Efficiency of positive event dependence models for self-controlled case series designs

Abstract

The self-controlled case series (SCCS) design is an outcome-dependent sampling design developed to investigate potential associations between time-varying exposures and adverse recurrent events in clinical settings. It is a case-only design, in which only cases, who are individuals experienced the event-of-interest at least one time, are included. The cases serve as their own controls. As a result, the SCCS design implicitly controls for all time-fixed confounders. The standard SCCS method is based on a conditional Poisson process model, which cannot be used to model the effects of past event occurrences. A new method, called positive event dependence self-controlled case series (PD-SCCS), has been proposed to deal with this issue. This method adjusts the baseline intensity function of an event process with the number of previous events, and maintains all features of the SCCS method. In this study, we consider rare recurrent events settings, where the events are generated from mixed nonhomogeneous pure birth models with immigration. We investigate the relative efficiency of the PD-SCCS design compared to other SCCS and cohort designs in the estimation of the relative incidence parameter, as well as impacts of some model misspecifications and violations of model assumptions required for the PD-SCCS design through extensive Monte Carlo simulation studies. We illustrate the methods by analyzing a dataset from a clinical vaccine study, as well as two synthetic datasets based on a postmarketing drug safety surveillance study.