Forest plots: how to read and interpret them

A forest plot displays every study in a meta-analysis on a single scale, with the pooled result beneath them. It is the standard way of presenting a meta-analysis because it shows the individual evidence and the summary at the same time, rather than replacing one with the other.

The name comes from the appearance of the plot: a row of vertical lines and squares resembling trees, against the vertical line of no effect. Its value is that it makes visible what a pooled number alone conceals — whether the studies agree, whether one dominates, and how precise each contribution is.

It is worth pausing on why the plot is laid out this way at all. A table of estimates and intervals contains identical information, and for a handful of studies is arguably easier to read precisely. What the plot adds is the ability to take in agreement and disagreement at a glance across twenty or thirty rows, where a table would require the reader to hold each interval in mind while comparing it against the others. That is the whole justification for the format, and it explains why the elements are chosen as they are — every one of them encodes something you would otherwise have to compute.

For each study, three things matter, and they are worth taking in order.

Applying this to the plot above: Ahmed 2021 sits at 0.30 with a short interval clear of the null — a precise, significant positive result. Dubois 2023 sits at 0.28, almost the same estimate, but its interval runs from −0.36 to 0.92 and crosses the null. Those two studies agree closely on the size of the effect; they differ entirely in how confidently they establish it.

Square area is proportional to the weight the study carries in the pooled estimate, and weight is determined by precision: 1 / variance. Since variance falls as sample size rises, larger studies get bigger squares.

The visual purpose of the squares is to stop your eye from treating all rows equally. Without them, five studies look like five equal votes; with them, it is immediately clear that one study accounts for over 40% of the result.

The diamond represents the pooled estimate. Its centre is the point estimate and its left and right points mark the confidence interval, so a narrow diamond means a precisely estimated pooled effect.

In the plot above the diamond sits at 0.39 with an interval of [0.24, 0.54], clear of the null line — a significant pooled effect of moderate size.

The diamond shape is not decorative. Because its widest point is at the centre and it tapers to the interval bounds, it draws the eye to the point estimate while still showing the range — and it is visually distinct from the study rows above, so nobody mistakes the summary for another study. Some plots additionally draw a vertical dashed line through the diamond's centre, extended up through the study rows, which makes it easy to see at a glance which individual studies sat above and below the pooled result.

Ordering matters more than it appears. Studies are most often listed alphabetically or by year, neither of which helps the reader. Sorting by effect size makes any gradient immediately visible; sorting by weight puts the influential studies at the top where they belong. Cochrane reviews conventionally use year order to show how the evidence accumulated, which is defensible — but whichever you choose, say so in the caption, because a reader who assumes one ordering and gets another will misread the pattern.

Which side is which

Always read the axis labels beneath the plot before interpreting direction. “Favours intervention” may be on the left or the right depending on how the outcome was coded, and for harmful outcomes such as mortality a lower value is the better result. Assuming right means good is a genuine and common source of error.

Before reading the statistics, look at how the intervals line up.

A useful habit is to look at the plot twice: once ignoring the diamond entirely, to form your own impression of what the studies show, and once with it. If your impression from the individual rows differs noticeably from the pooled result, that gap is worth understanding before citing the number. Usually it means one heavily weighted study is doing most of the work, or that a couple of imprecise studies pointing the other way are contributing almost nothing to the average despite occupying as much vertical space as the rest.

Statistics accompanying the plot usually appear beneath it as Q, I² and τ². Read them alongside the visual impression rather than instead of it: I² is a proportion and can be high simply because the studies were precise, so a plot of tightly estimated studies with small genuine differences may report a large I² that overstates the practical disagreement.

Judging by eye has a known weakness worth naming: the visual impression is driven by how wide the plotted axis happens to be. Software chooses that range to fit the most extreme interval, so a single very imprecise study stretches the axis and makes every other interval look narrow and tightly clustered. Two plots of the same data can therefore give quite different impressions of consistency. Check the axis values before trusting the visual impression.

Where an interval extends beyond the plotted axis, software draws an arrowhead rather than rescaling the whole chart around one imprecise study. That is the right visual choice, but it hides the true extent of that study's uncertainty. Always read the numeric column alongside the graphic, because a row with an arrow is telling you the study established very little.

A subgroup plot shows a diamond per subgroup plus an overall diamond, and is usually accompanied by a test for subgroup differences. Note that subgroup analyses are observational even within a randomised trial — the studies were not randomly assigned to subgroups — so a difference between them is a hypothesis, not a causal finding, and it must be pre-specified to carry much weight.

Leave-one-out plots deserve wider use. If removing a single study moves the pooled estimate substantially or changes its significance, the conclusion rests on that study and the review should say so plainly.

1. Treating every row as an equal vote

Counting how many studies were significant ignores the weights entirely. Look at the square sizes.

2. Reading a wide interval as a contradictory finding

An imprecise study that agrees on the estimate but crosses the null is not evidence against the effect.

3. Assuming right means favourable

Read the axis labels. For outcomes such as mortality or relapse, a result on the left is the better one.

4. Ignoring the scale on ratio plots

Odds and risk ratios are plotted on a log axis with the null at 1. Distances are not linear.

5. Reporting the diamond without the heterogeneity

A pooled estimate from studies that disagree needs that disagreement reported alongside it.

6. Overlooking a dominant study

If one study carries 60% of the weight, the meta-analysis is largely reporting that study. Run a leave-one-out analysis and say so.

Blank rows and section headings earn their space in longer plots. Grouping studies by design, setting or risk of bias with a labelled heading between the groups costs nothing and reveals structure that would otherwise be buried in an undifferentiated list of thirty rows. Most meta-analysis packages support this directly.

One practical caution on producing your own: the plot must be generated from the same analysis that produced your reported numbers, not assembled separately. Hand-drawn or spreadsheet-built plots drift out of step with the analysis as soon as a study is added or a model changed, and a figure that disagrees with the text is the kind of inconsistency reviewers notice immediately.

Standard software includes the metafor and meta packages in R, the metan suite in Stata, and RevMan for Cochrane reviews. All produce publication-standard plots directly; drawing one by hand in a graphics program is both laborious and prone to introducing errors of scale.