Using high-dimensional disease risk scores in comparative effectiveness research of new treatments.

摘要

Nonexperimental research using automated healthcare databases can supplement randomized trials to provide both clinicians and patients with timely information to optimize treatment decisions. These studies, however, are susceptible to confounding and require design and statistical methods to control for large numbers of confounding variables. The propensity score (PS), defined as the conditional probability of treatment given a set of covariates, has become increasingly popular for controlling large numbers of covariates in pharmacoepidemiologic studies. During early periods after the introduction of a new treatment, however, accurately modeling the PS can be difficult because of rapid change over time in drug prescribing patterns and few exposed individuals. A historically estimated disease risk score (DRS), which summarizes covariate associations with the outcome absent of exposure, has been proposed as an alternative to PSs for controlling large numbers of covariates during these periods. Little is known about the performance and potential benefits of using DRSs for confounding control when evaluating the comparative effectiveness of newly marketed drugs.;In this study, we examined the benefits and challenges of using historically estimated DRSs compared to PSs when controlling for large numbers of covariates during early periods of drug approval. We further evaluated novel strategies for determining the validity of fitted DRS models in their ability to control confounding. We investigated these methodological questions using Monte Carlo simulations and empirical data. The empirical analyses included 20% and 1% samples of Medicare claims data to compare the new oral anticoagulant dabigatran with warfarin in reducing the risk of combined ischemic stroke and all-cause mortality in older populations.;When PS distributions are separated, DRS matching can improve the precision of effect estimates and allow researchers to evaluate the treatment effect in a larger proportion of the treated population. However, accurately modeling the DRS can be challenging compared to the PS. When evaluating the validity of DRS models, measures of predictive performance do not always correspond well with reduced bias in treatment effect estimates. Calculating the pseudo bias within a "dry run" analysis can provide a more direct measure for assessing the ability of fitted DRS models to control confounding.