PII S0360-3016(98)00146-1 ● Clinical Investigation ANALYSIS OF CAUSE-SPECIFIC FAILURE ENDPOINTS USING SIMPLE PROPORTIONS: AN EXAMPLE FROM A RANDOMIZED CONTROLLED CLINICAL TRIAL IN EARLY BREAST CANCER TONY PANZARELLA, M.SC.* AND J. WILLIAM MEAKIN, M.D. † *Biostatistics Department, Princess Margaret Hospital, Toronto, Ontario M5G 2M9 Canada and † Cancer Care Ontario, Toronto, Ontario M5G 2L7 Canada Purpose: To describe a statistically valid method for analyzing cause-specific failure data based on simple proportions, that is easy to understand and apply, and outline under what conditions its implementation is well-suited. Methods and Materials: In the comparison of treatment groups, time to first failure (in any site) was analyzed first, followed by an analysis of the pattern of first failure, preferably at the latest complete follow-up time common to each group. Results: A retrospective analysis of time to contralateral breast cancer in 777 early breast cancer patients was undertaken. Patients previously treated by mastectomy plus radiation therapy to the chest wall and regional nodal areas were randomized to receive further radiation and prednisone (RP), radiation alone (R), or no further treatment (NT). Those randomized to RP had a statistically significantly delayed time to first failure compared to the group randomized to NT (p  0.0008). Patients randomized to R also experienced a delayed time to first failure compared to NT, but the difference was not statistically significant (p  0.14). At 14 years from the date of surgery (the latest common complete follow-up time) the distribution of first failures was statistically significantly different between RP and NT (p  0.005), but not between R and NT (p  0.09). The contralateral breast cancer first failure rate at 14 years from surgery was 7.2% for NT, 4.6% for R, and 3.7% for RP. The corresponding Kaplan–Meier estimates were 13.2%, 8.2%, and 5.4%, respectively. Conclusion: Analyzing cause-specific failure data using methods developed for survival endpoints is problematic. We encourage the use of the two-step analysis strategy described when, as in the example presented, competing causes of failure are not likely to be statistically independent, and when a treatment comparison at a single time-point is clinically relevant and feasible; that is, all patients have complete follow-up to this point. © 1998 Elsevier Science Inc. Cause-specific failure endpoints, Kaplan–Meier estimate, Gelman approach, Cumulative incidence, Contralat- eral breast cancer. INTRODUCTION Actuarial methods such as the Kaplan–Meier (K–M) esti- mate (1) and the log rank test (2) were developed to analyze censored survival data. A patient’s event time is said to be ‘‘censored’’ if, at the time of analysis the study event for that patient has not yet occurred. Examples of survival endpoints include overall survival, where death from any cause is considered an event, and disease-free survival, where events include disease occurrence or death without disease. An underlying assumption of these methods is that the cause of censoring is independent of the impending event. This would be true, for instance, if the censoring was due to the planned termination of follow-up. However, it is not uncommon to also find these methods applied to cause- specific failure endpoints, such as time to local recurrence and cause-specific survival. Whereas survival endpoints are characterized by the fact that every patient will eventually experience the study outcome (if the follow-up time is long enough), this is not the case with cause-specific failure data which, instead, are characterized by various risks competing for the same patient. Thus, the cause of censoring may not be independent of the event of interest, which could bias the analysis. Recently, several authors have described this problem (3–5) and proposed analysis strategies for cause-specific failure endpoints. We describe the application of one of these methods to data from a randomized controlled clinical trial (RCT) in early breast cancer, given the cause-specific Presented in part at the Sixteenth Annual Meeting of the Society For Clinical Trials, Seattle, Washington, May 1995. Reprint requests to: Tony Panzarella, Biostatistics Department, Princess Margaret Hospital, 610 University Avenue, Toronto, On- tario, M5G 2M9 Canada Acknowledgments—We thank Thomas Pajak and Richard Caplan for their encouragement and review of an early draft of the manu- script, and the reviewers for their helpful comments. Accepted for publication 27 March 1998. Int. J. Radiation Oncology Biol. Phys., Vol. 41, No. 5, pp. 1093–1097, 1998 Copyright © 1998 Elsevier Science Inc. Printed in the USA. All rights reserved 0360-3016/98 $19.00 + .00 1093