In short
A forest plot is the figure that summarises a meta-analysis. Each row is one study: the square marks its effect estimate, the horizontal line is its 95% confidence interval, and the size of the square is the weight that study carries in the pooled result. The vertical line marks the null value, 0 for differences such as Hedges' g and 1 for ratios such as an odds ratio or a relative risk; a study whose interval crosses it is not significant on its own. The diamond at the bottom is the pooled effect: its centre is the estimate and its width is the confidence interval. Lines that barely overlap each other are the visual signal of heterogeneity, which you then quantify with Q, I² and τ².
A forest plot is the chart that summarizes a meta-analysis. Each row is one study: the square marks its effect size, the horizontal line is its 95% confidence interval, and the size of the square shows how much that study weighs in the pooled result. The diamond at the bottom is the combined effect of all of them. This is what one looks like:
The rule that makes it readable in seconds is the vertical line of no effect. If a study's interval crosses that line, the study did not find a significant effect at the 5% level; if the interval stays clear of it, it did. The same goes for the diamond: here it runs from 0.14 to 0.46, clear of zero, so the pooled effect is positive and significant. Two things most readers skip: Chen 2021, the biggest square, carries 23% of that result on its own, and the I² = 54% at the foot is telling you the six studies do not really agree with each other.
If you want to generate your own without touching R, the forest plot generator draws it from your studies and computes the model, heterogeneity and pooled effect for you.
What is a forest plot?
A forest plot is the graphical representation of a meta-analysis: it shows the effect size of each included study with its 95% confidence interval and combines them into an overall estimate. It works with standardized mean differences (Cohen's d, Hedges' g), odds ratios, relative risks, correlations or any other effect measure, because the visual structure is identical whatever the metric. The name refers to the forest of confidence interval lines stacked one on top of another, which is also why you keep seeing it misspelled as "forrest plot".
Part by part:
- Square: the effect that study estimated. Its size is not decoration: it is the weight the study carries in the pooled estimate. Larger sample and less variability, larger square.
- Horizontal line: the 95% confidence interval of that estimate. A short line means the study pins the effect down; a long one means it barely narrows anything.
- Vertical line (line of no effect): the null value, 0 for mean differences, 1 for odds ratios and relative risks. Crossing it means non significant at the 5% level.
- Diamond: the pooled effect and its confidence interval. The center is the point estimate; the left and right tips are the confidence limits.
- Numeric columns: study and year on the left; effect, 95% CI and percentage weight on the right. At the foot you usually get the heterogeneity statistics (I², Q, τ²) and the model.
Read as a yes or no (does the diamond cross the line?), a forest plot gives you a fraction of what it holds. The rest of this guide is how an experienced reviewer reads the same chart: which studies are really driving the result, when heterogeneity makes the diamond uninformative, what changes between a fixed-effect and a random-effects model, and the five mistakes that keep turning up in peer review. If you want to put numbers on how much the studies disagree, the heterogeneity calculator (I², Q, τ²) does it from your own data.
How to read a forest plot in seven steps
When you open a forest plot, follow this order so you do not miss anything:
1. Identify the effect measure. Are you looking at standardized mean differences (d), odds ratios (OR), relative risks (RR), correlations (r)? The horizontal axis and the caption will tell you. This determines where the line of no effect sits (0 for differences, 1 for ratios).
2. Locate the line of no effect. Everything on one side favors one condition; everything on the other side favors the opposite. The labels at the bottom of the chart (for example, "Favors control" and "Favors treatment") orient you.
3. Go study by study. Does the square fall on one side or the other? Does its confidence interval cross the line? If it crosses, that individual study is not significant at the 5% level. If it does not, it is.
4. Look at the size of the squares. Big squares are the studies that dominate the pooled result. If the overall effect rests mostly on one or two studies with huge squares, the conclusion is sensitive to those specific studies. Check whether they have methodological limitations.
5. Read the diamond. Does it cross the line of no effect? If it does not, the pooled effect is statistically significant. The center is the point estimate and its width is the confidence interval: a narrow diamond means high precision; a wide one means the meta-analysis has not reduced uncertainty much. It is normally narrower than any single study's interval, precisely because it pools information from all of them; when it is not, heterogeneity is inflating it (under random effects, τ² widens the diamond).
6. Assess the visual spread. Do the squares cluster in the same zone or are they scattered? The more scattered they are, the higher the heterogeneity: the studies are not saying the same thing. In that case, look at I² (the percentage of variance attributable to real differences between studies rather than sampling error) and at Cochran's Q test.
7. Find I² and the model. At the foot of the chart (or in the text) you will usually see whether the model is fixed-effect or random-effects, along with the value of I². An I² above 50% is considered substantial heterogeneity; above 75%, high. If heterogeneity is high, the pooled effect may be an oversimplification and you should explore moderators (subgroup analysis or meta-regression).
Fixed-effect vs. random-effects: what changes in the forest plot
Something that goes unnoticed by many readers is that the same set of studies produces a visually different forest plot depending on whether you fit a fixed-effect or a random-effects model. The difference is more than cosmetic: it can change the conclusion.
