A comparative simulation study for estimating accelerated failure time models

Abstract

Semiparametric accelerated failure time (AFT) models directly relate the predicted failure times to covariates and are a useful alternative to Cox's proportional hazards models that work on the hazard function or the survival function. In this paper, we briey review di_erent approaches to estimate the AFT model and evaluate their performance with _nite samples via extensive simulation studies. Speci_cally, we compared (i) inverse probability of censoring weighted (IPCW) least squares, (ii) log-rank estimator, (iii) Gehan-type log-rank estimator, (iv) Buckley-James estimator, and (v) nonparametric maximum likelihood estimator (NPMLE). Overall, rank-based estimators and Buckley-James estimator are e_cient and relatively more robust to distributions of residual and censoring variables, whereas the IPCW estimator is very sensitive to distribution and amount of censoring. The NPMLE is asymptotically e_cient and useful as it allows for hazard-based formulation, and thus can be used to analyze more structured survival data.