What Would Prove This Literature Wrong

A 2026 randomised trial of 1,151 older adults living in poverty, alone, and digitally excluded found that eight telephone sessions of behavioural activation and mindfulness reduced loneliness at twelve months more than befriending calls did. Reported plainly, that is a positive result: an intervention worked. Reported one clause differently, it is a null result: befriending — the most widely deployed real-world response to loneliness, the thing social prescribing schemes actually fund — did not outperform its comparator. Both descriptions are accurate. The HEAL-HOA trial illustrates a problem that runs through this entire field: a null finding in loneliness research almost never means “nothing happened.” It means something happened relative to whatever else was in the room, and figuring out what would count as genuine disconfirmation requires knowing which layer of the evidence is being tested.

There are, in practice, two separate literatures here, and they would be falsified by different kinds of study.

The association layer is hard to break

The first layer is observational: does social disconnection predict worse health outcomes? This is the most replicated part of the field. Julianne Holt-Lunstad’s 2010 meta-analysis, pooling 148 studies and 308,849 participants, found that stronger social relationships were associated with a 50% increased likelihood of survival — an effect size in the range of established risk factors like smoking. Her 2015 follow-up, isolating isolation, loneliness, and living alone as distinct predictors, found odds ratios of 1.29, 1.26, and 1.32 for early mortality respectively, with the association holding after adjusting for baseline health. The American Heart Association’s 2022 scientific statement extended this into cardiovascular-specific terms: roughly 29% increased risk of heart attack or death from heart disease, 32% increased risk of stroke, worse prognosis for people who already have either condition.

What would a null result look like at this layer? A large, well-powered cohort study that measured objective network size or contact frequency, followed participants for a decade or more, adjusted for the standard confounders — age, income, baseline disease, health behaviours — and found no association with mortality or cardiovascular events. Nobody has published that study, across dozens of cohorts and multiple countries. It is theoretically possible that a sufficiently large null study exists and was never submitted, since null results are harder to publish than positive ones in most biomedical fields. But the consistency of effect sizes across such different populations and time periods argues against a file drawer full of contradicting data. This part of the literature would need many strong nulls, not one, to move.

The intervention layer is where a null would matter

The second layer is different in kind: if disconnection causes poor health, then fixing disconnection should improve health. This is the layer the AHA statement flags explicitly as the field’s weak point, stating outright that the absence of intervention evidence is the central research gap in cardiovascular outcomes specifically. Systematic reviews of social prescribing — the UK’s flagship policy response, in which a clinician refers a patient to a community group, exercise class, or befriending scheme rather than, or alongside, a prescription — report a pattern that should worry anyone using this evidence to justify budget lines. A 2021 review found all nine included studies reported positive individual-level impacts, with three showing reductions in GP or emergency service use. A companion review the same year documented gains in self-esteem and confidence but flagged limited trial evidence and heterogeneity across programmes. A 2025 systematic review protocol states, four years later, that the effectiveness of social prescribing for older adults specifically remains unclear, and that only one peer-reviewed randomised controlled trial exists in the area at all.

That last fact deserves emphasis. An entire national policy apparatus — funded, embedded in the NHS, exported as a model to other countries — rests on an evidence base with a single RCT. This is not a criticism of the policy; social prescribing may well work. It is a statement about what the evidence can and cannot bear. A single trial cannot be robustly disconfirmed any more than it can be robustly confirmed. If that one trial had come back null, it would have told researchers almost nothing about social prescribing in general, only something about that particular programme, in that particular place, with that particular comparator.

This is where HEAL-HOA becomes instructive rather than confusing. Its 2024 iteration tested prosocial engagement and volunteering against a control among lonely older adults in Hong Kong — one of the few genuinely randomised loneliness interventions in the literature, as opposed to an uncontrolled programme evaluation with a before-and-after comparison. Its 2026 iteration tested behavioural activation and mindfulness against befriending as the comparator, not against no intervention, and found the structured psychological approach won by a measurable margin on the UCLA Loneliness Scale. A separate 2025 randomised trial in residential aged care found befriending itself reduced loneliness scores by 2.39 points at eight weeks and 2.71 points at sixteen weeks against a true control. Read in isolation, HEAL-HOA’s 2026 result looks like evidence against befriending. Read alongside the 2025 aged-care trial, it looks like evidence that befriending works but is outperformed by a more structured alternative. Neither reading is wrong. The lesson is that “null” is only interpretable relative to a specified comparator, and most coverage of these trials drops the comparator from the headline.

What a genuine disconfirming study would need

A study capable of meaningfully undermining the intervention hypothesis would need several features that are currently rare in combination: a true no-treatment or attention-control comparator rather than an alternative active intervention; an outcome measured beyond self-reported loneliness scores — ideally an objective isolation measure, a health service utilisation record, or a mortality endpoint, given that isolation and loneliness are independently predictive and often diverge, as a 2024 study on their interplay by age demonstrates; a sample large enough and followed long enough to detect effects on hard outcomes rather than just questionnaire scores at short follow-up; and pre-registration, so a null result is reported rather than filed away. A qualitative synthesis of participant experience with social prescribing found something relevant here too: benefit tends to come from restored meaningful participation and purpose, not contact for its own part, which suggests that a trial testing unstructured contact against a control may be testing the wrong mechanism entirely and generating a misleading null.

A 2023 review of the field’s state of knowledge concluded that inconsistent measurement across studies is itself a barrier to comparing findings — which means that before the intervention literature can produce an interpretable null, it first needs a stable, shared way of describing what is being tested. Until that exists, most reported nulls in this space are not disconfirmations of the loneliness-health link. They are disconfirmations of one narrow delivery mechanism, tested against one particular alternative, on one particular questionnaire, at one particular time point. That is a much smaller claim than the headlines usually make it sound.