Abstract
Direct primary care (DPC), in which patients or employers pay a practice a recurring fee for primary care outside fee-for-service billing, is among the fastest growing models in American medicine. A peer-reviewed national study counted concierge and direct-pay practice sites rising 83.1%, from 1,658 to 3,036, between 2018 and 2023. The outcomes literature has not kept pace. A PubMed search for "direct primary care" restricted to 2021 to 2026 returns 38 papers, of which only about 14 study the US model directly, and no randomised trial of DPC exists. The strongest independent analysis, an actuarial study of a single employer, found risk-adjusted claims 12.64% lower and emergency department use 40.51% lower in a DPC cohort whose baseline risk score was 8.3% lower than the comparison group's, with the employer's net cost 1.3% higher once membership fees were counted. Two doctoral dissertations using difference-in-differences designs found that employees who chose DPC were younger and healthier and that total costs rose. The model's best documented figures, from an academic practice serving a vulnerable, largely Spanish-speaking panel, ended with that practice's closure when it could not break even. From January 2026, federal law makes qualifying DPC fees compatible with health savings accounts, a public subsidy with no outcomes-reporting condition attached. This commentary argues that a model asking for tax-advantaged growth, employer contracts and workforce exemptions owes purchasers and patients a standard accounting: denominator reporting, named risk adjustment, total cost including fees, patient out-of-pocket cost as an endpoint, and independent evaluation. The author writes as someone who ran a DPC clinic and believes the model can work. Belief is not evidence, and the field has had a decade of belief.
Keywords: direct primary care; primary care organisation; health care costs; selection bias; health policy; practice management
1. A model I ran, and a question I cannot answer from my own clinic
I bootstrapped a direct primary care clinic and ran it as chief executive. I believed in the model then and I believe in it now: small panels, same-day access, a doctor who knows your name and your history, a price a family can read on one page. Patients told us the care felt different, and I watched doctors who were burning out in fifteen-minute medicine remember why they trained.
If you asked me whether my clinic lowered total cost of care, I could not tell you, and that is the subject of this article. We did not track a denominator properly. We did not risk-adjust. Our happiest patients wrote the testimonials, and our unhappiest quietly stopped paying and disappeared from our numbers. Every DPC practice I know tells its story the way we told ours, with conviction and with data that would not survive peer review. The published record, examined closely, is the same story at larger scale.
This matters now because the model is no longer a cottage experiment. It has a growth curve, a trade association, employer contracts, venture-backed platforms and, from 2026, a place in the federal tax code. Growth on that scale normally brings an evidence burden with it. DPC has so far been excused from carrying one, partly because its advocates are sincere, partly because primary care is so visibly broken that any working alternative feels like relief. Sincerity and relief are not outcomes. What follows is an attempt to state, fairly and with citations, what is actually known about DPC outcomes, what the strongest criticisms are, where the genuine positives lie, and what accounting the field should now be required to produce.
2. What the growth numbers count
The headline figure comes from Zhu and colleagues in Health Affairs (2025), who tracked concierge and direct-pay primary care in national data from 2018 to 2023: practice sites rose from 1,658 to 3,036, an increase of 83.1%, and clinicians from 3,935 to 7,021, up 78.4%.[1] It is a careful study, and its caveats deserve as much attention as its headline. The cohort pools concierge medicine and DPC, two models with different prices and different patient populations, and the data cannot separate them. Roughly 60% of the clinicians billed Medicare, which points to hybrid practices that keep one foot in insurance billing rather than the pure subscription model described in DPC advocacy. Independent ownership fell from 84.0% to 59.7% over the period, while corporate-affiliated practices grew by 576%. The growth is real. A rising share of it is corporate, hybrid and concierge, which is a different story from the independent-practice revival the movement describes.
