Publication Bias in Antidepressant Trials: What the FDA Record Revealed

Publication bias can change a treatment's apparent effect without changing a single result inside a published paper. If favorable trials are published and unfavorable trials remain unavailable, the visible literature becomes a selected sample. Systematic reviews then pool what can be found and produce a precise answer to a distorted evidence set.

Erick Turner and colleagues demonstrated the mechanism by comparing US Food and Drug Administration reviews with journal publications for 74 antidepressant trials involving 12 drugs and 12,564 participants, and the FDA had judged 51% of the trials positive. Among the studies that appeared in journals, 94% were presented as positive. Published effect sizes were, on average, 32% larger than estimates derived from the FDA's complete trial set.

The missing denominator problem#

A conventional literature search begins with published reports. It cannot easily count trials that were completed but never published, because those trials leave no journal article to retrieve. This makes publication bias difficult to distinguish from ordinary variation.

Regulatory drug reviews offer a rare solution. Sponsors seeking approval submit a defined set of registration trials to the FDA, including trials with favorable and unfavorable results. Reviewers assess prespecified endpoints and document whether each study supports efficacy. That archive can be compared with what later reaches journals.

The regulator is not an infallible source, and a marketing application may not include every study ever conducted, though it is still much closer to a complete inception cohort than a set assembled by searching journal databases. Turner's design was powerful because the trial list was identified without regard to publication status or result.

What the 2008 comparison found#

The investigators studied 74 FDA-registered trials of 12 second-generation antidepressants approved between 1987 and 2004. According to FDA judgments, 38 trials were positive and 36 were negative or questionable.

Publication followed the direction of results. Nearly every positive trial was published, but many nonpositive trials were unpublished, while several that were published were written in a way that conveyed a positive outcome despite the FDA's less favorable judgment. Overall, 48 of the 51 published studies would look positive to you as a journal reader.

Twenty-three trials, 31% of the FDA trial set and 3,449 participants, did not appear as separate publications in the search used for the study, and the participants had contributed time and accepted research burdens, yet their results were largely absent from the ordinary evidence pathway.

The contrast is stark: 51% positive in the FDA record versus 94% positive among published studies. The publication filter did not merely change the count. It changed the estimated magnitude of efficacy.

How effect sizes changed#

Turner and colleagues calculated standardized mean differences using FDA data and published data. For every drug, the published estimate was equal to or larger than the estimate based on FDA reviews. Inflation ranged from 11% to 69% across individual drugs and averaged 32% across the class.

This does not mean each paper reported an incorrect calculation. Selection changes the distribution before meta-analysis begins: imagine ten fair trials around a small average effect, then imagine five weak or negative ones disappearing, and the remaining five can be pooled flawlessly and still overstate the truth.

Precision can make the distortion more persuasive. A meta-analysis may report narrow confidence intervals because the available studies are internally precise. Those intervals quantify sampling uncertainty within the selected set. They tell you nothing about the studies that never arrived.

Publication bias is broader than nonpublication#

Whole-study nonpublication is the clearest form, but several related processes can produce the same favorable literature:

• a trial is published but the prespecified primary outcome is replaced by a favorable secondary outcome;
• time points or subgroups are selected after results are known;
• a nonpositive trial is pooled with another study so its separate result is difficult to see;
• the abstract or conclusion uses positive language unsupported by the primary analysis;
• harms receive less complete reporting than benefits;
• multiple analyses are run and only the favorable specification appears.

These are usually grouped under reporting bias. Outcome-reporting bias occurs within a study. Publication bias concerns whether and how whole studies enter the visible literature. You usually need the protocol, the registry entry, the regulatory review, and the publication side by side to see the difference.

Why “negative” needs careful interpretation#

A negative trial is not proof that a drug has no effect. It may be too small, poorly executed, or unable to distinguish an effective treatment from placebo. In antidepressant trials, large placebo responses and measurement variability can reduce assay sensitivity. The FDA's positive, negative, or questionable judgment is also a regulatory classification, not a complete clinical interpretation.

Those nuances do not justify hiding the result. A failed or inconclusive trial is part of the uncertainty. Selectively publishing only trials that cross a significance threshold converts random variation and study limitations into a systematically favorable record.

The 2008 paper should not be used to claim that antidepressants are ineffective. When all FDA trials were included, the average effect still favored medication, but by less than the journal literature suggested. Questions about clinical importance, drug differences, patient variation, and harms require additional evidence.

Did transparency improve?#

A 2022 update examined 30 trials of four newer antidepressants approved by the FDA from 2008 through 2013. Fifteen were positive according to FDA reviews, and all 15 were transparently reported as positive. Of the 15 nonpositive trials, seven, or 47%, were transparently reported as nonpositive. Eight were unpublished or presented in a way that did not match the FDA result.

That 47% was an improvement over the 11% transparent reporting of nonpositive trials in the older set. It was still far from complete. Positive results remained more likely to enter the literature in a form consistent with FDA judgments. The update matters because transparency policy can improve behavior without eliminating selection, so you should neither assume the 2008 pattern is unchanged nor assume that registration solved it.

What trial registration and results reporting do#

Prospective registration creates a public record of a trial's existence, design, primary outcomes, and planned timing before results are known. Summary-results posting can make key data available even if no journal article is published. Linking registry records to publications helps reviewers identify missing studies and changed outcomes.

In the United States, FDAAA 801 and the Final Rule at 42 CFR Part 11 establish registration and results-submission requirements for certain applicable clinical trials; the NIH has a dissemination policy for NIH-funded clinical trials, and journal policies commonly require prospective registration. These frameworks overlap but do not cover every trial, and deadlines, certifications, extensions, and enforcement provisions are specific.

Legal obligation is not the same as complete compliance. In 2026 the FDA reported that 29.6% of studies it considered highly likely to be subject to mandatory results reporting lacked submitted results information and sent reminders to more than 2,200 responsible parties. That current enforcement action shows why you still have to audit the public record yourself.

Why funnel plots are not a cure#

A funnel plot compares study effect estimates with precision. Asymmetry can suggest missing small negative studies, but it also arises from true heterogeneity, methodological differences, outcome choice, or chance. With few studies, visual and statistical tests have low power.

Methods such as trim-and-fill or selection models rely on assumptions about why studies are missing. They can be useful sensitivity analyses, not machines that reconstruct unseen data: the strongest approach is to identify trials from a source created before outcomes were known, such as a registry, ethics cohort, funding cohort, or regulatory dossier.

A publication-bias audit for a systematic review#

1. Was the search limited to journal databases, or did it include registries and regulatory documents?
2. Were trials identified by protocol number, drug application, sponsor, and registry identifier?
3. Were registry outcomes and time points compared with the publication?
4. Are unpublished results included, and is their data source described?
5. Were conference abstracts, dissertations, and preprints assessed without assuming they are complete?
6. Is publication status related to effect direction or statistical significance?
7. Are funnel plots interpreted cautiously given the number and heterogeneity of studies?
8. Does the certainty rating address suspected reporting bias?
9. Are conclusions robust to plausible missing-study scenarios?
10. Is the review's search date recent enough for delayed publications and results postings?