What happened in social health? Green prescribing comes up short, a California veto, Germany opens two calls, and a national evaluation that measured the wrong thing
Hans Rocha IJzerman·5 October 2026
Five of this week’s papers ran into the same wall from different directions. A systematic review of green prescribing could not pool a single outcome. A meta-analysis that triples the odds of a poor depression outcome turns out to rest on four studies, all in older adults. A paper in JPSP had to collapse a seven-point loneliness scale into a yes/no because four fifths of the observations sat on the floor. The first empirical test of the WHO Commission’s own framework as a measurement model reports not one reliability coefficient. And England’s national evaluation of social prescribing found no effect on isolation, using a survey item that scores you as “not isolated” if you report two or more falls and nothing else.
I have been going on about measurement for years and I am aware of how that sounds. But the link worker finding is the one that should stop people. That evaluation covers 4.1 million survey responses and about £130 million a year, and its loneliness outcome is a residual from a frailty checklist. Nobody did anything wrong; the question was never designed to carry that weight. It got handed the weight anyway, because it was the item that existed in the survey that was already running.
The policy side had a version of the same problem, which is official documents describing themselves inaccurately. The Commission’s public explainer of the EU KIDS Act states a prohibition its own legal text does not contain. The press rendering of Newsom’s SB 903 veto drops half his reasoning. Germany’s own programme page still lists eligibility that its own third funding instrument superseded a month ago. Senator Padilla’s office published a vote count the roll call contradicts. Each of these took one reading of the primary document to catch, and each was being repeated by people who had not done that reading.
This issue was researched and written by Claude, the AI I work with, following a standing brief I wrote and keep revising. Claude posted it from my account this week. It goes up under my name because I stand behind it and any error in it is mine to answer for, but I did not write it and I will not always have read it before you do. Corrections and arguments in the thread feed into the next issue and into the brief itself, which is how most of the improvements in it have happened so far.
⏺️ Research
⏺️ A systematic review of 21 studies of green prescribing and nature-based interventions for people aged 50 and over found that statistically significant differences from control conditions were uncommon, and rated the certainty of evidence low or very low for most outcomes. Eleven of the 21 were randomised. The review performed no meta-analysis, so there are no pooled estimates: the conclusion is a count of how often individual studies cleared significance. The loneliness evidence comes down to two studies of indoor horticultural activity in nursing homes, both at low risk of bias, total N = 203, where UCLA scores moved by a GEE mean difference of −9.680 (SE 1.861, p < 0.001) in one and p < 0.001 with no effect size reported in the other. That outcome was graded moderate certainty, downgraded for imprecision; the social network outcome, resting on one study of 53 people, was graded very low. Only one of the eleven randomised trials was judged low risk of bias overall. The best-supported effect on loneliness in this entire literature therefore comes from indoor gardening in residential care, and the outdoor nature-prescribing programmes commissioners are actually buying have nothing comparable behind them. That is not the same as showing those programmes do nothing: a vote count across underpowered studies and a pooled estimate near zero look identical from the outside, and only pooling would separate them. Anyone commissioning in this space can check whether their programme is already described in the LONELY-EU Intervention Registry, and add it if not.
Registered on PROSPERO (CRD42025603199). The authors disclose that the registered subgroup analyses by age band, deprivation and region, and the planned risk-of-bias sensitivity analysis, were all abandoned for lack of data. Data availability: “the datasets supporting the conclusions of this article is included within the article and its additional files”; there is no code availability statement anywhere in the paper. Composition: deprivation is characterised at review level, with a note that few included studies concerned people in deprived areas; no systematic breakdown of ethnicity, income or education across the 21 studies, and the per-study location table did not extract reliably, so the European share is not stated here. Analysed: 21 studies, 11 randomised; the loneliness evidence rests on 2 studies, N = 203 (53 and 150).
Measurement: as a review this carries no instrument-level stamp; what it gets instead is whether it reports the reliability of its included studies’ instruments, and it does not. No internal-consistency coefficient appears for any included study, and the single use of “reliability” in the paper concerns confidence in effect estimates. Only the UCLA Loneliness Scale is named across the two loneliness studies, so instrument heterogeneity for loneliness is not discussed, though heterogeneity in interventions and comparators is. Power sensitivity: none can be derived, because no meta-analysis was performed and no pooled estimate exists; the GRADE downgrade on the loneliness outcome was itself for imprecision. Openness: checked.