Practice counts from other sources need their labels kept on. The DPC Frontier Mapper, an advocacy-run directory of self-listed practices, showed 3,240 practices when checked in October 2026. A vendor platform has claimed more than 3,600. These are marketing surfaces as much as measurements, and they move. Commercial market-size estimates are worse: published figures for the "direct primary care market" span from about $4.2 billion to about $70 billion for overlapping periods, a sixteen-fold spread that tells the reader nothing except that the firms selling the reports do not agree on what the market is. This article cites none of those figures as fact. The peer-reviewed count, with its caveats, is the defensible one: about three thousand sites, growing fast, increasingly corporate.
The demand side is easier to read. Primary care in the United States is short of doctors, short of time and short of patience, and both patients and physicians are looking for an exit from fifteen-minute, billing-driven encounters. DPC fees in published sources cluster around $50 to $100 per adult per month (Eskew and Klink's 2015 survey found a mean of $93.26 and a median of $75 across direct-pay practices, with DPC proper averaging $77.38 against $182.76 for concierge;[2] a 2023 dissertation put the national average adult DPC price at $81.33[3]). Panels run around 400 patients against roughly 2,200 to 2,500 in traditional practice. The offer is legible. The question is what it delivers, and to whom, and at whose expense.
3. The outcomes literature, all of it
A PubMed search for "direct primary care" limited to 2021 to 2026 returns 38 papers. Reading the titles, about 14 concern the US model directly; the rest match incidentally. None is a randomised trial, and the PubMed randomised-trial filter on the same term returns nothing. There is no published study of standard quality measures, preventive screening rates or chronic disease control from independent DPC practices that this author could find. What exists is a small set of claims analyses, two dissertations, a handful of experience studies and a large quantity of commentary. Here is the rigorous core, in full.
The actuarial study. The Society of Actuaries commissioned Milliman to evaluate an employer-sponsored DPC programme, published in 2020 (Busch and colleagues).[4] About 900 DPC members were compared with about 1,100 in a PPO. On risk-adjusted claims, the DPC cohort cost 12.64% less; emergency department use was 40.51% lower (p<.001); inpatient admissions were 19.9% lower and not statistically significant. Two qualifications sit inside the same report. The DPC cohort's risk score was 8.3% lower at baseline, meaning the people who chose DPC were healthier before the model touched them. And once the membership fees were added to the claims, the employer's net cost was 1.3% higher than the PPO. Quality and satisfaction could not be assessed from the data available. This is the best independent evidence the field has, and its honest summary is: utilisation shifted in the right direction, selection explains an unknown share of it, and the purchaser did not save money.
The dissertations. Doctoral work is where null and negative results go to be published, and DPC has two careful examples. Brekke studied the Paladina (later DaVita) worksite clinic programme used by Union County, North Carolina employees, using difference-in-differences.[5] Employees who chose the DPC option were younger, healthier and more likely to have had no usual source of care. Specialist visits rose significantly among choosers. Expenditures rose. The emergency department result was inconclusive. Schwartzman, in a 2023 Washington University dissertation on Nextera Healthcare's contract with the St. Vrain Valley school district, used instrumental variables and difference-in-differences and found selection by younger, cheaper employees, higher patient out-of-pocket costs, and higher total costs among low spenders, alongside two genuinely encouraging process signals: mammography rose and low-value imaging fell.[3]
Union County as a case study in contested arithmetic. Union County is quoted constantly in DPC advocacy, usually as a 26% saving. The advocacy analysis itself acknowledged that the DPC group carried a lower risk score. The actuarial break-even in the Milliman-style reading rested on an assumed fee near $61 per member per month, while the actual contract fee was closer to $96 to $106; corrected for the fee, the net position moves to roughly $33 per member per month in the wrong direction. The county plan also carried a $0 deductible, which pushed specialist use toward twice national rates, a benefit-design effect that has nothing to do with the clinic model. Neither the savings claim nor the full re-analysis is peer-reviewed, which is precisely the problem: a decade into the employer DPC experiment, its flagship case is an argument between an advocacy brief and a blog-grade rebuttal, with a dissertation quietly agreeing with the rebuttal's direction.
