Publication of Clinical Trials Supporting Successful New Drug ApplicationsA Literature Analysis

The Food and Drug Administration approved those drugs based on evidence that you cannot read. It's not because it's classified in any dramatic sense, but because more than half of the clinical trials submitted to gain approval were simply never published. Let that sink in. Every clinician who looked up a drug on PubMed, every patient who asked their doctor what the studies show, and every policy maker who assumed that the medical literature tells the whole story were all working from an incomplete record. That's not a bug in one corner of the system; it's the baseline. Here's how the system is supposed to work. When a company or research institution wants a new drug approved, they file a new drug application with the FDA. That application must include everything: all animal and human trial data, complete protocols, data from failed trials, and manufacturing details. The FDA then assembles a team of clinicians, statisticians, chemists, and clinical pharmacologists to review the submission and confirm the sponsor's conclusions about safety and efficacy. It's a rigorous process. The problem is that it happens largely behind closed doors. The FDA does release a Summary Basis of Approval, but those documents present only selected results and can redact data protected under the Freedom of Information Act as commercially confidential. The drug label summarizes clinical studies too, but with even less detail. For the full picture—protocols, deviations, additional analyses, subgroup data—clinicians and patients have always depended on peer-reviewed journal publications. And that's exactly where the gap opens. Kirby Lee, Peter Bacchetti, and Ida Sim set out to measure that gap precisely. They built a cohort of every new molecular entity the FDA approved between January 1998 and December 2000, then went into the FDA review documents and drug labels to identify all the clinical trials those approvals relied on. They ended up with nine hundred and nine trials supporting ninety approved drugs. Then they searched the literature—PubMed, the Cochrane Library, CINAHL, and The Medical Letter—the same sources a typical clinician would use, through August 2006. That gave them a follow-up window of roughly five and a half to eight and a half years after approval. Only full journal articles counted; abstracts and review articles were excluded. The headline finding is stark. Of those nine hundred and nine trials, only forty-three percent, or three hundred and ninety-four of them, were ever matched to a full publication. More than half of the trials that convinced the FDA to approve new drugs never appeared in the medical literature. Now, not all trials are equal. The ones that actually mattered most for prescribers—the so-called pivotal trials, those described in the FDA-approved label's clinical studies sections, the ones demonstrating efficacy and safety most directly—those published at a higher rate: seventy-six percent, or two hundred and fifty-seven of three hundred and forty. Better, but still, one in four of the most important trials for any drug approved in this era was invisible to the clinician trying to understand what the evidence actually showed. For one of the ninety approved drugs, Lee and colleagues couldn't find a single published supporting trial anywhere. The next question is obvious: which trials got published, and which didn't? This is where the analysis gets uncomfortable. Lee and colleagues ran multivariable mixed-effects logistic regression—a statistical model that lets you isolate the independent contribution of each factor to the outcome—and what they found is a clear picture of selective reporting. Trials with statistically significant primary results had about three times the odds of being published compared to trials without significant results. The odds ratio was three point zero three. In raw numbers, sixty-six percent of trials with significant results were published, versus thirty-six percent of trials without. Larger trials also published more frequently—the odds of publication increased by about thirty-three percent for every doubling of sample size. And pivotal status was the strongest single predictor: pivotal trials had more than five times the odds of publication compared to non-pivotal ones. When Lee and colleagues restricted their analysis to just the three hundred and forty pivotal trials—the ones you'd most want to see—the bias didn't disappear. Statistically significant results still nearly tripled the odds of publication among that subset, with an odds ratio of two point nine six. Larger sample size still predicted publication. The bias runs even through the trials that matter most. Cox proportional hazards models examining time to publication confirmed the same pattern: significant results, larger samples, and pivotal status predicted not just whether a trial published, but how fast. Most published trials appeared quickly—ninety-two percent of all published supporting trials, and ninety-five percent of published pivotal trials, were in print within three years of FDA approval. So, the unpublished trials weren't just slow; they were gone. Consider what this means for the medical literature as an information system. The FDA saw nine hundred and nine trials. The literature showed three hundred and ninety-four of them—and those three hundred and ninety-four were systematically the ones with positive results and larger samples. That's not a random sample of the evidence; it's a curated highlight reel. A clinician reading the published literature for any drug approved in this era was getting a picture that overrepresented positive findings and underrepresented trials that didn't work out. Lee and colleagues stated it plainly: publication bias may lead to an inappropriately favorable record in the medical literature and may thus lead to preferential prescribing of newer and more expensive treatments. The distortion isn't just academic; it shapes which drugs get prescribed and at what scale. The authors are careful about what they can and can't claim. They acknowledge they didn't search EMBASE, didn't contact investigators directly, and may have missed some matches. They note that determining statistical significance wasn't possible for every trial in the dataset. And critically, their cohort covers only drugs that were approved—they have no window into trials supporting applications that failed, or trials that were registered but never submitted to the FDA at all. The baseline they've measured is probably the best-case version of the problem. Lee and colleagues also point to the policy tools available to close the gap. The FDA Amendments Act of 2007, passed after this study's data collection but during its publication, now requires that trials be registered at inception and that basic results be posted publicly, typically within one year after trial completion or approval. Those posted results include participant demographics, dropout numbers, and the numeric outcomes for all primary and secondary endpoints declared at registration. That's real progress. However, the authors are cautious: basic results reporting doesn't substitute for full publications. The detailed information clinicians need—protocols, deviations, subgroup analyses, conflicts of interest—will for the foreseeable future primarily live in journals. They also endorse broader adoption of Consolidated Standards of Reporting Trials, or CONSORT, guidelines to improve the completeness of what does get published, and they flag a possible unintended consequence of mandatory results posting: if sponsors believe that basic results satisfy the disclosure obligation, they may feel less compelled to submit full manuscripts. The law could paradoxically concentrate the remaining publication bias in exactly the trials that most need scrutiny. The deeper point is that this is a policy choice, not an inevitability. The trials exist. The data exists. The FDA has seen it. The decision about whether the rest of us see it is made downstream by sponsors, journals, and regulators, and those decisions have been consistently skewed in one direction. Every informed consent conversation, every prescribing guideline, and every formulary decision rests on the assumption that the relevant evidence is visible. Lee, Bacchetti, and Sim put a number on how far that assumption was from true: forty-three percent. That's the share of the evidentiary foundation for a cohort of approved drugs that the public could actually see. Knowing that number is the first step toward demanding a better one. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.