Tuesday, July 25, 2017
12:30 PM – 2:30 PM
KRESGE 201
Lunch will be provided
Sufficient Cause Interaction for Ordinal OutcomesÂ
Jaffer Zaidi
Abstract: VanderWeele and Robins (Biometrika 2008) derived counterfactual and empirical conditions for sufficient cause interaction between two binary exposures for an event. Sufficient cause interaction can be shown present if there exists a subpopulation for whom the binary outcome occurs if both exposures are present, but will not occur if either of the two exposures is absent. We extend the sufficient cause framework from binary outcomes to ordinal outcomes. Novel empirical conditions, in the form of inequality constraints on the observed data distribution, are derived for detecting sufficient cause interaction for ordinal outcomes. These inequality constraints cannot be derived through first dichotomizing the ordinal outcome, then applying the earlier inequality tests from the framework for binary outcomes. Inference to test the null hypothesis that there is no sufficient cause interaction using these novel inequality constraints is developed. Using the Stanford HIV drug resistance database, we discover mutations that mechanistically interact to confer resistance (none, partial, full) to particular HIV drugs.
Estimating Average Treatment Effects with a Response-Informed Calibrated Propensity Score
David Cheng
Abstract: Adjusting for the propensity score (PS) is a common approach to estimate treatment effects in observational studies. The performance of inverse probability weighting (IPW) and doubly-robust (DR) estimators deteriorate when underlying parametric models for the PS and response are mis-specified or when adjusting for high-dimensional covariates. We propose a response-informed calibrated PS approach that is more robust to model mis-specification and accommodates a large number of covariates while preserving the double-robustness and semi-parametric efficiency properties under correct model specification. Our approach achieves additional robustness and efficiency gain by estimating the PS using a two-dimensional smoothing over an initial parametric PS and another parametric response score. Both of the scores are estimated via regularized regression to accommodate a large number of covariates. Simulations confirm these favorable properties in finite samples. We illustrate the method by estimating the effect of statins on colorectal cancer risk in an electronic medical record (EMR) study and the effect of smoking on C-reactive protein in the Framingham Offspring Study.
