Ask Mallard
Core concept

Choosing a meta-analysis model

5 min read

Common-effect pooling targets one shared effect; random-effects pooling targets a mean across varying effects and needs a separate account of heterogeneity.

Fit

When it matters, and when it does not

It matters when

  • You plan to pool comparable estimates from several studies.
  • You need to distinguish uncertainty in a pooled mean from variation between underlying study effects.

It matters less when

Before you start

What to check in your study

Decide these first
  1. What the summary should meanDecide whether a shared effect or a mean across varying effects answers the review question.
  2. Whether studies can be combinedCompare populations, interventions, outcome definitions and effect scales.
  3. How heterogeneity and uncertainty will be estimatedSpecify the variance estimator, interval method and sensitivity analyses.
It rests on these
  1. A common-effect interpretation assumes a shared underlying effect.How to check: Clinical differences can challenge that assumption even when a heterogeneity test is not significant.
  2. Random-effects inference needs a plausible distribution of effects.How to check: Few studies make its spread and prediction uncertain. Examine outliers and influential small studies.

Plain language

How it works

Common-effect inverse-variance pooling weights each estimate by its precision. Under a shared-effect model, it estimates that common effect. Random effects adds between-study variance, called tau-squared and measured in squared units of the effect, to each study variance before forming weights.

The choice at a glance
One shared effect: Shared-effect interpretation is defensible leads to Common-effect pooling; Distribution of effects: Comparable effects may differ across settings leads to Random-effects mean; Interpret variation: Spread differs from precision of the mean leads to Tau-squared, I-squared and prediction.One shared effectShared-effect interpretation is defensibleCommon-effect poolingUse the study question and assumptions together.Distribution of effectsComparable effects may differ across settingsRandom-effects meanUse the study question and assumptions together.Interpret variationSpread differs from precision of the meanTau-squared, I-squared and predictionUse the study question and assumptions together.
Conceptual diagram. It shows the structure of the idea, not data from any study.
Reading the choice

I-squared describes the share of observed variation attributed to heterogeneity rather than sampling error. It is not the spread of effects in clinical units. Both measures can be uncertain, especially with few studies.

What decides it

Choose the target from the review question before testing heterogeneity. Report the mean, its confidence interval and the heterogeneity estimates. Consider a prediction interval when justified. Random effects gives smaller studies more relative weight and does not repair bias.

One shared effect or a distribution
Two panels. In a common-effect model, study estimates scatter around one true effect. In a random-effects model, the true effects themselves form a distribution around an average; a short bar shows the interval for the average and a longer dashed bar shows where a new study's effect may fall.Common effect: one true effectEffectThe one true effectStudies differ only by chance.Random effects: a distribution of trueeffectsEffectAverage effectDots: study estimates. Solid bar: interval for theaverage. Dashed bar: where a new study's effect mayfall.
Conceptual diagram. It shows the structure of the idea, not data from any study.

Worked example

A worked example

Invented for illustration
Setting
Two comparable studies estimate mean differences in the same units.
Question
Why can adding between-study variance change the summary?
Approach
Study estimates are 2 and 8, with variances 1 and 9. For illustration, set between-study variance to 9.
Result
Inverse-variance pooling gives 2.6 under common effect and about 4.14 under random effects. The latter gives the less precise study more relative weight, a change in the summary of about 1.54 units.
What it means
The two studies illustrate weights only. A between-study variance of 9 was supplied, not estimated from these data.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

Do not choose a pooling model from a heterogeneity P value. Random effects cannot make incompatible studies comparable.

  • Reading I-squared as the size of clinical differences.Instead: I-squared describes inconsistency relative to sampling error. Tau-squared measures between-study variance in the squared units of the chosen effect scale.
  • Reading a confidence interval for the mean as a prediction interval.Instead: A prediction interval describes an underlying effect in a comparable new setting, allowing for heterogeneity and its uncertainty.

Inside Mallard

How Mallard uses this

Mallard lists Choosing a meta-analysis model under Core research concepts.

No Mallard code example points to this page.
Atlas node concept.choosing_meta_models_heterogeneity, the id a plan cue records.

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. Measuring inconsistency in meta-analyses (opens in a new tab) · BMJ · doi 10.1136/bmj.327.7414.557
  2. A re-evaluation of random-effects meta-analysis (opens in a new tab) · J R Stat Soc Ser A Stat Soc · doi 10.1111/j.1467-985X.2008.00552.x
Prepared with AI assistance. Source titles, authors, publication years and available DOIs were verified against PubMed on 2026-11-06. No independent professional approval is claimed. Higgins and colleagues explain I-squared and its limits. Higgins, Thompson and Spiegelhalter distinguish the random-effects mean from the distribution and prediction of study effects.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.