|
Abstract:
|
In the clinical trial setting, common survival analysis regression models parameterize the treatment effect as being multiplicative on some scale, such as the hazard scale in the case of the Cox model, or the survival time scale in the case of the accelerated failure time model. Regardless of the model assumed, the integrity of the treatment effect estimate relies on the fact that the model and estimator must be pre-specified. It is therefore of interest to understand the properties of the chosen estimator when the model is misspecified. In this paper, we examine the properties of the most common semiparametric estimators, and show that they in general are consistent for a parameter that depends on the censoring distribution. We propose re-weighted estimating equations that remove this dependence, along with methods for interval estimation.
|
