Quantile regression for longitudinal data: unobserved heterogeneity and informative missingness

Abstract: Linear quantile regression models aim at providing a detailed and robust picture of the (conditional) response distribution as function of a set of observed covariates. Longitudinal data represent an interesting field of application of such models; due to their peculiar features, they represent a substantial challenge, in that the standard, cross-sectional, model representation needs to be extended for dealing with such kind of data. In fact, repeated observations from the same statistical unit poses a problem of dependence; in a conditional perspective, this dependence could be ascribed to sources of unobserved, individual-specific, heterogeneity. Along these lines, quantile regression models have recently been extended to the analysis of longitudinal, continuous, responses, by modelling dependence via time-constant or time-varying random effects. In this manuscript, we introduce a general quantile regression model for longitudinal, continuous, responses where time-varying and time-constant random parameters are jointly taken into account. A further feature of longitudinal designs is the presence of partially incomplete sequences, due to some individuals leaving the study before its designed end. The missing data process may produce a selection of units which can be informative with respect to the parameters of the longitudinal data model. To deal with the case of irretrievable drop-out, we introduce a pattern mixture version of the linear quantile hidden Markov model, where we account for time-varying heterogeneity and for changes in the fixed effect vector due to differential propensities to stay in the study. The proposed models are illustrated using a well known benchmark dataset on longitudinal dynamics of CD4 cells and by means of a large scale simulation study, entailing different quantiles and both complete and partially complete (ie subject to drop-out) individual sequences.