Clinical Trial Diversity Action Plans: The Data and Systems Behind the Commitment

Where the Requirement Actually Stands

Start with the facts, because a lot of what circulates about this requirement is out of date in one direction or the other.

On December 29, 2022, the Consolidated Appropriations Act, 2023 (Public Law 117-328) was enacted. It included the Food and Drug Omnibus Reform Act of 2022, known as FDORA. Section 3601 of FDORA amended the Federal Food, Drug, and Cosmetic Act to require a Diversity Action Plan for three categories of study: a phase 3 investigation of a new drug (or another pivotal study other than a bioavailability or bioequivalence study), a device study requiring an investigational device exemption, and certain device studies that do not require one but that support a 510(k), a request for classification, or a premarket approval application.²

The plan itself has three required parts, and they are short enough to quote in substance: the sponsor’s goals for enrollment in the clinical study, the sponsor’s rationale for those goals, and an explanation of how the sponsor intends to meet them. FDORA further specifies that the goals must be disaggregated by the race, ethnicity, sex, and age group characteristics of the clinically relevant population.²

Section 3602 then directed FDA to issue guidance on the form and content of those plans. FDA published the draft, Diversity Action Plans to Improve Enrollment of Participants from Underrepresented Populations in Clinical Studies, in June 2024 under docket FDA-2021-D-0789. The comment period closed on September 26, 2024, and the statute gave FDA nine months from that close to publish final guidance, which put the deadline around late June 2025.¹³

What happened next, and why it matters operationally

In January 2025, following executive orders on diversity programs and gender ideology, FDA removed the draft guidance page from its website without notice.⁴ On February 11, 2025, a federal district court in Doctors for America v. Office of Personnel Management ordered HHS to restore removed health webpages, and the guidance page came back.⁵ The page carries that history on its face. Checked in August 2026, it still reads “Draft Guidance for Industry, June 2024” and “Draft. Not for implementation. Contains non-binding recommendations,” and it displays a notice that HHS is required by court order to restore the page to its January 29, 2025 version, with the caveat that information on the page may be modified or removed in the future subject to the terms of that order. The page footer shows content current as of July 25, 2025.¹

One further detail is worth stating precisely, because sponsors keep tripping over it. FDA’s Report to Congress footnotes the draft guidance document at a specific FDA media link. At the time of writing, that link does not resolve.² The landing page exists; the document it points to is not reliably retrievable from FDA. If your regulatory affairs team is working from a locally saved copy of the June 2024 draft, that is currently the more dependable source, and it is worth recording where and when you obtained it.

Sponsors did not wait

The most useful evidence about industry’s actual posture is in the same Report to Congress. FDA counted the voluntary diversity plans it received between October 1, 2022 and September 30, 2024, before any obligation existed. The Center for Drug Evaluation and Research received 124 plans in fiscal year 2023 and 161 in fiscal year 2024. The Center for Biologics Evaluation and Research received 9 and then 21. The Center for Devices and Radiological Health received 6 and then 24.²

Volumes roughly doubled in a year for two of the three centers, with no legal trigger. That is the practical answer to the question senior leaders keep asking, which is whether to pause this work until the regulatory picture clears. Sponsors are already writing these plans into investigational new drug applications and pre-submissions. Review divisions are already reading them. A plan that states a goal and cannot demonstrate how the sponsor will know whether it is being met is a weak document in that conversation regardless of what the guidance says.

The Measurement Problem Underneath Every Target

Almost every article on this topic treats race and ethnicity as if they were fields on a form that simply get filled in. They are not. They are constructed categories with a federal standard behind them, a self-identification rule, a multi-response structure that breaks clean arithmetic, and no coherent meaning outside the United States. Every downstream number depends on how you handle that, so it is worth being exact.

