Analysis of uncontrolled treatment changes in HIV clinical trials

Some clinical trials involving subjects with the human immunodeficiency virus (HIV) allow a subject to change treatment when a marker of disease progression, for example, the CD4 cell count, falls below a defined threshold. Although the change does not involve a randomized control, there is considerable interest in assessing whether it might have beneficial effects on the marker levels. However, the effects of regression to the mean will confound a simple analysis of the repeated measurements obtained before and after the cross-over. In this paper, we present a likelihood-based methodology for this problem and apply it to assess a cross-over from placebo to treatment with a drug, zidovudine, in a cohort of HIV infected individuals. This cross-over occurred when a subject first had two successive CD4 cell counts below 500/mm³. As well as providing important information about the effects of introducing zidovudine among subjects with early asymptomatic HIV infection, this paper illustrates methods for verifying model assumptions, an important component of such analyses which are necessarily model dependent.