The choice between a fixed-effect and a random-effects model changes three things at once: how wide your confidence interval is, how much influence small studies get, and how far you can generalize the pooled result. Both models take the same set of study-level effect estimates and produce a single pooled number, but they start from opposite assumptions about what that pooled number is supposed to represent. Picking the wrong one doesn’t just shift a decimal — it can flip whether a result looks precise and generalizable or uncertain and context-bound.
The two models, side by side
Both models are inverse-variance weighted averages of the study effect estimates. They differ only in what goes into the weight and, as a result, in what the pooled estimate is meant to estimate.
| Dimension | Fixed-effect model | Random-effects model |
|---|---|---|
| Assumption about the true effect | One single true effect underlies every study; observed differences between studies are entirely sampling error. | Each study estimates its own true effect, and those true effects are themselves drawn from a distribution across a population of possible settings. |
| What’s being pooled | The common effect. | The mean of the distribution of effects. |
| Study weight | 1 / within-study variance only. | 1 / (within-study variance + between-study variance, τ²). |
| Effect of adding τ² | Not applicable — τ² is assumed to be zero. | Compresses the gap between large and small studies’ weights, since τ² is added to every study’s variance. |
| Confidence interval width | Narrower, and gets narrower still as more studies are added, because it reflects only within-study sampling error. | Wider under heterogeneity, and does not shrink toward zero just by adding more heterogeneous studies, because τ² is a property of the studies, not a sample-size problem. |
| Relative weight given to small studies | Low — a large, precise study can dominate the pooled estimate. | Higher than under the fixed-effect model, since adding τ² shrinks the gap between large and small studies’ variances. |
| What the result generalizes to | The specific set of conditions the included studies share. | A broader population of settings the included studies are treated as a sample from. |
| When it’s defensible | Studies are functionally replicates of each other — same population, intervention, comparator, and outcome definition — and heterogeneity is negligible. | Studies vary in ways expected to produce genuinely different true effects (different populations, dosages, follow-up lengths, settings), which is the ordinary case in most applied research syntheses. |
Both are described in Chapter 10 of the Cochrane Handbook: the fixed-effect model assumes “there is one true effect size … which underlies every study,” while the random-effects model assumes the true effects “could vary from study to study” and estimates the mean of that distribution. The standard random-effects estimator for the between-study variance, τ², is the DerSimonian–Laird method (DerSimonian & Laird, Controlled Clinical Trials, 1986), still the most widely implemented default in meta-analysis software, though restricted maximum likelihood (REML) is increasingly preferred where confidence-interval accuracy at small study counts matters.
The misconception this term invites
“Random effects” here does not mean the studies were randomly sampled from some registry of all possible studies on the topic — they weren’t, and no meta-analysis’s included-studies list is a random sample in that sense. It means the model treats the true study effects as random variables drawn from a distribution, purely as a statistical device for allowing them to differ. This is a standard source of confusion because “random effects” is also used in an unrelated sense in mixed-effects (multilevel) regression models, where it describes group-varying intercepts or slopes in nested data such as repeated patient measurements or clustered trial sites. That is a related but distinct statistical concept from the fixed-effect/random-effects choice in meta-analysis discussed on this page — the two ideas share vocabulary, not application.
One dataset, analyzed both ways
Illustrative worked example — hypothetical data, not drawn from any published study or real trial. Five studies report a standardized mean difference (SMD) for the same intervention, each with its own standard error:
| Study | SMD | SE | Variance | Fixed weight (1/var) | Fixed weight share |
|---|---|---|---|---|---|
| A | 0.10 | 0.10 | 0.0100 | 100.00 | 16.6% |
| B | 0.15 | 0.08 | 0.0064 | 156.25 | 25.9% |
| C | 0.55 | 0.15 | 0.0225 | 44.44 | 7.4% |
| D | 0.05 | 0.06 | 0.0036 | 277.78 | 46.0% |
| E | 0.40 | 0.20 | 0.0400 | 25.00 | 4.1% |
Fixed-effect model: summing the inverse-variance weights (603.47) and the weighted effects gives a pooled SMD of 0.135, 95% CI 0.056 to 0.215. Study D, the smallest-variance (most precise) study, supplies 46% of the total weight on its own — more than double its nearest competitor — and pulls the pooled estimate down toward its own low value of 0.05.
Testing for heterogeneity: Cochran’s Q for this set of five studies is 11.57 on 4 degrees of freedom, giving I² ≈ 65% — per the Cochrane Handbook’s overlapping interpretation bands, a value in the range that “may represent substantial heterogeneity.” That’s the signal that a fixed-effect model’s core assumption (one shared true effect) doesn’t hold well here, and that a random-effects model is the more defensible choice for this dataset. See Heterogeneity in Meta-Analysis for how I², τ², and prediction intervals are calculated and interpreted in full.
