- This event has passed.
Designing Observational Studies using Propensity Score Cardinality Matching and the SUPPORT Right Heart Catheterization Study
May 12 @ 9:00 am - 10:30 am
We will demonstrate the use of modern matching techniques to address the issue of treatment selection bias (both overt and hidden) in the design and analysis of observational studies.
Confronted with the problem of finding an appropriate matched sample to compare two exposures while accounting for imbalances in covariates, several approaches exist. A new method called cardinality matching (developed by Zubizarreta et al. 2013) has some very attractive properties, including a requirement of meaningful input on both the clinical and statistical/epidemiological level.
Our motivating example uses data from the SUPPORT right heart catheterization (RHC) analyses (Connors et al. 1996 JAMA). After some background and discussion of the use and abuse of propensity score methods, we present a practical demonstration of the use of mixed integer programming to implement cardinality matching.
This approach allows the research team to find the largest matched sample that is balanced “by design.” We illustrate the direct balance of several features of the covariates, including their marginal distributions – without requiring exact matching, which can be a very stringent constraint in practice. Cardinality matching also lets us address the problem of limited overlap in covariate distributions using the original covariates, as opposed to a summary of them (like the propensity score), balancing covariates directly in a flexible way. Using the RHC data, we will show that this design approach reduces sensitivity to hidden bias (estimated using Rosenbaum bounds) relative to more typical designs using 1:1 greedy matching on the propensity score.
David Ngendahimana is a Ph.D. candidate in the Department of Epidemiology and Biostatistics at CWRU. He holds a Masters degree in Statistics from Youngstown State University. He is a pre-doctoral fellow at the Prevention Research Center for Healthy Neighborhoods working on the Building Capacity for Prevention of Obesity (BCOP) project. He is developing a dissertation with Professor Love in the design of observational studies with a focus on sensitivity analysis.
Thomas E. Love, Ph.D., is Director of Biostatistics and Evaluation at the Center for Health Care Research and Policy, Professor of Medicine, Epidemiology & Biostatistics at CWRU, Chief Data Scientist for Better Health Partnership, and a Fellow of the American Statistical Association. He was a Finalist for this year’s John S. Diekman Award for Outstanding Graduate Teaching at CWRU, and has won several awards (including some with his CHRP colleagues) for short courses in the design and analysis of observational studies and of cluster-randomized trials.