Controlling for Unmeasured Confounding in Time-to-Event Analysis of Longitudinal Observational Studies

• Activity Number: 454
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): James Troendle*+ and Zhiwei Zhang and Eric Leifer and Song Yang and Heather Jerry
• Companies: NIH and FDA/DHHS and NHLBI/NIH and NHLBI/NIH and Nebraska Department of Health and Human Services
• Keywords: Bias ; Hazard ; Frailty ; Matching ; Propensity

Abstract

In observational studies of incident events a major challenge is accounting for confounding which could come through unknown or unmeasured factors. Standard methods such as a naïve proportional hazard model or propensity matching 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 initiation and incident events according to a proportional hazard model. The key information used in the proposed model is the time to treatment initiation. Since this information is itself censored, the resulting model is only able to partially account for confounding. However, by partitioning the study period into a run-in phase followed by an analysis phase, the resulting estimate of the effect of starting treatment becomes less biased. If a run-in phase is used, simulations from a model with residual unmeasured confounding indicates that the bias is reduced as the maximal length of the run-in phase and the sample size simultaneously increase. The method is illustrated by analyzing the Women's Health Initiative observational study of hormone replacement for post-menopausal women.