Quantile regression via the EM algorithm for joint modeling of mixed discrete and continuous data based on Gaussian copula

In this paper, we develop a joint quantile regression model for correlated mixed discrete and continuous data using Gaussian copula. Our approach entails specifying marginal quantile regression models for the responses, and combining them via a copula to form a joint model. For modeling the quantiles of continuous response an asymmetric Laplace (AL) distribution is assigned to the error terms in both continuous and discrete models. For modeling the discrete response an underlying latent variable model and the threshold concept are used. Quantile regression for discrete responses can be fitted using monotone equivariance property of quantiles. By assuming a latent variable framework to describe discrete responses, the applied proposed copula still uniquely determines the joint distribution. The likelihood function of the joint model have also a tractable form but it is not differentiable in some points of the parameter space. However, by using the stochastic representation of AL distribution, the maximum likelihood estimate of parameters are obtained using an EM algorithm and also in order to carry out inference about parameters Bootstrap confidence intervals are specified using a Monte Carlo technique. Some simulation studies are performed to illustrate the performance of the model. Finally, we illustrate applications of the proposed approach using burn injuries data.