Research Workshop in Applied Statistics presentation by Francesca Dominici, Harvard T H Chan School of Public Health

Joint with: Matt Cefalu, Giovanni Parmigiani, Nils Arvold. 

ABSTRACT. Researchers are increasingly being challenged with decisions on how to best control for a high-dimensional set of potential confounders when estimating causal eects. Typically, a single propensity score model is used to adjust for confounding, while the uncertainty surrounding the procedure to arrive at this propensity score model is often ignored and failure to include even one important confounder will results in bias. We propose a general causal framework that overcomes the limitations described above through the use of model averaging. We illustrate the proposed framework in the context of double robust estimation.The MA-DR estimator is defined as a weighted average of double robust estimators, where each double robust estimator corresponds to a specific choice for the outcome model and the propensity score. The MA-DR estimator extend the desirable double robustness property by achieving consistency under the much weaker assumption that either the true propensity score model or the true outcome model be within a specified, possibly large, class of models. Importantly, using simulation studies, we found that our MA-DR estimator dramatically reduces mean squared error by the largest percentage in the realistic situation where the set of potential confounders is large relative to the sample size. We apply the methodology to estimate the comparative effectiveness of the oral chemotherapy temozolomide on 1-year survival in a cohort of 1887 Medicare enrollees who were diagnosed with glioblastoma between June 2005 and December 2009.