Music Therapy for Dementia: What the 2025 Cochrane Review Actually Found

The 2025 Cochrane review of music based therapeutic interventions in dementia reports a small effect on depressive symptoms and moderate certainty evidence of no meaningful effect on agitation, which is the outcome the intervention is most often recommended for. It covers 30 studies and 1,720 participants randomised, it was published on 7 March 2025 as CD003477.pub5, and it is quoted a great deal more often than it is opened. A large share of what circulates about it is a description of the previous version.

I went to it wanting it to say something else. My father’s home had told me, sincerely and more than once, that music settled him in the evenings, and by the time I read the review I had built a working week on the assumption that this was a known thing rather than a hoped one. The row I had to accept is the agitation row. It does not say that nothing happened in his room at seven o’clock, and it does not say the staff who told me that were making it up. It says that when trials randomised people with dementia and measured agitation on a scale, the pooled difference between the groups was SMD -0.05, with a narrow interval sitting on zero, and that the reviewers had reasonable confidence in that. I did not stop playing in the evenings. I stopped telling families it would settle anybody.

Every claim about effect on this page carries a label: Tier 1 supported, Tier 2 promising but limited, Tier 3 practice and anecdote. The third label means two quite different things, and this page says which one it means each time it uses it. The checks that produce all three are set out on reading arts in health research.

What is actually being reviewed, which is a category and not a thing

The review is of music based therapeutic interventions, and that phrase is a container rather than a synonym for one activity. Reading it as though it meant “a session with a registered music therapist” is the first and most consequential mistake available here.

The authors chose that wording in preference to “music therapy”, and the choice is doing work. A category assembled that way holds together practices that look nothing like each other in a corridor: active music making and receptive listening, individual work and group work, sessions run by a registered therapist and structured music activity delivered by somebody trained to deliver it. Pooling those is a defensible research decision, because the alternative is a set of literatures too small to say anything about. It also means the pooled answer belongs to the category and not automatically to any one member of it.

That cuts in both directions, and this is the part that gets skipped. A reader who takes the depressive symptoms result as a finding about registered music therapy specifically has overclaimed. A reader who takes the agitation result as a refutation of a particular practice, say a personalised playlist delivered through headphones by a care assistant, has done the same thing in reverse. What the review supports is a statement about the category as pooled. Where a specific practice needs its own answer, it needs its own evidence, which is why personalised playlists and dementia choirs and singing groups have separate pages with separate labels rather than inheriting this one.

If you want the shape of the clinical version, what happens in a music therapy session describes it, and music therapy against community music sets out the boundary that the review’s title deliberately declines to draw.

The review in full, including the row nobody quotes

Here is the whole of what the 2025 review supports, in the units it reports, with the tier attached to each line¹.

Outcome	Result	Certainty	Tier
Depressive symptoms	SMD -0.23 (95% CI -0.42 to -0.04)	Moderate	Tier 1
Agitation and aggression	SMD -0.05 (95% CI -0.27 to 0.17)	Moderate	Tier 1, null result
Any effect persisting 4 weeks after treatment ends	No evidence of one		Tier 3
Quality of life	No separate estimate exists in the review		Nothing to quote

The last row is the one nobody quotes, and it is the reason this table has four lines rather than three. The review contains no separate quality of life estimate. Not a small one, not an uncertain one, not one the reviewers downgraded. There is no figure there to report. So when a service description, a funding bid or a news piece attributes a quality of life improvement to this review, the number has come from somewhere else and been attached to a citation that does not hold it. The useful question in that situation is not whether the number is big; it is which document it came from, and whether that document was a trial or an evaluation.

The depressive symptoms row is the one honest positive. SMD -0.23, interval -0.42 to -0.04, moderate certainty. The interval does not cross zero, the certainty rating is reasonable, and the effect is small. All three of those facts belong in any sentence that quotes it. A page reporting only that music produced a significant improvement in depressive symptoms would have said something true and left the reader with a false picture of the size.

