A key principle in the design of observational studies for comparative clinical effectiveness research (CER) is to approximate a randomized controlled experiment — specifically, the experiment that would have been conducted under ideal circumstances to estimate the causal effect of the treatment. To date, inverse propensity score weighting methods have been extensively used in the pursuit of this principle for CER.
At present, the seminal article by Rosenbaum and Rubin (1983) on the propensity score has been cited more than 30,000 times and the terms “propensity score” and “weighting” appear in nearly 70,000 articles. However, in most settings the true propensity score (PS) is unknown, and thus it must be estimated from the data. Often, this is done by explicitly modeling the PS and inverting the predicted probabilities to form the weights. In practice, this approach may suffer from two important problems: (i) covariate balance may be inadequate due to misspecification of the PS model, small samples, or sparse covariates; and (ii) the estimated weights can be highly variable and in turn produce estimators that are highly unstable.
In previous work, the PI has developed a novel optimization approach to weighting that overcomes these problems by finding the weights of minimum variance that directly balance functions of the covariates. The proposed methods will substantially improve causal inference in CER relative to existing weighting methods.
In Aim 1, new weighting methods will be developed that can handle large data sets quickly, produce robust and interpretable estimators in difference-in-differences (DiD) settings, and facilitate targeted CER. In particular, in Aim 1.1 the research team will develop new weighting methods to facilitate the study of heterogeneity of treatment effects (HTE), patient-targeted estimation, and personalized medicine. In Aim 1.2, the team will develop new weighting methods for DiD and related designs, which are ubiquitous in CER studies. In Aim 1.3, cutting-edge algorithms for weighting in massive electronic medical record (EMR) data sets will be implemented that are increasingly used in CER studies.
In Aim 2, the team will apply and evaluate the performance of the proposed methods in both simulated and real EMR data sets from the Veterans Health Administration (VHA). The data sets will allow the team to illustrate the value of the new methods by addressing questions in mental health research that are of independent interest to physicians, policymakers, and other stakeholders.
Finally, dissemination is a critical objective to make this new call of matching methods widely available to practitioners in CER and PCOR. Therefore, through Aim 3, the research team will disseminate the new weighting methods to a wide audience of CER and PCOR investigators with open-source software and step-by-step tutorials.