How Selective Publication Inflated Antidepressant Efficacy

Selective publication made a class of antidepressants look more effective than the complete evidence supported. When Erick Turner and colleagues compared the trials the FDA held on file against what actually reached medical journals, nearly all the published studies read as positive, while the FDA's full record showed only about half were. That filtering inflated the apparent effect size by roughly a third across the drug class. The distortion did not come from bad statistics inside any single trial. It came from which trials the literature quietly left out.

What the FDA record made visible

Before a new drug reaches the market, its sponsor submits every registration trial to the FDA, positive or not. That regulatory dossier is a rare thing in medicine: a near-complete, pre-specified set of studies assembled before anyone knew how the results would turn out. It offers a baseline the published literature cannot, because journals see only the trials that get submitted to them.

In their 2008 New England Journal of Medicine analysis, Turner and colleagues obtained FDA reviews for 74 registration trials covering 12 antidepressant agents and 12,564 patients. Using the FDA's own judgment of each trial's outcome, about half were positive. The published record told a very different story. Of the studies the FDA viewed as positive, essentially all appeared in print. Of those the FDA viewed as negative or questionable, most were either never published or, in a smaller set, written up in a way that presented a negative trial as though it were positive.

The arithmetic of that filtering is stark. Roughly 94 percent of the published literature conveyed a positive result, against about 51 percent in the FDA's complete set. A reader relying only on journals would conclude these drugs almost never failed. The regulator's files showed they failed about as often as they succeeded.

How missing trials inflate an effect

Publication bias is not only about counting positive and negative studies. It changes the size of the effect a meta-analysis will report. When Turner's team pooled the trials, the effect size derived from the published literature exceeded the effect size derived from the full FDA data for every one of the 12 drugs. The inflation ranged from 11 to 69 percent depending on the agent, and came to 32 percent for the class as a whole.

The mechanism is intuitive once you picture it. Negative and marginal trials tend to be the ones that vanish. Removing them lifts the average of what remains. So the surviving, visible evidence is not a random sample of what was studied; it is the flattering tail of the distribution. A clinician weighing benefit against side effects, or a guideline panel setting a treatment threshold, is working from a number that has been shifted upward by absence rather than by data.

None of this establishes that the drugs do not work. Antidepressants have real, measurable effects for many patients. The point is narrower and more unsettling: the magnitude of benefit that reached practitioners was systematically larger than the magnitude the complete evidence supported. When the true effect is modest, a one-third inflation can be the difference between a drug that looks clearly worth it and one whose value is genuinely arguable.

Turner and colleagues were careful about attribution. They could not determine whether the bias came from sponsors and authors declining to submit manuscripts, from editors and reviewers declining to publish them, or from both. The distortion is a property of the system, not proof of intent by any one party.

How registries and regulatory review correct the record

The fix is structural. If every trial is registered before it begins and its results are reported regardless of outcome, the negative studies can no longer quietly disappear. The 2007 FDA Amendments Act built much of that scaffolding in the United States, requiring registration of applicable trials on ClinicalTrials.gov and reporting of results, with the reach of those obligations later clarified so that reporting became due within set windows after a trial's completion, as the current ClinicalTrials.gov reporting requirements describe.

There is evidence the scaffolding is working. In a 2022 PLOS Medicine update, Turner and colleagues compared an older cohort of antidepressants against a newer one approved after registration norms took hold. Transparent reporting of negative trials rose sharply between the two eras, and the effect-size inflation attributable to selective publication shrank in the newer set. Reporting bias had diminished, though it had not disappeared, and positive trials still reached print more reliably than negative ones.

That residual gap is why the authors argued that nothing short of full transparency should count as acceptable, and why they pointed to designs like Registered Reports, where a journal accepts a study on the strength of its question and methods before the results exist. When the decision to publish is locked in before the outcome is known, the outcome can no longer decide whether the study is seen.

What this means for reading evidence

The practical lesson generalizes well past psychiatry. A meta-analysis is only as trustworthy as the completeness of the studies feeding it, and completeness is exactly what publication bias erodes. When you encounter a pooled estimate, ask whether the underlying trials were pre-registered, whether the negative ones are accounted for, and whether the analysis checked for signs of missing data. A regulatory review or a registry-based search can serve as a reality check against the polished version in the journals.

The Turner comparison endures because it quantified something the field had long suspected but rarely measured: the distance between what was studied and what was seen. Closing that distance is not a matter of trusting researchers more. It is a matter of building systems where the unflattering results cannot be left on the shelf.

This article is educational and not medical advice.