Flexible template matching for observational study design.

Area of Science:

• Causal inference
• Observational data analysis
• Biostatistics

Background:

• Matching is a nonparametric method for causal inference with observational data, aiming to simulate randomization.
• Existing matching designs face limitations regarding the causal estimand and unequal sample sizes between treatment arms.

Purpose of the Study:

• To propose a flexible template matching design to overcome limitations of traditional matching methods.
• To provide theoretical justification for unbiased estimation of average treatment effect and average treatment effect on the treated.
• To enhance matching quality and develop a practical template size selection strategy.

Main Methods:

• Introduced a template matching approach by identifying a representative template group and matching subjects to it.
• Proposed a triplet matching algorithm for improved matching quality.
• Developed a strategy for selecting the template size.
• Utilized a randomization inference framework for binary outcomes in matched data, incorporating sensitivity analysis.

Main Results:

• Provided theoretical justification for unbiased estimation of average treatment effect and average treatment effect on the treated, particularly when the treatment group is larger.
• Demonstrated the flexibility of the design in handling different causal estimands and sample sizes.
• The proposed methods were applied to a trauma care evaluation study.

Conclusions:

• The proposed template matching design offers a flexible and robust approach for causal effect inference from observational data.
• The method allows for both randomization-based and model-based inference, with enhanced robustness through randomization inference.
• The framework supports heterogeneous effects and sensitivity analysis for unmeasured confounding.