Somewhere on your phone there is a group chat that used to be the most important professional relationship in your life.
It was your residency class. Four of you, or eight, or twelve. You watched each other work at three in the morning for three to seven years. You know exactly who is calm in a crisis and who talks too fast when they are unsure. You know who reads everything, who is brilliant and disorganized, who you would want operating on your mother, and who you would very politely not. You know these things at a resolution that no CV, no board certification, no press-ganey score, and no directory listing will ever approach.
Then you all graduated and scattered. The chat was busy for a year. Then it was birthdays. Then it was nothing.
Here is what makes that unremarkable story worth an entire article: that dead group chat is the single strongest measured lever on referral quality in American medicine, and virtually nobody is using it.
Not a metaphor. Measured, published, and replicated. And in the most careful study of the question, 92 percent of the opportunities to use it were missed.
The study that should have changed referral practice
In 2021, researchers Maximilian Pany and J. Michael McWilliams published an analysis in Health Services Research examining 40,495 referrals inside a large health system. The question was simple: does it matter whether the referring physician and the specialist trained together?
The folk belief in medicine has always been yes. Everyone assumes doctors refer to their friends. What nobody had done was measure the size of the effect, and, more importantly, measure how often it actually happens when it could.
Their findings, in order of how surprising they are:
First, the effect is real. Physicians referred to co-trainees at a rate of 27.0 percent versus a 21.4 percent baseline (p < 0.001). Having trained together meaningfully shifts where patients get sent.
Second, and this is the sharp part, medical school does not count. Sharing a medical school produced a referral rate of 21.6 percent against the 21.4 percent baseline, with a p-value of 0.965. That is not a weak effect. That is no effect whatsoever.
Only residency and fellowship co-training produced the shift.
Sit with that for a moment, because it tells you precisely what kind of trust this is. Medical school is where you sit next to someone in lectures for two years and then rotate separately. Residency is where you watch someone make decisions under fatigue, uncertainty, and consequence, for years. The first produces acquaintance. The second produces calibrated judgment about a specific person's clinical character, and only the second one changes behavior.
This is the most useful single finding about professional trust in medicine, and it is remarkably underused. Trust that changes behavior is not built by proximity. It is built by observed performance under pressure.
Third, and this is the number in the title. Of those 40,495 referrals, 22.5 percent had an eligible co-trainee available to receive them. Of that eligible group, only 8.0 percent actually went to the co-trainee. That is 1.8 percent of all referrals.
Which means that 92 percent of the time when a physician could have referred to someone they trained with, they did not.
Not because they chose otherwise. Because they had no idea the option existed.
The follow-up that should have been on the front page
If the story ended there, you could argue reasonably that co-training preference is just homophily. People like their friends. Interesting sociology, minimal clinical relevance.
Then in 2023 the same line of research produced a finding in JAMA Internal Medicine that is genuinely hard to explain away.
Across 9,920 visits involving 502 specialists, when the referring physician and the specialist had trained together, patients rated the specialist 9.0 percentage points higher on a composite experience measure.
The authors put that effect size in context in a sentence worth quoting for its clarity: it is the equivalent of moving a specialist from the median to roughly the 91st percentile.
Stop and consider what that means. The same specialist, seeing a comparable patient for a comparable problem, is rated as though they were a substantially better doctor, depending on whether they happen to have trained with the person who sent the patient.
The authors noted the effect exceeded what has been demonstrated for public reporting programs, for accountable care organization participation, and for hospital characteristics. Those are the levers health policy has spent two decades and enormous sums pulling.
And this one costs nothing. It is a database lookup.
Why would that possibly work?
A finding this large demands a mechanism, and "doctors are nicer to their friends' patients" is not adequate. The literature suggests something more structural and, once you see it, more obvious.
When you refer a patient to someone you trained with, you have re-established accountability across an institutional boundary.
The specialist knows the referring physician will hear how the visit went. Not through a formal quality mechanism, but because they know each other and the patient will report back and the referrer might text about it. The researchers found associated behavioral changes consistent with this: clearer explanations, more shared decision-making, and even a measurable rise in prescribing activity, consistent with a physician who is engaging more thoroughly with the case.
The mechanism is that physicians behave differently when their work may be scrutinized or recognized by a peer whose judgment they respect.
This should not be surprising. It is how professionals have always worked. What is surprising is that medicine has systematically dismantled the conditions for it. In a fragmented, employed, multi-system landscape, the overwhelming majority of referrals now travel between people who have never met and never will. The consult note goes into a void. The referrer never learns what happened. Nobody is watching.
Co-training referral works because it accidentally restores something the system removed: a named human being on the other end who will know how you did.
There is a supporting finding worth noting from the JAMA curbside study, which measured something adjacent. 77.2 percent of subspecialists said curbside consultations were essential to maintaining professional relationships, against 38.6 percent of primary care physicians. The specialists understood something the referrers underrated. The informal exchange is not overhead. It is how the relationship that makes the referral work is maintained.
