Balancing Confounding and Generalizability Using Observational, Real-world Data: 17-gene Genomic Prostate Score Assay Effect on Active Surveillance.

Abstract

Randomized, controlled trials can provide high-quality, unbiased evidence for therapeutic interventions but are not always a practical or viable study design for certain healthcare decisions, such as those involving prognostic or predictive testing. Studies using large, real-world databases may be more appropriate and more generalizable to the intended target population of physicians and patients to answer these questions but carry potential for hidden bias. We illustrate several emerging methods of analyzing observational studies using propensity score matching (PSM) and coarsened exact matching (CEM). These advanced statistical methods are intended to reveal a "hidden experiment" within an observational database, and so refute or confirm a potential causal effect of assignment to an intervention and study outcome. We applied these methods to the Optum™ Research Database (ORD; Eden Prairie, MN) of electronic health records and administrative claims data to assess the effect of the 17-gene Genomic Prostate Score® (GPS™; Genomic Health, Redwood City, CA) assay on use of active surveillance (AS). In a traditional multivariable logistic regression, the GPS assay increased the use of AS by 29% (95% CI, 24%-33%). Upon applying the matching methods, the effect of the GPS assay on AS use varied between 27% and 80% and the matched data were significant among all algorithms. All matching algorithms performed well in identifying matched data that improved the imbalance in baseline covariates. By using different matching methods to assess causal inference in an observational database, we provide further confidence that the effect of the GPS assay on AS use is statistically significant and unlikely to be a result of confounding due to differences in baseline characteristics of the patients or the settings in which they were seen.