In January 2013, the US Food and Drug Administration told manufacturers of zolpidem, the sleeping pill best known as Ambien, to halve the recommended starting dose for women: from 10 milligrams to 5 for immediate-release products. The reason was simple. Women clear the drug more slowly, so more of it is still in the blood the next morning, when they get in a car and drive to work.
Zolpidem had been on the US market for about two decades by then.
That story is usually told as a cautionary tale about the bad old days when women were left out of research. It is better told as a warning about the present. Getting women into trials was the first fight, and it was largely won. The fight that matters now is getting anyone to look at what the data says about them.
How the exclusion happened
The gap was not an accident. In 1977, in the long shadow of thalidomide, the FDA issued guidance recommending that women of childbearing potential be excluded from early-phase drug trials. It was widely interpreted to cover almost any premenopausal woman, including those using contraception.
The intention was protective. The effect was that the first human data on many drugs came almost entirely from men, and dosing assumptions were built on those bodies.
The consequences surfaced later. In 2001, the US Government Accountability Office reviewed the ten prescription drugs withdrawn from the American market since 1997 and found that eight posed greater health risks for women than for men. Some of that was because women took them more often. But four of the eight caused more adverse events in women even though they were widely prescribed to both sexes.
The reforms worked, partly
By 1993 the policy had reversed. The NIH Revitalization Act wrote into federal law that women and minorities must be included in NIH-funded clinical research, and required that trials be designed to allow valid analysis of whether effects differ by sex. A few weeks later the FDA issued new guidance encouraging women’s inclusion from the earliest phases.
Enrollment improved. And that is where the conventional story usually stops, with a satisfying arc from exclusion to inclusion.
But look at the second half of that 1993 law: the part about analysis.
Inclusion without analysis is decoration
In 2018, Stacie Geller and colleagues published a study in Academic Medicine that checked NIH-funded randomized trials in 14 leading US journals in 2015 against the Revitalization Act’s requirements. Of the 107 trials assessed on this measure, only 26 percent reported at least one outcome by sex or included sex as a covariate in the analysis. The authors compared this with their earlier reviews from 2004 and 2009 and found no statistically significant improvement.
Sit with that. More than two decades after federal law required trials to be able to tell us whether a drug works differently in women, roughly three in four trials in that sample did not report whether it did.
This matters because enrolling women and then pooling their results with men’s does not answer the question anyone cares about. A trial can be 50 percent female and still tell a woman nothing about whether her dose is right, her side effects are different, or her benefit is smaller. Inclusion is the price of admission. Analysis is the product.
The incentive nobody names
Why does analysis lag so far behind enrollment? The honest answer is that sex-specific analysis is costly to the people who would have to do it.
- Statistical power. A trial powered to detect an overall effect is rarely powered to detect a difference between subgroups. Doing it properly means larger, longer, more expensive trials.
- Regulatory risk. A sponsor who finds a sex difference may face labelling changes, dose adjustments or narrower approvals. Not looking is, commercially, the safer choice.
- Journal norms. Subgroup analyses are treated with suspicion, often rightly, because they can produce false positives. That caution has hardened into a habit of not reporting them at all.
Each of those reasons is defensible on its own. Together they produce a system where nobody is obliged to find out, and so, mostly, nobody does.
The problem starts before humans
The bias begins even earlier, in the lab. Preclinical research has long relied mainly on male animals and cells, partly on the assumption that female hormonal cycles add noise. In 2016 the NIH began requiring grant applicants to account for sex as a biological variable in vertebrate animal and human studies, and to justify single-sex designs.
That was the right move. But it repeats the pattern of 1993: a requirement at the design stage, with far less scrutiny of whether results are actually analysed and reported by sex at the end.
What would change the numbers
If the bottleneck is analysis, the fixes are unglamorous and specific.
- Funders should audit outputs, not just enrollment tables. Grant renewal could depend on whether sex-disaggregated results were reported, even where they show no difference.
- Journals should require disaggregated data, at minimum in supplementary tables, so others can check. A null finding by sex is still a finding.
- Regulators should ask for it routinely, and treat “we did not look” as a gap in the dossier.
- Clinicians should ask the question. When a new drug is presented, “was it analysed by sex?” is a reasonable thing to raise at journal club.
A worked example: how a sex difference hides in plain sight
Consider a hypothetical trial of a new blood pressure drug. It enrolls 2,000 people, half of them women, and finds a clear benefit overall. The paper reports the headline result, a table of baseline characteristics showing a 50:50 sex split, and a line about adverse events. It passes peer review. The drug is approved.
Now suppose the drug actually works slightly less well in women and causes a particular side effect, say dizziness, somewhat more often, because women on average have lower body weight and clear the drug differently. Neither difference is large enough to change the overall result. Both would matter to an individual woman deciding whether to take it, or to her doctor choosing the dose.
