate the performance of the estimators through a simulation study.