The four week row deserves its own sentence because it is easy to misread. No evidence of any effect persisting four weeks after treatment ends is Tier 3, and it is the not adequately demonstrated kind rather than the demonstrated null kind. The distinction is exact: on agitation the review reports a measured absence of effect, whereas on persistence it reports that it could not find evidence of one, which is the weaker of the two statements. That matters for how services are commissioned. A block of sessions funded on the assumption that it leaves something behind is being funded on a premise the evidence does not currently support in either direction.

The agitation result, which is a finding and not a gap

Agitation is the outcome music is most often offered for in dementia care, and it is the outcome where the pooled evidence is clearest that nothing measurable is happening. That is an unusual and strong thing to be able to say.

Look at what the row consists of. SMD -0.05 is essentially nothing. The interval, -0.27 to 0.17, is narrow, which means the trials were consistent enough with each other to rule out a large effect in either direction. And the certainty rating is moderate, which means the reviewers thought the underlying trials were good enough to be worth believing. Put those three together and you do not get “we do not know yet”. You get a positive finding of absence: on the scales used, across the trials pooled, music based interventions do not meaningfully change agitation. That is Tier 1, null result, and it is the single most important number on this site.

I want to be precise about the boundary of that claim, because the temptation to over read it is as strong as the temptation to over read a positive. It is a statement about pooled group averages on agitation scales in trials of this category. It is not a statement that a distressed person cannot be helped by somebody singing to them, and it is not a licence to tell a care home to stop. Scales measure what they were built to measure. A man who is calmer for eleven minutes and rated the same at the weekly review has had a better eleven minutes, and the review would not detect it.

What the row does forbid is a specific and very common sales pitch: that arranging music will reduce agitation in a home, on a ward, or in one person’s evenings. That claim now has moderate certainty evidence against it, and repeating it after March 2025 is not optimism, it is a citation failure. When a provider makes it, ask which review they are quoting. If the answer is a Cochrane review of music in dementia, the answer to give back is the interval.

What SMD -0.23 means, and why it is not about the person you are asking for

A standardised mean difference expresses a gap between group averages in standard deviations, because the pooled trials measured the same idea on different questionnaires. SMD -0.23 is roughly a fifth of a standard deviation, which is conventionally a small effect.

The important property of that number is not its size but its type. It is a difference between the average of one group and the average of another. Inside the trials that produced it, some participants improved a great deal, some did not change, and some got worse, and the average sits where it sits because those experiences cancelled and combined. Nothing in the figure predicts which of those a named person will be, and no honest reading of it can promise an individual outcome. That caveat is not a hedge attached to weak findings; it applies with exactly the same force to the strongest evidence in this whole field.

There is a second thing an SMD conceals, which is what a fifth of a standard deviation looks like on a Tuesday. On most depression scales used in dementia research it is a shift of a point or two, distributed unevenly. It is the kind of change a relative might notice over a month and might not, and it is nothing like the transformation that a clip of somebody singing along suggests. Both things are real. The clip is real, and the average is small, and the gap between them is what this page exists to hold open.

None of that makes the finding unimportant. A small effect on low mood, at moderate certainty, from an intervention with essentially no harms attached, is a perfectly good reason to offer somebody music. It is not a good reason to promise their family anything.

The version trap: 22 studies and 1,097 participants

If a source tells you the Cochrane review of music in dementia covered 22 studies and 1,097 participants, that source is describing the 2018 version, which has been superseded.

The current version is CD003477.pub5, published 7 March 2025, covering 30 studies and 1,720 participants randomised. The earlier version reached different conclusions, notably on agitation, which is precisely the claim most likely to be repeated from it. This is not a pedantic distinction between two similar documents. It is the difference between a source that supports a widely made claim and a source that contradicts it.

The trap persists for two reasons. Secondary coverage compounds: a 2023 article quotes the 2018 review, a 2026 page quotes the 2023 article, and the numbers arrive with a fresh date and no version number at all. And the update runs against the direction of the sector’s own interest, so nobody has been eager to carry it. I have seen the older figures in material published in 2026, in good faith, by people who would correct them immediately if anyone pointed at the record.

