Use and Misuse of Multivariable Approaches in Interventional Cardiology Studies on Drug‐Eluting Stents: A Systematic Review

Abstract

Randomized clinical trials (RCTs) are the most reliable evidence, even if they require important resource and logistic efforts. Large, cost-free and real-world datasets may be easily accessed yielding to observational studies, but such analyses often lead to problematic results in the absence of careful methods, especially from a statistic point of view. We aimed to appraise the performance of current multivariable approaches in the estimation of causal treatment and effects in studies focusing on drug-eluting stents (DES). Pertinent studies published in the literature were searched, selected, abstracted, and appraised for quality and validity features. Six studies with a logistic regression were included, all of them reporting more than 10 events for covariates and different length of follow-up, with an overall low risk of bias. Most of the 15 studies with a Cox proportional hazard analysis had a different follow-up, with less than 10 events for covariates, yielding an overall low or moderate risk of bias. Sixteen studies with propensity score were included: the most frequent method for variable selection was logistic regression, with underlying differences in follow-up and less than 10 events for covariate in most of them. Most frequently, calibration appraisal was not reported in the studies, on the contrary of discrimination appraisal, which was more frequently performed. In seventeen studies with propensity and matching, the latter was most commonly performed with a nearest neighbor-matching algorithm yet without appraisal in most of the studies of calibration or discrimination. Balance was evaluated in 46% of the studies, being obtained for all variables in 48% of them. Better exploitation and methodological appraisal of multivariable analysis is needed to improve the clinical and research impact and reliability of nonrandomized studies.