How to Read (and Build) a Forest Plot

The anatomy of a forest plot

**Each horizontal line represents one study.** The square (or diamond, depending on software) marks that study's point estimate -- its measured effect size. The horizontal line through the square is the confidence interval, usually 95%. A wider line means more uncertainty, typically because the study had a smaller sample.

**The size of the square reflects the study's weight** in the pooled analysis. Larger studies, or studies with more precise estimates, get bigger squares and pull the pooled result more strongly toward their own findings.

**The vertical line down the middle is the "line of no effect."** For ratio measures like risk ratio or odds ratio, this line sits at 1. For difference measures like mean difference, it sits at 0. If a study's confidence interval crosses this line, that study's result alone did not reach statistical significance.

**The diamond at the bottom is the pooled estimate** -- the overall effect size across all included studies, combined according to your chosen model. Its width represents the confidence interval for that pooled result.

Fixed-effect vs. random-effects: the choice that changes everything

Before you can build a forest plot, you need to choose a model, and this decision should be based on your actual heterogeneity, not on habit.

A **fixed-effect model** assumes every included study is estimating the exact same true effect, and any differences between them are just sampling error. This assumption rarely holds in practice unless your included studies are unusually similar in population, intervention, and setting.

A **random-effects model** assumes the true effect varies across studies -- which is realistic for most real-world evidence synthesis -- and produces wider confidence intervals to reflect that added uncertainty.

Choosing fixed-effect when your studies are genuinely heterogeneous will make your pooled estimate look more precise than it actually is. This is one of the more common statistical missteps a methods reviewer flags.

Reading heterogeneity: I² and tau²

Most forest plots report an I² statistic below the pooled diamond -- the percentage of variation across studies attributable to real differences rather than chance. Rough conventional benchmarks: under 25% is low heterogeneity, 25-75% is moderate, above 75% is substantial. High heterogeneity doesn't necessarily mean you shouldn't pool your studies, but it does mean you should investigate why -- through subgroup analysis, meta-regression, or at minimum, a transparent discussion in your manuscript.

Common mistakes when building your own

**Pooling clinically dissimilar studies just because they measured the same outcome.** A forest plot can be built on any set of studies mechanically, but that doesn't mean the pooled estimate means anything if the interventions or populations differ substantially.

**Ignoring publication bias.** A funnel plot, alongside tests like Egger's regression, should accompany your forest plot when you have enough studies (generally 10 or more) to check whether smaller studies with null results are systematically missing from your analysis.

**Reporting the pooled estimate without reporting certainty.** A forest plot shows you the numbers, but a GRADE Summary of Findings table tells your reader how much confidence to place in them -- and journals increasingly expect both.

Getting it right

Building a defensible forest plot means the statistical decisions behind it -- model choice, heterogeneity handling, publication bias checks -- are made deliberately and documented, not defaulted to whatever your software's out-of-the-box setting happens to be. If you want a second opinion on your model selection or your heterogeneity results before you finalize your meta-analysis, that's core to the statistical consulting we do.

**[Get a quote](/systematic-reviews#meta-analysis-support)** for a second opinion on your meta-analysis.