The economics, run carefully
Let us be careful here, because this is where enthusiasm usually outruns evidence.
There are roughly 100 million specialist referrals a year in the United States. If 22.5 percent have an eligible co-trainee, that is about 22 million referrals a year with the option available. At the observed 8 percent capture, roughly 1.8 million fire. If capture rose to even 25 percent, an additional 3.7 million referrals a year would land on a co-trained pair.
What is that worth? Two lines of evidence.
Patient experience. A 9 percentage point improvement in specialist ratings across millions of visits is enormous by the standards of any quality intervention, and patient experience scores carry direct financial consequences under value-based payment.
Spending. Separate research in Management Science found that a one standard deviation increase in referral concentration, meaning referrers sending to a tighter set of specialists they know well, was associated with 7.4 percent lower spending with no measurable quality decline. Relationship-based referral appears to be cheaper as well as better rated, which is not the usual tradeoff.
Apply even a conservative version of that to modest downstream spending per referral and the aggregate runs to hundreds of millions of dollars annually. From a database lookup.
For an individual specialist, the arithmetic is more personal. Fifty additional referrals a year from people who trained with you, at a plausible downstream revenue figure, is a meaningful annual difference. Multiply across a thousand specialists and you are describing a substantial redistribution of clinical work toward relationships that demonstrably produce better-rated care.
So why does nobody do this?
Here is the question that matters. The finding is published. The effect is large. The intervention is nearly free. Why has no health system, no EHR vendor, and no referral platform implemented "check for co-trainees first"?
The answer is one of the cleanest examples of structural failure in healthcare, and it is worth stating precisely.
Nobody knows who trained with whom, because nobody has any reason to.
The co-training graph exists in three places. It exists in individual memory, which decays and does not scale. It exists in state licensing databases and national provider records, in scattered and non-queryable form. And it exists in residency programs' own records, which stop being maintained the day people graduate.
No clinical system encodes it. No referral platform queries it. And here is the reason, stated as plainly as possible:
The person you trained with almost certainly works for a competitor.
That single fact explains everything. A health system's referral infrastructure exists to keep referrals inside the system. Every dollar spent on referral technology is spent on capture and leakage reduction. Building a tool that says "the best person for this patient is someone you trained with, who works at the health system across town" is directly contrary to the interest of the organization paying for the tool.
So the strongest known lever on referral quality goes unused, not because it is hard, and not because it is unproven, but because the entity that would build it loses money if it works.
Alumni networks are built for the wrong customer
The reasonable objection: residency programs and medical schools have alumni networks. Why do those not fill this gap?
Because they are built for the institution, not for you.
Look honestly at what a medical alumni association does. It sends a magazine. It organizes reunions. It maintains a roster of some accuracy. And it asks for money, repeatedly. Data from the higher education advancement sector is unambiguous about the dynamic: the overwhelming majority of alumni professionals report that leadership prioritizes fundraising over alumni experience, alumni receive multiple solicitations in their first year alone, and a meaningful share of institutions see substantial numbers opting onto do-not-contact lists.
Even the best of them are institution-serving by design. Cleveland Clinic's alumni program, roughly 27,000 members and genuinely well run, exists to drive referrals to Cleveland Clinic, support recruitment, and generate philanthropy. That is a legitimate purpose. It is simply not your purpose.
Some residency programs, remarkably, outsource alumni networking to Doximity entirely.
And Doximity, to be fair, holds the residency field on millions of physician profiles. It could add "search by people I trained with" tomorrow. What it cannot easily add is the part that makes the mechanism work: the norm of reporting back, the expectation of an answer, the mutual accountability that produced the 9 point effect in the first place. A search field returns names. The effect came from a relationship where someone knew you would find out how it went.
The graph is copyable. The obligation is not.
The uncomfortable implication about scale
One more property of this graph deserves attention, because it determines what any solution has to look like.
A co-training edge only fires if both people are reachable. Your residency classmate has to be findable, current, and open to receiving the referral. Which means the usefulness of a co-training graph scales roughly with the square of participation within any given program.
If ten percent of a residency program's graduates are in a given system, only about one percent of the possible pairs are live. At fifty percent, you get a quarter. At ninety percent, nearly everything works.
This has a strong practical consequence: you do not build this graph by signing up individual physicians. You build it class by class, program by program, because a half-populated class is nearly worthless while a fully populated one is immediately useful to everyone in it.
That is a familiar pattern to anyone who has studied network cold starts. Facebook did not launch to the world. It launched at one college and did not move until that college was saturated. The residency class is an unusually good version of this: it is small, already densely connected, already has a group chat, and its members already trust each other in exactly the way the data says matters.
The atomic unit is the class. Not the specialty, not the hospital, not the individual.
What you can do about it, starting now
This is one of the rare structural problems in medicine where an individual can capture most of the value alone, this month, with no technology whatsoever.
Rebuild your cohort roster
Sit down with your residency and fellowship classes and reconstruct where everyone is. Not a group chat. An actual roster, with three fields that matter and that no directory holds:
- Where they are now, and whether they are taking outside referrals.