In the fixed-effect model, the assumption is that every study estimates exactly the same true effect and that the differences between them are nothing but sampling noise. Weights are assigned purely by within-study precision (the larger the sample and the smaller the variance, the more weight). Large studies dominate and the diamond tends to be narrow.
In the random-effects model (the standard across health, social and behavioral sciences), each study is assumed to estimate a different true effect, all drawn from a distribution. On top of the within-study variance, the weights incorporate the between-study variance (τ²). This redistributes the weights: small studies get slightly more, large studies slightly less, and the squares become more similar in size. The diamond is usually wider because the model is more conservative once it accounts for that extra source of uncertainty.
The practical consequence: a fixed-effect meta-analysis can yield a significant effect that vanishes when you switch to random effects, because the diamond widens enough to cross the null line. If you see a fixed-effect forest plot in a context where studies come from heterogeneous populations, settings or interventions, be suspicious: the model is probably underestimating uncertainty. In practice, the vast majority of meta-analyses published in Q1 journals use random effects (REML estimation in metafor).
Heterogeneity: what the forest plot shows and what it hides
The forest plot gives you a first visual signal of heterogeneity: if the squares are scattered and the confidence intervals barely overlap, something is explaining why the studies disagree. But the visual check is only the first step. The quantitative indicators that accompany a forest plot are:
- Cochran's Q: a significance test for whether the observed variability between studies exceeds what you would expect by chance. If p < .05 there is significant heterogeneity, but with few studies the test has low power.
- I²: the percentage of total variability attributable to real differences between studies (not sampling error). Rough benchmarks: < 25% low, 25-50% moderate, 50-75% substantial, > 75% high (Higgins et al., 2003).
- τ² (tau squared): the between-study variance on the effect scale. It is the component added to the weights in the random-effects model.
For a deeper walkthrough of how to interpret I² without over-reading it, see the article on heterogeneity in meta-analysis.
When heterogeneity is high, the pooled effect (the diamond) is still mathematically correct, but it may not be substantively informative: saying that "on average the effect is d = 0.40" means little if some studies find 0.80 and others find −0.10. In that scenario the next step is to explore what explains the variation, either through subgroup analysis or meta-regression.
Forest plot with subgroup analysis
When a categorical moderator (for example, delivery format: in person vs. online) can explain part of the heterogeneity, meta-analyses present a subgroup forest plot. Studies are grouped under their category, each subgroup gets its own partial diamond, and the overall diamond appears at the bottom. It is the most visual way of showing whether the effect varies with a moderating variable.
To read a subgroup forest plot, apply the same seven steps within each subgroup and then compare the partial diamonds: if they overlap substantially, the moderator probably does not explain the heterogeneity. If they are clearly separated, the moderator does influence the effect. The test for subgroup differences (usually reported as Qbetween) tells you whether that separation is statistically reliable.
Five common mistakes when interpreting a forest plot
After reviewing dozens of meta-analyses as a statistical consultant, these are the mistakes I see over and over:
1. Judging significance without looking at magnitude. "The diamond does not cross the line, so the treatment works." No: a significant but trivial effect (d = 0.08 with N = 10,000) has no practical relevance. Always look at how big the effect is, not only whether it crosses the line, and anchor that magnitude with the usual effect size benchmarks.
2. Ignoring the size of the squares. If a positive overall effect rests on a single large study with an enormous square, the conclusion is fragile. All it takes is for that study to have limitations (biased sample, questionable measurement) for the whole meta-analysis to wobble. A leave-one-out sensitivity analysis exposes that fragility.
3. Confusing study-level significance with meta-analytic significance. An individual study crossing the null line does not cancel the evidence: the power of a meta-analysis lies precisely in pooling individually non-significant studies to obtain a more precise estimate. Do not dismiss a meta-analysis because some of its studies do not reach p < .05.
4. Assuming the diamond applies to every population when I² is high. An I² of 80% tells you the studies are finding very different effects. In that case the pooled effect is an "average" that may not represent any actual population. Explore the sources of heterogeneity before you quote that diamond.
5. Not checking publication bias. A forest plot does not show you the studies that were never published. If only positive results made it in (because the null ones stayed in the file drawer), the pooled effect is inflated. The funnel plot, Egger's test and the p-curve are the indispensable complements for assessing that risk.
Don't want to install R or a paid package for a one-off chart? The forest plot and funnel plot generator computes fixed and random effects, heterogeneity, and lets you download the chart as a high-resolution PNG with no watermark, right from your browser.
Checklist before you submit
Before accepting a meta-analytic conclusion (as a reader) or sending it to a journal (as an author), check that the forest plot includes, and that you have reviewed: the effect measure clearly labeled on the horizontal axis; the null line in the right position (0 for differences, 1 for ratios); every study with a square proportional to its weight and its 95% CI; the model stated (fixed effect or random effects); the diamond with the pooled effect and its 95% CI; the heterogeneity indicators (I², Q, p); a subgroup analysis if heterogeneity is substantial; and a publication bias check with a funnel plot.
