What does a funnel plot show?

A funnel plot displays each study by its effect estimate against its precision. Precise studies sit near the top and imprecise ones spread out below. A symmetric inverted funnel suggests no small-study effects, while asymmetry can indicate publication bias or other causes.

What does an asymmetric funnel plot mean?

Asymmetry suggests small-study effects, often from publication bias where small non-significant studies were never published. But it can also come from genuine heterogeneity, differences in study quality, or chance, so asymmetry is a prompt to investigate rather than proof of bias.

How many studies do you need for a funnel plot?

Formal asymmetry tests such as Egger's regression are unreliable with fewer than about ten studies. With a small number of studies, the plot may be uninformative, and it is more honest to report that asymmetry could not be reliably assessed.

How do you test funnel plot asymmetry?

Egger's regression test gives a p-value for whether smaller studies report systematically different effects, and Begg's rank-correlation test is a lower-power alternative. For binary outcomes the Harbord and Peters tests correct a false-positive tendency in the original Egger's test. Related methods such as trim-and-fill estimate how the pooled result would shift if missing studies were filled in. These tests supplement, but do not replace, careful interpretation of the plot.

What is a contour-enhanced funnel plot?

A contour-enhanced funnel plot overlays the regions of statistical significance onto the usual axes. If the studies appear to be missing from a non-significant region, that is the signature of publication bias, because null results went unpublished. If they are missing from a region that would have been significant, the asymmetry is more likely caused by heterogeneity or other factors rather than suppression.

What should be on the axes of a funnel plot?

The horizontal axis carries the effect estimate, shown on a logarithmic scale for ratio measures so the plot is not artificially skewed. The vertical axis should be the standard error, with the most precise studies at the top, because that makes the expected funnel boundaries straight lines and a missing corner easier to judge than on a sample-size axis.

Funnel Plot Interpretation: Spotting Bias

Funnel plot interpretation is a way of checking a meta-analysis for small-study effects and possible publication bias. Each study is plotted with its effect estimate on the horizontal axis and its precision on the vertical axis, so precise large studies sit near the top and imprecise small studies spread out below. If no bias is present, the points form a symmetric inverted funnel around the pooled effect; a lopsided funnel is a warning that something is missing from the evidence base.

Why the shape carries meaning

Precise studies estimate the true effect tightly, so they cluster near the apex. Imprecise studies scatter widely by chance, so without bias they fan out evenly on both sides, producing the symmetric cone. The logic of the plot is that scatter should be random. When the bottom-left or bottom-right corner is conspicuously empty, the most common explanation is that small studies with unwelcome results, often small and non-significant ones, were never published or never found. That gap is the visual signature of publication bias.

A funnel plot: precise studies cluster near the top around the pooled effect; an asymmetric base, with one corner missing, suggests small-study effects.

Reading symmetry and what breaks it

A symmetric funnel is reassuring but not proof of unbiased evidence, and an asymmetric one does not automatically mean studies were suppressed. Asymmetry has several innocent causes worth ruling out first: heterogeneity, where true effects genuinely differ by study size; differences in methodological quality between small and large studies; and simple chance when only a few studies are plotted. Treating any lopsided funnel as automatic proof of bias is a classic over-reading, which is why the plot is a prompt for investigation, not a verdict.

Choosing the axes correctly

How you build the funnel changes what it shows. The horizontal axis carries the effect estimate, and for ratio measures such as the odds ratio or risk ratio it should be on a logarithmic scale, because the sampling distribution of a log ratio is roughly symmetric while the raw ratio is not; plotting raw ratios manufactures asymmetry that is not really there. The vertical axis should be the standard error rather than the sample size or its inverse, with the most precise studies at the top. Standard error is the recommended choice because it makes the expected funnel boundaries straight lines, so a missing corner is easier to judge by eye than on a sample-size axis where the shape curves.

Contour-enhanced funnel plots

A plain funnel cannot separate “studies are missing because they were null” from “studies are missing for some other reason”. A contour-enhanced funnel plot overlays the regions of statistical significance, typically the p less than 0.05 and p less than 0.01 bands, on the same axes. If the gap in the funnel falls in an area of non-significance, that is the signature of publication bias, because it is precisely the unwelcome null results that went unpublished. If instead studies are missing from a region that would have been significant, the asymmetry is more likely driven by something other than suppression, such as genuine heterogeneity. This single addition turns the plot from a vague prompt into a far more specific diagnostic.

Putting numbers on the asymmetry

Because eyeballing symmetry is subjective, you back it with a formal test. Egger’s regression test regresses the standardised effect on precision and reads the intercept: an intercept significantly different from zero flags asymmetry. Begg’s rank-correlation test is an alternative with lower power. For binary outcomes the Harbord and Peters tests correct a false-positive tendency that the original Egger’s test shows when the effect and its variance are mathematically linked. Related approaches such as trim-and-fill estimate how the pooled result would change if the missing studies were filled back in. You can run these checks with our publication bias calculator. A key caveat: these tests are unreliable with fewer than about ten studies, so a small review often cannot assess asymmetry at all, and saying so plainly is better than over-claiming.

The link to the wider analysis

A funnel plot never stands alone. It is read alongside the forest plot, because the same small studies that scatter in the funnel are the ones whose weighting depends on your fixed-effect versus random-effects choice. If asymmetry is present, the pooled estimate may overstate the effect, and that belongs in your discussion of publication bias and in the GRADE rating of certainty of evidence.

A step-by-step way to read one

A disciplined reading follows the same order every time, which stops the eye from jumping to a conclusion:

Count the studies. With fewer than about ten, note that asymmetry cannot be reliably assessed and stop there.
Check the axes are sensible: a log scale for ratio effects and standard error on the vertical.
Look for a missing corner at the wide base, where small studies should sit, rather than scatter near the precise apex.
Read a contour-enhanced version to see whether any gap falls in the non-significant region.
Confirm the visual impression with a formal test matched to the effect type.
Rule out heterogeneity and quality differences before attributing the pattern to suppressed studies.

Common misreadings to avoid

Three errors recur. The first is over-reading a small funnel: with a handful of studies, almost any shape can appear, and chance alone produces lopsided-looking plots. The second is declaring publication bias from asymmetry alone, skipping the innocent explanations and the contour check, when ordinary between-study heterogeneity fits the data just as well. The third is plotting raw ratios on a linear axis, which fabricates an asymmetry that disappears the moment the axis is logged. Each comes down to letting the picture make a claim the data cannot support.

Reporting it responsibly

State how many studies the funnel contains, whether a formal test was feasible, and what the result suggests without overreaching. If asymmetry is found, explore the innocent explanations through a subgroup analysis or meta-regression before concluding bias, and confirm the headline result survives a sensitivity analysis. Used this way, the funnel plot is a genuine safeguard rather than a decorative figure, and it fits naturally into the full workflow of how to do a meta-analysis. If you would rather have the small-study assessment built and reported properly, our statistics service runs it as part of the synthesis.