How to make a funnel plot
- Enter one row per study in the same format as the forest plot generator: effect + 95% CI, events and totals per arm, or means, SDs and sample sizes. Paste from Excel, drop an .xlsx or CSV file, or reuse your forest plot data.
- Choose the effect measure and summary model. Ratios are plotted on a log scale. The random-effects summary uses the DerSimonian–Laird estimator, as in the forest plot.
- Read the tests above the figure: Egger’s test, Begg’s test and trim and fill. With fewer than 10 studies they are shown with a warning.
- Choose what to draw: significance contours, pseudo 95% limits, imputed studies, Egger’s line.
- Apply your journal’s rules and export a vector PDF or SVG, or a PNG/TIFF up to 1200 dpi. Copy the methods text and the metafor script for your supplement.
Reading a funnel plot
- Top of the funnel: large, precise studies close to the summary estimate. Bottom: small studies, which scatter more widely by chance.
- Symmetric funnel: no sign of small-study effects. Missing corner: small studies with unfavourable or non-significant results may be missing.
- Asymmetry is not proof of publication bias. It can also come from heterogeneity, poorer methods in small studies, or chance (Sterne et al., 2011). The contours help tell these apart.
- Points far outside the pseudo 95% limits on both sides point to heterogeneity rather than bias, as in the BCG vaccine example.
Data formats
The funnel plot takes exactly the same tables as the forest plot, so you can move between the two tools without reformatting. Subgroup columns are ignored.
| Study | OR | Lower | Upper |
|---|---|---|---|
| Garfinkel 1981 | 1.18 | 0.90 | 1.54 |
Frequently asked questions
What does a funnel plot show?
Each study is plotted by its effect estimate (horizontal axis) against its standard error (vertical axis, with the most precise studies at the top). Without small-study effects the points form a symmetric inverted funnel around the summary estimate, because small studies scatter more widely. A gap on one side at the bottom, where small studies with unfavourable results would be, suggests small-study effects such as publication bias.
How many studies do I need?
The plot can be drawn from 3 studies, but tests for funnel plot asymmetry are only recommended with at least 10 studies, and not when the studies are of similar size (Sterne et al., BMJ 2011; Cochrane Handbook chapter 13). With fewer studies the tests have too little power to rule asymmetry out. CleanFigure shows a warning when there are fewer than 10 studies.
Which Egger’s test does CleanFigure use?
The classical test of Egger et al. (BMJ 1997): a weighted linear regression of the effect estimates on their standard errors, with weights equal to the inverse variances, which is equivalent to regressing the standard normal deviate on precision. The P value is from a t test of the slope with k − 2 degrees of freedom. It matches regtest(res, model = "lm") in the metafor R package and metabias in Stata.
What is the difference between Egger’s and Begg’s test?
Both test for an association between effect size and precision. Egger’s test uses a regression and is usually more powerful. Begg and Mazumdar’s test uses the rank correlation (Kendall’s tau) between the standardised effects and their variances; it is less powerful with few studies. CleanFigure reports both, calculated as in metafor (regtest and ranktest).
How does trim and fill work, and should I trust the adjusted estimate?
Trim and fill (Duval and Tweedie, 2000) estimates how many studies are missing from the sparse side of the funnel (L0 estimator), removes the most extreme studies on the other side, re-estimates the centre, and then adds mirror images of the removed studies. The adjusted estimate is a sensitivity analysis, not a corrected result: it assumes asymmetry is caused by publication bias, which may not be true.
What does the contour-enhanced funnel plot add?
The shaded regions show where a study would be statistically significant (P < 0.01, 0.01 to 0.05, 0.05 to 0.10) or not (white). If the missing studies lie in the white region, publication bias is a plausible explanation; if they lie in shaded regions, other causes of asymmetry, such as differences in study quality, are more likely (Peters et al., 2008).
Can I use the data from my forest plot?
Yes. The funnel plot uses the same three formats as the forest plot generator (effect + 95% CI, events and totals, means and SDs). If you made a forest plot in this browser, click “Use the data from my forest plot”. Nothing is uploaded: both tools keep the data in your browser.