- What they actually focus on now, which is frequently not what their title says.
- Whether they would take a call from you about a hard case.
This takes an afternoon and a dozen text messages. It will produce a more useful referral resource than any directory you have access to, because it is current, it is specific, and every entry comes with your own calibrated judgment about that person's clinical character.
Make the report-back a habit
The mechanism behind the 9 point effect was mutual accountability. You can create that unilaterally. When you refer to someone you trained with, tell them. When you receive a referral from someone you trained with, close the loop.
This is a thirty-second act that reconstructs the exact condition the research identified as causal. It also, not incidentally, makes you the person others want to refer to.
Look up the co-trainee before you look up the directory
Change the order of operations. When you need a specialist, ask first whether anyone you trained with does this. The Pany data says the answer is yes about 22.5 percent of the time and that you will fail to think of it 92 percent of those times.
That is a habit change, not a technology project, and it captures a large share of the available effect.
If you run a program, own graduation day
Residency programs sit on the single most valuable professional graph in medicine and hand it to their advancement office.
Graduation is the one moment the cohort is assembled, emotional, and about to scatter. Right now most programs use it to start the donation pipeline. A program that instead used it to establish a durable professional covenant among its graduates, a real roster, a norm of answering each other, an expectation of reporting back, would be creating something the evidence says improves patient experience by nine percentage points for decades.
It costs one hour on one day.
If you are early career, this is your most underrated asset
You are told to build a network. You already have one. It is the eight people who saw you work at 3 a.m. for four years, and it is worth more than any conference contact you will ever make.
The mistake almost everyone makes is treating it as friendship rather than infrastructure. Friendship decays without maintenance. Infrastructure gets maintained deliberately.
Frequently asked questions
Do patients do better when their doctor and specialist trained together? Patients rate the specialist substantially better. Research in JAMA Internal Medicine covering 9,920 visits and 502 specialists found a 9.0 percentage point higher composite patient experience rating when the referring physician and specialist had trained together, equivalent to moving from the median to roughly the 91st percentile. The proposed mechanism is mutual professional accountability: the specialist knows a respected peer will hear how the visit went.
How often do physicians refer to people they trained with? Less than you would expect. In a study of 40,495 referrals, 22.5 percent had an eligible co-trainee available, but only 8.0 percent of those eligible referrals actually went to the co-trainee, about 1.8 percent of all referrals. Roughly 92 percent of the opportunities went unused, mainly because referrers had no way to know the option existed.
Does sharing a medical school have the same effect? No, and this is one of the most instructive findings in the literature. Shared medical school produced a referral rate of 21.6 percent against a 21.4 percent baseline, statistically indistinguishable from chance. Only residency and fellowship co-training produced the effect. Trust that changes behavior appears to come from observing someone work under pressure, not from shared classroom time.
Why don't hospitals build a co-training referral tool? Because the person you trained with usually works for a competing organization. Health system referral technology exists to reduce leakage and keep referrals in network. A tool that routes patients to the best-matched co-trainee at a rival system runs directly against the interest of whoever is paying for the tool.
Is referral concentration good or bad for costs? The evidence suggests it lowers them. Research in Management Science found that a one standard deviation increase in referral concentration was associated with 7.4 percent lower spending with no measurable decline in quality. Sending patients to a tighter set of specialists you know well appears to be both cheaper and better rated.
Can Doximity or LinkedIn solve this? They hold the raw data and could add a search field. What a search field cannot reproduce is the accountability loop that generated the measured effect: the expectation that you will hear how it went, and that a peer whose judgment you respect will know how you performed. That is a norm inside a community, not a feature inside a database.
The bottom line
Medicine has spent two decades and enormous sums on levers to improve specialty care: public reporting, accountable care organizations, quality dashboards, network design. The measured effects have generally been modest.
Meanwhile there is a lever sitting in plain sight with a larger effect size than any of them. It requires no new technology, no new payment model, and no new training. It requires knowing who trained with whom.
In the best study we have, it went unused 92 percent of the time it was available.
The reason is not mystery or complexity. It is that the graph belongs to individuals, the benefit accrues to patients, and the institution that would have to build it loses money when it works. So it sits there, in a thousand dead group chats, quietly not being used.
Your residency class already knows who is good. That was never the hard part.
The hard part is that nobody wrote it down.
Part of a series on the missing professional infrastructure of healthcare. Previously: How Long Does It Take a Doctor to Find the Right Doctor?
Evidence note: primary findings are drawn from Pany and McWilliams, Health Services Research (2021), analyzing 40,495 referrals; the associated JAMA Internal Medicine analysis (2023) of 9,920 visits and 502 specialists; Management Science (2021) on referral concentration and spending; and JAMA (1998) on curbside consultation and professional relationships. Single-system referral studies may not generalize to all practice environments, and the co-training effect on patient experience is an observational association rather than a randomized finding. National referral volume figures are approximate.
