Prediction Intervals in Meta-Analysis: The Range a Confidence Interval Hides

In a random-effects meta-analysis, the confidence interval tells you how precisely the average effect across studies has been estimated. The prediction interval answers a different and more practical question: given that studies genuinely differ, what range of true effects might you expect in a new setting? When studies are heterogeneous the prediction interval is much wider than the confidence interval, and it often reaches past no effect even when the average looks clearly beneficial.

Two intervals, two questions

A random-effects meta-analysis usually reports a diamond on a forest plot and a confidence interval around it. That interval is easy to misread. It describes how precisely the average effect across the included studies has been pinned down. Gather more studies and it narrows, because the mean is better estimated.

But clinicians rarely want the average of past studies. They want to know what might happen in the next patient, the next clinic, the next setting. That is a different question, and the prediction interval is what answers it. It estimates the range in which the true effect is likely to fall for a new, comparable study.

Where the extra width comes from

The gap between the two intervals is heterogeneity. If every study were estimating the exact same true effect, there would be nothing to predict, and the two intervals would nearly coincide. In reality studies differ, in their populations, their delivery, and their settings, and a random-effects model captures that spread with a between-study variance, written tau squared.

The prediction interval folds that spread back in. It widens the picture to include not just uncertainty about the average, but the genuine variation of true effects from one setting to another. The more the studies disagree, the wider it gets.

The finding that should change how you read forest plots

How much does this matter in practice? An analysis of heterogeneous meta-analyses drawn from the Cochrane database found that in roughly seventy percent of statistically significant results, the prediction interval reached past no effect. In other words, the average pointed to benefit, but the range of plausible true effects in a new setting included no benefit at all. In about a fifth of cases, the prediction interval even included effects in the opposite direction.

That is a sobering correction to a common reflex. A tight confidence interval and a significant pooled effect can sit directly on top of a prediction interval that admits the next study might show nothing.

When a prediction interval is not trustworthy

The prediction interval has an honest weakness. It depends on estimating the between-study variance, and that quantity is hard to pin down when there are only a few studies. With a small number of trials the interval can be so wide as to be uninformative, or unstable enough that it should be read with caution.

Methodologists therefore treat it as most meaningful when a reasonable number of studies contribute, and warn against leaning on it in small syntheses. Seeing a prediction interval is good practice, but its width still has to be read in light of how many studies produced it.

How to use it as a reader

When a meta-analysis reports a prediction interval, read it as the more realistic promise. The confidence interval tells you how well the past has been summarized; the prediction interval tells you what a new setting might actually deliver.

When the two diverge sharply, that divergence is itself the signal: the studies disagree, and a single average is a thin description of the evidence. When a review reports only the confidence interval and the studies are clearly heterogeneous, you are entitled to ask what the prediction interval would have shown, because that is often where the honest uncertainty lives.