Score Test in a Pseudo-Value Regression Setting with Correlated Survival Data

• Activity Number: 266
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Statistics Without Borders
• Author(s): Yanzhi Wang*+ and Brent Logan
• Companies: Medical College of Wisconsin and Medical College of Wisconsin
• Keywords: pseudo-value regression ; score test ; center effect ; correlated survival data

Abstract

With clustered time to event data, score tests such as Commenges and Andersen (CA,1995) are often used to detect a random cluster effect on the hazard function. However, researchers may be interested in regression models for survival, cumulative incidence, or multistate model probabilities at specific time points directly, which can be done using pseudo-value regression. We propose a score test for a cluster effect on state probabilities at a specific time point, using a generalized linear mixed model framework for pseudovalue regression. The test statistic is the sum of within cluster pairwise correlations based on the pseudo-value of the response variable, similar to Jacqmin-Gadda and Commenges (1995). The test is asymptotically normal under the null hypothesis, and a method of moments variance estimator is proposed. A simulation study was conducted to compare the performance of our score test and the CA score test in the survival setting over a variety of types of random effects. When the random effects are time dependent, our test at appropriate time points has higher power than the CA test based on the frailty proportional hazards model.