|
Abstract:
|
We present a method for constructing confidence bounds for the true survival curve as estimated by the Kaplan-Meier survival function, derived as an extension of the Clopper-Pearson method for Binomial proportions. As a consequence of the definition of the Kaplan-Meier survival function, the upper bound should decrease only when events occur, but the lower bound should decrease when either events or censorings occur. The confidence interval width is proportional to the current weights of the observations still present in the sample under Efron's redistribute-to-the-right algorithm. Simulations show that the proposed method gives at least nominal coverage in a variety of situations, even for small sample sizes, except in the right tail of the survival distribution with heavy censoring. Finally, the proposed method is illustrated with an application from a leukemia clinical trial.
|
