Article
Estimation of survival probabilities in the presence of ties
Search Medline for
Authors
Published: | September 6, 2007 |
---|
Outline
Text
In the real world we usually face the problem of discrete survival times, typically associated with the presence of ties between events and censored observations. However, the conventional Kaplan-Meier approaches, as well as Greenwood's variance estimator, do not adequately consider this fact, which leads to underestimation of true survival probabilities and variance estimates. We therefore present a modified Kaplan-Meier approach, by explicitly considering the presence of ties called the Tie survival function. A variance estimator based on tie survival function approach is developed. In absence of ties the new variance estimator equals to Greenwood variance estimator, while in uncensored data, it reduces to the binomial variance estimator. A simulation study was conducted in order to compare the performance of the conventional Kaplan-Meier estimator and the tie survivor estimator for different censoring rates. Our simulation results suggest a significant improvement, in terms of bias of the tie survivor approach compared to the conventional Kaplan-Meier estimator. Similarly, the results of variance simulation favor the proposed, tie variance estimator. Our new approaches are illustrated on a leukaemia data set.