Qliance and the canonical miscitation. Qliance, the Seattle DPC pioneer backed by prominent investors, quoted striking figures for years: emergency visits down by figures in the range of 35% to 65% depending on the telling, hospitalisations and specialist use sharply lower. A University of Washington brief noted plainly that there had been no independent evaluations; the numbers were the company's own claims analyses, circulated from a 2013 BMJ news feature (Page) into hundreds of secondary sources as if they were study findings.[6] Qliance closed its clinics in June 2017 and entered Chapter 7 bankruptcy in May 2018. A model can fail commercially while working clinically, and closure proves nothing about outcomes. What the episode proves is about the literature: the field's most repeated numbers were never independently produced.
Experience and engagement. Patient experience evidence is thin and mixed in an informative way. Tou and colleagues (2022) surveyed patients at a single DPC clinic (31 patients, 10 interviewed in depth): access and communication scored well, and difficulty getting specialist referrals scored badly, a finding any honest DPC doctor will recognise, because a model built on keeping care in-house has no natural owner for the referral problem.[7] Ashaye and colleagues (2025) reported a Houston practice with intense engagement, about 13 documented interactions and 7 appointments per patient per year, rising with tenure, in a panel that was 71% Black; there was no comparison group, and 44% of that panel paid more than $125 a month, which sits uneasily beside equity claims made for the model.[8]
Physician outcomes. Here the evidence is at its most consistent. Boylan and Hurley (2025), using the validated Professional Fulfillment Index, found DPC physicians reporting lower burnout (p=.002) and higher fulfilment (p=.013) than traditional-practice peers at similar hours.[9] American Academy of Family Physicians survey briefs, self-selected and to be read as such, report satisfaction of 94% in DPC against 57% in traditional practice. Doctor wellbeing is a real outcome, and a workforce haemorrhaging physicians has reason to care about it. It is also the outcome the literature measures best, because doctors answer surveys about themselves more readily than systems publish claims about patients.
That is the whole rigorous record: one actuarial study with a null net-cost result, two dissertations pointing the same way, process signals worth following, engagement data without controls, and physician wellbeing data that is genuinely encouraging. Cole saw this coming. Writing in the Journal of the American Board of Family Medicine in 2018, she concluded that "there is little evidence in the peer-reviewed literature to support these claims", and called for "descriptive and comparative research of included and excluded patients", a phrase that names the field's central omission: the denominator, the people the model does not reach, is never in the paper.[10]
4. The case against, stated fairly
Selection is documented, not alleged. At every level of rigour the same pattern appears: Milliman's 8.3% baseline risk gap, Brekke's younger and healthier choosers, Schwartzman's younger and cheaper ones. The concierge literature, DPC's older sibling, shows where unexamined selection leads. Alexander and colleagues (Journal of General Internal Medicine, 2005) found that when practices converted to concierge, they kept about 12% of their patients.[11] Leive and colleagues (Journal of Health Economics, 2023), studying concierge conversions with Medicare data, found spending 30% to 50% higher among switchers and no mortality benefit. A model that selects healthier, wealthier patients will post beautiful unadjusted numbers indefinitely. That is why the demand in this article is not cynicism about DPC. It is arithmetic.
The workforce arithmetic is stark. Zhu and Song, writing in the New England Journal of Medicine in 2025, put it in one sentence: "With panels reduced by 80% or more, a single full-time physician's departure… could mean that roughly 2000 patients lose their PCP."[12] Weisbart calculated in Family Practice Management in 2016 that universal DPC at prevailing panel sizes would need roughly three times the current primary care workforce. DPC doctors answer, reasonably, that they were leaving medicine altogether and a 400-patient panel retained beats a resignation. That answer is true at the level of the individual doctor and incomplete at the level of the system, where every conversion redistributes two thousand patients onto colleagues whose panels are already full.