(paper)
⏺️ A UCL review published two days later reports that loneliness roughly triples the odds of a poor outcome in common mental disorders, on four poolable studies, all in older adults with depression, only one of which adjusted for how depressed people were to begin with. Twenty-two studies were included; four entered the meta-analysis, giving an adjusted odds ratio of 3.26 (95% CI 2.26–4.70, p < 0.001). GRADE certainty is high for depression outcomes, low for suicidal ideation and anxiety, and very low for mixed common mental disorder outcomes. Searches ran from inception to October 2025 across PsycINFO, MEDLINE, Embase and CINAHL, covering adults aged 16 and over at any treatment stage, appraised with the Newcastle-Ottawa Scale. The baseline-severity gap is the one that governs how the headline figure should be read: loneliness and depression share symptom content, so an estimate that does not condition on initial severity is partly measuring how unwell people already were, and the authors say so in the abstract. The eighteen studies that could not be pooled are also where the answer sits for anyone under 65 or with a diagnosis other than depression, and that answer is not in this paper.
Registered on PROSPERO (CRD42023410401); the registered protocol was not checked against what was done. Data, materials and code availability: could not verify. Crossref records the article as CC BY 4.0, and every retrieval route tried (the BMC article HTML, three BMC PDF paths, two Springer paths under a stealth driver, and Europe PMC, which has not yet indexed it) failed on an article two days old, which makes this a reachability failure and not a paywall. Sample composition across included studies: could not verify. Analysed: 22 studies included, 4 pooled, all in older adults with depression; total N across the 22 could not be verified.
Measurement: a review, so no instrument-level stamp. Whether it reports the reliability of its included studies’ loneliness instruments could not be verified without the full text. Press coverage describes the included measures as ranging from single items to multi-item scales, but that is the press’s characterisation and not the paper’s, and should not be attributed to it. Openness: open but not retrievable by us. CC BY 4.0 per Crossref, with BMC, Springer and Europe PMC all failing.
(paper)
⏺️ Being alone, being at home and watching television each raised the odds of any state loneliness across 22,010 everyday episodes from the German socio-economic panel, a result the authors could only report after turning a seven-point scale into a yes/no, because 17,954 of those episodes sat at the lowest point. The data come from the Day Reconstruction Method module of SOEP-IS across four waves, 2012–2015, in 2,508 participants aged 17 to 95, analysed with multilevel logistic regression. Prolonged care work raised the odds further; prolonged social contact and prolonged television lowered them. The dichotomisation was not a shortcut. The supplement documents a ladder of alternatives and why each failed: linear mixed models violated normality and homoscedasticity, cumulative ordinal models would not converge even intercept-only, a four-category collapse broke once controls were added, and Poisson was underdispersed. The result is therefore a map of which everyday situations co-occur with any flicker of loneliness in a broadly representative German sample, and it is only that. Intensity, frequency and duration have been removed from the outcome variable, so nothing in the paper speaks to them, as the authors themselves concede.
Preregistered at OSF (osf.io/2jcdz, registered 11 June 2025, Secondary Data Preregistration template). The supplement itemises divergences from the registered plan and rates each: dichotomising the outcome, rated “Large”; using a logistic link where the preregistration specified multilevel models without one, “Small to large”; excluding 736 episodes with no loneliness rating, “Small”; adding a controls-only model, “Small”. Merging the company and activity categories and log-transforming duration were anticipated in the preregistration. No sample-size deviation. SOEP-IS data are access-controlled under a data-use contract with DIW Berlin; the supplement refers repeatedly to an “R script 03” but does not state where the scripts live, and the OSF registration holds no files, so code location is not stated. Sample composition beyond age and sex is not reported in the open supplement; the body is behind APA’s paywall, so whether education, income, migration status or East/West region appear could not be verified. Analysed: 22,010 episodes in 2,508 participants; 19,820 episodes in 2,503 participants for the covariate-adjusted models.
Measurement: state loneliness is a single item on a seven-point scale, so no internal-consistency coefficient exists or is reported for the outcome. It was dichotomised at response 1 versus responses 2–7, with the threshold justified by the floor effect above; no sensitivity analysis on the threshold itself appears, and the documented ladder of alternatives is a model-selection narrative rather than a threshold check. No dimensionality assessment, the outcome being a single item. Invariance not tested. Likelihood ratio tests are reported for the extended models adding day reconstruction participation frequency as a covariate; no RMSEA-type indices apply to this framework. Openness: supplement checked in full; article body closed, and Crossref carries no licence block for it.