Before you hit submit: if the forest plot is already inside your manuscript, run the draft through the AI paper reviewer. It is a free Reviewer 2 style pre-review that writes down what a real reviewer would object to: an unjustified model, high heterogeneity nobody explores, publication bias left unchecked, or a diamond interpreted as if it held for every population. Better you read it now than in the rejection letter.
Need a meta-analysis with reviewer-proof forest plots?
I am a PhD in psychology and I run complete meta-analyses: search, coding, random-effects model, forest plots, heterogeneity and publication bias. APA 7 write-up ready to submit.
Get a free assessment →Frequently asked questions
What does a forest plot show?
The result of every study included in a meta-analysis and the pooled result, on the same scale. Each row is a study, with its effect estimate (the square), its 95% confidence interval (the horizontal line) and its weight (the size of the square). The diamond at the bottom is the combined effect. It is the standard figure of any meta-analysis and, in APA 7, it is treated like any other figure: bold number, italic title, labelled axis and a self-explanatory note.
What does the diamond at the bottom of a forest plot mean?
It is the pooled effect of all the studies. The centre of the diamond is the point estimate and its width is the confidence interval: if it crosses the line of no effect, the pooled effect is not statistically significant. Some plots add a wider horizontal bar around it, the prediction interval, which is a different thing: the confidence interval describes the precision of the mean effect, while the prediction interval describes the range in which the true effect of a new study would be expected to fall.
Why are the squares in a forest plot different sizes?
Because they represent the weight of each study in the pooled estimate, which comes from the inverse of its variance. A study with a large sample and a small standard error gets a big square; a small, imprecise study gets a small one and a long line. Under a random-effects model the weighting also incorporates the between-study variance (τ²), which is why small studies get relatively more weight there than under a fixed-effect model.
What is the vertical line in a forest plot?
The line of no effect, the value at which there is nothing to see. It sits at 0 when the effect is a difference (mean difference, Cohen's d, Hedges' g) and at 1 when it is a ratio (odds ratio, relative risk, hazard ratio), which is also why ratio plots are usually drawn on a logarithmic scale, so that halving and doubling look symmetrical.
How do I know from a forest plot whether there is heterogeneity?
Look at how much the confidence intervals overlap: lines pointing in different directions with little overlap are the visual sign that the studies are not estimating the same effect. Then quantify it, because the eye is not enough: Cochran's Q with its p, I² as the proportion of variability attributable to heterogeneity rather than sampling error, and τ², which is the one expressed in the units of your effect. I² does not tell you how large the heterogeneity is, only what share of the variability it accounts for, so read it alongside τ².
Can I read a single study's row in a forest plot?
You can read that study's row, yes, but a study whose interval crosses the null line is not "negative": it is imprecise. That is exactly the point of meta-analysis, and it is the mistake most often made when reading these figures. Interpret each row as an estimate with its uncertainty, and reserve the conclusion for the diamond.
How to create a forest plot
In R, the metafor package (Viechtbauer, 2010) is the reference for meta-analytic analyses, and its forest() function produces high-quality and highly customizable forest plots. Three lines are enough for a publication-ready figure:
library(metafor)
# Fit a random-effects model (REML)
res <- rma(yi = effect, vi = variance, data = my_data)
# Draw the forest plot
forest(res, slab = my_data$study, header = TRUE,
xlab = "Standardized mean difference")If you start from raw means and standard deviations rather than precomputed effects, run escalc() first. With measure = "SMD" you get Hedges' g (corrected for small-sample bias); with "OR" or "RR", odds ratios or relative risks. The meta package is another simpler option that also generates forest plots suitable for publication. If you want the full process (from raw data to a figure exported at 300 dpi), follow the step-by-step tutorial on how to make a forest plot in R with metafor. In both cases, you can control aspects such as the order of studies (by year, by effect size, by weight), the type of model (fixed effects or random effects), the inclusion of subgroups, and the visual elements of the chart.
For those who do not use R, RevMan (the Cochrane Collaboration software) produces standard forest plots automatically when performing a meta-analysis, although with fewer customization options. JASP also offers the ability to perform meta-analyses with forest plots from its specific module, with an accessible interface that does not require programming: import your data as .csv (one column for the effect, one for its standard error or variance), pick Meta-Analysis in the menu, choose the model and tick the Forest Plot box. CMA (Comprehensive Meta-Analysis) is another paid option with a graphical interface. Regardless of the tool, what is essential is that the forest plot you present in your article is readable, includes all the necessary information (individual studies, weights, confidence intervals, combined effect, heterogeneity), and is accompanied by a substantive interpretation that goes beyond simply stating whether the effect is significant or not.
The forest plot does not exist in a vacuum: it is part of a family of charts that evaluate meta-analytic evidence. Once you can read it, the next step is checking publication bias with the funnel plot and Egger's test. To spot outlying studies, the Galbraith plot and the L'Abbé plot are useful companions, and if your question compares several treatments at once, network meta-analysis is where you go next. For the write-up itself, the guide on how to report results in APA 7 covers the formatting. If you are preparing a meta-analysis for your dissertation or for a Q1 journal submission and need support with the analyses or forest plot interpretation, my statistical consulting service guides you from systematic search to the final write-up. See how it works on the journal publication page.