How to Read and Report a Forest Plot in a Meta-Analysis

Open almost any systematic review in a medical journal and the forest plot is where the eye lands. Editors use it to judge whether the pooled result is believable. Peer reviewers use it to check whether your text matches your figure. Clinicians use it to decide whether the finding applies to the patient in front of them. If you are writing a meta-analysis, the forest plot is the single figure most likely to be scrutinized, and also the one most often drawn or described badly.

This guide is for two readers. The first is the researcher or student who has been handed a forest plot and needs to understand what it says. The second is the author who has run the analysis and now has to draw the figure and write the results paragraph. The skills overlap more than you might expect, because good reporting is mostly a matter of telling the reader what you would want to be told if you were looking at someone else's plot.

What a Forest Plot Actually Shows

A forest plot has a simple structure. Each row is one study. The square in that row marks the study's effect estimate, and the horizontal line through it is the confidence interval, usually 95 percent. The size of the square reflects the weight the study carries in the pooled result, so larger squares come from larger or more precise studies. At the bottom sits a diamond. Its centre is the pooled estimate, and its left and right tips are the ends of the confidence interval around it.

Running down the middle is a vertical line of no effect. For ratio measures such as odds ratios, risk ratios and hazard ratios, that line sits at 1, and the horizontal axis is almost always on a logarithmic scale. For differences, such as mean difference or risk difference, the line sits at 0 and the axis is linear. Which side of the line favours treatment depends on the outcome, and this is where many readers go wrong. A hazard ratio below 1 favours treatment when the outcome is death, but a risk ratio below 1 would be bad news if the outcome were recovery. Most software prints labels such as "favours experimental" and "favours control" under the axis. If yours does not, add them by hand.

Beside the plot, a table usually lists the events and totals in each arm, the weight, and the numerical estimate with its interval. That table matters more than it looks. Anyone checking your work will want to compare it against the original papers, and a figure with squares but no numbers forces them to guess.

How to Read a Forest Plot, Step by Step

Start with the axis and labels, before you look at any square. Confirm which outcome is being measured, which effect measure is used, and which direction is good. This takes ten seconds and prevents the most embarrassing misreadings.

Next, look at the diamond. If it sits entirely on one side of the line of no effect, the pooled result is statistically significant at the chosen level. If it crosses the line, the data are compatible with no effect, but that is not the same as showing there is none. A diamond that crosses 1 with a wide interval often means the evidence is thin, not that the treatment fails. Then ask how large the effect is and whether that size matters clinically. A statistically significant pooled risk ratio of 0.97 with a tight interval can be real and still trivial for patients.

Then scan the individual studies. Do the squares cluster, or are they scattered? Does one large square dominate the weight? If a single trial contributes the majority of the weight, the pooled estimate is largely that trial in disguise, and it is worth asking what happens when it is removed. Do the confidence intervals overlap with each other and with the diamond? Intervals that barely touch are a visual warning of heterogeneity before any statistic confirms it.

Finally, read the heterogeneity statistics, which brings us to the part of the plot most often misunderstood.

Heterogeneity: What I², Tau² and the Q Test Tell You

Three statistics usually appear under the diamond. Cochran's Q test asks whether the variation between studies is larger than chance alone would produce. It has low power when there are few studies and can flag trivial differences when there are many, so a non-significant Q is weak reassurance. The I² statistic expresses the share of total variation attributable to between-study differences rather than sampling error. Tau² estimates the actual variance of the true effects, in the units of the effect measure.

The habit of reading I² against fixed cut-offs (low, moderate, high) is widespread, but the Cochrane Handbook itself treats such thresholds as rough guides that depend on the size and direction of effects and on the strength of evidence for heterogeneity. An I² of 50 percent from a handful of small trials carries a wide uncertainty range of its own, and many software packages will give you that interval if you ask. Report it. I² also says nothing about how much the true effects differ in clinical terms, only about the proportion of variability. Two meta-analyses can share an I² and have very different practical spread.

Fixed, Random and the Prediction Interval

A fixed-effect (or common-effect) model assumes every study estimates one underlying true effect and that differences are sampling noise. A random-effects model assumes the true effect varies from study to study and estimates its average. In clinical medicine, where populations, doses, settings and follow-up differ, the second assumption is usually more defensible, and it is the more common choice. What you should avoid is choosing the model after seeing which one gives the nicer p-value, or switching based on a significant Q test. State the model in your protocol and methods, and explain the reason.

Two details have shifted in recent years. First, the Cochrane Handbook's updated guidance on random-effects analysis now favours restricted maximum likelihood (REML) for estimating between-study variance, rather than the older DerSimonian and Laird method that many authors still default to. Second, for the confidence interval around the pooled estimate, Cochrane's methods group points to the Hartung-Knapp-Sidik-Jonkman adjustment when there are more than two studies and between-study variance is above zero, because the traditional Wald-type interval can be too narrow when the number of studies is small. If you are working in RevMan or the R package metafor, check which defaults your version applies and report them, since the choice can change your interval.

The more useful addition for readers is the prediction interval. The confidence interval around the diamond describes how precisely you have estimated the average effect. The prediction interval describes the range in which the true effect in a new, similar study might fall. When heterogeneity is present the two can tell very different stories. A 2016 analysis in BMJ Open of Cochrane meta-analyses found that the wider prediction interval would have altered the conclusion in a large share of statistically significant results with heterogeneity, which is why its authors recommended routine reporting. Authors can draw it as a separate bar or hollow diamond beneath the summary diamond, so that readers do not mistake it for the confidence interval.

A caution is warranted here. Prediction intervals computed from a small number of studies are themselves imprecise, and I would not lean on one from three or four trials. In that situation it is more honest to say so than to present a tidy interval that implies more certainty than the data hold.