The check takes ten seconds. Look at the pub suffix on the identifier and the publication date on the Cochrane record itself, never the date on the page quoting it. If the page does not name a version at all, treat the figures as unverified. The general version of that habit, applied to every kind of source, is on reading arts in health research.

What NICE guidance does and does not say

NICE does not recommend music therapy for dementia. A full text search of NG97, the guideline on dementia assessment, management and support for people living with dementia and their carers, returns zero occurrences of music, art, dance, drama or creative activity².

It matters to state what the guideline does contain rather than only what it lacks, because the bare absence invites a misreading in the opposite direction. NG97 does address non pharmacological approaches to distress and agitation, in general terms, and it does so before it addresses medication. Nothing in it discourages music. A home offering a singing afternoon is not acting against guidance, and nobody should be told that it is.

What the guideline does not do is name any arts intervention, anywhere, in any recommendation. So “NICE recommends music therapy” is not a strong claim phrased loosely. It is a claim about the contents of a specific document, and the document does not contain it. That is a different category of error from an overstatement, and it is worth naming as such, because it cannot be fixed by softening the wording. There is no weaker true version of it.

It survives because it is useful in funding bids and because verifying it requires opening a guideline that nobody in the room has a copy of. If you are writing a bid, the honest form of the sentence is that NG97 supports non pharmacological approaches to distress in general terms and names none.

The scale that keeps the claims alive

57 million people worldwide were living with dementia in 2021, according to the WHO fact sheet updated on 3 July 2026³. That number is the reason a small effect on one outcome is worth arguing about at all, and it is also the reason the overclaiming is so hard to dislodge.

Two things follow from a figure that size. The first is that even a small average effect, spread across a population like that, represents a great deal of low mood shifted, and it is entirely rational for services to want it. The second is less comfortable: a population that large, with families desperate for something to do, is an enormous market for confident claims, and confident claims travel faster than intervals do. Most of the people repeating the agitation claim are not selling anything. They are quoting somebody who was.

One narrow point about the fact sheet, since it is quoted constantly. The current version carries no 2050 projection. If you need a projection you have to name one and its source and its year, rather than attributing it to the fact sheet, which no longer has one to attribute. Any page that writes “WHO projects” and then a 2050 figure is citing something other than the document it names.

Against that scale, the supply side is small. In July 2026 the HCPC register listed 6,103 arts therapists in total across all four protected titles⁴, which covers art psychotherapists, art therapists, dramatherapists and music therapists between them. Availability, rather than evidence or cost, is what usually determines whether anyone with dementia in the UK gets access to a registered music therapist. How to find a music therapist sets out the registers and the realistic routes.

What to do with this if you are deciding whether to ask

If you are choosing whether to ask for music for someone with dementia, the evidence supports a modest expectation about mood, no expectation about agitation, and no expectation of anything lasting once it stops. That is a thinner set of promises than you will be offered, and it is not a reason to decline.

Three practical consequences. First, ask what is actually being offered and by whom, because the review’s category holds several very different things and the arrangement, the cost and the waiting list differ enormously between them. A registered therapist’s course of work, an activities coordinator’s singing hour and a set of headphones are three distinct offers, described respectively on music therapy, arts in care homes and personalised playlists. Second, if a provider justifies the offer by saying it reduces agitation, that is now a checkable error rather than a matter of emphasis, and it tells you something about how the rest of their claims were assembled. Third, whatever is arranged, the person’s own reaction over two or three sessions is better evidence about them than any pooled average, including the ones on this page.

What happens in the room also has to be organised properly. Anything delivered to somebody who cannot readily consent, in a home or on a ward, runs straight into consent and safeguarding, and what good practice looks like day to day is on creative activity in dementia care.

None of this is a reason to change anybody’s medication, treatment or care plan, and music is not a substitute for any of them. What it is, on the current evidence, is a low risk thing with a small measured benefit for low mood and an honest blank where the biggest claim used to be. My father’s home was not lying to me. They were describing something they had watched happen, in a language borrowed from research that had not found it. Both halves of that are worth keeping.