Two questions, five races, self-reported

FDA’s operative final guidance on collecting these data is still the October 2016 document. The January 2024 revision, Collection of Race and Ethnicity Data in Clinical Trials and Clinical Studies for FDA-Regulated Medical Products, remains a draft under docket FDA-2016-D-3561.⁸ That draft recommends a two-question format. Ethnicity is asked separately, with minimum choices of “Hispanic or Latino” and “Not Hispanic or Latino.” Race is asked with five minimum choices: American Indian or Alaska Native, Asian, Black or African American, Native Hawaiian or Other Pacific Islander, and White.⁷

Three rules in that draft carry real operational weight. First, the data are self-reported. FDA is explicit that race and ethnicity “should not be assigned by the study team conducting the trial,” and that where the information exists in a medical record, site staff should verify it with the participant rather than copy it.⁷ Second, participants may select more than one race, and FDA recommends the instruction be worded “Mark one or more” or “Select one or more.” Third, when results are presented, sponsors should report counts for participants who selected only one category, separately for each of the five races, and are encouraged to report the full distribution of multiple responses. If those are condensed, at minimum the total reporting more than one race must be reported.⁷

The standards under the standards just moved

On March 28, 2024, the Office of Management and Budget published final revisions to Statistical Policy Directive No. 15, the standard that governs how federal agencies collect and present race and ethnicity data.⁹ The revision made three changes that matter here. It requires a single combined question for race and ethnicity rather than two. It adds Middle Eastern or North African as a required minimum category. And it requires collection of more detailed categories beyond the minimum set in most situations.¹⁰¹¹ Agencies have until March 28, 2029 to comply.

So the federal statistical standard is now a one-question, seven-category instrument, while the FDA guidance sponsors actually build case report forms against is a two-question, five-category instrument that is either eight years old (the 2016 final) or unfinalized (the 2024 draft). Any sponsor benchmarking enrollment against United States census data collected under the new standard, while collecting trial data under the old one, is comparing two things that were not measured the same way. Middle Eastern and North African participants in particular will appear in the census denominator as their own category and in the trial numerator distributed across White and other selections.

The categories do not travel

FDA says this itself. The January 2024 draft states that for trials enrolling participants outside the United States, the agency “recognizes that the recommended categories for race and ethnicity were developed in the United States and that these categories may not adequately describe racial and ethnic groups in other countries,” and it recommends using more detailed categories organized by geographic region, mapped back to the five races and two ethnicities where possible.⁷

That mapping is where global trials come apart. In the European Union, racial or ethnic origin is a special category of personal data under Article 9 of the General Data Protection Regulation, prohibited from processing by default and permitted only when a specific condition applies on top of a lawful basis. The European Commission’s guidance on equality data based on racial or ethnic origin stresses self-identification, voluntary participation, and the fact that member states differ substantially in what they collect and how.¹³ France in particular frames collection around geographic and cultural origin rather than race. A single global case report form field with five United States race categories will produce large volumes of missing or “other” data across Europe, and in some jurisdictions it will produce a data protection question before it produces a number.

ICH E17, the guideline on multiregional clinical trials, offers the framework most sponsors should be using here instead. It treats region as an indicator for underlying intrinsic factors (genetic and physiological) and extrinsic factors (medical practice, diet, standard of care) that may drive differences in treatment effect.¹⁴ Region is not a substitute for race, but for the scientific question a Diversity Action Plan is ultimately about, which is whether the evidence generalizes to the people who will take the product, region and intrinsic and extrinsic factors are often the more defensible variables. The strongest plans we see hold both: a United States demographic target expressed in FDA categories, and a regional and intrinsic-factor rationale that explains what those categories are standing in for.

Where the data actually live

One more layer. In the standard clinical data model, the demographics domain carries a single RACE variable. When a participant selects more than one race, the common convention is to set RACE to “MULTIPLE” and put the individual selections in the supplemental qualifiers dataset as RACE1, RACE2, and so on.¹² That is a perfectly sensible modeling choice for submission datasets. It is a poor structure for operational monitoring, because the detail a sponsor needs to track progress against a subgroup goal has been pushed out of the main table and into a supplemental one that most reporting tools do not join by default.

