Publication bias is the tendency for studies with statistically significant or positive results to be published, and to be published faster and cited more often, while studies with null or unfavourable findings are delayed, hidden in file drawers, or never written up at all. Because a meta-analysis can only pool the studies it can find, this selective availability distorts the evidence base: the visible literature leans toward larger and more flattering effects than the full set of conducted studies would support. The result is a pooled estimate that is too optimistic, and a synthesis that quietly misleads clinicians, policymakers, and patients who trust it.
Why publication bias inflates the pooled effect size
The damage happens during the search itself. When the studies that found nothing are systematically missing, the remaining sample is no longer a fair picture of the truth. A meta-analysis weights and averages the studies it locates, so if the absent studies are precisely those with small or negative effects, the weighted average drifts upward and the pooled effect size is biased away from the null. The problem is worst for small studies, which need a large observed effect to reach statistical significance; a small study that finds nothing is the most likely of all to vanish, while a small study that happens to find a big effect, often by chance, sails into print. This pattern is part of a broader phenomenon called small-study effects, where smaller studies report systematically different results from larger ones.
Confidence in a synthesis therefore depends not only on the studies you have, but on a reasoned judgement about the studies you may be missing. A pooled estimate built on a truncated literature can look precise and authoritative while pointing in the wrong direction. That is why every credible systematic review treats the detection of publication bias as a mandatory step rather than an optional extra, and why our overview of how a meta-analysis combines study results stresses comprehensive searching before any number is pooled.
How to detect publication bias
No method proves that bias exists, but several tools together build a defensible case. The starting point is almost always the funnel plot, a scatter of each study's effect estimate against a measure of its precision, usually the standard error on a reversed vertical axis. Large, precise studies sit near the top and cluster tightly around the pooled effect; small, imprecise studies spread out across the bottom. With no bias the cloud is a symmetric inverted funnel. A gap in one lower corner, where small studies with unwelcome results should appear but do not, is the classic signature of funnel plot asymmetry. The contrast between this precision-based diagram and the study-by-study summary is laid out in our guide to the difference between a forest plot and a funnel plot, and you can build the diagram itself with our funnel plot builder for asymmetry checks.
Because eyeballing a funnel is subjective, formal tests quantify the asymmetry. Egger's regression test regresses the standardised effect on its precision and asks whether the intercept differs from zero; a significant intercept signals asymmetry consistent with small-study effects. Begg's rank correlation test instead measures the rank correlation between effect size and variance, a non-parametric alternative that is less powerful but makes fewer assumptions. The trim-and-fill method goes a step further: it estimates how many studies appear to be missing from the sparse side, imputes mirror-image studies to restore symmetry, and recomputes the pooled estimate so you can see how much the result might shift if the suppressed studies were present. Finally, contour-enhanced funnel plots overlay regions of statistical significance onto the funnel, helping you judge whether a gap sits in a non-significant region (pointing to publication bias) or in a significant one (pointing to other causes).
The dangers of relying on a biased evidence base
When publication bias goes undetected, the consequences ripple outward. A treatment can look more effective or safer than it truly is, leading to clinical guidelines built on inflated benefits and understated harms. Resources are then directed toward interventions whose real effect is modest or absent, and patients may be exposed to risks that the published record conveniently downplayed. Even when researchers suspect bias, an overstated pooled estimate undermines the central promise of a systematic review, which is to give an unbiased summary of all the relevant evidence rather than a curated highlight reel.
Why asymmetry is not proof of publication bias
A funnel that looks lopsided is a clue, not a verdict, and treating it as proof is a serious error. Funnel plot asymmetry has several plausible causes beyond suppressed studies. True heterogeneity is the most important rival explanation: if smaller studies were conducted in higher-risk populations or used more intensive versions of an intervention, they will genuinely show larger effects, and the funnel will tilt for an honest reason. The amount of between-study variability is itself worth quantifying, which is why our explainer on measuring heterogeneity with the I-squared statistic belongs in the same analysis. Other causes include differences in study quality, where poorly conducted small studies exaggerate effects, and pure chance when only a handful of studies are available. For this reason most guidance discourages asymmetry tests when fewer than roughly ten studies are pooled, because the tests have too little power and too high a false-positive rate to be trustworthy.
Preventing publication bias before it starts
Detection methods only estimate the damage after the fact; prevention is far stronger. Prospective registration of every trial in a public registry, ideally before the first participant is enrolled, creates a permanent record so that a study cannot simply disappear when its results are unwelcome. At the review stage, a thorough search must reach beyond indexed journals: searching grey literature such as theses, conference abstracts, and unpublished reports, querying trial registries directly, and contacting authors for unpublished data all widen the net so that the missing studies are found rather than imputed. Reporting standards reinforce this discipline, with the PRISMA guidelines requiring authors to describe their search strategy in full and to assess the risk of bias across studies.
In practice, a defensible synthesis combines all of these moves: a comprehensive prospective search, a funnel plot, at least one formal asymmetry test when enough studies exist, a trim-and-fill sensitivity analysis, and a candid discussion of heterogeneity and study quality as alternative explanations. When you are ready to pool your extracted effect sizes and inspect how robust the summary is, our pooled effect size calculator lets you run the synthesis and stress-test it against the missing-study problem rather than assuming it away.