In 2016 Scotland stopped paying its GPs through the Quality and Outcomes Framework, the incentive scheme that had rewarded practices for hitting targets on chronic disease care since 2004. England kept it. That created one of the cleanest natural experiments in health financing.
A 2023 study in the BMJ compared the two. One year after the incentives ended, recorded performance in Scotland had fallen significantly on 12 of 16 quality indicators relative to England. Three years on, the widest gap was in tick-box recording of mental health care planning, about 40 percentage points below England. But it was not only paperwork. Blood pressure control among patients with hypertension, stroke and diabetes fell too.
People read that result in opposite ways. Enthusiasts say it proves incentives work. Sceptics say it proves the incentives were buying documentation. I think it shows something more awkward than either: when you pay for a result, you are renting behaviour, and the rent never stops.
That is the real choice underneath the fashionable slogan of “paying for outcomes instead of procedures.” It is not a choice between paying for the wrong thing and the right thing. It is a choice about who carries risk, what can be measured honestly, and how long you are willing to keep paying.
Providers respond to what they control
The best evidence on performance pay in low-income settings still comes from Rwanda. A randomised evaluation published in The Lancet in 2011 compared 80 facilities that received performance-based financing with 86 that received an equivalent increase in ordinary budget. The researchers isolated the effect of the incentive from the effect of simply having more money.
Facilities paid for performance saw a 23% increase in institutional deliveries and large increases in preventive care visits for young children. But there was no improvement in the number of women completing four prenatal visits, or in full immunisation.
That split is the most useful lesson in the study. A delivery in a facility is something staff can influence directly: outreach, a welcoming midwife, a clinic that is open. Four prenatal visits depend on a pregnant woman returning again and again, through distance, cost and work. Incentives moved the outcomes within the provider’s reach and did little for those that were not.
“Paying for outcomes” sounds like paying for health. In practice, payers end up paying for whatever part of health a provider can control and a data system can capture. Those are narrower things.
Outcome metrics can hurt the people they measure
The American experience adds a darker note. Medicare’s Hospital Readmissions Reduction Program, created by the Affordable Care Act, penalises hospitals with higher than expected 30-day readmissions for conditions including heart failure and pneumonia. Readmissions fell. That looked like an outcome-based success.
A 2018 study in JAMA, covering 8.3 million hospitalisations, found that after the programme was announced and implemented, 30-day post-discharge mortality rose for heart failure and pneumonia patients relative to earlier trends, and that much of the increase was among patients who died without being readmitted. The authors were careful: the design could not prove the policy caused the deaths, and mortality within 45 days of admission did not rise significantly. But the finding haunted the debate for good reason. A readmission is a measurable event. A patient who should have been readmitted and wasn’t does not show up as a failure on the metric.
This is the core hazard of outcome payment. The outcome you can count is rarely the outcome you care about. It is a proxy, and proxies bend.
The risk transfer nobody prices
Paying for outcomes also moves financial risk from the payer to the provider. If a clinic serves a poorer, sicker, harder-to-reach population, its outcomes will look worse regardless of effort. Risk adjustment helps, but only for the risks the data can see.
Providers understand this even when payment designers do not. The rational response to outcome payment is to select patients whose outcomes are likely to be good, and to steer away from the rest. In a fee-for-service system, a complicated patient is a source of revenue. In a poorly designed outcomes system, the same patient is a liability.
That does not mean fee-for-service is better. Paying per procedure has its own well-documented pull toward overtreatment. It means that each payment model has a characteristic way of failing, and the choice is really about which failure a system can best detect and contain.
The mechanism: attention is the thing being bought
Why would incentives change recorded quality so quickly, and why would it fall back so quickly when they are removed? The answer is that performance payment mostly buys attention.
A general practice or a health centre has far more possible tasks than time. Clinicians and managers constantly decide which ones get done thoroughly, which get done quickly and which get skipped. An incentive tells them which tasks the payer cares about. Staff build templates, reminders and recall systems around those tasks. The receptionist knows which patients are overdue for a check. The practice manager watches a dashboard.
Remove the payment and nothing forces those systems to be dismantled. But nothing keeps them maintained either. The template is not updated. The recall list lapses. Staff turnover removes the person who knew how it worked. Within a year or two, attention has moved elsewhere, often to whatever the next incentive or inspection is focused on.
This is why the “rented behaviour” framing matters. The incentive does not change what providers value or know. It changes what they prioritise under pressure. That can be valuable. It is also inherently temporary unless the priority is embedded some other way, through professional norms, clinical software, training, or contracts that make the task a standard expectation.