The practical result is that a sponsor asking “how many Black or African American participants have we enrolled” gets a different answer depending on whether the query reads the demographics domain, joins the supplemental dataset, or reads the raw case report form. None of the three is wrong. They answer different questions. If your Diversity Action Plan does not say which one it means, the number in your progress report and the number in your submission will not match, and someone will ask why.

Setting a Number You Can Defend

FDORA requires a goal, a rationale, and a plan to meet it. The rationale is the part that most often gets written last and reviewed hardest, and it is where the honest constraint lives: the epidemiology of a disease frequently does not match the demographics of the general population. A target set to the census will sometimes be scientifically wrong.

When disease burden and population share diverge

Multiple myeloma is the clearest published example. The disease occurs at roughly twice the rate in Black Americans as in White Americans, yet Black patients have historically made up under five percent of participants in myeloma trials. A pooled analysis presented through the American Society of Hematology found that four percent of patients screened for eligibility were Black while 83 percent were White. Among Black patients who were screened out, the most common reasons were failure to meet hematologic laboratory criteria (19 percent) and failure to meet treatment-related criteria (17 percent).¹⁷

That second finding is the operationally important one. A meaningful share of the gap was created by eligibility criteria interacting with population biology, including the Duffy-null phenotype, which is associated with lower baseline neutrophil counts in people of African ancestry without any underlying disorder. A protocol with a conventional absolute neutrophil count floor will exclude healthy people at different rates across populations. No amount of recruitment outreach fixes that. The fix is in the protocol.

What the current baseline actually looks like

FDA’s Drug Trials Snapshots program publishes the demographic composition of pivotal trials supporting each novel approval, which makes it the best available baseline. In 2024, CDER approved 50 novel drugs, with about 31,000 participants across the supporting pivotal trials. Across therapeutic areas, White participants ranged from 19 percent to 95 percent of enrollment and made up more than half the population in a majority of programs. Asian participants were the second largest category at 0 to 80 percent, with oncology programs averaging 44 percent. Black participants were the lowest, ranging from 0 to 68 percent. Hispanic or Latino participants ranged from 0 to 46 percent.¹⁵

Two numbers from the same report deserve more attention than they get. Oncology programs enrolled 86 percent of participants outside the United States, and reproductive, urologic, and rare metabolic programs enrolled 97 percent outside the United States.¹⁵ A sponsor writing a United States-referenced demographic target for a trial where nine in ten participants will be enrolled abroad is setting a goal that a small minority of the study population can move.

An independent analysis of 341 phase 3 pivotal trials supporting FDA approvals from 2017 to 2023 reached a similar place from a different direction. Only six percent of trials achieved enrollment aligned with United States census distribution across all four racial and ethnic groups examined. In the 2023 cohort of 55 trials, 23 percent had representative numbers of Black participants, 30 percent for Hispanic participants, and 81 percent for Asian participants. Black representation was positively correlated with United States-based participation, while the United States share of participants fell from a peak of 47 percent in 2020 to 25 percent by 2023. Cancer trials showed among the lowest Black representation at four percent.¹⁶

Documenting a target that is deliberately not proportional

Sometimes the right target is not the census. If a condition is concentrated in one population, proportional enrollment understates the group that matters most. If a condition is rare and concentrated in a specific ancestry group, the census is close to irrelevant. FDORA asks for the rationale precisely so that sponsors can make this argument rather than defaulting to a number that looks fair and is scientifically empty.

A defensible non-proportional target has four written components. State the reference population you are using and why: disease incidence or prevalence in the intended use population, not general demographics. Cite the source for that epidemiology with a date, because incidence estimates get revised. Show the arithmetic that converts the epidemiology into an enrollment number given your planned sample size and site footprint. And state what you will do if the actual enrolled population diverges from the target, including at what threshold and by when you would change the plan.

That last item is the one that separates a document from a commitment, and it is the one that cannot be written at all unless the monitoring capability described in the next two sections exists.

