Cautionary tales of survival analysis: conflicting analyses from a clinical trial in breast cancer

Data from a completed randomized trial in breast cancer are used to demonstrate and quantify the variation in estimated survival curves and log-rank statistics at different times throughout a trial. False 'plateaux' are common, as are wide fluctuations in chi(2) values obtained from the log-rank test when there are few events. We show how analyses conducted at different times can demonstrate different effects. Long follow-up is often necessary to allow correct interpretation of results. We discuss the assumption of proportional hazards and the consequences of making that assumption inappropriately. We show how checking whether hazards are proportional can help in avoiding erroneous conclusions.