Equity cuts both ways, and the closure case matters. Liaw and colleagues (JABFM, 2024) described an academic DPC practice in Austin that did the thing critics say DPC cannot do: it reached a vulnerable panel, 37% Spanish-speaking, with a social vulnerability index of 0.60 against 0.47 for its surroundings.[13] Then it closed, in November 2022, because it could not break even. The lesson is double-edged. Demand among low-income patients for this kind of care is real, and the model's standard price point does not survive contact with the populations who need primary care most, unless someone subsidises the difference. Geography tells a similar story: Goldstein and Yerkes (Annals of Family Medicine, 2024) found DPC practices present in shortage areas at roughly population rates (44% against 47%) but less likely to sit in the highest-priority shortage areas.[14]
Policy has now subsidised the model without asking for the accounting. State "not insurance" statutes, which exempt DPC agreements from insurance regulation, number somewhere between 33 and "35 or more" depending on who counts and how Oregon's certification law is classified; the disagreement itself shows how loosely the sector is mapped. Pure DPC sits largely outside Medicare, with 77% to 81% of practices in surveys reporting they do not bill it. The largest change is federal. Before 2026, an IRS position treated a DPC arrangement as disqualifying coverage for health savings account purposes. The One Big Beautiful Bill Act (P.L. 119-21, section 71308) added Internal Revenue Code section 223(c)(1)(E), effective 1 January 2026: a qualifying direct primary care service arrangement, solely primary care for a fixed fee capped at $150 a month for an individual and $300 for a family, indexed after 2026, is no longer treated as a health plan, and its fees are HSA-payable.[15] Two limits deserve print: the fees do not count toward the high-deductible plan's deductible or out-of-pocket maximum, and the plan may not pay them before the deductible. What deserves more print is what the statute does not contain. No outcomes reporting, no quality measurement, no evaluation requirement of any kind attaches to the new tax advantage. Congress has, in effect, decided that the model works. The literature reviewed above does not support that confidence, and the statute asks for none of the data that would create it.
A terminological warning belongs here, because purchasers trip on it: the CMS "Direct Contracting" programme and its successor ACO REACH are Medicare risk contracts with entities, and have nothing to do with direct primary care despite the near-identical names. Any vendor deck that borrows evidence across that boundary should be read with suspicion.
5. The case for, stated fairly
The strongest version of the DPC case does not rest on the contested savings claims. It rests on four things the record does support.
First, access and continuity are the model's design, and the fragments of evidence point the right way: Milliman's emergency department reduction survived risk adjustment even if its size is uncertain; Ashaye's engagement counts show patients actually using the access they pay for, more with longer tenure; Schwartzman's mammography rise and low-value imaging fall are exactly the process pattern a relationship-based model predicts. Second, the physician wellbeing data are the most consistent findings in the whole literature, and in a country projecting primary care shortages they are not a soft outcome; a model that keeps experienced family doctors in practice has a workforce value that panel arithmetic alone misses. Third, Liaw's Austin experience proves demand: vulnerable patients chose this care, stayed, and used it, which kills the lazy version of the equity critique that says DPC only appeals to the worried well. Fourth, DPC's presence in rural shortage areas at roughly population rates, achieved without programme funding, compares well with many subsidised interventions.
What none of this establishes is the claim the market actually sells, which is total cost of care. The honest position, and the one this author held while running a clinic, is that DPC is a bet: on continuity, on access, on doctor retention, with cost savings unproven and possibly absent once fees are counted. Bets can be worth taking. They should be labelled as bets, priced as bets, and evaluated like bets, with the denominator published.
