Estimating relative risks and risk differences in randomised controlled trials: a systematic review of current...

Guidelines for randomised controlled trials (RCTs) recommend reporting relative and absolute measures of effect for binary outcomes while adjusting for covariates. There are a number of different ways covariate-adjusted relative risks and risk differences can be estimated.

Our goal was to identify methods used to estimate covariate-adjusted relative risk and risk differences in RCTs published in high-impact journals with binary outcomes. Other secondary objectives included the identification of how covariates are chosen for adjustment and whether covariate adjustment results in an increase in statistical precision in practice.

We included two-arm parallel RCTs published in JAMA, NEJM, Lancet, or the BMJ between January 1, 2018, and March 11, 2023, reporting relative risks or risk differences as a summary measure for a binary primary outcome. The search was conducted in Ovid-MEDLINE.

Of the 308 RCTs identified, around half (49%; 95% CI: 43-54%) reported a covariate-adjusted relative risk or risk difference. Of these, 82 reported an adjusted relative risk. When the reporting was clear (n = 65, 79%), the log-binomial model (used in 65% of studies; 95% CI: 52-76%) and modified Poisson (29%; 95% CI: 19-42%) were most commonly used. Of the 92 studies that reported an adjusted risk difference, when the reporting was clear (n = 56, 61%), the binomial model (used in 48% of studies; 95% CI: 35-62%) and marginal standardisation (21%; 95% CI: 12-35%) were the common approaches used.

Approximately half of the RCTs report either a covariate-adjusted relative risk or risk difference. Many RCTs lack adequate details on the methods used to estimate covariate-adjusted effects. Of those that do report the approaches used, the binomial model, modified Poisson and to a lesser extent marginal standardisation are the approaches used.