A meta-analysis takes the effect sizes from several studies and combines them into one estimate, weighting each by how precisely it was measured.
A meta-analysis takes the effect sizes from several studies and combines them into one estimate, weighting each by how precisely it was measured. Larger and better-controlled studies move the result more than small ones. It is arithmetic applied to findings rather than a summary of expert opinion, which is both its strength and its limitation.
If the included studies disagree with each other more than chance explains, pooling them produces an average of incompatible things. Reviews report this as a heterogeneity statistic, usually near the forest plot. A high value is a reason to distrust the combined number rather than a technical footnote, and good reviews say so in the text.
Two reviews covering the same literature can reach different numbers purely by weighting differently. That is not misconduct; it is a modelling choice. A transparent review states which model it used and why, and a reader who sees no such statement should treat the pooled figure as one of several defensible answers.
A fixed-effect model assumes every study estimates one true value and that differences between them are noise. A random-effects model assumes the true value itself varies between settings and populations. The second is usually the more realistic and produces wider intervals, which is why a review choosing the first owes the reader a reason.
Studies finding nothing are published less often, so the available literature is already a selected sample before anyone pools it. Funnel plots are the usual check. A review that does not look for this is combining a biased set without saying so, and the resulting estimate will lean toward whatever was publishable.
A careful review re-runs its estimate with the weakest studies removed, or with the single largest trial excluded, and reports whether the conclusion survives. If a finding depends on one trial, that is worth knowing, and a good review states it plainly rather than leaving it to be discovered.
Pooling tightens the confidence interval, which makes the result look more certain than any individual study. If the inputs shared a common flaw, the tighter interval is simply a more confident wrong answer, and nothing in the arithmetic can detect that.
The direction and rough magnitude of an effect across a field, read together with the heterogeneity figure. Not a verdict, and not a substitute for reading what the largest included trial actually did and to whom.
The claims on this page about combining results statistically, and the conditions under which that is legitimate can be checked against sources outside this site: