A prediction interval in a random-effects meta-analysis is the range within which the true effect of a new, future study is expected to fall. It answers a different question from the familiar confidence interval: where the confidence interval captures uncertainty about the average effect across studies, the prediction interval captures the full spread of true effects that the random-effects model assumes exists in the population of studies. Because it folds in the between-study variance, written as tau-squared, it tells you what a clinician or policymaker should actually expect the next time the intervention is applied in a fresh setting.
Confidence interval versus prediction interval
These two intervals are routinely confused, and the difference is not cosmetic. The confidence interval of the pooled estimate describes how precisely you have located the mean of the distribution of true effects. Add more studies, or larger ones, and that confidence interval shrinks toward a point, because the average is being estimated ever more sharply. The prediction interval does not shrink toward a point in the same way, because it is describing the dispersion of the true effects themselves, not the precision of their mean. Even with a thousand studies, if the intervention genuinely works differently in different populations, the prediction interval stays wide.
Put plainly: the confidence interval answers "where is the average effect?" and the prediction interval answers "where will the effect land in my study?" This distinction matters most when you are deciding whether the body of evidence supports acting in a new context; if the idea of the interval around the average is still hazy, the explainer on what a confidence interval actually represents lays the groundwork. For background on why random-effects models posit a distribution of effects in the first place, see the contrast between the two pooling models and their assumptions.
The intuition behind the formula
The prediction interval is built by widening the pooled estimate to account for two sources of variation at once. The first is the standard error of the mean effect, the same quantity that drives the confidence interval. The second is tau-squared, the estimated variance of the true effects between studies. The interval combines these by taking the square root of their sum, so that both the imprecision of the average and the genuine heterogeneity of effects are represented in a single range. When tau-squared is large, the second term dominates and the interval balloons outward.
The multiplier is not the familiar value from a normal distribution. The prediction interval uses a t distribution with k minus 2 degrees of freedom, where k is the number of studies. The subtraction of two reflects that you have estimated two quantities from the data, the mean effect and the between-study variance, before predicting a new value. With only a handful of studies that t multiplier is substantially larger than the normal value, which is the model's honest way of admitting that tau-squared itself is poorly estimated when the evidence base is thin.
Why it is usually much wider than the confidence interval
Because the prediction interval adds tau-squared on top of the squared standard error and then applies a larger t multiplier, it is almost always considerably wider than the corresponding confidence interval. In meta-analyses with appreciable heterogeneity it can be several times wider. This is not a flaw; it is the interval doing its job. A narrow confidence interval around a beneficial pooled effect can sit inside a prediction interval that still stretches into harm, and that gap is precisely the information a careful reader needs. The degree of heterogeneity driving this width is summarized by the I-squared statistic and its companions.
When a prediction interval is meaningful
A prediction interval is only worth reporting under specific conditions, and presenting one outside them misleads readers. Two requirements stand out:
- At least three studies. With k below three, the degrees of freedom for the t distribution collapse and the between-study variance cannot be estimated with any stability. A prediction interval from two studies is effectively meaningless and should not be shown.
- Genuine heterogeneity. If tau-squared is essentially zero, the studies are estimating one common effect and the prediction interval collapses toward the confidence interval, adding little. The interval earns its place when there is real variation in true effects to communicate.
Even when both conditions are met, treat a prediction interval estimated from a small number of studies with caution. The interval inherits all the instability of a tau-squared estimate built on sparse data, so a handful of studies can produce an interval that looks authoritative but rests on a fragile variance estimate. Report the number of studies alongside it so readers can judge for themselves.
How it changes the clinical interpretation
The most consequential moment arrives when the prediction interval crosses the line of no effect while the confidence interval does not. Imagine a pooled risk ratio whose confidence interval sits entirely below one, signaling benefit on average. If the prediction interval spans from clear benefit through to no effect or even harm, the honest conclusion is that the intervention reliably helps on average but may not help, and could plausibly harm, in some future settings. A reader who saw only the confidence interval would walk away overconfident.
This is why the prediction interval belongs on the plot itself, usually drawn as an extended bar beneath the pooled diamond. Learning to read that extra bar is part of reading the whole figure, which we cover in the guide to interpreting the complete display. Seeing the diamond and the prediction bar together stops a reader from mistaking a precise average for a dependable guarantee.
A random-effects concept only
The prediction interval is defined strictly within the random-effects framework, because it relies on the model positing a distribution of true effects governed by tau-squared. A fixed-effect model assumes a single common true effect shared by every study, so there is no distribution to predict from and a prediction interval is simply not defined. If you find yourself wanting to report one, that desire is a signal you should be working under a random-effects model in the first place. The deeper reasoning behind that choice runs through the logic of synthesizing studies quantitatively.
Why leading methodologists recommend reporting it
Higgins and colleagues have argued for years that a random-effects meta-analysis is incompletely reported when it shows only a pooled estimate and its confidence interval. Their case is that decision makers care about the effect in a new setting, and only the prediction interval speaks to that question directly. Reporting it routinely guards against the common error of reading a tight confidence interval as a promise of consistent benefit, and it forces the analyst to confront how much real heterogeneity the synthesis actually contains.
In practice this means computing the interval whenever you have three or more studies and non-trivial tau-squared, displaying it on the forest plot, and discussing what it implies for application. You can build the underlying random-effects synthesis directly with the live forest plot editor, which keeps the pooled estimate, heterogeneity statistics, and the prediction interval in view together. Treated this way, the prediction interval stops being an afterthought and becomes the part of the analysis that most honestly answers what the evidence means for the next study.