Are You Evaluating a Service or a Community? Pick One

A programme that runs a weekly walking group for twelve weeks and a programme that runs a weekly walking group indefinitely, with the same members showing up for years, are not the same intervention. They will look identical on an attendance sheet. They should not be evaluated the same way, and most funders currently ask them to be.

This matters because the choice of what you are building — a service or a community — determines what “success” looks like, and getting it backwards produces exactly the wrong reading of your own programme. A service that succeeds discharges people. A community that succeeds retains them. If you apply service logic to a community programme, retention looks like failure to move people on. If you apply community logic to a service, low retention looks like it should, but you’ve built no case for anything lasting.

The distinction, stated plainly

A service delivers a defined input to a defined need and closes the loop. Referral, session, outcome, discharge. Social prescribing, as usually designed, is a service: a link worker connects someone to an activity, the activity addresses a specific goal, and the relationship with the service ends. Systematic review evidence on social prescribing supports this framing — gains reported are in self-esteem, confidence, and reduced use of GP or emergency services, measured over a defined period with a defined endpoint.

A community has no discharge point. Its value is precisely that people keep coming back without a programmatic reason to. Robert Putnam’s account of associational decline and Ray Oldenburg’s concept of the third place are both describing this: informal, recurring, unscheduled contact that accumulates into what Putnam calls social capital. Eric Klinenberg’s argument about social infrastructure makes the same point about physical space — libraries and parks work because they are used repeatedly, by the same people, without anyone booking a session.

Most funded programmes are, in practice, a service wearing community language. That mismatch is the actual subject of this piece.

Why the qualitative evidence points at the seam

A 2022 meta-synthesis of how people experience social prescribing found something the outcome metrics miss: participants describe the benefit as extending beyond social contact itself, toward restored meaningful participation and a sense of purpose. Structured, purposeful group activity reads, in their accounts, as more effective than contact alone. That is a community-shaped benefit showing up inside a service-shaped programme. The people running the intervention measured the service outcome; the people receiving it experienced something closer to belonging.

This is the seam where most evaluation goes wrong. The programme was commissioned and funded as a service — twelve weeks, a referral pathway, a discharge criterion — but the actual mechanism of benefit, if the qualitative accounts are right, is the ongoing, undirected relationship. Measuring only the service metric (did symptoms improve at week 12) will show a modest effect and miss the reason it happened.

Why the trial evidence cuts the other way

Here the guidance has to be blunt, because the two most careful randomised trials in this literature both found structured, service-like intervention outperforming unstructured, community-like contact.

The HEAL-HOA volunteering trial and its follow-up behavioural activation trial both used befriending — unstructured, ongoing, relationship-first contact, the closest thing in the intervention literature to “community” — as the comparison arm, not the intervention being tested. In the 2026 trial, telephone-delivered behavioural activation and mindfulness, delivered in eight structured half-hour sessions over one month, reduced loneliness at twelve months significantly more than befriending did, in a sample of 1,151 older adults living in poverty, alone, and digitally excluded. A separate 2025 randomised trial found that befriending in residential aged care did reduce loneliness scores meaningfully at eight and sixteen weeks — it is not that unstructured contact does nothing. It is that head-to-head against a structured intervention with a specific psychological mechanism, it lost.

Read together, this says something uncomfortable for community-building as a stated goal: if the outcome you are funded on is a validated loneliness score at a fixed follow-up point, structure beats openness. Communities are slow and diffuse by design. Trials reward fast and targeted.

What this means for evaluation design

You cannot resolve this by better wording of your theory of change. You have to decide, before you write the logic model, which one you are running — and then measure accordingly.

If you are running a…	Success looks like	Right metric	Wrong metric
Service	Defined need addressed, participant exits	Score change at a fixed endpoint; discharge rate; reduction in downstream service use	Long-term retention (retention is not the goal — resolution is)
Community	Recurring, self-sustaining contact with no exit criterion	Return rate over 6–12+ months; ratio of unprompted to programmed contact; member-initiated activity	Symptom score at week 12 (too early, and not the mechanism)

What this means in practice: before you write a single outcome measure, answer one question in the funding proposal itself — is there a discharge criterion, or not? If there is, you are running a service: measure change against a validated scale at a defined follow-up, the way the HEAL-HOA trials did. If there isn’t, say so explicitly, and measure durability and self-generated activity instead. Do not let a funder impose a discharge-shaped metric on a programme that was never designed to discharge anyone.

The named failure mode: the graduation trap

Programmes fall into what can usefully be called the graduation trap when they design for community — open-ended, member-led, no fixed endpoint — but report to funders using service metrics, because service metrics are what commissioning templates ask for. The result is a programme that shows a disappointing twelve-week outcome (because twelve weeks was never the unit of change) while quietly running a group that has met every month for three years. The funder sees stagnation. The members see the thing working exactly as intended. Nobody has the numbers to show either.

The fix is not better spin. It’s picking one evaluation frame per programme and stating it in the commissioning conversation, before the first cohort starts, so a low twelve-week score isn’t read as failure when the actual claim being made is about month twenty-four.

The evidence-status table

Claim	Evidence status
Social prescribing produces measurable service-level gains (self-esteem, confidence, reduced GP/A&E use)	Reasonably supported by systematic review, though heterogeneous and short-follow-up
Structured intervention outperforms unstructured befriending on loneliness scores at fixed follow-up	Supported by two RCTs — but this is one comparison, not a general law
Unstructured, ongoing communal contact has independent long-term value distinct from symptom reduction	Argued persuasively in the qualitative and social-infrastructure literature; not tested in a trial designed to detect it
Retention is a meaningful success metric for a service	Weak — retention in a service usually signals the service isn’t working
Retention is a meaningful success metric for a community	Reasonable, but still under-specified — no agreed threshold for what counts as durable

What this does not solve

None of this tells you which model to build. That’s a resourcing and mission decision, not an evaluation one. It also does nothing for the deeper measurement gap in this field: there is no validated instrument for “durable, self-sustaining community” the way there is for loneliness scores, so even a well-designed community evaluation will lean on proxies — return rate, member-initiated contact — that funders may find less legible than a twelve-week score, however misleading that score is for this purpose. And as with nearly everything in this literature, both models measure people who showed up. Neither tells you anything about the people who didn’t.