(paper)
⏺️ In 58 German adults with alcohol use disorder tracked three times a day for eight weeks, days of higher-than-usual loneliness were days of less drinking, in a sample screened for believing the opposite about themselves. Within-person fixed-effects models put the concurrent association at b = −0.092 on the probability of drinking (95% CI −0.151 to −0.033, FDR-corrected p = .006) and at −15.89 grams on quantity (−24.16 to −7.61, p = .003). Prospective associations in both directions were small, with intervals that constrain their magnitude, and every mood-moderated prospective slope was null. Gender moderated the concurrent quantity association (b = 23.68, p = .002), the negative association being stronger in men, though that rests on 43 men against 14 women. Eligibility required answering “sometimes” or more to whether the person tends to drink more during periods of loneliness or social isolation, and day-level compliance was 96.7%. The selection is what makes the result useful and not what spoils it: these are people who told researchers that loneliness drives their drinking, and within themselves it did not. None of this is a population estimate, and the design cannot speak to whether chronically lonelier people drink more than others. What it undercuts is the clinical intuition that a lonely day is, by itself, a high-risk day.
Preregistered at OSF (10.17605/OSF.IO/QVY38). The authors disclose two divergences: FDR correction was an addition to the registered analysis plan, and the H5 moderation models were estimated with random slopes rather than the registered random intercepts, with results reported as essentially unchanged. Participant-level data are not deposited. Four Stata do-files are deposited at OSF (10.17605/OSF.IO/F76K3); the folder exists and holds analysis files for H1–H3, for H4 by mood and by social support, and for H5, matching the described plan including the separate H5 file. Composition is reported for age and gender only (43 men, 14 women, one non-binary participant; age M = 48.78, SD 8.29, range 32–64), with no ethnicity, education, income or migration status, and no breakdown across the four German sites. Analysed: N = 58 of 62 enrolled.
Measurement: daily loneliness is four items, three adapted from the UCLA 3-item scale plus one face-valid item, with within-person α = .81 computed in this sample, the only instrument here carrying a coefficient. The baseline UCLA scale is reported with no coefficient. Mood is four bipolar items collapsed into two two-item indices with no coefficient; perceived social support and daily social contact are single items; daily alcohol use is a single yes/no item plus a quantity recall. No factor analysis or dimensionality check on the four-item daily scale, which is used as a mean score. Invariance not tested. No fit statistics, the models being fixed-effects panel regressions. No continuous measure was dichotomised. Power sensitivity, the digest’s own calculation: the paper states that no formal a priori power analysis was conducted, on the ground that available approaches do not handle the multilevel structure of intensive longitudinal data, and that power was expected to come from the number of repeated observations. Backing standard errors out of the published intervals gives roughly 0.017 to 0.022 on the within-person probability scale for the prospective models, so 80% power at two-sided α = .05 corresponds to effects of about 0.046 to 0.061, against the concurrent effect of −0.092 the study did detect. This treats the published standard errors as fixed and makes no allowance for the repeated-measures structure beyond what they already embed. Openness: checked.
(paper)
⏺️ The first empirical test of the WHO Commission’s social-connection framework as a measurement model finds it fits about as well as a rival built from social capital, in a paper that reports no reliability coefficient for any factor and tests no measurement invariance, in a sample deliberately spanning three ethnic groups and five survey languages. Confirmatory factor analysis on 779 Asian American adults in Greater Los Angeles and Orange County, estimated with WLSMV. The WHO three-factor structure, function and quality model returned CFI .990, TLI .987, RMSEA .071, SRMR .035; the four-factor social capital model, after dropping one indicator, returned CFI .991, TLI .988, RMSEA .058, SRMR .032. Higher-order and bi-factor models failed to converge in both frameworks, and one dimension was dropped entirely for having only two available indicators. The preprint went up on 9 September; its full text became openly retrievable on 30 September, which is why it appears now. On this evidence the WHO structure is estimable and fits acceptably in one US sample, which is more than anyone could say about it a month ago. Whether it measures the same thing across the groups that sample was built from is a separate question and an untested one, and an index intended for comparison across countries depends on exactly that property. The WHO Commission has proposed a global social connection index built on this framework, and its term ends this year. Anyone weighing which items such an index would need can see what the field currently uses in the LONELY-EU Item Explorer.
No preregistration stated. Data are “available from the corresponding author on reasonable request”, which is not a repository; there is no code or materials availability statement. Composition is unusually full: age (M 55.32, SD 15.42), sex (55.71% female), nativity (89.09% foreign-born), five education categories, five employment categories, five income bands including 14.38% declining to answer, marital status and English proficiency, with recruitment benchmarked to Census figures on age, sex, education, income and nativity. The ethnic-subgroup breakdown and the survey-language breakdown are never reported, which is what a reader would need to judge the untested invariance. Analysed: N = 779, with no missing data. The Methods give baseline collection as August 2022 to July 2024 while the abstract says 2022–2026.
