Meta-analysis

In one sentence

Meta-analysis combines estimates from separate studies using a specified statistical method to answer a defined question.

The intuition

Several small windows can offer a broader view. They help only if you know where each window looks and whether its glass distorts the scene. Combining studies can improve precision, but a larger combined result can still answer the wrong question or combine biased evidence.

How it works

A systematic review uses explicit methods to find, select and appraise relevant studies. Meta-analysis is its possible statistical component. Some reviews should not pool results because the populations, interventions, comparators or endpoints differ too much.

Pooling commonly uses a weighted combination of estimates. A common-effect model targets a shared underlying effect; a random-effects model allows a distribution of study effects and summarizes its mean. Neither removes bias. Cochrane Statistical Methods Group: meta-analysis.

Heterogeneity means study results differ beyond the variation addressed by a particular analysis. Investigate differences in setting, method and risk of bias. A confidence interval for an average and a prediction interval describing possible study effects answer different questions. Prediction intervals also have assumptions and can be unreliable with few studies. Higgins, Thompson and Spiegelhalter: random-effects interpretation.

Why it matters in cancer

Trials may compare different treatment sequences or count different events. Preserve each study's question before pooling. Several papers can report the same participants; a later follow-up is additional information, not an independent trial. Count unique studies and check overlapping cohorts.

A pooled observational association does not acquire randomized treatment assignment by becoming larger. A combined estimate also does not identify one person's benefit.

Worked example

Two fictional studies report risk differences of minus 4 and minus 8 percentage points for the same defined event and time horizon. Suppose a specified teaching calculation gives them normalized weights of 0.75 and 0.25. Their weighted mean is 0.75 × (−4) + 0.25 × (−8) = −5 percentage points. These weights are supplied only to illustrate averaging; they are not derived estimates of study precision.

Now discover that the first study started follow-up at randomization and the second at a later response-selected landmark. The arithmetic is still correct, but the studies answer different questions. Revisit inclusion and the target estimate before presenting the mean as a treatment effect.

Common confusions

A systematic review need not include a meta-analysis.
More papers do not necessarily mean more independent participants.
A random-effects model does not make clinical differences or reporting bias disappear.
A precise pooled mean does not mean effects are similar in every setting.
A prediction interval for study effects is not an individual's prognosis interval.

Try it

An abstract, full trial report and later follow-up all appear in a review. Should they receive three independent study weights?

Answer: No. Link the reports to their shared trial and choose the appropriate information for the defined analysis, avoiding duplicate participants.

Sources and scope

Source check: October 9, 2026. Studies, effects and weights are fictional. Expert and learner review remain pending.