In the published paper, none of this appears. The women were enrolled. Their data was collected. It was simply never split out. Years later, post-marketing reports of dizziness accumulate, somebody runs the analysis, and a label change follows. That is roughly the shape of the zolpidem story, and it is the shape the Geller data suggests is still common.
The point of the example is that nothing in the trial was hidden. The failure was a decision not to ask a question the data could have answered.
The strongest counterargument
The best case against routine sex-specific analysis comes from statisticians, and it deserves a hearing.
Subgroup analyses are notorious for producing false positives. If you split a trial into enough groups, you will find apparent differences by chance. The medical literature has examples of subgroup findings that looked important and turned out to be noise. A trial that is not powered to detect a sex difference may produce a misleading one, and a misleading sex difference could lead doctors to under-treat women with a drug that actually works for them. On this view, pooling is not negligence. It is discipline.
This argument partly holds. Underpowered subgroup analyses should not be used to change practice on their own. A single trial showing a sex difference should be treated as a signal to investigate, not a conclusion.
But the argument proves less than it seems to. First, the problem with subgroup analyses is over-interpretation, not reporting. Publishing sex-disaggregated results, clearly labelled as exploratory, lets later meta-analyses combine data across trials and reach adequate power. Not publishing them makes that impossible. Second, sex is not an arbitrary subgroup chosen after the fact. There are well-understood biological reasons, in body composition, hormones and drug metabolism, to expect differences for some drugs. Pre-specifying sex as a factor of interest is quite different from fishing through dozens of post hoc splits. Third, the alternative, assuming that results in a mostly pooled population apply equally to everyone, is itself a strong assumption that is rarely tested.
The discipline argument is really an argument for doing sex-specific analysis properly: pre-specified, adequately reported and combined across studies. It is not an argument for leaving it out.
Second-order effects
The gap in analysis has consequences beyond individual drugs.
Clinical guidelines inherit the blind spot. Guidelines are built from trial evidence. If trials do not report by sex, guidelines cannot make sex-specific recommendations, and the default becomes a single dose for everyone.
Devices and diagnostics are affected too. The same logic applies to medical devices, diagnostic thresholds and algorithms. A threshold calibrated on predominantly male data may perform differently in women. When training data or validation studies are not analysed by sex, those differences stay invisible.
The evidence gap compounds for other groups. Race, ethnicity, age and pregnancy face the same pattern of inclusion without analysis, and often worse. Geller’s study found an even lower rate of reporting by race or ethnicity. Pregnant and breastfeeding women remain largely excluded from trials, leaving clinicians to prescribe with little evidence.
Trust erodes. When sex differences emerge after approval, as with zolpidem, it feeds a perception that medicine was not designed with women in mind. That perception is not unreasonable given the history, and it affects willingness to participate in future research.
What this means for different actors
For funders, the most powerful lever is to connect money to outputs. Grant reports could be required to include sex-disaggregated results tables, even for null findings. Funders can also support meta-analyses that pool existing trial data by sex, which is often the only way to get adequate power.
For journal editors, reporting guidelines already exist. The Sex and Gender Equity in Research guidelines, known as SAGER, were developed to encourage consistent reporting of sex and gender in research. Editors can make adherence a condition of publication rather than a suggestion.
For regulators, the question is whether sex-specific data is routinely requested and scrutinised in submissions, and whether labels reflect what is known. Regulators can also require sponsors to plan for sex-specific analysis before trials begin.
For hospital pharmacy and therapeutics committees, there is a practical role: when adding a new drug to a formulary, ask whether sex-specific data exists and flag where it does not.
For clinicians, the habit worth building is scepticism about one-size dosing, especially for drugs with narrow therapeutic windows or known pharmacokinetic differences, and attention to post-marketing safety communications.
What to watch next
Watch whether funders and journals move from encouraging sex-disaggregated reporting to requiring it. Watch for more meta-analyses that pool trial data by sex, which can answer questions that individual trials cannot. And watch for sex-specific findings in areas where the history of exclusion was longest, such as cardiovascular disease, where differences in presentation and treatment response are increasingly recognised.
The measure of progress is not how many women are in trials. It is how often a woman can find out whether the results apply to her.
The takeaway
The history of women in clinical research is usually told as a story of exclusion that has since been fixed. The zolpidem label change and the GAO’s withdrawn-drug findings belong to that story. But the Geller data points to a quieter failure: we included women, and then largely stopped asking what the results meant for them.
The next gains will not come from recruitment campaigns. They will come from making sex-specific analysis a default that someone is accountable for.
Have you seen a drug or device where a sex difference only emerged after approval? We’d like to hear about it in the comments.