Site Selection With Data, and Its Limits

The standard recommendation is to select sites where the target population actually receives care, using epidemiology, census data, registries, and claims. That recommendation is correct. It is also frequently oversold, and the published evidence is unusually clear about where it stops working.

What the data can tell you

The inputs worth assembling before a site list is drafted are straightforward and mostly public. Disease incidence and prevalence by geography and demographic group, from cancer registries, surveillance programs, and disease-specific registries. Census and American Community Survey data at the tract level around candidate sites, which is far more informative than county or metropolitan-level figures. Claims or electronic health record derived counts of diagnosed patients in a catchment area. Historical enrollment performance at candidate sites, broken out by subgroup rather than reported as a single accrual number. And the practical access factors: distance, public transit, clinic hours, interpretation services, and whether the site has staff who can consent in the languages its catchment speaks.

What the data cannot do on its own

Here the evidence is worth reading carefully. Alliance A191901, a national randomized trial on breast cancer endocrine therapy adherence, was designed with deliberate oversampling targets of 30 percent for Black women and for women under 50. Analysis of site-level recruitment found that neighborhood racial composition was strongly associated with recruitment of Black participants: sites in the highest-composition neighborhoods enrolled 16.2 percent Black participants versus 1.4 percent at sites in the lowest. Geographic region also predicted. Notably, site type and historical accrual volume did not predict Black participant enrollment at all. And the trial, despite patient-engaged design and deliberate targets, recruited Black and younger patients at population-representative proportions but not at its oversampling targets during the period studied. Recruitment trajectories for Black participants ran slower than target rates while non-Black recruitment ran faster.¹⁸

A second study makes the same point from the site side. An academic cancer center analyzed 2,317 trial participants enrolled between 2015 and 2023 and compared them against its own catchment area. Hispanic participants were 20.5 percent of enrollment against 34.0 percent of the surrounding county population. Asian participants were 20.1 percent against 23.3 percent of the county. Both figures were far better than national comparators, and the authors concluded that aligning the trial portfolio to catchment burden does improve accrual. But neither matched the catchment.¹⁹

The two findings that should change how you build a site list

First, historical accrual volume is a bad proxy for subgroup accrual. A high-enrolling site is a site that enrolls quickly from the population it already reaches. If that population has never been diverse, past performance predicts future homogeneity. Selecting on total accrual is how sponsors reproduce the same participant profile study after study while genuinely believing they are optimizing.

Second, neighborhood-level data outperforms institution-level data. The predictive signal in A191901 came from the composition of the area around the site, not from what kind of institution it was.¹⁸ That is a tractable change: geocode candidate sites, pull tract-level demographics for a realistic travel radius, and score on that rather than on institutional reputation.

Seeing the Gap While Enrollment Is Still Open

This is the section that almost no coverage of Diversity Action Plans includes, and it is the one that determines whether any of the rest matters. A target missed is only actionable if the sponsor can see the gap while enrollment is open. Once the last participant is randomized, the number is the number.

Where demographic data sits, and how stale it is

Race, ethnicity, sex, and age are collected at screening or baseline, entered into the electronic data capture system by site staff, and then generally sit there. They are not typically part of the operational enrollment feed that program leadership watches, which is usually a count of screened, enrolled, and randomized participants by site pulled from the clinical trial management system.

The delay between a participant’s visit and the moment that participant’s demographic record becomes visible to the sponsor is measurable and larger than most executives assume. The Tufts Center for the Study of Drug Development’s eClinical Landscape Study found an average of five days from patient visit to data entry when the study database was released before first patient first visit, and ten days when it was released afterward. Database build and release averaged 68 days. Database lock averaged 31 days for the early-release group and 54 days for the late-release group.²³

Regional variation makes it worse in exactly the trials where it matters. In the Altair multinational study, the median time to data entry was 7 days overall, but 3 days at sites in Asia and 13 days at sites in North America.²⁴ That is not a rounding difference. It means a sponsor’s view of subgroup enrollment is systematically fresher in some regions than others, which biases every mid-study comparison drawn from it. If your United States sites report ten days later than your Asian sites, your dashboard will consistently understate United States enrollment and, with it, the enrollment of the exact subgroups your plan is about.

