Uma classe de modelos de regressão bivariados para respostas discreta e contínua

Abstract

In this thesis, a wide general class of models for mixed responses is proposed in which joint distributions are constructed by the conditional approach (probability density functions, (pdf), as the product of a marginal pdf and a conditional pdf). It is assumed that the distribution of the discrete response and the conditional distribution of the continuous response given the discrete variable belong to one- or two-parameter exponential family of distributions. Furthermore, the marginal means are related to the covariates by link functions using linear and/or nonlinear and/or non-parametric predictors and a dependency structure between the responses is introduced into the model via the conditional mean. Estimation methods, diagnostic analysis and in uence techniques are presented as well as a simulation study considering the Bernoulli-exponential and Poisson-normal semiparametric models, two particular cases of the proposed class. Finally, one of the proposed models is used in a real data set involving the total cost of care for each patient during hospitalization, the use or not of the intensive treatment units and the age of the patient.