Methods for deriving risk difference (absolute risk reduction) from a meta-analysis

Trading off benefits and harms requires knowledge of the absolute risk reduction or risk difference, making risk difference a critical measure for decision making. The confidence interval of risk difference is the basis for imprecision judgments made by guideline developers. Estimating risk difference is not straightforward, however, and the available methods have various limitations. Four methods are discussed in this article. The major limitation of the first method, pooling of risk differences generated from multiple studies in a meta-analysis, is the inconsistency of risk difference across baseline risks. The major limitation of the second method, transforming a pooled relative effect (such as a risk ratio or odds ratio) into a risk difference, is that its confidence interval does not incorporate uncertainty in the baseline risk. This confidence interval widens in a linear fashion as the baseline risk increases, making risk difference estimates in higher risk populations imprecise, and can lead to misleadingly precise risk differences when the baseline risk is low. Two alternative methods can reduce some of these limitations. A simple microsimulation approach can model uncertainties in both the relative effect and the baseline risk. The bivariate random effects model jointly analyzes the risks in treatment and control groups and computes conditional effects based on baseline risks. We apply these four methods to a case study and provide recommendations on when to use each approach. This article also provides practical advice and open source R coding to facilitate risk difference estimation by meta-analysts and guideline developers.