A Kernel-Based Metric for Balance Assessment.

Area of Science:

• Statistics
• Causal Inference
• Machine Learning

Background:

• Achieving covariate balance across treatment groups is crucial in causal inference.
• Traditional balance metrics focus on finite moments, potentially missing distributional differences.

Purpose of the Study:

• Introduce distributional balance preserving as a requirement for covariate distributions.
• Propose kernel distance, a novel balance measure, to capture multivariate distribution differences.
• Evaluate kernel distance's efficacy in identifying bias in causal effect estimation.

Main Methods:

• Defined distributional balance preserving and kernel distance based on reproducing kernel Hilbert spaces.
• Compared kernel distance with traditional balance measures using simulations.
• Integrated kernel distance into the genetic matching algorithm.
• Applied the enhanced genetic matching to analyze the Early Dieting in Girls study.

Main Results:

• Kernel distance outperformed common balance measures in indicating bias of estimated causal effects.
• The proposed kernel distance-integrated genetic matching was applied to a real-world study.
• Analysis of the Early Dieting in Girls study suggested a non-significant causal effect of maternal weight concern on daughters' early dieting.

Conclusions:

• Kernel distance offers a superior approach to assessing covariate balance in causal inference.
• The method provides a more robust assessment of bias compared to traditional metrics.
• The application highlights the importance of advanced balance measures in observational studies.