The literature is a selected record
A review can search carefully, summarize every paper it finds, and still miss important evidence. Some studies never become publicly available. Some papers omit a measured outcome. Others provide a conclusion without the numerical details needed for a synthesis. These are different forms of missing information.
The concern becomes bias when availability is related to the result. The Cochrane Handbook describes how missing results can distort a synthesis when the visible findings differ systematically from those that remain unavailable. Missingness alone does not prove the direction or size of that distortion, but it changes how complete the evidence can be assumed to be.
A fictional set of ten studies
Imagine ten equally sized studies evaluating the same question. Five produce estimates near no difference, three suggest a modest benefit, and two suggest a larger benefit. Suppose only the five studies suggesting benefit are published. A reader of the published literature sees consistent favorable evidence and has no obvious reason to count the missing five.
This example is deliberately simplified. Real studies differ in design, precision, and populations, and unpublished results are not necessarily null or negative. The example shows the mechanism: selecting which results become visible can change the apparent pattern without changing any individual study’s data.
The number of citations can then reinforce the impression. A review, commentary, and news article may all discuss the same five studies. More discussion does not create additional independent evidence or reveal the studies that were not available.
A paper can be published while an outcome disappears
Suppose a fictional trial measures pain, mobility, sleep, and adverse effects. Its paper emphasizes pain because that result was favorable, briefly mentions mobility, and omits the other two outcomes. The study exists in the literature, but the evidence needed to understand its full result is incomplete.
Another form of selectivity concerns analyses rather than outcomes. Researchers may have several reasonable ways to define a variable, handle missing data, or select a follow-up time. If only the most favorable analysis is presented without explaining the alternatives, the apparent certainty can exceed what the full analytical process supports.
Not every omission is motivated by the result. Space limits, data problems, and changes in a study can also affect reporting. Comparing protocols, registrations, and publications can help distinguish what was planned from what was reported, while leaving room for legitimate, transparently explained changes.
What registration can reveal
A dated protocol or registration can identify intended outcomes and analyses before the results were known. That makes discrepancies easier to see. Registration does not guarantee a good design, complete reporting, or an unbiased analysis. A vague plan may leave substantial flexibility, and an unavailable result remains unavailable even when its existence is known.
Registered Reports add another mechanism: a journal reviews the research plan before results are known and can offer in-principle acceptance subject to stated conditions. The Center for Open Science describes this format as a way to make publication less dependent on whether a result is exciting or statistically significant. It changes an incentive; it does not make research infallible.
Why a funnel plot is not a verdict
In a meta-analysis, a funnel plot can display study estimates against a measure of precision. Certain patterns may raise concern that smaller studies with some results are missing. But asymmetry can have other explanations, including differences in study methods or populations and chance.
The Cochrane guidance therefore treats such patterns as evidence to interpret in context. A plot with few studies may be especially difficult to read. A symmetrical plot is not a certificate that the literature is complete, and an asymmetrical one is not proof of misconduct or a specific number of missing studies.
What remains possible when evidence is incomplete
Reviewers can search registries and additional sources, seek unavailable results, compare publications with protocols, and examine how conclusions change under plausible assumptions. Those steps can improve an assessment without recovering everything. Sensitivity analyses describe dependence on assumptions; they do not reveal the exact values of unknown results.
For a reader, the useful question is whether the review recognized the possibility of missing evidence and explained how it affects confidence. A conclusion can remain supported while becoming less certain. Another may depend heavily on assumptions about unavailable findings.
Publication bias is therefore a reason to examine how the evidence became visible, not a reason to assume that every published result is wrong. The distinction helps preserve what research can tell us while being honest about the part of the record that cannot be seen.
