Does Arts on Prescription Save Money? Following the £216 Back to Its Arithmetic

The savings figures quoted for arts on prescription come from a single uncontrolled evaluation of a single scheme, and following their arithmetic is the fastest way to understand why they cannot carry the weight put on them. The document is a cost benefit summary of the Artlift scheme in Gloucestershire, written by a GP in December 2011 and covering 2009 to 2012. It describes itself as “a simple observational study” and states in its own conclusion that it “does not imply causality”¹.

I should admit at the start that I have repeated one of these figures myself. It was a conversation with the manager of a care home, a few years ago, about whether her budget could stretch to a weekly session, and I told her that arts programmes had been shown to save the health service a couple of hundred pounds a head. She asked me where that came from. I said a report, and heard myself say it, and realised that I had no idea which report, who wrote it, when, or on how many people. I had absorbed the number from the air in the sector and used it to persuade somebody to spend money. This page is the result of going and finding out, and the awkward part of what I found is that the author of the original document had already written the caveat I should have quoted.

Where a claim about effect appears below it carries a tier (Tier 1 supported, Tier 2 promising but limited, Tier 3 practice and anecdote), explained on reading arts in health research.

The document itself

One scheme, one county, three years, ninety patients analysed, no control group.

The essentials, before any of the figures: the evaluation covered the Artlift scheme in Gloucestershire; the analysis of healthcare spend was based on 90 patients out of roughly 500 referred over the period; there was no control group; and the author’s own description of the design is “a simple observational study”¹. Everything below sits inside those constraints, and none of the arithmetic that follows can escape them.

It is worth being clear about what kind of criticism this is. The report is not dishonest, it is not badly done for what it is, and the author printed the caveat himself. What has happened since is entirely downstream of the document: the figures were lifted out, the design was left behind, and the caveat did not travel. That is a failure of everybody who quoted it, including me.

The 37%, and the 24% printed beside it

The consultation figure most often quoted is the second half year figure, and the report’s own full year figure is smaller.

Here are the report’s own cells. GP consultations were 1,014 in the year before referral, 454 in the first six months after, and 320 in the second six months after. The report states that six months after seeing an artist, consultation rates dropped by 37%. In the same section it prints its own full year calculation and labels it: “Percentage reduction = 24%“¹.

So both numbers are in the document, describing different windows of the same data, and the larger one is the one that travels. That is not fabrication. It is selection, and it is the sort of selection that happens without anybody deciding to do it, because 37% is the more quotable half of a table.

Underneath the selection is the deeper problem, which applies to both figures equally. In a before and after design with no comparison group, a fall in consultations cannot be attributed to the scheme. People are referred at a bad point, and consultations peak around bad points; a year later they would very likely have fallen anyway. That is regression to the mean, and in this population it is not a small correction.

The 27%, and the admissions row it gets confused with

The 27% is a reduction in overall NHS spend. It is not a reduction in admissions, and it is quoted as one constantly.

The report gives spend as £157,473 before and £115,050 after, and calls that “a 27% reduction in overall spend”. Separately, it gives admissions as 54 before and 33 after, with no percentage attached and no significance test applied¹.

Look at those two rows together and the origin of the misquote is obvious. Two numbers on the same page, one of which has a percentage and one of which does not, and the percentage has migrated across to the more dramatic sounding row. “27% fewer admissions” is a sentence that exists nowhere in the source document. It is the easiest of these errors to check and the most widely repeated.

For completeness: 54 to 33 is a fall of about 39%, which nobody quotes, and which would be no more supportable if they did, because 21 admissions in an uncontrolled sample of this size, with no significance test, is not a finding.

The £216, one step at a time

The £216 is a subtraction the report never prints, and its two halves are computed on different populations.

The saving side, which totals £576 per patient.

• £42,423 is the observed reduction in healthcare spend, and the report attributes it explicitly: “Reduction observed in Healthcare spend in 90 patients = £42,423”.
• £42,423 divided by 90 patients gives £471 per patient.
• £105 is added for GP time, calculated as 2.7 fewer consultations per patient at £39 a consultation, across 500 patients.
• £471 plus £105 is £576.

The cost side, which totals £360 per patient.

• The cost of the Artlift scheme is given as £180,000.
• Spread across the 500 patients referred, that is £180,000 divided by 500, or £360 per patient.

£576 minus £360 is £216¹.

Every one of those calculations is arithmetically correct. The problem is structural. The £471 is a per patient figure derived from the 90 people whose healthcare spend was analysed. The £360 is a per patient figure derived from the 500 people referred. Subtracting one from the other produces a number that describes no population at all, because the numerator and the denominator come from different groups of people.

Consider what happens if the 410 people who were referred but not analysed saved nothing. Then the total saving is still £42,423, but the cost is still £180,000, and the scheme costs more than it saves. Nothing in the report rules that out, because the report never claims to. It is a summary of an observational study, and the £216 is what happens when somebody takes two of its figures and puts a minus sign between them.

