Evaluation of Methods for Analyzing Multiple Time-to-failure Data

In many clinical trials the event of interest is nonfatal and can recur. The standard analysis models the time to the first occurrence of the endpoint; however, this analysis ignores recurrences which might provide additional information and enhance power. Six models for analyzing time-to-event data where events are nonfatal and can recur are evaluated using simulations and examples. The models differ in their adjustments for dependency among endpoints, in their definition of event time and in their risk sets. Three of the models make implicit assumptions about the structure of the within subject correlations and two of these actually introduce correlation between event times within subjects by modelling total time since study start instead of gap time between events (Andersen and Gill 1982; Prentice, Williams and Peterson 1981 ). The true significance levels of tests based on these models are shown to exceed nominal levels in many cases. A fourth model estimates the correlation of the observations without specifying the structure and this model preforms well (Wei, Lin and Weissfeld 1989). The size of the test remains near nominal levels in most situations; however, this method models total time since •