Random-effects model (DerSimonian–Laird): the between-study variance works out to τ² ≈ 0.0183. Adding τ² to every study’s variance before inverting compresses the weights: Study D’s share drops from 46.0% to 28.0%, while Study E’s share rises from 4.1% to 10.5% and Study C’s from 7.4% to 15.0%. The pooled estimate moves to 0.198, 95% CI 0.044 to 0.351 — a noticeably higher point estimate and a confidence interval roughly 90% wider than the fixed-effect model’s.
| Result | Fixed-effect | Random-effects |
|---|---|---|
| Pooled SMD | 0.135 | 0.198 |
| 95% CI | 0.056 to 0.215 | 0.044 to 0.351 |
| CI width | 0.159 | 0.307 |
| Largest single study’s weight share | 46.0% (Study D) | 28.0% (Study D) |
Same five numbers, same standard errors — a materially different pooled estimate, a wider interval, and a different picture of which study dominates the conclusion. That divergence is the entire reason the choice of model matters and isn’t a formality.
Why the divergence happens
Everything traces back to the weight formula. Under the fixed-effect model, weight is purely a function of a study’s own precision, so a large, tightly-estimated study can outweigh several smaller studies combined — exactly what happened with Study D above. Under the random-effects model, the constant τ² term is added to every study’s variance before the weight is calculated. Since τ² is the same number added to a small study’s already-large variance and a large study’s already-small variance, it shrinks a smaller fraction of the small study’s variance proportionally — the net effect is that weights move toward equality across studies as τ² grows relative to the within-study variances. When τ² is genuinely zero (no detectable heterogeneity), the random-effects model’s weights and result converge exactly to the fixed-effect model’s; the two models only diverge once heterogeneity is present, and diverge more as it increases.
Choosing a model: the decision isn’t about which p-value you’d prefer
- Start from clinical and methodological similarity, not from Q or I² alone. The Cochrane Handbook is explicit that I² interpretation bands are not mechanical cutoffs and can be unreliable with few included studies — a low I² from three studies is weak evidence of true homogeneity, not proof of it.
- Default to random-effects unless there’s a specific reason to expect one shared true effect. Genuine replication studies (same protocol, same population, same measurement) are the main case where a fixed-effect model’s assumption is realistic. Most applied research syntheses — pooling studies run in different populations, settings, or with different intervention intensities — are not that case.
- Don’t pick a model after seeing which one gives a “better” result. The model choice should be pre-specified in the review protocol based on the expected similarity of the included studies, not chosen post hoc from among the pooled estimates it produces — picking whichever model narrows the CI or crosses a significance threshold is a form of outcome-driven analysis.
- If heterogeneity is severe enough that even a random-effects mean is hard to interpret, consider not pooling at all and reporting a narrative or structured synthesis instead. A single pooled number implies the studies are estimating a common (if variable) quantity worth averaging; sometimes they aren’t.
- Report a prediction interval alongside the random-effects confidence interval when the model is random-effects — the CI describes uncertainty in the mean effect, while the prediction interval describes the range a new study’s true effect could plausibly fall in, which is usually the more honest answer to “what should I expect from this intervention in a new setting?”
Frequently asked questions
Does a random-effects model always give a wider confidence interval than a fixed-effect model?
When there is real between-study heterogeneity (τ² > 0), yes — adding τ² to every study’s variance can only widen the pooled interval relative to the fixed-effect calculation. When τ² is exactly zero, the two models produce identical weights, estimates, and intervals; there is no case where random-effects produces a narrower interval than fixed-effect on the same data.
Can the pooled point estimate move in either direction, or does random-effects always push it a particular way?
Either direction. Because random-effects weighting moves every study’s weight toward equality, the pooled estimate shifts toward whatever the more equally-weighted average would be — which direction that pulls the estimate depends entirely on which studies were being under- or over-weighted under the fixed-effect model. In the worked example above it moved the estimate up; with a different mix of studies it can move it down.
Which model do systematic review tools default to?
Defaults vary by tool and have shifted over time as the field’s default recommendation has moved toward random-effects for most applied reviews; check the specific software’s current documentation rather than assuming a default, and always report which model and estimator (e.g. DerSimonian–Laird vs. REML) were used, per PRISMA 2020 reporting requirements.
Is a random-effects model the same thing as a mixed-effects (multilevel) model?
No. A meta-analysis random-effects model treats the true effect sizes across studies as a random sample from a distribution, for the specific purpose of pooling summary effect estimates. A mixed-effects regression model is a different tool used on individual-participant (not study-level summary) data to account for correlation within nested groups, such as repeated measurements on the same patient. Both use the word “random effects,” and that shared vocabulary is the single most common source of confusion between the two.
What if I only have two or three studies — can I still estimate τ² reliably?
Not well. The Cochrane Handbook cautions that I² and, by extension, τ² estimates are unstable with a small number of included studies. With very few studies, some methodologists recommend reporting both models’ results, being transparent about the instability, or avoiding an automated τ² estimator in favor of a documented sensitivity analysis rather than presenting a single point estimate of between-study variance as precise.
Related CASRAI resources
- Heterogeneity in Meta-Analysis: I², τ², and Prediction Intervals — the full mechanics of quantifying and investigating heterogeneity, which is the input this page’s model choice depends on.
- Systematic Review vs. Meta-Analysis — how pooling fits into the broader systematic review process.
- Meta-analysis — operational definition and worked examples.
- Network Meta-Analysis (NMA) — extending pairwise pooling to three or more interventions compared at once.
- Forest plot — the standard visualization for exactly this kind of study-by-study weight and pooled-estimate comparison.
- Effect size — what’s actually being pooled in either model.
- How to Peer Review a Systematic Review or Meta-Analysis — includes checking whether the pooling-model choice was pre-specified and justified.
