How to Control for Unmeasured Confounding in an Observational Time-to-Event Study with Exposure Incidence Information: The Treatment Choice Cox Model

• Activity Number: 700
• Type: Contributed
• Date/Time: Thursday, August 4, 2016 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): James Troendle* and Zhiwei Zhang and Eric Leifer and Song Yang and Michael Sklar and Heather Jerry
• Companies: National Institutes of Health and FDA/CDRH and National Heart, Lung, and Blood Institute and National Heart, Lung, and Blood Institute and University of Pennsylvania and Nebraska Department of Health and Human Services
• Keywords: Bias ; Frailty ; Hazard ; Matching ; Propensity

Abstract

In observational studies of incident events a major problem is accounting for confounding which could come through unknown or unmeasured factors. Standard methods based on propensity matching or adjustment using a model of the observed time to first event can only address known and measured confounding factors. We propose a method that can account for unknown or unmeasured factors that are related to both treatment changes and incident events according to a Cox model. The key information used in the proposed model is the treatment choice (TC) process itself. To do this one needs to have available the information on incidence of treatment changes. Since this information comes as a time varying process, the resulting model can only account for confounding to the degree that the TC process is able to be modeled. Modeling the TC process well requires large sample sizes. Three TC models are developed. Simulations indicate the bias is reduced as the sample size is increased, allowing for more points where the TC covariates change value. The methods are illustrated by analyzing the Women's Health Initiative observational study of hormone replacement for post-menopausal women.