The Statistical Analysis of Multivariate Failure Time Data: A Marginal Modeling Approach.

Abstract

In large studies of public health interventions, individuals are typically at risk of many different types of events. Despite the multitude of events that may arise, a single failure time is often chosen and the primary study objective is then to assess the effect of an intervention on reducing the risk of this event. The effects of the intervention on the risk of other events are then explored in secondary analyses. When selection of a primary event (endpoint) is difficult, composite endpoints are often adopted, but these lead to estimands that are generally difficult to interpret. Methods that facilitate inferences about intervention effects on two or more events are needed to provide a rigorous basis for investigators and health policy scientists to weigh the evidence and anticipate the impact of any decision. Robustness of these inferences is naturally critically important.