Background: Choice of fixed or random-effects models in meta-analysis is controversial. We present a conceptual framework for considering the relative merits of the two approaches, and highlight results of a recent meta-analysis that identifies one situation in which fixed effects is clearly superior. One can think of fixed and random effects in terms of underlying theory, computational implications, and practical consequences (see Table). Some find random effects appealing in that we are less confident in pooled estimates when there is large variability across studies; thus, incorporating among-study variability may be desirable; and wider confidence intervals in the face of heterogeneity encourage conservative interpretations. However, assumptions of randomness in random effect models are open to serious question, and poor estimates of among-study variability are problematic. Random effects become most highly questionable when a single high-quality study is far larger than other studies, and appreciable heterogeneity exists. In this situation, random-effects models may yield non-intuitive results that place insufficient weight on the large study. Results: We encountered this situation in a meta-analysis of betablockers in non-cardiac surgery. The POISE trial showed increased deaths with beta blockers (266 events in 8351 patients, RR 1.33, 95% CI 1.03 to 1.72). The other seven trials found a total of 69 events in 2260 patients.With appreciable variability present (I² 36%), the fixed effects model showed an intuitively sensible pooled effect and confidence interval (RR 1.24, 95% CI 0.99 to 1.56); the random effects did not (RR 1.05, 95% CI 0.65 to 1.68).