Referral leakage, measured

The number is mostly an artifact of four choices

Two teams inside the same organization will produce leakage rates that differ by a factor of two, both correctly, because they made different decisions about four things that nobody wrote down. Until those decisions are explicit and shared, the rate is not a measurement of the world — it is a measurement of the analyst's assumptions, and every subsequent conversation about improvement is unfalsifiable.

What counts as a referral. An order placed in the record is the cleanest event, and it undercounts badly, because a great deal of real referral activity happens as a verbal recommendation, a note in an after-visit summary, or a front-desk handoff that never becomes an order. If your organization does not have disciplined referral ordering, your denominator is a subset of referrals shaped by which departments happen to document well.

What counts as leaving. The patient saw an outside specialist. But there are gradations: seen outside once and returned, seen outside for the entire episode, seen outside for a service you do not offer, seen outside because your own clinician sent them there deliberately. The last of those is not leakage in any meaningful sense, and it is frequently counted as leakage because the data cannot tell intent from outcome.

The attribution window. Thirty days, ninety, one hundred eighty. The rate rises with the window because more of the world gets included, and comparisons across windows are meaningless. Choose one, write it down, and never quote a figure without it.

What counts as in. Employed clinicians only, or the wider affiliated group, or everyone in the clinically integrated network. This choice alone can swing the result substantially, and it tends to get changed quietly when the number is unflattering.

What each data source can and cannot see

Every leakage measurement is limited by the same thing: you observe care delivered inside your walls in full, and care delivered elsewhere only through whatever channel happens to tell you.

Read the second and third rows together. For a fee-for-service commercially covered patient with no risk arrangement and no exchange participation in your region, you have essentially no visibility into whether they were seen elsewhere. They are not in your leakage numerator, not because they stayed, but because you cannot see them.

That sentence is the whole discipline. A leakage figure without a stated observability share is not a finding; it is a number of unknown provenance. Report it as of the referrals where we could observe an outcome, this many were completed outside — and report separately what share of all referrals that observable subset represents. If the observable share is small, say so before the rate rather than in a footnote, because the two figures together support a decision and the rate alone does not.

The largest bucket is often not leakage

When organizations first get outcome visibility on referrals, the finding that surprises them is not where patients went. It is how many never went anywhere at all.

A referral is placed. No appointment is made, or one is made and not kept, and there is no record of care with anyone, inside or outside, for that specialty in the following months. In the analyses we have seen built carefully, this group is frequently comparable in size to the group that was seen outside, and sometimes larger. It gets buried because a report that only classifies inside versus outside has nowhere to put it, so it lands in whichever bucket the join defaults to.

Separate it out, and it changes the priority order. Referrals that were never completed are a care gap first and a revenue question second, and they are the easiest of the three groups to act on: the patient wanted the referral, nothing about competition or network adequacy is involved, and the interventions are ordinary operational ones — closing the loop on the order, making the appointment before the patient leaves, following up when it is not kept. Any organization that finds a large never-completed bucket should work that before anything else on this page.

What the analysis should produce instead of a rate

A single organization-wide percentage supports no decision. The useful output is a ranked list, and each row is a place where something could be done.

Cut the data by specialty, by geography at a granularity where travel time is meaningful, by referring clinician or practice, and by the destination. Attach the volume and, where the arrangement makes it meaningful, the value. Then attach the most likely reason, because the reason determines whether anything can be done at all.

The reasons are a short list and they are separable with data you have or can get cheaply:

No available appointment. Compare referral dates against actual appointment availability for that specialty and location. If the wait is long, the referral pattern is explained and the fix is capacity, not messaging. This is the most common actionable driver we see, and it is the one measured least often.

Coverage. The destination accepts the patient's plan and the internal option does not, or the patient believes so. This is directly checkable against contract data, and it overlaps heavily with directory accuracy problems.

Geography. The outside option is materially closer. Compute the travel-time difference rather than assuming, and accept that a large share of what gets called leakage is a patient making a sensible choice.

