What Social Prescribing Evaluations Actually Measure, and What to Copy

Most loneliness evaluations fail before the programme even starts, because someone picks the outcome measure in a meeting that also covered three other agenda items. A systematic review of social prescribing initiatives found nine studies claiming positive impacts on loneliness, and every one of them used a different combination of tools, timeframes, and comparison groups. That is not a criticism of the field. It is what happens when an entire delivery model — link workers referring people to community activity — gets built faster than the measurement infrastructure to support it.

Social prescribing services are a useful case to study precisely because they have already made every mistake and a few good choices, in public, at scale. What they’ve settled on, imperfectly, is a smaller set of workable measurement moves than the size of the literature suggests.

What the published evaluations actually did

A 2021 systematic review of social prescribing and wellbeing found that the outcome most consistently reported was not loneliness itself but self-esteem and self-confidence. That is a tell. Programmes that set out to reduce loneliness ended up measuring something adjacent because it was easier to detect change in over a typical delivery window, and because the qualitative accounts kept surfacing it. A separate review focused specifically on loneliness found all nine included studies reported positive individual impact, but only three reported any reduction in service use — GP visits, emergency attendance, social worker contact, inpatient admission. The loneliness reduction claims mostly rested on pre/post self-report using whatever scale the local team had to hand.

A 2022 qualitative meta-synthesis of how people receiving social prescribing describe the benefit adds a specific and useful finding: participants describe the value as extending beyond social contact to something closer to restored purpose and meaningful participation. Structured, purposeful activity reads as more effective in these accounts than unstructured social contact. That is a claim about mechanism, not just outcome, and it should change what a programme decides to measure — not just “did loneliness go down” but “did this person get something to do that mattered to them.”

None of this is a controlled evidence base. The 2023 review mapping the state of loneliness and social isolation research names inconsistent measurement as a structural barrier to comparing findings across studies, and it is right. But inconsistency across the field is different from inconsistency inside your own programme, which is the thing you actually control.

The single-instrument trap

Call this the single-instrument trap: a programme picks one scale, administers it before and after, and reports the difference as the effect of the intervention. It is the most common design in the published social prescribing literature and it is nearly worthless as evidence, for three reasons that compound.

First, self-report loneliness scales regress toward the mean on their own — people who score unusually high when they first show up to a service tend to score lower next time regardless of what happens in between. Second, without a comparison group there is no way to separate the activity’s effect from the effect of simply being asked about your loneliness by someone who seems to care, which the Cigna and AARP survey literature both suggest matters on its own. Third, attrition is not random: people who found the programme unhelpful are the ones who stop turning up to be measured again, which biases every remaining before/after pair toward looking successful.

None of this means abandon self-report loneliness measurement. It means stop treating a single pre/post administration of it as an evaluation, and start treating it as one data point inside a design that also captures something else.

Evidence-status table

Claim	Evidence status
Social prescribing improves self-esteem and self-confidence	Reported consistently across included studies in the 2021 wellbeing review; measurement instruments vary widely between studies
Social prescribing reduces loneliness	Reported positively in all nine studies in a 2021 loneliness-specific review, but via heterogeneous, mostly uncontrolled designs
Social prescribing reduces service utilisation (GP, A&E, inpatient)	Reported in only 3 of 9 studies in the same review; the weakest and most inconsistently measured claim in this set
Purposeful, structured activity works better than unstructured social contact	Consistent with participant accounts in a 2022 qualitative meta-synthesis; not established through outcome-measure comparison
Loneliness measurement across the field is comparable enough to benchmark against	Contradicted; a 2023 review names inconsistent measurement as a structural barrier across the research base

Designing something you can actually run

Given all that, here is a design that a link worker programme or a community organisation with no research staff can realistically execute, drawing on what has and hasn’t worked in the published record.

1. Pick one validated loneliness instrument and commit to it for at least a year. The UCLA Loneliness Scale is the instrument the AARP national survey used precisely because it lets results sit alongside the wider academic literature rather than living in isolation. A bespoke five-question survey your team wrote last month cannot do that.

2. Add one functional or purpose-related item, not just an affective one. Given the meta-synthesis finding that participants describe benefit in terms of restored participation, ask something like “in the past two weeks, did you do something you found meaningful with another person” alongside the loneliness scale. This captures the mechanism the qualitative evidence points to, not only the headline outcome.

3. Build in a comparison, even an imperfect one. A waitlist comparison — measuring people referred but not yet started against people further along — is not a randomised trial, but it is worlds better than pre/post alone, and it costs nothing extra to administer if referral already produces a waiting period.

4. Track one service-use indicator only if you already have reliable baseline data for it. The 2021 review’s finding that only a third of included studies could report service-use change is largely a data-access problem, not a measurement-design problem. If your local health system will not share GP attendance data reliably, do not build your evaluation around a number you cannot get.

5. Report attrition, not just endpoint scores. Say how many people you measured at follow-up out of how many you measured at intake, and say who dropped off. The National Academies’ clinician-facing commentary on isolation assessment makes the point that assessment embedded in ordinary practice, done consistently, beats an ambitious instrument administered inconsistently.

What this means in practice: if you can run only one thing well, run a validated loneliness scale plus a waitlist comparison, administered at intake and at a fixed follow-up point, with attrition reported honestly. That single design outperforms most of what appears in the published social prescribing literature, and it is well within what a two-person team can sustain for a year.

What this does not solve

Everything above assumes people already inside a referral pathway. It says nothing about who never gets referred — the person whose GP doesn’t ask, the person who doesn’t trust the system enough to accept a link worker’s call, the person isolated enough that no professional touchpoint exists to catch them. The UK’s 2018 loneliness strategy and the National Academies’ 2020 report on older adults both flag this as the harder problem: assessment only works on people already in front of someone with the standing to assess them. A well-designed evaluation of a social prescribing service tells you whether the service worked for the people it reached. It cannot tell you anything about the much larger group it never touched, and no measurement instrument fixes that. That is a reach problem, and it sits upstream of anything discussed here.