October 24, 2018

Biostatistics Journal Club: Evaluating the Bias of Two-Stage Instrumental Variable Models in Cancer Research

Presented by: Fei Wan, PhD.

Tuesday, November 6th, 2018
12:00 – 1:00pm.


Observational studies are commonly used by cancer researchers to evaluate the relative effectiveness of different treatment options. However, confounding due to unmeasured or unknown variables can pose a serious issue when estimating treatment effects using observational data. To address this issue of unmeasured confounding, instrumental variable (IV) methods have become increasingly popular.
Specifically, the IV approach using two-stage residual inclusion (2SRI) has become a common analytic tool in studies of cancer therapies where the outcome of interest is overall or cancer-specific survival. However, despite its popularity, a compelling theoretical rationale has not been postulated nor have the limitations underlying the use of 2SRI in the context of survival outcomes been carefully laid out.
In this study we first provide a brief description of the concept of instrumental variables and their underlying assumptions. We then describe the 2SRI approach and provide examples from the literature of cancer comparative effectiveness studies that have used this approach. We show that the previous conclusion on the consistency of 2SRI in proportional hazards model relies on the unrealistic assumption that the effects of unmeasured confounders on treatment and outcome are proportional. Given a perfect instrumental variable, extension of 2SRI to proportional hazards model can generally result in biased estimates of treatment effect. We present a simple approach of assessing the bias of 2SRI as an omitted- variable-bias problem.
We conclude with our recommendations for the analysis of comparative effectiveness studies from observational data.