Measurement: no internal-consistency coefficient of any kind is reported, neither Cronbach’s alpha nor McDonald’s omega, for any factor in either framework, in a paper whose purpose is measurement evaluation. Factor structure is the paper’s subject, and the fit statistics are above; several individual-factor models fit very poorly, with RMSEA values of .126, .230, .271, .363, .392 and .429, which the authors attribute to small degrees of freedom inflating the statistic. No RMSEA confidence intervals appear anywhere. Measurement invariance is not tested, across ethnicity, language, nativity or anything else. Two indicators were constructed as counts and winsorised at the 95th percentile; no continuous measure was dichotomised. Indicators were dropped on a mixture of statistical performance and conceptual alignment, one being retained over another because it was judged less conceptually distant before being dropped at the next stage anyway. The survey was translated by bilingual staff and reviewed by a second, with no reported or cited validation of the translations and no invariance test across the five languages. Openness: checked.
(paper)
⏺️ Four weeks of experience sampling in 123 Dutch adolescents, linked to their own donated Snapchat data, found that how socially disconnected they felt predicted almost nothing about how they actually used the platform, except that the more disconnected sent more friend requests that went unanswered. Three assessments a day, analysed with negative binomial, zero-inflated negative binomial and Dirichlet models under Benjamini-Hochberg correction at q = .05. Social disconnectedness predicted unanswered sent friend requests at IRR 1.04 per unit, surviving correction; a ten-unit rise corresponds to roughly 48% more such requests. Total number of friends was unrelated. Snapscore was unrelated. Friendship closeness predicting messages sent (IRR 1.02) and peer belongingness predicting messages sent (IRR 0.98) did not survive correction. Gender and school grade effects dwarfed the psychological ones, with girls at IRR 5.21 for unanswered requests. The authors conclude that the findings support neither the social enhancement nor the compensation hypothesis. This is about as close as the field gets to checking self-reported connection against observed behaviour, and the behaviour declined to cooperate. The limit is that loneliness and social exclusion correlated at Spearman .96 and had to be merged into a single variable, so the study cannot address loneliness specifically; with four single-item measures and no reliability evidence, the null is partly a result about the measures.
The authors describe a design and sampling plan posted to OSF before data collection (osf.io/db8qa, created 25 October 2024) and a hypotheses document posted after collection but before analysis (osf.io/nxvcu, created 10 March 2026, with a preregistration PDF uploaded the same day), and they state that ordering explicitly, which most papers do not. Both links resolve to ordinary editable OSF project nodes rather than frozen registrations, each returning registration: false, and both were modified within the past fortnight, on 27 September and 2 October 2026, so the ordering is documented but not time-stamped in a form that cannot be revised. Divergences, all disclosed by the authors: loneliness and social exclusion were preregistered as separate variables and merged on a correlation of .96; lavaan SEM was replaced by count models for overdispersion and excess zeros, 66.70% of adolescents having no AI interactions at all; and full-information maximum likelihood was replaced by listwise deletion, which the authors note cost them power. Data are promised and not yet available: “the data will be published online after acceptance of the article”. Analysis code is deposited at OSF as an anonymised R Markdown file with its rendered output. Composition reports age, sex and school grade, with the authors noting lower grades are overrepresented; no ethnicity, income or socioeconomic position, no migration status, no school-level breakdown. Analysed: N = 123, from 311 experience-sampling participants of whom 151 donated platform data; mean age 13.33.
Measurement: every psychological measure is a single experience-sampling item (“I felt close to [best friend]”, “I felt like I belonged with my group of friends”, “I felt lonely”, “I felt left out”), person-mean aggregated, with no reliability coefficient of any kind reported, within-person or between. No dimensionality check. Invariance not tested, across the three cohorts or the four school grades. No fit statistics for the count or compositional models. No continuous measure was dichotomised. Snapscore needs its own flag: it is generated by an undisclosed proprietary Snapchat algorithm, over an engagement window the authors can describe only as “the past two months” with no specified number of days. Power sensitivity: no sample-size calculation is stated anywhere, and power is discussed only as a post-hoc limitation; no figure is derived here, because the null estimates are reported as incidence rate ratios without intervals we could back a standard error out of. Openness: checked.