Why the whole cluster is Tier 3 rather than a contested Tier 2

Because there is nothing better anywhere. This is not one weak study alongside stronger ones; it is one weak study in an empty room.

The 2024 systematic review of arts on prescription screened 7,805 records and included 25 studies: 10 qualitative, 7 mixed methods and 8 quantitative studies using uncontrolled before and after designs. In the review’s own words, no randomised controlled trials were identified in the search. The quantitative studies were all rated medium quality, most commonly for an absence of comparator groups, and exactly one, by Bergman and colleagues, included a treatment as usual comparison. The included programmes ran in Australia, Denmark, Sweden, the United Kingdom and the United States, in community settings, arts venues, GP surgeries, primary healthcare settings and one school, and most interventions ran for 8 to 10 weeks².

That review’s meta analysis did find a statistically significant improvement in wellbeing. That result rests entirely on uncontrolled before and after designs, so it stays Tier 3 on this site and is never described as a trial finding. And note that wellbeing is not money: a wellbeing improvement, even a real one, does not establish a saving.

A Tier 3 label is not a verdict that arts on prescription is a waste of money. It is a statement about the literature. The reasons the trials do not exist are structural: these programmes are hard to randomise, funding runs in short cycles, and the organisations delivering them put their money into delivery. The honest position is that the question is open.

The four things an uncontrolled design cannot rule out

Worth setting out in one place, because they apply to almost every economic claim in this sector.

Time passing. Many of the conditions people are referred with fluctuate. A year later, some will be better whatever happened in between.

Regression to the mean. People are referred at a bad point, and consultation rates and healthcare spend both peak around bad points. The following period is very likely to be quieter on its own.

Selection. The people who accept a referral, turn up and stay are not the same people as those who do not. They are typically more mobile, less unwell and better supported, and their subsequent healthcare use would differ for those reasons alone.

Attrition. Whoever is present at the second measurement is a different group from the one that started. In this evaluation, 90 people out of roughly 500 had their spend analysed, and the report does not establish that those 90 were representative of the 500.

All four push in the same direction, towards an apparent improvement. That is what makes a before and after design in this population so treacherous: the biases do not cancel out.

How economic claims go wrong more generally

The other classic failure is treating cheaper as better value, and there is a clean example of it in this field.

The MATISSE trial of group art therapy in schizophrenia found that total costs in the art therapy arm were lower than in the two comparison arms, and that referral to group art therapy was nonetheless judged not to be a cost effective use of resources³. Both things at once. Cost effectiveness compares what you get against what you spend, so a lower cost with no better outcome is not value, it is just less spending. That pairing gets misreported often enough to be worth stating explicitly, and the trial is covered on art therapy for psychosis.

A third failure is the number that turns out to belong to something else. The American social prescribing figure of more than 300 cultural organisations belongs to a statewide programme rather than the pilot it is attached to, and the independent peer reviewed evaluation of that pilot reports 12 cultural organisations, 20 healthcare providers, 414 referrals, 363 participants and participant data on 84⁴. More on that at arts on prescription around the world.

What a real economic answer would require

Stated plainly, so that it is clear this page is not simply refusing to be satisfied.

• A control group, so that the counterfactual is measured rather than assumed. Without it, none of the arithmetic above means anything, however carefully it is done.
• The same population on both sides of the sum. Savings and costs computed on the same denominator, and stated as such.
• Costs counted fully, including the artists’ time, venue, coordination, the referral process and the link worker time, not only the grant that paid for the sessions.
• A stated time horizon. A saving over twelve months is a different claim from a saving that persists, and the second one has never been demonstrated here.
• A stated perspective. Savings to the health service, to the local authority, to the participant and to society are four different sums, and reports routinely slide between them.
• Sensitivity analysis, showing what happens to the conclusion when the assumptions move.

Until something like that exists, the honest sentence for a funding bid is that participants value these courses, that the wellbeing signals are consistent though uncontrolled, and that no saving has been demonstrated. That is a weaker case and it is one that will not fall over when somebody checks it.

What this changes for a reader

Very little about whether to attend, and quite a lot about what to believe.

If you have been offered a place, nothing here is a reason to turn it down. People who attend these courses very often value them, and the qualitative half of the literature is consistent about social connection, psychological benefit and routes onwards². That is worth having and it does not depend on a saving.

If you are commissioning, funding or running a scheme, the practical implication is sharper. Build the case on what is documented, which is that people value the courses and that a service exists to connect them, and not on a per patient saving that came from a subtraction across two populations. The National Academy for Social Prescribing describes how the referral machinery is meant to work⁵, and the sector’s own reports are gathered in one place⁶, where they should be read as descriptions of practice rather than as evidence of effect. That distinction is worked through on creative health policy and reports, and the funding cycles behind all of it on who pays for arts on prescription.

And if somebody quotes you a figure from this cluster, the question that settles it in about a minute is: is that the 37% or the 24%, and was the 27% about spending or about admissions. Nothing on this page is a reason to change, delay or decline any part of anybody’s treatment, and no arts course substitutes for one.