Analysis of multivariate failure-time data from HIV clinical trials.

We illustrate the use of marginal methods for the analysis of multivariate failure-time data using a large trial in HIV infection in which the composite endpoint of AIDS or death incorporates more than 20 events with varying severity. Multivariate failure-time methods are required to investigate whether treatment delays development of new AIDS events. AIDS events can be grouped and treatment effects estimated using only the first event to occur in each group for each individual. Alternatively, all events can be included by fitting a separate baseline hazard for development of each event, and restricting treatment effects to be common within groups of events. In either case, model-based or minimum-variance estimates of the overall effect of treatment can be constructed. The covariance matrix for the treatment-effect estimates can be used in multiple testing procedures. Results from the Delta trial suggest that combination antiretroviral therapy with AZT plus either ddI or ddC may delay progression to more severe AIDS events compared to AZT monotherapy. These late events are generally untreatable and prophylaxis is not available. Trials are not generally powered to detect treatment effects on individual events making up a composite endpoint, and therefore all analyses are exploratory rather than providing definitive evidence. However, marginal multivariate models provide an easily available approach for modeling the effect of covariates on multiple disease processes, and allow the likely effects of treatment to be presented in a manner which is easily understood. They can be used in a variety of ways to explore different patterns of treatment effects and are also useful for testing multiple hypotheses regarding treatment effects on several different composite endpoints.