Dr. Atul Sharma is a pediatric nephrologist and biostatistician, who recently joined the Biostatistical Consulting Group as a senior consultant. As such, he brings a unique combination of clinical and statistical insights to statistical topics of interest to clinical investigators.
Dr. Atul Sharma | Propensity Scores: Making Sense of Non-Randomized Observational DataDownload Workshop Files Here Workshop preparation info Intro to R
Background: Randomized, controlled trials remain the gold standard in medical research. Nevertheless, with the increasing availability of administrative data and ‘natural experiments’ in health policy, clinical investigators must be prepared to interpret and analyze observational data arising from non-randomized trials and quasi-experimental study designs. This workshop is intended to support investigators as they design and interpret such trials. In it, we will focus on the application of propensity score methods to account for the selection biases that can confound such studies, even making it possible to analyze observational data as if they arose from a randomized trial.
Divided into two sessions, the first will provide a 1h introduction to the theory and application of propensity score methods, concentrating on real-world examples from the medical literature. The second session will be a 1h computer lab, to review the use of specialized software needed to perform propensity score analysis, including the R statistical language and specialized libraries for propensity score matching, assessment of post-match balance, and sensitivity analysis (e.g. the Matching and rbounds packages).
- Understand the role of randomization in ensuring ‘covariate balance’ between experimental groups
- Recognize the implications of selection biases as confounders in non-randomized observational data
- Understand how the ‘balancing property’ makes it possible to condition on propensity scores to balance experimental groups in non-randomized trials
- Regression adjustment
- Weighted regression
- Propensity score matching
- Appropriate methods for assessing covariate balance
- Sensitivity analysis to test the robustness of conclusions to hidden biases from unobserved confounders (Rosenbaum bounds)
- Devices for testing the robustness of study conclusions, including multiple control groups, coherence, and dose-response relationships.
- Understand the various ways that of propensity scores (PS) can be used to analyze quasi-experimental designs, including:
- Recognize the limitations of propensity score methods and understand the importance of post-analysis assessment of modeling assumptions, including: