Directional penalties for optimal matching in observational studies

Area of Science:

• Statistics
• Observational Studies
• Biostatistics

Background:

• Multivariate matching in observational studies often treats covariate differences symmetrically.
• This symmetric approach can be suboptimal when specific biases, like treated subjects being consistently older than controls, need correction.

Purpose of the Study:

• To introduce and evaluate easily used, asymmetric, directional penalties for improving covariate balance in matched samples.
• To demonstrate how these penalties can refine existing matched samples by addressing residual imbalances.

Main Methods:

• The approach involves starting with a conventionally matched sample and then applying directional penalties to the distance matrix.
• These penalties are adjusted based on balance diagnostics to reduce specific covariate imbalances.
• The connection between directional penalties and Lagrangian relaxation in integer programming is explored.

Main Results:

• Directional penalties can substantially improve covariate balance in matched samples.
• The method requires minimal computational effort, with significant improvements often achieved in a few adjustments.
• The magnitude of the directional penalty is crucial to avoid reversing the bias.

Conclusions:

• Asymmetric, directional penalties offer an effective and efficient way to enhance covariate balance in observational studies.
• This technique refines matching by specifically targeting and correcting residual imbalances.
• The R package DiPs is available for implementing these methods.