How to Make a Forest Plot in R and Other Software

Most medical authors produce forest plots with one of four tools. RevMan, from Cochrane, is the natural choice for Cochrane-style intervention reviews and exports clean figures. Stata's meta suite and the older metan command are common in epidemiology. In R, the meta package (forest function) and metafor (forest function) are the usual routes for anyone who wants scripted, reproducible output, and the survival and clinical communities often work there. Review Manager aside, all of these will let you set the model, the estimator, and the interval method explicitly, which is exactly what you want.

The R route has one advantage beyond flexibility: you can share the script. Journals and reviewers increasingly look kindly on analyses that can be reproduced, and a short appendix with the code and the extracted data table answers many questions before they are asked. Whichever tool you use, resist the default styling. Check that the axis is on the right scale, that the labels name the direction of benefit, that the font is legible at the column width of the journal, and that the exported resolution meets the publisher's figure specification. Many journals ask for vector or high-resolution files, and a blurry screenshot of a plot is a quick way to earn a technical revision request.

Reporting the Forest Plot in Your Manuscript

PRISMA 2020 expects a review to present, for each study, the summary data and the effect estimate with its precision, ideally in structured tables or plots, and to report the results of each synthesis including measures of statistical heterogeneity. The methods side is covered by the items on how heterogeneity was identified and quantified, which the reporting guidance explicitly links to methods such as prediction intervals. Our guide to PRISMA 2020 reporting walks through those items in order, and it is worth reading alongside this one.

In the results text, do not simply say "see Figure 2." State the finding in a sentence that a reader could quote without opening the figure. Name the outcome, the number of studies and participants, the pooled estimate with its confidence interval, the model, and the heterogeneity. A workable pattern looks like this: pooled across eleven randomized trials (2,814 participants), the intervention reduced the risk of the outcome compared with control (risk ratio 0.78, 95% CI 0.66 to 0.92; random-effects, REML with Hartung-Knapp adjustment; I² = 41%, tau² = 0.02; 95% prediction interval 0.55 to 1.10). The numbers there are illustrative, not drawn from a real study, but the structure is the point.

Then interpret. If the prediction interval crosses 1 while the confidence interval does not, say plainly that the average effect favours treatment but a future setting could see little or no benefit. If the result depends on one trial, say that and show the leave-one-out analysis. Note the certainty of the evidence too; many journals now expect a GRADE assessment alongside the pooled estimate, and a precise-looking diamond built from high-risk-of-bias trials deserves a caveat.

Common Mistakes That Draw Reviewer Comments

The most frequent problem is pooling things that should not be pooled. Combining a randomized trial with an observational study, or mixing different doses and comparators in a single diamond, produces a number that answers no clinical question. The distinction between a systematic review and a meta-analysis matters here, and we cover it in our comparison of the two. A review can be sound and still decide that pooling is inappropriate.

Several smaller errors recur. Counting the same participants twice by including multiple arms against a shared control without adjustment is one. Reading a non-significant pooled result as evidence of no effect is another. Interpreting a funnel plot as proof of publication bias when there are fewer than about ten studies, a limit the Cochrane guidance suggests, is a third. Reporting subgroup differences without a formal test for interaction is a fourth, and one that reviewers catch quickly. Last, there is the quieter problem of data extraction errors, which are more common than most authors admit. Having two people extract data independently, and saying so, is cheap protection.

There is also a wider integrity concern worth keeping in mind. Meta-analyses are only as good as the trials they pool, and journals have grown more alert to the contamination of the literature by fabricated or unreliable studies. If a suspiciously clean trial drives your result, check it against retraction databases before the reviewers do.

A Practical Checklist Before You Submit

Before the manuscript leaves your desk, read the forest plot as a stranger would. Cover the text and see whether the figure alone tells you the outcome, the comparison, the direction of benefit, and the size of the effect. Then read the text with the figure hidden and see whether the paragraph stands up by itself. If either fails, fix it.

Check that the numbers in the figure, the table, the abstract and the results paragraph agree to the last decimal. Check that the model and estimator in the methods match what the software actually ran. Confirm that heterogeneity is described with Q, I² and tau², and a prediction interval where it is sensible. Make sure every pooled analysis you report was pre-specified or is clearly labelled as post hoc. Finally, read your target journal's instructions for authors on figure format and statistical reporting, since requirements for resolution, file type and supplementary data differ. Our broader note on statistical reporting in medical manuscripts covers the general rules that sit behind these choices.

The practical takeaway is simple. A forest plot earns trust when it is readable without effort and when the text beside it admits what the data cannot say. Report the interval as well as the average, describe heterogeneity as more than a percentage, and let the reader see where your conclusion is firm and where it is not.

Frequently Asked Questions

How do I interpret a forest plot if the diamond crosses the line?

It means the pooled interval includes no effect, so the result is not statistically significant at that level. It does not prove the treatment is ineffective. Look at the width of the interval to see whether the data rule out a clinically meaningful benefit or harm, or simply lack the precision to say.

What is a good I² value for a meta-analysis?

There is no universal cut-off. Lower values mean less inconsistency, but the same I² can reflect very different spreads of true effects. Report I² with tau², Q and ideally a prediction interval, and explain any substantial heterogeneity rather than quoting a threshold.

Should I use a fixed-effect or random-effects model?

Decide in advance, based on whether it is reasonable to believe the studies estimate one common effect. In most clinical reviews they do not, so random-effects is the usual choice. Do not switch models after seeing the results.

How many studies do I need to make a forest plot?

Two studies are enough to compute a pooled estimate, but very small meta-analyses give unstable heterogeneity estimates and wide intervals. If you have only a few studies, say so, favour cautious interpretation, and avoid funnel plots and prediction intervals that the data cannot support.

Further Reading