Experience with Using a Random-Effects Cox Model for Genetic Data

• Activity Number: 192
• Type: Invited
• Date/Time: Tuesday, August 5, 2003 : 8:30 AM to 10:20 AM
• Sponsor: IMS
• Author(s): Terry M. Therneau*+
• Companies: Mayo Clinic

Abstract

We have implemented a random-effects Cox model function coxme in Splus, based on the Gaussian random effects framework of Ripatti and Palmgren (2000). The model is \lambda_i(t) = \lambda_0(t) e^{X_i \beta + b_i}, where $\lambda_i$ is the hazard function for subject $i$, $\lambda_0$ is an unspecified baseline hazard, $\beta$ and $X$ represent the fixed effects and covariates, respectively, and $b_i$ is a per-subject random effect b ~ N(0, \sigma_1^2 A + \sigma_2^2 B + .)with A, B, etc., as arbitrary correlation matrices. Commonly, A will be the kinship matrix for the study. We have applied the model to analyses from the Minnesota Breast Cancer Family Study, a large cohort effort containing data on 26,050 multi-generational subjects from 426 families, using familial (one b per family, A = identity), and correlated (A = kinship) random-effects models. In other datasets, a series of fits with A = the kinship matrix and B a sequential series of IBD matrices along a chromosome has been used to create a LOD score/linkage map. The model framework has some advantages, including direct utilization of age-of-onset data and case-cohort methods to account for proband selection.