Primary Submission Category: Matching, Weighting
Quantifying Overlap in Causal Inference: A Framework for Early-Stage Assessment
Authors: Geondo Park, Juyeon Kim, Kwonsang Lee,
Presenting Author: Geondo Park*
Overlap between treated and control groups is a critical assumption in causal inference, and the most common way to assess overlap is through visual diagnostics, such as propensity score histograms or density plots. We propose a novel framework to quantify overlap magnitude prior to estimation, offering insight during the initial assessment phase. To address the challenges of existing methods, we develop a computationally efficient approach that uses a weak distributional assumption, inspired by two-component mixture models, to directly quantify overlap. Indirect approaches, such as cardinality matching or effective sample size (ESS), may be used to quantify overlap, though their primary goal is not to do so. Cardinality matching maximizes the size of matched controls within specified constraints, and its maximum size can, roughly speaking, correspond to the maximum overlap. However, this indirect measure is unreliable due to the variability introduced by changing constraint conditions. ESS, commonly used in weighting approaches, can indirectly reflect overlap but is highly variable and depends heavily on the choice of weighting methods. By providing a robust and interpretable measure of overlap, our approach enables researchers to make informed decisions early in the analysis, ensuring more efficient and reliable causal estimators. This work fills a critical gap in causal inference methodology, offering a practical and scalable tool for early-stage assessment of overlap.