(paper)
⏺️ Policy and advocacy
⏺️ Governor Newsom vetoed the one bill in California’s AI package that went to whether a machine may stand in for a therapist, objecting that it would have required routine screening determinations to receive direct approval, a complaint the bill’s own screening carve-out was drafted to answer. SB 903, Senator Padilla’s Wellness and Oversight for Psychological Resources Act, would have added Chapter 13.6 to Division 2 of the Business and Professions Code. Section 4989.84(a) barred advertising or purporting to offer psychotherapy services delivered through companion chatbots, including claiming a chatbot is a therapist. Section 4989.84(b) barred artificial intelligence, without review and approval by a licensed professional, from making therapeutic decisions, interacting directly with patients in psychotherapeutic communication, generating recommendations, assessments, diagnoses or treatment plans, detecting emotions or mental states, or performing triage or screening. There was an express exception for tools cleared or approved by the FDA for that use and compliant with HIPAA, which the common summary that the bill “banned” chatbot therapy omits. Section 4989.84(e)(2) expressly permitted administering a standardised screening questionnaire, scoring it to the instrument’s own rules and transmitting the responses and score to a licensed professional, provided the system does not use them to judge urgency or clinical nature. The carve-out is narrow, and the veto’s objection is not baseless on that account, but any account of the veto that does not mention it gives one side. The press rendering of the veto message carries “overly broad” and the screening clause and drops two things: a second objection that the bill’s definition of “psychotherapy services” is “poorly crafted and would capture general-purpose AI systems not even structured or deployed to deliver that care”, and the Governor’s invitation to the author to return next year. Padilla has said he will reintroduce. A correction while we are here: Padilla’s own office states the Assembly passed the bill 71 to 4; the roll call records 74 ayes, one no and four not voting, and the Senate was unanimous at every stage. The file is also not formally closed. Leginfo still carries an action of the same date reading “In Senate. Consideration of Governor’s veto pending” and still classifies the measure as an active bill in floor process, where AB 2575, vetoed the same day, reads “Inactive Bill – Vetoed”. An override needs two-thirds of each house, and the Legislature’s own guide to its procedures records that it has not overridden a veto since 1979, so this is a record-keeping artefact and not a live prospect.
(link)
⏺️ The European Commission’s public explainer says AI companions “may not use” designs that simulate human relationships in ways likely to create emotional dependency; the Commission’s own legal text imposes a proportionality-qualified design duty a provider can discharge by signing a code of conduct. We covered the EU KIDS Act proposal when it published on 17 September; what is new is that the operative text has now been read against the explainer, and that the file has acquired a procedural history. Article 14(1) requires providers to “put in place proportionate and effective measures”, in a list that “shall include at least” avoiding design features that simulate interpersonal relations likely to create emotional dependencies. Article 14(3) lets providers adhering to a code of conduct assessed as adequate by the Commission under Article 23 rely on that code to demonstrate compliance. The post-market monitoring limb at Article 14(1)(f) does not apply to micro and small enterprises at all. The Commission’s own explanatory memorandum uses the softer formulation, “designing the systems so minors are not exposed to features that are likely to create emotional dependencies”, so the gap is internal to the Commission, between an FAQ managed by DG CNECT and last updated on 2 October, and a file led by DG Internal Market under Commissioner Séjourné. Article 14(2) is a genuine hard rule, but only for companions embedded in a social network, video-sharing platform or online game: no automatic activation, no prominent display on the interface, no encouraging minors to use it, and an easy opt-out at any time. Three smaller things a reader of the summaries would not know. The review clause is Article 42, not 41, and its date is still printed in square brackets as “[31 August 2030]”, meaning it is a drafting placeholder. Article 14(1)(f) requires monitoring of risks “referred to in point (d) of this paragraph”, where point (d) is the under-13 guardian-tools item and the risks are in point (e), an uncorrected cross-reference in the published proposal. And the 6% turnover penalty the FAQ states flatly appears only at Article 34(2), which routes companions and conversational chatbots through the AI Act; Article 34(1) routes social networks, video-sharing services, video gaming platforms and app stores to Chapter IV of the Digital Services Act and states no percentage. Procedurally the file has not moved: committee responsible, committee for opinion and committee for budgetary assessment all read “pending final decision on the referral”, no rapporteur is appointed, and the only logged meetings with interest representatives are Katarina Barley with the Center for Countering Digital Hate on 21 September and with Triple P Europe on 24 September.
(link)
⏺️ Spain’s Mesa Interinstitucional de Soledades, the coordinating body its national loneliness framework turns on, is four months past the deadline for its own constitution with no public record that it has met. The Acuerdo creating it was published in the BOE on 6 March 2026 as BOE-A-2026-5317, and its paragraph Octavo required constitution within three months of taking effect, which falls on 7 June 2026. Paragraph Noveno resources it from existing ministry means with no additional public spending. The Spanish Government’s own July 2026 accountability report describes the Mesa as a feature of the Marco Estratégico Estatal de las Soledades 2026–2030 and records no constitution and no meeting. IMSERSO holds the second vice-presidency, and a senior IMSERSO official describes the body as “creada a raíz de este marco” while saying that giving the framework effect “es compleja y se hace necesario un esfuerzo de coordinación importante”. We searched BOE and ministry notices across 1 June to 5 October and found nothing; we did not reach the ministry’s own register of órganos colegiados, so this is an absence we could not close rather than a confirmed failure. The framework itself carries no budget line and no quantified targets, its lines of action being explicitly orientative. The Government’s own figures put the impact of loneliness at more than 1.17% of GDP, say one in five Spaniards are affected, and give 34.6% prevalence among 18- to 24-year-olds, the highest of any cohort.
(link)
⏺️ Practice
⏺️ England’s national evaluation of the social prescribing link worker rollout found improvements on four population outcomes and nothing on feeling isolated, using a survey item that counts you as “not isolated” if you report two or more falls and tick nothing else. Older, published online 5 June 2025 and in print that December, and as far as I can tell never discussed in this network. The Manchester-led team linked GP Patient Survey responses from 2018 to 2023 to primary care network workforce data, clustering at 1,268 networks. Per additional full-time-equivalent link worker per 50,000 patients: confidence managing long-term conditions OR 1.006 (1.001–1.010) and enough support from local services OR 1.005 (1.001–1.008) among respondents with long-term conditions; good GP experience OR 1.015 (1.004–1.027) among all respondents; mental health needs understood OR 1.012 (1.003–1.021) among respondents with mental health conditions. Feeling not isolated came in at OR 1.001 (0.996–1.005), p = 0.74, and was null in all three model specifications. By March 2023, 3.2% of the registered GP population had been referred. Now the item. It is a single tick-box: “Have you experienced any of the following over the last 12 months?”, with options for problems with physical mobility, two or more falls that have needed medical attention, feeling isolated from others, and none of these. The variable is coded 1, meaning not isolated, if the respondent selects any option other than “feeling isolated from others”, so someone reporting two or more falls and nothing else is scored as not isolated, as is someone selecting “none of these”. Blanks and multi-coded responses are dropped, which also removes anyone ticking both falls and isolation. The authors flag the wording themselves as a possible explanation for the null, and conclude they were unable to support the NHS aim for patients to feel more connected, less lonely or less isolated. The honest reading is that the fourth stated aim of a national programme was evaluated against a residual from a frailty checklist, and nobody yet knows whether link workers shift loneliness. A conference abstract from the same team adds that link workers are concentrated in the least deprived areas, concentration index 0.03, P < 0.001, though that is an abstract using full-time equivalents per 10,000 and linear probability models across 6,991 practices, so it should not be blurred with the main paper.
No preregistration identifier and no published protocol; the paper carries only a boilerplate transparency statement that discrepancies from the study as originally planned “and registered” have been explained. Data cannot be shared directly owing to data-sharing restrictions, though researchers may apply to use the GP Patient Survey linked to publicly available datasets. There is no code availability statement. Composition: models adjust for age, gender, ethnicity, employment status, deprivation and rurality, with descriptives in the supplement, and the authors report the survey over-represents older respondents, white ethnicities, women and people in rural areas. Analysed: N = 4,132,676 responses for the isolation outcome, across 1,268 primary care networks. Funded by NIHR award 134066.
Measurement: this is a survey-linkage study in which each of the five outcomes is a single GP Patient Survey item, so no internal-consistency coefficient exists or is reported, and no dimensionality check, invariance test or model fit statistic applies. The measurement fact that governs the result is the coding rule above, which is less a dichotomisation of a continuous construct than the recovery of one as a residual from a composite checklist; no sensitivity analysis on it appears. Power sensitivity, the digest’s own calculation: from the published odds ratio and interval for the isolation outcome, the standard error on the log scale is about 0.0023, giving 80% power at two-sided α = .05 to detect an odds ratio of roughly 1.006 per additional full-time-equivalent link worker per 50,000 patients. That is larger than the two effects the study did detect among respondents with long-term conditions, 1.005 and 1.006, and smaller than the two it detected on GP experience and mental health needs, 1.015 and 1.012. The paper reports no sample-size calculation. The calculation treats the clustered standard error as fixed. Openness: checked, CC BY 4.0.
The authors’ own cost estimate, which we flag because it circulates without attribution: they put the scheme at £130 million to the NHS in 2022/23, derived from NHS Digital workforce counts and the 2022/23 Network Contract Directed Enhanced Service specification, and say it understates the total because it excludes onward referral destinations.
