Rank Estimation of Accelerated Lifetime Models With Dependent Censoring

Under independent censoring, estimation of the covariate effects in the accelerated lifetime model may be based on censored data rank tests. Similar rank methodology has been developed with bivariate accelerated lifetime models for dependent censoring but uses artificial censoring, which may lead to substantial information loss. We present a new artificial censoring technique using pairwise ranking and establish the asymptotic properties of a pairwise rank estimator. Simulations show that the pairwise approach achieves large reductions in artificial censoring and large efficiency gains over the existing rank estimator. The simulations evidence moderate efficiency gains under independent censoring over a rank estimator that is semiparametric efficient under independent censoring. An AIDS data analysis illustrates the practical utility of the inferential procedures.