6. The accounting the field owes: a research agenda
The calls for this work are older than the growth curve. The American College of Physicians' 2015 position paper (Doherty) devoted its Recommendation 9 to a seven-item research list for direct patient contracting practices: growth patterns, burnout and sustainability, access for underserved populations, workforce effects, out-of-pocket and system costs, experience, quality and outcomes, and the consequences of sitting outside quality programmes and standard EHR reporting.[16] A decade on, most of that list is still open. The National Academies' 2021 primary care report stated the position in one sentence: "the literature assessing outcomes is small. In addition, studies are plagued by the difficulty of adjusting for the selection of healthy patients into these practices."[17] Cole's 2018 call for research on "included and excluded patients" remains unanswered in the specific form she posed it.[10]
What follows is the minimum standard this author would set, each item tied to a documented failure in the record above.
- Publish the denominator. Every outcomes report states how many patients were eligible, how many enrolled, how many left, and who the leavers were. Qliance-style reporting dies the day this is normal.
- Name the risk adjustment. Method, variables and baseline scores for both arms, printed. Milliman published its 8.3% baseline gap; advocacy reports that omit the equivalent number should be assumed to be hiding it.
- Count total cost, fees included, benefit design stated. Claims plus membership fees plus the deductible structure of the comparison plan. Union County's argument happened because the fee assumption and the $0-deductible design were buried.
- Report patient out-of-pocket cost as an endpoint. Schwartzman found it rose. Patients experience cost at the till, not in the employer's ledger, and a patient-centred model should be judged partly on what patients pay in total.
- Firewall the evaluation. Analyses funded by the vendor or platform being evaluated should be labelled as such in the abstract, and purchaser contracts should require independent analysis of claims as a condition of renewal. The MDVIP arc in concierge medicine, where vendor-authored savings claims converged over a decade to a well-designed study (Morefield 2020, n=30,727) finding admissions down 10% but total Medicare spending not significantly different,[18] is the template of what happens without this rule: the truth arrives eventually, late, after the marketing has spent it.
- Stratify by equity variables as standard. Income proxy, race and ethnicity, language, rurality and baseline morbidity, so that Liaw-type reach and Ashaye-type pricing can be seen in the same table.
- Use a shared taxonomy. Pure DPC, hybrid, concierge and employer onsite clinics are different interventions. Zhu's pooled cohort shows the cost of blurring them: the field's flagship growth statistic cannot say which model grew.
- Register evaluations before results are known. Employer DPC evaluations are natural experiments with known start dates. A simple registry, run by a specialty society or a journal, would end the file-drawer problem that leaves dissertations carrying the null results.
None of this needs a randomised trial to begin, though a pragmatic trial in an employer or Medicaid managed-care setting is feasible and overdue. It needs editors willing to ask for denominators, purchasers willing to contract for independent analysis, and a specialty willing to hold its own favourite model to the standard it applies to everyone else's.
7. What to ask now, if you are buying or building
For employers and purchasers, the practical version of the standard above fits in one contract conversation. Ask for the baseline risk scores of joiners and non-joiners. Ask for total cost with fees in, over at least two years, against a comparison that shares your benefit design. Ask what happens to specialist referrals, because Tou's referral finding and Brekke's specialist increase point the same way. Ask who owns the claims analysis and whether you may publish it either way. A vendor who answers all five cleanly is selling something real.
For doctors considering the model, the clinical case and the business case are both honest ones to make: smaller panels, saner days, patients who chose you. Go in with your eyes open about the equity criticism, because your critics have read Alexander's 12% figure even if your future patients have not, and consider the Liaw lesson before you set your fee: the populations with the worst access will want what you are building, and your price decides whether they get it.
For policymakers, one sentence: the HSA change of 2026 can be defended as access policy, but attaching a light reporting condition to the tax advantage, a registry entry and a denominators table, would cost the sector little and buy the public the evidence the next decade of growth will otherwise spend without producing.
