Publication Bias

A scientific record becomes systematically unrepresentative when the probability that a study, result, or outcome becomes publicly available depends on its direction, magnitude, statistical significance, novelty, or sponsor-favoredness.

Core Idea

Publication bias is systematic distortion of the accessible scientific record caused by result-dependent visibility. A population of studies, analyses, or measured outcomes exists before dissemination. Between that population and the literature encountered by readers lies a publication, reporting, indexing, or disclosure gate. When the probability of crossing that gate depends on direction, magnitude, statistical significance, novelty, or sponsor favorability, the visible record no longer represents the research that was conducted.

The abstraction is broader than completed null studies left in file drawers. It also includes preferential visibility of striking signs or magnitudes, sponsor-favored results, selected outcomes within studies, and indexing or reporting decisions that make some results discoverable and others practically absent. The downstream risk is the same: synthesis treats the visible sample as if it were the evidence population.

Structural Signature

• The conducted evidence population — studies, results, analyses, or outcomes that exist before public dissemination.
• The result-bearing property — sign, size, significance, novelty, or alignment that affects visibility.
• The visibility gate — publication, acceptance, reporting, indexing, or disclosure.
• Differential admission — access probability changes systematically with the result.
• The distorted public record — available evidence differs from the conducted population beyond random sampling.
• The downstream synthesis — readers, reviews, or meta-analyses risk treating the selected record as complete.
• Gate-directed prevention — registration and results-independent publication decisions.

What It Is Not

Publication bias is not established merely because small studies report larger effects or a funnel plot is asymmetric. Genuine heterogeneity, quality differences, analytic choices, and chance can create the same observable. Those patterns are diagnostic prompts, not a verdict about the visibility mechanism.

It is not p-hacking or HARKing. Those practices alter analysis or hypothesis construction before the visibility decision. They can interact with publication selection, but a cleanly analyzed null study can still disappear, and a p-hacked study can be published through a separate gate.

Scope of Application

• Journal and conference publication: significant, positive, or novel findings receive preferential acceptance or submission.
• Clinical research: trials or outcomes unfavorable to an intervention remain unpublished or incompletely reported.
• Meta-analysis: an accessible sample of studies is synthesized as if it represented all conducted studies.
• Within-study reporting: outcomes and analyses become visible selectively after results are known.
• Indexing and discoverability: published outputs differ in whether databases expose them to systematic search.
• Sponsor disclosure: commercial or institutional interests affect which completed evidence enters the public record.

Clarity

The abstraction separates research quality from record representativeness. Every visible study may be competently executed while the literature as a whole is biased because entry into visibility depended on result. It also separates “not published” from “not conducted,” a distinction essential to interpreting apparent replication or consensus.

Manages Complexity

The mechanism compresses a complicated research ecosystem into a selection model: define the conducted population, identify the result-bearing property, locate the visibility decision, estimate differential admission, and determine what downstream estimand is distorted. Interventions then sort cleanly. Prospective registration, registered reports, and mandatory result disclosure change the gate; trim-and-fill, selection models, and sensitivity analyses attempt to infer what an already selected record omits.

Abstract Reasoning

Publication bias is conditional sampling. If visibility V depends on estimate E, the distribution of E among visible studies differs from its distribution among conducted studies. Increasing the number of visible studies does not restore representativeness when the same admission rule persists. Precision can increase around the wrong selected distribution.

The mechanism can also combine with selection on noisy estimates. When a noisy effect estimate determines whether a result becomes visible, the visible effects can be exaggerated even if individual estimates were unbiased before the gate. Publication Bias names the scientific visibility setting; Selection on Noisy Estimates carries the broader conditional-error mechanism.

Knowledge Transfer

The underlying selection logic transfers from the prime Selection Bias, but the named abstraction remains tied to the institutions of research communication. Within that domain, the framework transfers across medicine, psychology, economics, ecology, and any field in which conducted results pass through selective dissemination.

Examples

Canonical

Twenty teams test a weak effect. Sampling noise gives two teams large positive estimates that cross a significance threshold; those studies are submitted and published. The other eighteen studies remain unavailable. A later reviewer sees two concordant positive reports rather than twenty noisy attempts and overestimates both effect size and evidential consistency.

Within-study

A trial measures ten outcomes. Only the two favorable outcomes appear in the article, while the trial itself remains visible. This is selective outcome reporting, a sibling route through the broader result-dependent visibility umbrella rather than the whole-study File Drawer species.

Structural Tensions

• Unobservability: the missing evidence is difficult to characterize precisely because it is missing.
• Diagnosis versus cause: asymmetry can raise suspicion without identifying a publication gate.
• Study-level versus outcome-level selection: the selected unit changes which corrective model is appropriate.
• Prevention versus reconstruction: prospective controls can change future visibility; retrospective adjustments depend on assumptions about unseen evidence.
• Novelty value versus representativeness: legitimate editorial interest in novelty can still distort cumulative inference.

Structural–Framed Character

Publication Bias is a domain-specific statistical and institutional frame over Selection Bias. The observed-sample distortion is structural, while studies, journals, indexing, registration, and literature synthesis are constitutive domain roles.

Structural Core vs. Domain Accent

Remove journals, studies, registries, and synthesis. What remains is differential entry into an observed sample based on a variable connected to the outcome, which is Selection Bias. The domain accent specifies the selected units as research outputs and the gate as public scientific visibility.

Instantiates / Related Primes

• Selection Bias is the strict genus.
• Selection on Noisy Estimates can be a constituent mechanism when a noisy estimated result determines visibility.
• File Drawer Problem is the whole-study null-suppression species.
• Small-Study Effects and Funnel Plot Asymmetry are diagnostic neighbors, not taxonomic children.

Relationships to Other Abstractions

Current abstraction Publication Bias Domain-specific

Parents (2) — more general patterns this builds on

• Publication Bias is a kind of Selection Bias Prime

  Publication bias is selection bias specialized to admission of scientific results into the visible literature.

• Publication Bias is part of, conditional Selection on Noisy Estimates Prime

  Significance-filtered publication bias contains selection on noisy estimates when visibility depends on a threshold-crossing estimated effect.

  Condition / exception Publication selection can instead depend on direction, novelty, sponsor preference, or whole-study availability without selecting on estimate extremeness.

Children (1) — more specific cases that build on this

• File Drawer Problem Domain-specific is a kind of Publication Bias

  Drawer Problem is the whole-study null-suppression species of the broader results-dependent publication and reporting filter.

Hierarchy paths (12) — routes to 6 parentless roots

• Publication Bias → Selection Bias → Bias

Show alternative paths (11)

Not to Be Confused With

• File Drawer Problem fixes the selected unit to an entire completed study and typically the missing result to null or non-significant findings.
• Selective outcome reporting hides parts of a study that otherwise remains visible.
• Small-Study Effects are a size-dependent empirical pattern with multiple possible causes.
• Funnel Plot Asymmetry is a diagnostic display, not the visibility process.
• P-hacking and HARKing alter analyses or hypotheses rather than directly deciding public availability.

References

<!– TODO: Claude re-authoring pass must add and verify authoritative meta-research and registered-report sources. –>

Notes

Created from workspace/mixed_dag_2026/missing_node_adjudications/publication_bias.yaml. Identity and hierarchy are adjudicated; final voice and citations remain editorial tasks.