Abstract:
|
In healthcare research with observational data, researchers often have concern about the possibility that unmeasured confounding would bias the estimated effect under investigation. Existing sensitivity methods are useful rules of thumb in assessing how robust is the estimated association to unmeasured confounders. However, If we have some knowledge of the magnitude of the effects of an unmeasured confounder to treatment and outcome, can we use it to form a plausible range of possible values for the true causal effect of the exposure on the outcome? We propose an intuitive method L^2 which is straightforward to understand, easy to use, easy to report, and allow the researcher to identify a plausible range of true effects based on prior knowledge of correlations between unmeasured confounders with both outcome and exposure. We will use an empirical example to illustrate the concept and application of our method: how effectiveness of Percutaneous Coronary Intervention (PCI) vs. Coronary Artery Bypass Grafting (CABG) in patients with stable Ischemic heart disease can be affected by unmeasured confounders and how we address this issue.
|