Bootstrapping the Kaplan-Meier Estimator

From MaRDI portal





Randomly censored data consist of iid pairs of observations \((X_ i,\delta_ i)\), \(i=1,...,n\). The observed random variable is \(X_ i=\min \{X_ i^ 0,Y_ i\}\), and \(\delta_ i=I[X_ i=X^ 0_ i]\). The random censoring variables \(Y_ 1,...,Y_ n\) are iid from a distribution G, and the random samples \(X_ 1,...,X_ n\) are iid from a distribution F. The Kaplan-Meier product limit estimator \(F_ n\) for F has been studied extensively and has been shown to enjoy a number of optimality properties. Two methods of bootstrapping are discussed in this paper. The first method takes a random sample with replacement from \((X_ 1,\delta_ 1),...,(X_ n,\delta_ n)\), whereas the second method takes a random sample from the Kaplan-Meier estimator \(F_ n\). With the first method, it is shown that confidence bands for the survival distribution can be obtained which are asymptotically correct even when discrete components are present and when the data are grouped. With the second method, the bootstrap limit distribution is shown to be different from the limit distribution of \(F_ n\). Hence, the second method can not produce asymptotically correct confidence bands for the survival curve. Simulation studies in the paper compare the bootstrap bands for the discrete and continuous data.




Cited in
(53)








This page was built for publication: Bootstrapping the Kaplan-Meier Estimator

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3774719)