stratamatch: Prognostic Score Stratification Using a Pilot Design.

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Block-randomized trials subdivide participants by baseline characteristics to reduce heterogeneity and improve treatment effect estimation.
• Observational studies often struggle with confounding factors, leading to less precise treatment effect estimates.
• Existing methods for observational studies may not fully capture the benefits of stratification seen in randomized trials.

Purpose of the Study:

• To introduce the stratamatch package, which extends block-randomized trial principles to observational studies.
• To provide functions for data stratification and assessment of stratification schemes in observational data.
• To enable propensity score matching within strata for more robust treatment effect estimation.

Main Methods:

• The stratamatch package implements a "pilot design" approach using prognostic scores to stratify observational data.
• It estimates prognostic scores (Hansen 2008) to divide individuals into strata, informing the matching process.
• Propensity score matching is then performed within these strata to emulate block-randomized trial designs.

Main Results:

• Stratification enhances computational efficiency for matching in large observational datasets.
• Using prognostic scores for stratification increases the precision of treatment effect estimates.
• This approach reduces sensitivity to bias from unmeasured confounding factors.

Conclusions:

• The stratamatch approach offers a robust design strategy for observational studies, improving the reliability of findings.
• Clever use of data in the design phase, via stratification, yields significant benefits in study robustness.
• This methodology allows observational studies to better approximate the rigorous design of block-randomized controlled trials.