Time-Dependent Covariates in Recurrent Event Models

• Activity Number: 18 - Survival Analysis Developments for Improving Medical Decision Making
• Type: Topic Contributed
• Date/Time: Sunday, July 29, 2018 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Xianghua Luo* and Tianmeng Lyu and Yifei Sun and Chiung-Yu Huang
• Companies: University of Minnesota, School of Public Health and University of Minnesota and Columbia University and University of California at San Francisco
• Keywords: Imputation for missing data; Kernel smoothing; Recurrent events; Time-dependent covariates

Abstract

In follow-up studies, investigators are often interested in evaluating the effects of time-dependent covariates on recurrent event risks. Theoretically, the usual proportional or additive rates model (PRM/ARM) allows time-dependent covariates, yet the estimation requires such covariates to be observed for all at-risk individuals at all event times (in PRM) or more unrealistically, throughout the follow-up (in ARM). In practice, simple methods such as the last value carried forward method have been used to approximate individual covariate trajectories until recently kernel-smoothed estimating functions were proposed for the PRM. Currently, our team is extending the ARM using similar techniques. In this talk, we focus on some frequently overlooked design/analysis issues shared by these models: 1) both time-dependent and -independent covariates being in the model; 2) non-synchronized measurements of different time-dependent covariates; 3) whether to use baseline measurements of time-dependent covariates in estimation; 4) time-dependent covariates not being measured at event times. We hope that the insights we share here can help inform the design of longitudinal studies in the future.