Parametric modeling of dependence of bivariate quantile regression residuals' signs

In this thesis, we propose a non-parametric method to study the dependence of the quantiles of a multivariate response conditional on a set of covariates. We define a statistic that measures the conditional probability of concordance of the signs of the residuals of the conditional quantiles of each univariate response. The probability of concordance is bounded from below by the value of largest possible negative dependence and from above by that of largest possible positive dependence. The value corresponding to the case of independence is contained in the interior of that interval. We recommend two distinct regression methods to model the conditional probability of concordance. The first is a logistic regression with a logit link modified. The second one is a nonlinear regression method, where the outcome is modeled as a polynomial function of the linear predictor. Both are conceived to constrain the predicted probabilities to lie within the feasible range. The estimated probabilities can be tested against the values of largest possible dependence and independence. The method permits to capture important aspects of the dependence of multivariate responses and assess possible effects of covariates on such dependence. We use data on pulmonary disfunctions to illustrate the potential of the proposed method. We suggest also graphical tools for a correct interpretation of results.