Clinician preference. A referring clinician sends nearly everything to one destination, often a long-standing relationship. Visible immediately in a by-clinician cut. What to do about it is a leadership and compliance question, not an analytics one.

Service does not exist internally. Trivially explained, frequently mislabeled as leakage, and worth removing from the number entirely so the remaining rows are real.

Once those five are subtracted, what remains is usually much smaller than the headline figure and much more actionable, because every remaining row has a name and a cause.

Where models help, and where they mislead

The common pitch is a model that predicts which patients will leave. It is technically feasible and usually the wrong instrument, for two reasons.

The first is that the label is contaminated by observability. You are training on patients whose outside care you can see, which is a non-random subset defined by coverage type and region. A model trained on that subset learns the shape of your data-sharing arrangements as much as it learns patient behavior, and it will generalize badly to exactly the patients you cannot see — who are the ones you were trying to learn about.

The second is that a per-patient prediction usually has no attached action. Knowing that a specific patient is likely to be seen elsewhere does not tell anyone what to change, and the space of legitimate interventions is narrower than in most industries.

Two modeling problems in this area are worth real effort. Capacity and access analysis — forecasting appointment availability by specialty and location, and quantifying how availability relates to referral completion. That is a forecasting problem with a clean signal and an obvious lever. And attribution and matching — deciding whether an outside claim and an internal referral concern the same patient and the same episode, across datasets with no shared key. That is entity resolution, it is genuinely difficult, and getting it wrong corrupts every downstream number without announcing itself.

Mistakes we see

• A leakage rate with no stated attribution window, compared against a figure computed on a different one
• The observable population reported as the whole population, with no observability share given
• Never-completed referrals counted as leakage, or silently dropped by the join
• Services the organization does not offer left in the numerator
• The definition of “in-network” changed between reporting periods
• No availability analysis, so a capacity problem is treated as a loyalty problem
• Patient matching across datasets done loosely, with no measured error rate
• A predictive model where a ranked list by specialty and geography would have answered the question

When you do not need help with this

If you are a single practice referring to a handful of specialists, the answer is in a conversation with your front desk, not in an analysis. They know which referrals do not get appointments and why, usually to the week.

If you already have referral orders and internal encounters in a warehouse, the never-completed analysis — orders with no subsequent internal encounter and no outside record — is a day of SQL and it is likely the most valuable output on this page. Do it before commissioning anything.

And if the only data you have is your own records, the honest answer is that you cannot measure leakage, and no vendor can measure it for you from that data either. What you can measure is referral completion inside your organization, which is a real and useful number with no inference in it. Buying a leakage product before there is any outside visibility purchases a confident figure derived from an eight-percent view.

The work worth doing with outside help is the harder end: assembling several outside data feeds into one longitudinal view, resolving patients across them with a measured error rate, and building the availability analysis that tells you whether the pattern is a preference or a queue.

What a defensible measurement has

• A written definition of referral event, completion, window and network boundary
• The observability share reported alongside every rate, in the same sentence
• Three outcome categories, not two: completed inside, completed outside, never completed
• Services not offered internally excluded from the numerator
• Results cut by specialty, geography, referring practice and destination
• Appointment availability measured beside the referral data, on the same cuts
• Patient matching with a measured error rate in both directions
• Definitions frozen across periods, with changes versioned and disclosed
• Value attached only where an arrangement makes it meaningful
• Counsel involved before any intervention is designed, not after

Bottom line

Referral leakage is measurable, but only for the fraction of patients whose outside care you can observe, and the first honest act is to state that fraction. Fix the definitions before computing anything, because they move the number more than any program will. Split never-completed referrals out as their own category — they are often the largest group, they are a care gap rather than a competitive loss, and they are the easiest thing on this page to act on. Then produce a ranked list with a probable reason attached to each row, and measure appointment availability beside it, because a queue and a preference look identical in the data and require entirely different responses.

Frequently asked questions