Mixture models for quantitative HIV RNA data

Clinical investigators are increasing their use of quantitative determinations of HIV viral load in their study populations. The distributions of these measures may be highly skewed, left-censored, and with an extra spike below the detection limit of the assay. We recommend use of a mixture model in this situation, with two sets of explanatory covariates. We extend this model to incorporate multiple measures across time, and to employ shared parameters as a way of increasing model efficiency and parsimony. Data from a cohort of HIV-infected men are used to illustrate these features, and simulations are performed to assess the utility of shared parameters.