8. Limitations
This is a commentary, not a systematic review. The literature counts come from a single PubMed search string run in October 2026; "direct primary care" misses employer-clinic and concierge work published under other names, and the classification of the 38 results was by title and abstract. Two of the most informative sources are doctoral dissertations, which carry less editorial scrutiny than journal articles, and figures drawn from one critical secondary re-analysis of the Union County case are flagged in the text as not peer-reviewed. The author's position is declared: he ran a DPC clinic, believes in the model, and has a commercial and professional history with it that readers should weigh. Absence claims in this article ("no study could be found") rest on targeted searching, not exhaustive review.
9. Conclusion
Direct primary care has grown 83% on the strength of a promise its literature has not yet tested: that paying directly for primary care produces healthier patients at lower total cost. What the evidence supports today is narrower and still worth having: patients who get access and use it, process signals pointing the right way, doctors who stay in practice, and demand that crosses income and language lines. The gap between the promise and the record is not a reason to abandon the model. It is a reason to stop excusing it from the accounting every other growing sector of American medicine is asked to produce, now with a tax advantage in hand. I ran one of these clinics and could not have passed the standard set out above. The next generation of practices should be built so that they can.
References
- Zhu JM, et al. Growth of concierge and direct primary care practices. Health Affairs. 2025;44(12):1473-1481. doi:10.1377/hlthaff.2025.00656. PMID: 41329882.
- Eskew PM, Klink K. Direct primary care: practice distribution and cost across the nation. Journal of the American Board of Family Medicine. 2015;28(6):793-801.
- Schwartzman J. Doctoral dissertation, Washington University in St. Louis. 2023. doi:10.7936/q3q8-cj86.
- Busch F, et al. Direct primary care: evaluating a new model of delivery and financing. Society of Actuaries / Milliman. 2020.
- Brekke G. Doctoral dissertation, University of Kansas (Paladina / Union County employer clinic evaluation).
- Page L. The doctor will see you now: Qliance. BMJ. 2013;347:f6465. (News feature reporting company data.)
- Tou LC, et al. Patient experience in direct primary care. Journal of Patient Experience. 2022. PMID: 35936512.
- Ashaye T, et al. Engagement and access in an urban direct primary care practice. Journal of Primary Care & Community Health. 2025. doi:10.1177/21501319251391011. PMID: 41369101.
- Boylan M, Hurley (2025). Professional fulfilment and burnout in direct primary care. Southern Medical Journal. 2025. PMID: 40316271.
- Cole ES. Direct primary care: applying theory to potential changes in delivery and outcomes. Journal of the American Board of Family Medicine. 2018;31(4):605-611. PMID: 29986986.
- Alexander GC, et al. Physicians transitioning to concierge practice: patient retention. Journal of General Internal Medicine. 2005.
- Zhu JM, Song (2025). Concierge and direct primary care and the primary care workforce. New England Journal of Medicine. 2025;392:1977-1979.
- Liaw W, et al. An academic direct primary care practice serving a vulnerable population, and its closure. Journal of the American Board of Family Medicine. 2024. PMID: 39142864.
- Goldstein, Yerkes (2024). Direct primary care in health professional shortage areas. Annals of Family Medicine. 2024. PMID: 39586711.
- One Big Beautiful Bill Act, P.L. 119-21, section 71308, adding IRC section 223(c)(1)(E), effective 1 January 2026; IRS Notice 2026-5.
- Doherty R, American College of Physicians. Assessing the patient care implications of "concierge" and other direct patient contracting practices. Annals of Internal Medicine. 2015;163:949-952. doi:10.7326/M15-0366.
- National Academies of Sciences, Engineering, and Medicine. Implementing High-Quality Primary Care: Rebuilding the Foundation of Health Care. Washington, DC: National Academies Press; 2021. Chapter 9.
- Morefield B, et al. MDVIP and Medicare spending: a difference-in-differences analysis. American Journal of Managed Care. 2020.
- KFF. Employer Health Benefits Survey 2025 (7% of firms with 50+ workers contracting directly for primary care).