Forest plots and meta-analysis: a practical guide
A forest plot shows each study’s effect size with its confidence interval and a diamond for the pooled estimate, so you can see in one figure whether the studies agree and how big the combined effect is. This guide covers the data you need, fixed versus random effects, heterogeneity and small-study effects, and how to draw the plot in LabTools for free.
In this guide
- What a forest plot shows
- The data you need
- Fixed effect or random effects
- Make a forest plot in LabTools
- Small-study effects and publication bias
- What to report
- Frequently asked questions
Eight studies, odds ratios with 95% CIs; the pooled fixed-effect odds ratio is 0.74 (95% CI 0.68–0.80). See the full example
What a forest plot shows
Each row is one study: a marker at its effect estimate and a horizontal line for its confidence interval. A vertical no-effect line sits at 1 for ratios (odds ratio, risk ratio, hazard ratio) or 0 for differences (mean difference, standardised mean difference, risk difference). At the bottom, a diamond shows the pooled estimate: its centre is the combined effect and its width the confidence interval.
Ratios are drawn on a log axis, because an odds ratio of 0.5 is as far from no effect as 2 is, and pooling is done on the log scale.
The data you need
One row per study with the effect size and the lower and upper bounds of its confidence interval, as reported in each paper. A weight column is optional: without one, the weights are calculated from the interval widths (inverse variance), so precise studies count more.
The forest plot example has eight studies with odds ratios and a downloadable CSV you can use to follow along.
Fixed effect or random effects
A fixed-effect model assumes all studies estimate one true effect and differ only by chance. A random-effects model assumes the true effect varies between studies, estimates that variance (τ²) and gives a wider, more cautious interval. When the studies differ in population, dose or design, random effects is usually the honest choice. Drawing both diamonds shows how much the choice matters.
Heterogeneity is summarised by Cochran’s Q (a test of whether the studies disagree more than chance allows), I² (the share of variation due to real differences, 0–100%) and τ² (the between-study variance itself).
Make a forest plot in LabTools
- In Charts & Plots click Upload File and load your table (CSV or Excel), or paste it with Paste Data.
- On the Chart chip choose Forest Plot.
- In Columns choose the Study/Label Column, Effect Size Column, CI Lower Bound Column and CI Upper Bound Column (and optionally a Weight Column). The study drop-down currently lists numeric columns only, so number your studies.
- In Statistical Analysis set the Effect Size Type (OR, RR and HR switch to a log axis with the line at 1), the Meta-Analysis Model (Fixed Effects, Random Effects or Both) and the Confidence Level.
- In Layout & Typography set the “Favours” labels and Sort Studies By; in Publication Features pick a Journal Style (APA, Nature, JAMA or Cochrane) and where the statistics table goes.
- Export (Ctrl+E) as PNG, SVG or PDF.
The plot pools random effects with DerSimonian–Laird. For REML, a prediction interval and Egger’s test, run labtools stats meta from the command line, or ask an AI assistant connected to LabTools over MCP.
Small-study effects and publication bias
If small studies report larger effects than big ones, the pooled estimate may be inflated, often because small null results go unpublished. A funnel plot shows this visually; a rank correlation between study size and effect, or Egger’s regression test, puts a number on it. In the example, Kendall’s τ between study size and log(OR) is 0.143 (p = 0.62): no sign of small-study effects, though eight studies give these tests little power.
What to report
- The effect measure and scale (for example odds ratio, pooled on the log scale).
- The model (fixed or random effects) and the τ² estimator.
- The pooled estimate with its confidence interval, and Q, I² and τ².
- Any small-study or publication-bias assessment, and how many studies it was based on.
Frequently asked questions
What is a forest plot used for?
To show the results of several studies on the same question side by side: each study’s effect size and confidence interval on one row, and the pooled (meta-analytic) estimate as a diamond at the bottom. It is the standard figure of a meta-analysis and of systematic reviews.
How do I read a forest plot?
Each row is a study: the marker is its estimate and the horizontal line its confidence interval. The vertical line is “no effect” (1 for ratios, 0 for differences); an interval that crosses it is not significant on its own. The diamond is the pooled estimate, and its width is its confidence interval.
How do I get the standard error from a confidence interval?
For a 95% CI on a difference: SE = (upper − lower) / (2 × 1.96). For a ratio (OR, RR, HR) do it on the log scale: SE = (ln(upper) − ln(lower)) / (2 × 1.96). LabTools accepts the interval directly and does this for you.
What does a high I² mean?
That most of the spread between study results reflects real differences between the studies (populations, doses, methods) rather than chance. Around 75% or more is usually called high. Report the random-effects estimate, look for the source of the heterogeneity, and consider a prediction interval, which shows where the effect of a new study would likely fall.
DerSimonian–Laird or REML?
Both estimate the between-study variance τ² for a random-effects model. DerSimonian–Laird is the classic default (used by RevMan); REML is often preferred, especially with few studies. The LabTools forest plot uses DerSimonian–Laird; labtools stats meta offers both, with a prediction interval and Egger’s test.