Primary Submission Category: Matching
Comparing Coarsened Exact Matching (CEM) and Almost Matching Exactly (AME) in Real-World Settings: Insights from Cigna’s Annual Value of Integration Study
Authors: Aran Canes,
Presenting Author: Aran Canes*
Coarsened Exact Matching (CEM) allows for the comparison of treatment and control populations by leveraging subject matter expertise to identify confounders and determine appropriate intervals for matching. Despite its effectiveness in case/control studies, CEM often encounters issues in real-world applications. Including a variable that does not influence the outcome biases the results. The appropriate coarsening, determined by subject matter expertise, may not be accurate enough to avoid introducing confounder bias. Additionally, the selection of confounders is susceptible to errors in judgment.
Almost Matching Exactly (AME), developed at Duke University, offers alternatives to mitigate these issues. AME employs machine learning to identify relevant confounders and determine appropriate coarsening for continuous variables, reducing the need for subject matter expertise.
However, CEM has a distinctive advantage over AME. By using an outcome variable to determine binning, AME requires a unique match for each distinct outcome. In contrast, CEM can be applied once to a treatment/control population with multiple outcomes.
This paper utilizes Cigna’s Annual Value of Integration Study to illustrate these advantages and disadvantages and their application in our 2024 study. By using real-world data, we aim to contribute to the causal inference community by demonstrating the utility of these methods in a non-academic setting.