A forest plot packs a whole meta-analysis onto one figure: every included study’s result, how much each one counts, and where the combined answer lands. Once you know what each element encodes, reading one takes under a minute. This guide walks through that process step by step, using a worked example, and covers the mistakes that trip up first-time readers. For the formal definition of the chart type itself, see the forest plot dictionary entry — this page assumes you already know what a forest plot is and focuses entirely on interpreting one.
The seven things to check, in order
Every forest plot, regardless of software (RevMan, R’s meta/metafor, Stata), encodes the same information in the same layout. Read it in this order rather than jumping straight to the diamond at the bottom.
- Find the effect measure and the line of no effect. The x-axis is an effect-size scale, and a vertical reference line marks "no difference between groups." For ratio measures — odds ratio (OR), risk ratio (RR), hazard ratio (HR) — that line sits at 1.0, because a ratio of 1 means the two groups performed identically. For difference measures — mean difference, risk difference — it sits at 0. Check which measure and which line you’re looking at before reading anything else; misreading the null value is the single most common error.
- Read one study row. Each row is one included study. The square (or diamond, circle — conventions vary by software) marks that study’s own point estimate on the x-axis. The horizontal line through it is that study’s confidence interval, almost always the 95% CI.
- Read the whiskers relative to the null line. If a study’s CI line crosses the line of no effect, that study’s result is not statistically significant on its own — its true effect could plausibly be zero (or 1.0, for ratios). If the entire CI sits on one side of the line, that study’s result is significant by itself, favoring whichever side it sits on.
- Read the square size, and the weight column. Square size is proportional to how much a study counts toward the pooled estimate — usually driven by sample size and precision (narrower CI = more weight). Most plots also print an exact percentage weight per study alongside the numeric estimate. A study can have a dramatic-looking point estimate and still barely move the pooled result if its weight is small.
- Scan down the whole column of point estimates before reading the diamond. Are the squares roughly lined up, or scattered across the axis with little overlap between confidence intervals? Scatter with little overlap is the visual signature of statistical heterogeneity — a genuine question about whether it’s valid to average these studies together at all, not just a footnote.
- Read the diamond. The diamond at the bottom is the pooled estimate: its horizontal center is the pooled point estimate, and its left and right tips are the pooled confidence interval — not a fourth study, a combined answer. If the diamond doesn’t touch the null line, the pooled result is statistically significant. Where the diamond sits relative to the individual study rows also tells you whether the pooled answer looks like a plausible average of what you just read, or an outlier relative to the individual studies.
- Check I² (and, if the plot includes one, the prediction interval). I² is normally printed near the diamond and estimates the percentage of total variation across studies attributable to real differences between them rather than chance. Per the Cochrane Handbook, treat I² interpretation bands as overlapping guidance, not a hard cutoff — use it alongside the visual scatter you already read in step 5, not instead of it. A prediction interval, when reported, is wider than the pooled CI and estimates the range a single new study’s true effect would plausibly fall in — it answers a different question than the pooled CI does.
Worked example
Illustrative example, not a real published study — numbers are constructed to demonstrate how the elements above combine.
Seven randomized trials compared a treatment to placebo on a binary outcome, pooled with a random-effects model and reported as risk ratios (RR). The line of no effect sits at RR = 1.0.
| Study | n | Risk ratio (95% CI) | Weight |
|---|---|---|---|
| Trial 1 | 142 | 0.58 (0.34–0.99) | 11.8% |
| Trial 2 | 96 | 0.71 (0.38–1.32) | 8.4% |
| Trial 3 | 310 | 0.66 (0.49–0.89) | 21.6% |
| Trial 4 | 88 | 1.05 (0.61–1.81) | 7.9% |
| Trial 5 | 204 | 0.74 (0.52–1.06) | 16.2% |
| Trial 6 | 151 | 0.60 (0.39–0.93) | 13.1% |
| Trial 7 | 267 | 0.69 (0.51–0.94) | 21.0% |
| Pooled (diamond) | 1,258 | 0.68 (0.58–0.81) | 100% |
Reading it row by row:
- Trial 4 is the one to notice first. Its point estimate (1.05) sits just past the null line on the "no benefit" side, and its CI (0.61–1.81) is wide and straddles 1.0 — not significant on its own, and it’s the only trial pointing the "wrong" direction. Before concluding anything about heterogeneity, check its weight: 7.9%, the second-smallest in the table. It’s a plausible chance result from a small trial, not evidence the whole synthesis is unreliable — though a full write-up would still investigate why, per the heterogeneity guide linked above.
- Trials 3 and 7 carry the most weight (21.6% and 21.0%), driven by their larger sample sizes and narrower CIs. Their estimates (0.66 and 0.69) sit close to the pooled estimate (0.68) — which is exactly what you’d expect, since heavily-weighted studies pull the pooled estimate toward themselves.
- Trials 1, 3, 6, and 7 are each individually significant — their entire CI sits below 1.0. Trials 2, 4, and 5 are not significant alone; their CIs cross 1.0.
