Fixed Effect Versus Random Effects Models in Meta-analysis: As Simple as It Gets

This article presents a simple explanation of fixed-effect and random-effects models in meta-analysis. In summary, we use a fixed effect model when the studies to be pooled are largely similar in design, methods, and sample characteristics, and when we can consequently assume that the values of an outcome in these studies represent sample-related variations of a single true value in the population. We use a random effects model when studies are meaningfully dissimilar in their design, methods, and/or samples, and when we can consequently assume that the values of the outcome in these studies represent sample-related variations of true values in different populations; that is, there are many true values. These conceptual issues should guide an a priori decision to favor one model or the other; the choice should not be made after eyeballing the results of analysis. This article also explains when a fixed effect model should be chosen, how the choice of model affects study weights in a forest plot, and how the choice of model affects the pooled estimate, its precision, and its chance of being statistically significant. Other practical notes are also provided.