Measuring Safety, Not Just Outcomes, in One-to-One Befriending

A befriending scheme sends one volunteer to one isolated older person’s home, on a repeating schedule, usually alone. That is the model. It is also, structurally, the least supervised form of social intervention that health and care systems currently fund. If you run or commission one of these programmes, the question you need answered is not “does this reduce loneliness scores” — the literature has something to say about that — but “how would we know if something had gone wrong.” On that second question, the published evidence has almost nothing to offer.

What the evidence actually covers

Systematic reviews of social prescribing report the outcomes you would expect: increases in self-esteem and self-confidence, and in some cases reductions in GP, emergency, or social worker contact. A qualitative meta-synthesis of participant experience goes further, finding that people describe benefit extending beyond contact itself to restored participation and purpose, and that structured, purposeful activity appears to work better than unstructured contact alone. These are genuine findings, and they are the right things to report to a funder.

None of them are safety measures. They tell you the intervention is landing for the people who stay in it. They tell you nothing about incidents, near-misses, exploitation, volunteer conduct, or what happens when a lone visitor is the only person who ever sees inside a particular flat.

The controlled trial evidence is thinner still, which matters because trials are where adverse-event reporting would normally live. A 2025 systematic review protocol on social prescribing for older adults notes that only one peer-reviewed randomised controlled trial exists in this space at all. Two trials do exist that test related one-to-one and group interventions directly: a randomised trial of befriending in residential aged care, which reduced UCLA Loneliness Scale scores by 2.4 points at eight weeks and 2.7 points at sixteen, and the HEAL-HOA trial in Hong Kong, which tested volunteering and prosocial engagement against a control among lonely older adults. Both are valuable precisely because they are controlled rather than uncontrolled programme evaluations. But as published, their findings are about efficacy on loneliness measures — not about what safeguarding processes were in place, whether anything went wrong, or how anyone would have found out.

The closed-door problem

Name the failure mode plainly: the closed-door problem. The isolation that makes someone eligible for a befriending referral is the same condition that removes anyone else from the room. A lone older adult with a shrinking network is, by definition, someone whose life has fewer independent witnesses. Send a single unsupervised volunteer into that space on a recurring basis and you have built an intervention whose central mechanism — one-to-one trust, low bureaucracy, informality — is also what strips out the oversight that group settings provide almost by accident. A community group, a walking club, a shared meal: these have several people present, which is itself a weak but real safeguard. Befriending has one.

This is not an argument against befriending. The research distinguishing isolation from loneliness as separate constructs — with different risk profiles and different predictors — makes clear that the most socially isolated people, who need contact most, are also the ones for whom a single visitor is often the only realistic entry point. Group programmes don’t reach them; that’s exactly why one-to-one models exist. The point is that the model’s strength and its blind spot are the same feature, and the published effectiveness literature is silent on the blind spot.

What programmes have measured instead of risk

Where safeguarding does appear in the published record, it appears one step removed. The National Academies’ 2020 consensus report calls on the health care system to routinely assess isolation and loneliness — a screening recommendation, not a safeguarding one. A clinician-facing commentary on that report discusses what routine assessment would require in practice, but the “risk” it addresses is clinical risk from unmet isolation, not operational risk from the intervention delivering contact. The distinction matters: screening for who is isolated is upstream of the problem this article is about. It tells you who to refer. It does not tell you whether the referral pathway itself is safe.

What this means in practice: track safety and track outcomes as two separate measurement systems, not one. A loneliness scale score, a service-use reduction, a self-esteem measure — these tell you the intervention works for people who complete it. None of them will surface a safeguarding incident. If your only dashboard is outcome scores, you have no visibility into the thing that would actually shut the programme down.

Claim	Evidence status
Structured group activity produces better outcomes than unstructured one-to-one contact	Supported — qualitative meta-synthesis and systematic review both report this
One-to-one befriending reduces loneliness scores	Supported, with rare trial evidence — two RCTs report measurable reductions
Routine screening for isolation should sit in health care settings	Recommended by the National Academies; adoption and effect on outcomes are separate open questions
One-to-one models carry higher unsupervised-contact risk than group models	Plausible from programme structure; no published incident data confirm or quantify it
Current social prescribing evaluations measure safeguarding incidents	Not supported — the published reviews and trials report outcomes, not adverse events or safety monitoring

Given that gap, the operational answer is not exotic: independent incident logging separate from outcome reporting, defined escalation routes that don’t depend on the same volunteer noticing a problem in themselves, supervision contact with volunteers on a schedule that isn’t triggered only by a complaint, and a policy for what happens when a volunteer is the last person to see someone before a decline is noticed. None of this is drawn from a trial, because no trial in this list reports it. It is drawn from the plain shape of the risk: a single point of contact has no redundancy, so the redundancy has to be built around it rather than assumed to exist.

What this does not solve

The published literature gives you outcome measures with real, if thin, evidential backing, and gives you almost nothing on safety measurement — which means any safeguarding framework you build for a befriending or one-to-one social prescribing programme right now is being built without a benchmark. You cannot yet say your incident rate is better or worse than typical, because there is no typical on record. And the reach problem compounds this: the people for whom one-to-one contact is the only workable model are, by the isolation-versus-loneliness research, the hardest to see into in the first place. The measurement gap and the safeguarding gap are the same gap.