Marginalized Two Part Models for Generalized Gamma Family of Distributions

Abstract: Positive continuous outcomes with a point mass at zero are prevalent in biomedical research. To model the point mass at zero and to provide marginalized covariate effect estimates, marginalized two part models (MTP) have been developed for outcomes with lognormal and log skew normal distributions. In this paper, we propose MTP models for outcomes from a generalized gamma (GG) family of distributions. In the proposed MTP-GG model, the conditional mean from a two-part model with a three-parameter GG distribution is parameterized to provide regression coefficients that have marginal interpretation. MTP-gamma and MTP-Weibull are developed as special cases of MTP-GG. We derive marginal covariate effect estimators from each model and through simulations assess their finite sample operating characteristics in terms of bias, standard errors, 95% coverage, and rate of convergence. We illustrate the models using data sets from The Medical Expenditure Survey (MEPS) and from a randomized trial of addictive disorders and we provide SAS code for implementation. The simulation results show that when the response distribution is unknown or mis-specified, which is usually the case in real data sets, the MTP-GG is preferable over other models.