Fixed-Effects Partial Likelihood for Repeated Events

Abstract

Methods of event history analysis are frequently applied to data in which individuals are observed for a fixed interval of time, with repeated events during that interval. It is common to analyze such data as though each spell represented a distinct observation, ignoring the fact that some individuals contribute multiple spells. This approach is likely to yield downwardly biased standard error estimates and possibly biased estimates of the parameters themselves. One alternative is fixed-effects partial likelihood, which is accomplished by applying Cox's partial likelihood method with each individual treated as a separate stratum. Fixed-effects partial likelihood controls for all constant characteristics of the individual (whether measured or unmeasured), while estimating effects of those variables that vary across spells or over time within spells. Using simulated data, the author shows that the method has good properties under a wide range of conditions. When censoring is severe, however, fixed-effects partial likelihood estimators may be badly biased for variables that describe the preceding event history.