Abstract: In many therapeutic HIV trials, patients or their providers may modify or switch their current antiretroviral therapy (ART) due to virologic failure, drug toxicity or intolerance, for example. Indeed, some drug changes are so common that they do not even constitute a true ART modification because of their similar chemical properties and are written-in as part of the clinical trial protocol. In an effort to improve ART treatment guidelines, it is important to understand what factors are associated with regimen modification. Modeling the time of regimen modification as a function of covariates is a principal question; but an equally important question is to model marks of regimen change as functions of covariates and this problem has received little attention. The problem is challenging because when time of regimen modification is censored in a clinical trial, the censoring mechanism on the time-scale induces a censoring mechanism on the mark-scale. Furthermore, the marked endpoints in HIV research can have skewed distributions and outliers, e.g. marks defined as functions of HIV-1 RNA. We propose a new regression estimator for mark-variable regression that possesses better operating characteristics than earlier proposals and whose asymptotic properties follow from a theory of U-statistics. We illustrate the method with data from ACTG 5095.
