Debashis Ghosh, University of Washington

Marginal Regression Models for Recurrent and Terminal Events

A major complication in the analysis of recurrent events data from medical studies is the presence of deaths. We propose semiparametric proportional rates and means models for the marginal mean number of recurrences, taking into account that death precludes further events. It is shown that when there is no loss to follow-up censoring, these models can be estimated using existing methods. Two new estimation procedures are proposed when censoring because of loss to follow-up exists. The first is based on inverse probability of censored weighting techniques, while the second is based on modeling survival. The asymptotic results of the estimators from these procedures are derived. Goodness of fit techniques for these models are considered. Some joint inference strategies for combining results from the recurrent events and deaths are examined. The finite-sample behavior of the proposed methodologies is assessed in simulation studies. The techniques are then applied to data from a cancer clinical trial.