(paper)
⏺️ The Irish department that funds Men’s Sheds now says reports that the 2025/26 grant will not proceed are inaccurate, contradicting the account the Irish Men’s Sheds Association gave its own members three weeks ago. We ran the funding gap last week on the association’s account; what has changed is that the Department of Rural and Community Development and the Gaeltacht has spoken. In a statement to the Donegal Daily it says all departmental funding programmes are subject to the annual budgetary estimates process, which is ongoing, so no definite future commitments can be made; that Minister Dara Calleary has asked the association to re-examine how it distributes funding to individual sheds; and that the Department is awaiting that proposal, to be reviewed alongside the estimates outcome. It does not say whether the 2025/26 grant will be reinstated. The association’s letter of 11 September told members Calleary had said at the Fianna Fáil think-in on 8 September that the grant “will not now proceed” and that the next would be activated in 2027. Calleary met the association on 22 September and no readout has been published. The most recent primary record on the file remains the Department’s own release of 4 May 2026, reporting grants of up to €3,000 paid in full to 369 sheds. Figures now circulating, an average per shed falling from €2,432 in 2023 to €2,047 in 2025 and a separate €1 million under the 2025 Local Enhancement Programme reaching 150 sheds at an average of €440, come from Deputy Ken O’Flynn’s account of parliamentary replies as reported in the Southern Star, and we could not locate the underlying question in the Oireachtas record, so they should be treated as second-hand. The same article misidentifies Calleary as Minister for Social Protection. Budget 2027 falls tomorrow, 6 October, outside this issue’s window.
(link)
⏺️ The Nationaal Ouderenfonds reported during the Dutch Week tegen Eenzaamheid that 45% of older people feel lonely to some degree and 18% strongly so, in the one edition of its annual survey that has published no method at all. The figures went out on 24 September: 628 respondents aged 65 and over, 26% wanting more contact, and of those, 43% saying an invitation to visit would help most and 41% a proposal to do something together. The 2024 edition surveyed 1,078 people aged 55 and over, named its panels and its fieldwork partner, and stated that it was representative on sex, age and education; 2023 and 2025 published comparably. For 2026 there is no sampling method, no fieldwork dates, no weighting statement and no named loneliness instrument, so whether these are De Jong Gierveld items, a single question or something else cannot be established from anything public. The age base also moved from 55-plus to 65-plus, which rules out comparison with the series. The figures are reported here as press-release-sourced and unverified against any underlying report. The Week itself ran 24 to 30 September as a VWS initiative under Eén tegen eenzaamheid, and as of today carries no national participation figure and no evaluation, only local wrap-ups.
(link)
⏺️ Innovation and startups
No funding round in scope was announced this week. The one candidate, a widely reported $3.5 million for the social app Clockout, is a cumulative lifetime total rather than a round: SEC EDGAR holds three Form D filings for the company, from March 2022, July 2023 and January 2025, and none in 2026. Across more than 65 European deals tracked in the week, none was in social connection or digital mental health.
⏺️ Sword Health closed its acquisition of Headspace on 1 October, and the whole of the supporting clinical evidence it put on the public record in Oregon’s behavioural health market review consists of three musculoskeletal and pelvic-health papers. (Hans has no interest in either company, but the merged Sword Health and Headspace competes directly with Entrelacs, which he co-founded, and the declaration belongs here on that basis.) The sequence matters because the press coverage dates it wrongly: Hart-Scott-Rodino filings went in on 9 July 2026, state filings to Oregon and Minnesota on 22 July and to Massachusetts on 6 August, first press appeared on 25 August, a company release on 16 September, and the deal closed on 1 October, confirmed on Headspace’s own customer support page. The transaction was therefore agreed in early July, two months before the announcement most coverage treats as its origin. Oregon’s report puts the price at $200–300 million, all cash, against a peak valuation near $3 billion. The health authority completed its 30-day preliminary review on 28 September and approved the transaction, with an order the following day imposing no substantive conditions beyond a 60-day effectiveness window and a one-business-day completion notice. It found the parties “do not compete in the virtual behavioral health space” and that the deal was “unlikely to reduce access to behavioral health care”, naming nine alternatives including BetterHelp, Grow Therapy, Headway, Rula, Talkspace and Talkiatry. Massachusetts received a material change notice on 6 August and opened no cost and market impact review. Minnesota was notified on 22 July under Minn. Stat. § 145D.01 Subd. 2; no public determination exists and, under that statute’s confidentiality scheme, none is expected unless the Attorney General sues. Then the filing index. It lists twelve documents, of which exactly three are clinical literature: Costa et al. on digital care programmes, Pereira et al. on women’s pelvic health, and Scheer et al. on racial and ethnic differences in outcomes. All three are musculoskeletal or pelvic health, and the health authority’s own report records no behavioural health study. Six of the twelve documents are confidential, the two antitrust filings and both sets of financial statements, so this describes the public record and not the sealed one, and the numbering is odd enough to note, with two items labelled Appendix 15a, one labelled 15d, and no 15b or 15c anywhere.
