Even though randomized experimentation is the sine qua non of causal inference, it is impractical or unethical to carry out controlled experiments to address many clinically important questions. Consequently, comparative effectiveness research (CER) based on observational data is needed as a best available alternative to controlled experimentation. The key concern in making causal inferences from nonexperimental data is nonrandom treatment assignment. Propensity score matching methods are widely used to address this problem in CER and broader types of causal inference. As of today, the original article by Rosenbaum and Rubin on the propensity score has been cited more than 20,000 times and the terms propensity score and matching appear in nearly 100,000 articles, according to Google Scholar. In observational studies, matching is used to approximate a randomized experiment by constructing samples that look alike across treatment arms in terms of observed co-variates. However, most matching methods involve a considerable amount of guesswork because they do not target co-variate balance directly, often leading to substantial residual imbalances in component co-variates. The PI previously developed a matching approach that resolves this problem via mathematical programming.
The proposed methods will dramatically improve over this and other existing matching methods. In aim 1, we will develop new statistical methods that encompass and improve fundamental aspects of matching and are tailored to big data. In aim 2, we will develop a new matching approach to estimate heterogeneity of treatment effects, by building optimal matched samples for specific patients that allow for a determination of which treatment is best for that patient’s medical condition. In aim 3, we will develop new matching methods for estimating treatment effects for interventions that have many options so as to allow assessment of treatment provider/setting quality.
In each of these aims, we will evaluate the performance of the new methods both in simulated data and in empirical data sets from the Veterans Health Administration. The latter data sets will allow us to illustrate the value of the methods for addressing questions in mental health research that are of independent interest to physicians, policy makers, and stakeholders. Dissemination is a critical objective of this project. To achieve this objective, in aim 4, we will develop open-source and easy-to-use software, case study vignettes, and YouTube video tutorials to make this new generation of matching methods widely available to practitioners in CER and patient-centered outcomes research.