Censored Regression and Lack-of-Fit Tests for Clustered Survival Data

• Activity Number: 388
• Type: Contributed
• Date/Time: Thursday, August 15, 2002 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section*
• Author(s): Song Yang*+
• Affiliation(s): Texas Tech University
• Keywords: Censored regression ; clustered data ; lack-of-fit test ; semiparametric regression

Abstract

In survival analysis, the log linear model is a useful model for assessing the covariate effect. When the data consist of clusters of correlated failure-times that are possibly censored, the marginal approach, in the spirit of generalized estimating equation, has been proposed. The regression parameter estimators proposed so far have asymptotic covariance matrices that are difficult to estimate, and simple lack-of-fit tests are not available. In this paper, we propose a class of estimators that are asymptotically normal, whose covariance matrices are easily estimated. We also derive some lack-of-fit tests that require little extra effort after the parameter estimators are obtained. These tests are asymptotically normal under the model and consistent against certain monotone or convex model misspecifications. In particular, in the k-sample case, the lack-of-fit tests are consistent when the alternative is the Cox model, the heteroscedastic errors model, or the proportional odds model, the three most common options when the censored regression model is inappropriate. The new methods are compared with previous methods and illustrated with examples.