The regulatory framework already supports this

Sponsors sometimes hesitate to build subgroup enrollment monitoring because it feels like an extra apparatus outside the trial’s quality system. It is not. ICH E6(R3), which reached Step 4 in January 2025 and which FDA has adopted, restructures good clinical practice around risk-based quality management, quality by design, and proportionate oversight, with substantially greater reliance on centralized monitoring rather than uniform on-site visits.²¹²² Enrollment composition against a stated goal is a textbook centralized monitoring metric: it is computable from data the sponsor already receives, it is meaningful at the site level, and it triggers a defined action when it drifts.

Framing it that way also solves the internal ownership question. Subgroup enrollment monitoring does not need a new function. It belongs in the monitoring plan, alongside screen failure rates and protocol deviation trends, with the same escalation path.

What a usable monitoring view contains

The mistake sponsors make is building a dashboard that shows the current enrollment percentage against the target percentage. That view tells you where you are and nothing about whether you will get there. Four elements make it actionable.

Retention Is a Separate Problem From Enrollment

A sponsor can hit every enrollment target and still produce an analysis population that does not resemble the enrolled population. Differential dropout is the mechanism, and it is rarely monitored by subgroup at all.

The STRRIDE exercise trials illustrate the size of the effect with unusually clean reporting. Of 947 enrolled participants, 295 (31 percent) dropped out. Loss to follow-up was 31.3 percent among Black men and 19.6 percent among Black women. The stated reasons cluster in ways that point directly at operational fixes: among Black men, work responsibilities accounted for 15.6 percent of dropouts, changing their mind 12.5 percent, transportation 6.3 percent, and lack of motivation 6.3 percent. Most dropout occurred during the ramp period of the intervention, with Black women showing a 50 percent dropout rate during that window.²⁰

Two things follow. First, the arithmetic. A trial that enrolls at target and then loses one subgroup at a materially higher rate ends with an analysis population that is less representative than the enrolled population, and the shortfall appears only in the final tables. Second, the timing. Dropout concentrated in an early, high-burden phase means the intervention window for retention is narrow and early, not spread evenly across the study.

What to monitor, and what it costs the site

Retention monitoring by subgroup requires three things that most trials do not currently produce as a routine output:

• Discontinuation rate by demographic group, by site, over time. Not a single end-of-study figure. A running rate, so a divergence is visible while there is still study left to run.
• Reason for discontinuation, coded consistently and analyzed by group. Withdrawal reasons are usually captured in a free-text or lightly-coded field that nobody analyzes until the clinical study report. Transportation and work conflict are fixable with visit windows, travel reimbursement, and evening or weekend hours. Lack of efficacy is not.
• Visit adherence and window compliance by group. Missed and out-of-window visits precede withdrawal. They are the leading indicator, and they are already collected.

There is a burden question here that deserves an honest answer. Every additional monitored metric creates site-facing follow-up. The way to keep this proportionate is to monitor centrally and escalate selectively: compute the subgroup retention metrics from data the site already enters, and only involve the site when a threshold is crossed. That is precisely the posture ICH E6(R3) describes for risk-based quality management.²¹

What the Systems Layer Actually Has to Do

Pull the operational requirements together and a fairly specific systems picture emerges. It is not exotic, but it is more than most sponsors have today.

Five capabilities

One. A single agreed definition of each demographic measure, written down. Which race categories, single-selection or any-selection, how “more than one race” is counted, how missing and declined are treated, and how non-United States sites map to the categories. This definition has to be identical in the Diversity Action Plan, the monitoring report, the statistical analysis plan, and the submission datasets. Getting four different numbers for the same question is the single most common failure, and it originates here rather than in any system.

