Fixed-effect vs random-effects meta-analysis
The two models answer different questions. Choosing between them by which one gives a narrower interval is choosing the answer before asking the question.
The choice is often presented as a technical preference. It is not. The two models estimate different quantities, and which one you want depends on what you believe about the studies.
What each model assumes
A fixed-effect model assumes every study is estimating the same single true effect, and that the differences between their results are sampling error alone. Under that assumption, larger studies deserve proportionally more weight, and the pooled estimate is the best estimate of that one true value.
A random-effects model assumes the true effect varies between studies — because populations, settings and implementations differ — and that the studies are a sample from a distribution of true effects. The pooled estimate is then the mean of that distribution.
What that changes in practice
- Weighting. Random effects gives comparatively more weight to smaller studies, because between-study variance is added to every study's variance. If small studies are biased, this makes the pooled estimate worse, not better.
- Interval width. Random-effects intervals are wider whenever heterogeneity is non-zero. That is the model being honest, not the model being conservative.
- Interpretation. A fixed-effect result says "this is the effect". A random-effects result says "this is the average effect across a set of settings that differ".
The wrong way to choose
Fitting both and reporting the one with the narrower interval. Choosing fixed effects because a heterogeneity test was not significant — that test has low power with few studies, so a non-significant result frequently means "we could not detect heterogeneity", not "there is none".
A defensible way to choose
Decide in the protocol, on the grounds of what you believe about the studies you expect to find.
For clinical questions where studies will differ in population, setting and delivery — which is most of them — a random-effects model is usually the honest default, because the assumption behind fixed effects is rarely credible in that context.
Where you genuinely expect one common effect, say why.
Report the prediction interval
A random-effects confidence interval describes uncertainty about the mean effect. It does not describe the range of effects you might see in a new setting.
The prediction interval does, and it is almost always considerably wider. For a clinician deciding whether an intervention will help their patients, it is the more relevant quantity, and it is routinely omitted.
When heterogeneity is substantial
Neither model is a fix. If the true effects vary enough that the mean is not a useful summary, the answer is to investigate the variation — subgroups specified in advance, meta-regression where there are enough studies — rather than to report a mean with a wide interval and move on.
Frequently asked questions
What is the difference between fixed-effect and random-effects meta-analysis?
A fixed-effect model assumes every study estimates the same single true effect and that differences between them are sampling error. A random-effects model assumes the true effect varies between studies and estimates the mean of that distribution.
Which model should I use?
Decide in the protocol, from what you believe about the studies you expect to find. For clinical questions where populations, settings and delivery differ — most of them — random effects is usually the honest default, because the fixed-effect assumption is rarely credible there.
What is a prediction interval in meta-analysis?
It describes the range of true effects you might see in a new setting, rather than uncertainty about the mean. It is almost always considerably wider than the confidence interval, and it is the more relevant quantity for a clinician deciding whether an intervention will help their patients.