Online Program

WITHDRAWN: Incorporating external information to assess robustness of comparative effectiveness estimates to unobserved confounding

MaryBeth Landrum, Harvard Medical School 
Alfa Ibrahim Yansane, Harvard Health Care Policy Statistics Section 

Keywords: sensitivity analysis,propensity score, iv regression,probit regression,comparative effectiveness

Successful reform of the health care delivery system relies on improved information about the effectiveness of therapies in real world practice. While comparative effectiveness research often relies on synthesis of evidence from randomized clinical trials to infer effectiveness of therapies, many rely on the analysis of observational data sources where patients are not randomized to treatment. Comparative effectiveness studies based on observational data are reliant on a set of assumptions that often cannot be tested empirically. The objectives of this paper are to compare the performance of various approaches in the context of an analysis of the comparative effectiveness of cancer therapies in elderly populations typically underrepresented in cancer RCTs. To this end, we review and apply logistic regression, propensity scores, IV regression, bivariate probit regression, and Bayesian bivariate probit regression to the data. Next, we assess the sensitivity of our conclusions in order to address the problem of unmeasured confounding and biased estimates. The results of each method produced a protective effect estimate for cancer therapy on mortality. The subsequent sensitivity analyses showed that the concluding effect of treatment was not heavily influenced by a range of both binary and continuous unmeasured confounders.