Two. Demographic data flowing out of EDC on a schedule, not on demand. A recurring extract of the demographics domain and its supplemental qualifiers, joined to enrollment status and site identifiers, landing somewhere the operational reporting layer can read. The frequency matters less than the fact that it is automatic and dated.

Three. Reference data for targets and benchmarks, versioned. The epidemiology and population figures behind your goals will be revised. Store the version and date you used, because the rationale in your plan is only defensible against the numbers that existed when you wrote it. This is the same discipline applied to any regulated reference data.

Four. The screening funnel captured as data. Referral and pre-screening activity often lives in site spreadsheets, recruitment vendor systems, or nowhere. Without it, you can see that you have a shortfall but not where in the funnel it originates, which means you cannot choose between the five very different responses in the table above.

Five. Traceability from the reported number back to the source. When a review division asks how a sponsor arrived at a stated enrollment figure or a progress claim, the answer needs to be reconstructible: which extract, which definition, which date. This is ordinary data lineage applied to an unfamiliar metric, and it is the part that turns a monitoring dashboard into something a regulator can rely on.

What this does not require

It is worth saying what is not needed, because the vendor conversation tends to expand quickly. This does not require a new platform, a purpose-built diversity module, or artificial intelligence. The data already exists in systems the sponsor already owns. What is missing is a definition, a scheduled extract, a joined view, and an owner. Sponsors who start by buying something usually end up with a tool that reports a number nobody has defined.

What to Build Now, Whatever the Guidance Does

The final guidance may publish. It may be narrowed. It may stay in its current state indefinitely. None of those outcomes changes the underlying scientific expectation, which existed before FDORA and is embedded in how review divisions read evidence: a marketing application should show that the product was studied in the people who will use it. That expectation is durable. The compliance packaging around it is not.

The work that holds its value across every scenario is the work that is hardest to do quickly under a deadline.

One further reason to build the capability rather than the paperwork: FDA convened a two-day public workshop under section 3603 of FDORA in November 2023, co-hosted with the Clinical Trials Transformation Initiative, specifically on increasing enrollment of historically underrepresented populations and on encouraging participation that reflects the prevalence of a condition among demographic subgroups.²⁵ The scientific interest that produced that workshop is not contingent on which draft guidance is posted in a given month. Review divisions ask about representativeness in meetings today. Sponsors who can answer with a number and a source are in a materially different conversation from those who cannot.

Conclusion

The honest summary of the regulatory position as of August 2026 is this: the statute stands, the requirement is not yet in force because it depends on a final guidance that has not published, the draft that would have shaped it has been through removal and court-ordered restoration and carries that history on its face, and the statutory deadline for finalizing it passed more than a year ago. Sponsors waiting for clarity have been waiting a long time and may wait longer.

What has not changed is the operational reality underneath. Setting enrollment goals by demographic subgroup commits a sponsor to measuring against them, and measurement here is genuinely hard for reasons that have nothing to do with politics. The categories are constructed, self-reported, multi-response, and United States-specific. The standard they derive from has just changed while the guidance sponsors build forms against has not. The data sit in a submission-oriented structure that is poorly suited to operational monitoring, arriving days to weeks after the visit, with regional differences in that lag large enough to distort mid-study comparisons. Site selection raises the ceiling and does not by itself lift the floor. And retention can undo enrolled diversity quietly, after the enrollment scoreboard has already been declared a success.

None of that is a reason to stop. It is a reason to build the measurement capability first and the commitment second, which is the reverse of how most organizations have approached this. A sponsor that can see subgroup enrollment against a planned trajectory, break it down by funnel stage and site, and act on it while enrollment is open is in a strong position whichever way the guidance goes. A sponsor with a well-written plan and no way to track it is exposed in exactly the scenario it was trying to prepare for.

Sakara Digital works with pharma and biotech organizations building the data foundations that make commitments like these measurable rather than aspirational. If you are setting enrollment goals for a study and want an independent view on whether your systems can actually tell you how you are doing against them, we are happy to have that conversation.

For Further Reading