Efficient algorithms for building representative matched pairs with enhanced generalizability

TL;DR

This paper introduces an efficient network-flow-based algorithm for creating well-matched observational study pairs that better resemble target populations, improving the comparability of real-world evidence with randomized controlled trials.

Contribution

The paper presents a novel, computationally efficient matching algorithm that enhances population similarity in observational studies, aiding in the validation of real-world evidence against RCTs.

Findings

The method successfully reduced discrepancies between observational studies and RCTs.

Adjusting for additional covariates further aligned study results.

The algorithm is implemented in the match2C R package.

Abstract

Many recent efforts center on assessing the ability of real-world evidence (RWE) generated from non-randomized, observational data to produce results compatible with those from randomized controlled trials (RCTs). One noticeable endeavor is the RCT DUPLICATE initiative (Franklin et al., 2020, 2021). To better reconcile findings from an observational study and an RCT, or two observational studies based on different databases, it is desirable to eliminate differences between study populations. We outline an efficient, network-flow-based statistical matching algorithm that designs well-matched pairs from observational data that resemble the covariate distributions of a target population, for instance, the target-RCT-eligible population in the RCT DUPLICATE initiative studies or a generic population of scientific interest. We demonstrate the usefulness of the method by revisiting the inconsistency regarding a cardioprotective effect of the hormone replacement therapy (HRT) in the Women's Health Initiative (WHI) clinical trial and corresponding observational study. We found that the discrepancy between the trial and observational study persisted in a design that adjusted for study populations' cardiovascular risk profile, but seemed to disappear in a study design that further adjusted for the HRT initiation age and previous estrogen-plus-progestin use. The proposed method is integrated into the R package match2C.