A scenario: designing a maternal health incentive
Imagine a ministry designing a performance scheme for rural health centres, with the goal of reducing maternal deaths. The ideal outcome to pay for is obvious: fewer women dying. It is also unusable. Maternal deaths in a single health centre’s catchment are rare, so the number jumps around by chance from year to year. The centre has limited control over many causes. And the data are unreliable, because deaths at home may never be recorded.
So the design team moves down the chain. They consider paying for facility deliveries, which a centre can influence and which are easy to count. They consider paying for four antenatal visits, which matter clinically but depend heavily on women’s ability to travel. They consider paying for the quality of each delivery, measured by a checklist, which requires expensive verification.
Each step away from the outcome makes the indicator easier to measure and more open to gaming. Facility deliveries can be recorded that never happened, or referred cases can be counted twice. Checklist scores can be filled in after the fact. The scheme then needs verifiers, counter-verifiers and community surveys to check the data, which can consume a real share of the programme’s budget.
By the end, the ministry is not paying for fewer maternal deaths. It is paying for a set of countable, verifiable activities that it believes are linked to fewer deaths, plus a verification apparatus to keep the counts honest. That can still be worth doing. The Rwanda evaluation suggests it can move real behaviour. But it is a long way from the slogan.
The strongest counterargument
The best defence of outcome payment is that the Scotland results prove my point in the other direction. If performance fell when incentives were removed, then the incentives were doing something real, including on blood pressure control, which is a clinical outcome and not just a tick box. Paying indefinitely for that may simply be the price of better care, and a reasonable one.
This is partly right. Some rented behaviour is worth renting forever. If a modest incentive keeps practices checking the blood pressure of patients with diabetes, and the checks prevent strokes, the payment may be excellent value even if it never becomes self-sustaining.
But it cuts against the usual pitch. Performance payment is often sold as a transitional tool: pay for a few years, change the culture, then withdraw and keep the gains. The evidence from the UK suggests that is optimistic. If the gains depend on continued payment, then performance schemes are a permanent budget commitment, and they should be designed and costed as one.
There is a second defence: that fee-for-service is worse, because it rewards volume without regard to value. That is also true. The argument here is not for fee-for-service. It is that outcome payment has its own characteristic failures, and that those failures are easier to hide because the metric itself looks like success.
Second-order effects
Performance schemes change more than the targeted indicators.
They can crowd out the unmeasured. Staff attention is finite. When some tasks carry payment and others do not, the unpaid tasks may suffer. The patient with a complex problem that does not fit any indicator gets less time than the patient whose visit can be ticked off.
They change professional motivation. Many clinicians work in health care because they care about patients, not because of piece rates. Some research on incentives suggests that external rewards can, in certain conditions, weaken intrinsic motivation. The effect is debated, but schemes that treat professionals purely as responders to money risk making it more true.
They reshape data systems. When payment depends on recorded data, the data become an object of negotiation, not only of care. Coding practices shift. Information systems are redesigned around what gets paid. That can improve data quality where it was poor. It can also make the data less useful for anything other than payment.
What this means for each actor
- For ministries and purchasers: treat every performance indicator as a permanent commitment unless there is a credible plan to embed it in standards or software. Budget for verification from the start.
- For hospital and clinic leaders: watch what happens to unincentivised work. If it is visibly declining, raise it with the purchaser before it becomes a safety problem.
- For funders: results-based financing is attractive because it looks accountable. Ask for evidence on what happened to unmeasured services and what happened after payments ended.
- For clinicians: push for indicators that reflect what matters clinically, and for companion measures that would reveal harm, such as mortality alongside readmissions.
What to watch next
Several countries are moving from fee-for-service toward capitation and bundled payments with performance components. The questions to watch are whether evaluations report what happens when incentives are withdrawn, whether they measure unintended consequences as seriously as intended ones, and whether the cost of verification is counted as part of the programme.
A more honest design
If outcome payment rents behaviour rather than buying it, a few design principles follow.
- Keep the incentive a minority of income. Enough to steer attention, not so much that survival depends on the metric.
- Pay for what providers control, and say so. Processes with strong evidence, like a foot examination for a diabetic patient, are often better targets than distant outcomes.
- Budget for permanence. Scotland shows that removing incentives can undo measured gains. If an indicator matters, plan to keep paying for it or to embed it in contracts and professional norms first.
- Watch the unmeasured. Every outcome metric needs a companion measure of what gaming would look like, such as mortality alongside readmissions.
- Blend. Capitation or global budgets for the base, with a small layer of targeted performance pay, usually beats either extreme.
The shift from paying for procedures to paying for outcomes is a good direction. It is not a destination. The payer who believes it has finally started paying for health has usually just stopped looking at what the metric misses.
Which performance incentive in your system has worked best, and what did it quietly cost somewhere else?
