Efficient and Robust Semi-Supervised Estimation of Average Treatment Effects in Electronic Medical Records Data
share
There is strong interest in conducting comparative effectiveness research (CER) in electronic medical records (EMR) to evaluate treatment strategies among real-world patients. Inferring causal effects in EMR data, however, is challenging due to the lack of direct observation on pre-specified gold-standard outcomes, in addition to the observational nature of the data. Extracting gold-standard outcomes often requires labor-intensive medical chart review, which is unfeasible for large studies. While one may impute outcomes and estimate average treatment effects (ATE) based on imputed data, naive imputations may lead to biased ATE estimators. In this paper, we frame the problem of estimating the ATE in a semi-supervised learning setting, where a small set of observations is labeled with the true outcome via manual chart review and a large set of unlabeled observations with features predictive of the outcome are available. We develop an imputation-based approach for estimating the ATE that is robust to misspecification of the imputation model. This allows information from the predictive features to be safely leveraged to improve the efficiency in estimating the ATE. The estimator is additionally doubly-robust in that it is consistent under correct specification of either an initial propensity score model or a baseline outcome model. We show that it is locally semiparametric efficient under an ideal semi-supervised model where the distribution of unlabeled data is known. Simulations exhibit the efficiency and robustness of the proposed method compared to existing approaches in finite samples. We illustrate the method to compare rates of treatment response to two biologic agents for treating inflammatory bowel disease using EMR data from Partner's Healthcare.
READ FULL TEXT