The impact of neglected confounding and interactions in mixed-effects meta-regression

TL;DR

This paper demonstrates that neglecting interaction terms in mixed-effects meta-regression can lead to biased results, emphasizing the importance of including plausible interactions to improve inference accuracy.

Contribution

It provides empirical evidence and simulation results showing the bias caused by ignoring interactions and recommends always including them when plausible.

Findings

Ignoring interactions can bias meta-regression results

Including interactions improves inference accuracy

Simulation confirms the importance of modeling interactions

Abstract

Analysts seldom include interaction terms in meta-regression model, what can introduce bias if an interaction is present. We illustrate this in the current paper by re-analyzing an example from research on acute heart failure, where neglecting an interaction might have led to erroneous inference and conclusions. Moreover, we perform a brief simulation study based on this example highlighting the effects caused by omitting or unnecessarily including interaction terms. Based on our results, we recommend to always include interaction terms in mixed-effects meta-regression models, when such interactions are plausible.