- The diamond (0.68, 95% CI 0.58–0.81) does not cross RR = 1.0, so the pooled result is statistically significant in favor of treatment — a roughly 32% relative risk reduction — even though three of the seven individual trials weren’t significant by themselves. That gap is the entire point of pooling: individually underpowered trials can combine into a precise, significant answer.
- An I² reported alongside this plot in the 30–40% range would indicate mild-to-moderate heterogeneity, consistent with the mostly-overlapping (with one exception) pattern seen row to row — not enough on its own to abandon pooling, but worth a sentence of explanation in the results.
Common misreadings
- Treating an individually non-significant study as contradicting a significant pooled result. These answer different questions. A study with a wide CI crossing the null line is simply underpowered on its own — it is not evidence against the pooled effect, and several such studies can still pool into a precise, significant estimate (see Trials 2, 4, and 5 above).
- Reading square size as statistical significance. Square size encodes weight (how much the study counts toward the pooled estimate), not p-value or significance. A large, heavy square can still have a CI that crosses the null line.
- Treating I² as a strict cutoff. The commonly-quoted bands (roughly "low/moderate/substantial/considerable") are explicitly described by the Cochrane Handbook as overlapping ranges to be interpreted alongside the plot and the context of the analysis, not a threshold that flips a decision at, say, exactly 50%.
- Assuming the diamond is just another study. It’s a computed pooled result, not an eighth (or nth) data point — its width reflects the pooled CI, not any single study’s precision.
- Missing which side is "good." Direction of benefit depends entirely on how the outcome and groups were defined. Always check the plot’s own labeling (often printed under the x-axis, e.g. "Favours treatment" / "Favours control") rather than assuming left or right always means the same thing.
Reading plots for different effect measures
The reading process in the seven steps above doesn’t change across effect measures — only where the line of no effect sits and how to phrase the conclusion does:
| Effect measure | Line of no effect | "No difference" means |
|---|---|---|
| Odds ratio (OR), risk ratio (RR), hazard ratio (HR) | 1.0 | Equal odds/risk/hazard in both groups |
| Mean difference (MD), standardized mean difference (SMD) | 0 | Equal mean outcome in both groups |
| Risk difference (RD) | 0 | Equal absolute risk in both groups |
Ratio measures (OR/RR/HR) are typically plotted on a logarithmic x-axis, which is why the CI whiskers for ratio-measure studies often look visually asymmetric around the point estimate even though the underlying interval is symmetric on the log scale — that’s a plotting convention, not an error in the data.
Where you’ll encounter forest plots
Forest plots are the standard results figure for a meta-analysis, whether that meta-analysis stands alone or is embedded within a systematic review. The PRISMA 2020 reporting guideline exemplifies forest plots as the expected format for presenting synthesized results, and the Cochrane Handbook devotes its chapter on analysing data and undertaking meta-analyses to how they’re constructed. If you’re producing one rather than reading one, see the PRISMA and systematic review methodology guide for the reporting requirements, and the heterogeneity in meta-analysis guide for how to investigate and report the I²/τ² statistics that typically appear next to the plot.
Frequently asked questions
What does it mean when a forest plot’s diamond touches the line of no effect?
It means the pooled confidence interval includes the null value, so the pooled result is not statistically significant at the conventional threshold (typically 95% CI / α=0.05) — the combined evidence does not rule out "no difference," even if the point estimate itself favors one side.
Why do some studies get bigger squares than others?
Square size is scaled to each study’s weight in the pooled estimate, which is generally driven by sample size and the precision of its effect estimate (narrower confidence interval = more weight, all else equal). It has nothing to do with study quality or risk of bias unless the specific meta-analysis explicitly incorporates a quality weighting scheme.
Can a forest plot show a significant pooled result even if most individual studies weren’t significant?
Yes, and it’s common. Pooling combines statistical power across underpowered individual studies; several studies each too small to reach significance alone can still combine into a precise, significant pooled estimate, as in the worked example above.
What’s the difference between a forest plot and a funnel plot?
A forest plot shows each included study’s effect estimate and confidence interval plus the pooled diamond, and is used to interpret the meta-analysis’s actual results. A funnel plot plots each study’s effect estimate against its precision (e.g. standard error) and is used to visually assess publication bias or small-study effects — a different diagnostic question, not an alternative way of showing the same results.
Related CASRAI content
- Forest plot — the full definition and anatomy of the chart type.
- Heterogeneity in meta-analysis — how to quantify and investigate the scatter you read across a forest plot’s rows.
- Systematic review vs. meta-analysis — comparison page.
- PRISMA and systematic review methodology — reporting requirements for the studies that produce these plots.
- Confidence interval — the statistic behind every whisker on the plot.
- Effect size — what the x-axis position of each square actually represents.
- Publication bias — what a funnel plot, the forest plot’s companion chart, is used to detect.