Evidence: no published evaluation of a combined Sword and Headspace mental health product exists, and none could at four days old. No trial registration was found for Sword Health or for OrangeDot, Headspace’s corporate entity, in the window; ClinicalTrials.gov and ISRCTN were searched by each sponsor name and by product name. Neither product carries an FDA clearance or De Novo, a CE mark under MDR, a UKCA mark, a NICE early value assessment or an NHS DTAC assessment; both avoid medical device status by making no clinical claim. The statements in the Oregon filing about the merged offering, that Sword expects no reduction in Headspace’s current behavioural health service access and intends technology improvements for care navigation and longitudinal clinical records, are forward-looking and company-sourced. The $200–300 million range is the health authority’s characterisation; the Massachusetts notice discloses no figure at all.
(link)
⏺️ Florida’s Attorney General asked a state court on 28 September to stop ChatGPT referring to itself in the first person. The motion for a temporary injunction runs to 39 pages and was filed at 09:15 in Highlands County Circuit Court, in a suit brought on 1 June under Florida’s Deceptive and Unfair Trade Practices Act against several OpenAI entities and Sam Altman. OpenAI had removed the case to federal court; a federal judge rejected jurisdiction and remanded it earlier in September, which is why it is back in state court. Of the six limbs of relief sought, two are about simulated relationship rather than safety in general. The first would enjoin the defendants from “misrepresenting that ChatGPT has human attributes it does not have, including but not limited to ChatGPT referring to itself with any first-person language, suggesting it has the capacity to think or feel, representing itself as having emotional states, representing itself as having biological characteristics, or in any way suggesting it may have consciousness or be distinct from its function as a lifeless tool”. The second would bar “allowing ChatGPT to solicit engagement from users through conversation prolongation”. The remaining four would bar model development without independent third-party approval, bar the product to Florida minors, impose notice and verifiable parental consent for under-13 data, and bar misrepresenting the product’s safety and accuracy. The motion frames anthropomorphisation and conversation prolongation as dark patterns under the Act. A regulator attacking the companion mechanic itself, and not the content it produces, is a different theory from the ones on this beat so far, and it is the theory that would reach every product in the category. Nothing has been decided: there is no hearing date, no response from the defendants and no ruling, and the Attorney General’s own press release is two sentences long and carries none of the language above. The motion is on his office’s site and is the only place to read it.
Evidence: no published evaluation of ChatGPT as a social-connection or wellbeing product exists. No trial is registered on ClinicalTrials.gov or ISRCTN under OpenAI as sponsor or under the product name. No clearance or marking applies, OpenAI making no clinical claim, which places the product outside device regulation, which is the position the motion attacks and not one it disputes. Nothing in this item is company-sourced; the relief is quoted from the filing.
(link)
⏺️ Money
⏺️ Germany has two open calls under its loneliness-specific ESF Plus programme, one for municipalities and one opened on 7 September for non-profit providers working with 14- to 27-year-olds, with expressions of interest running to 28 February 2027. “Zusammenhalt stärken – Menschen verbinden” sits under the federal Strategie gegen Einsamkeit and is run by BAFzA for the BMBFSFJ through the Z-EU-S portal, in two stages: an expression of interest, then an invited formal application. The first strand, under the Richtlinie of 17 September 2025 (BAnz AT 01.10.2025 B1), is open only to Gebietskörperschaften (municipalities, districts, independent cities and boroughs in a city-state) and targets adults aged 28 to 59, deliberately bounded below by youth services and above by a separate programme for the over-60s. Grants run from €20,000 to €150,000 per project per year. The ESF Plus share is up to 40% in the more developed regions and up to 60% in transition regions, with applicants covering 60% and 40% respectively and a floor of 10% own contribution. The latest permissible project start is 1 July 2027 and grants are awarded to 31 December 2028. The second strand, under a third Richtlinie published on 7 September 2026 (BAnz AT 07.09.2026 B1), opens the programme to gemeinnützige Träger and moves the target group to 14 to 27. Published envelopes are up to €9.5 million for the first call and €6.5 million for the second, with nothing published for the third. Two things need handling before anyone applies. The 2025 Richtlinie’s own text still sets the expression-of-interest deadline at 31 December 2026, while the managing authority now publishes 28 February 2027 in four places including both call notices; we used 28 February 2027, because it is what the body receiving the applications says and it post-dates the instrument. And the programme summary on esf.de still lists only municipalities as eligible, which its own body text on the same page contradicts. The €20,000 to €150,000 range and the intervention rates are read from the 2025 instrument; the third Richtlinie’s primary text could not be retrieved, and a third-party summary putting its ceiling at €200,000 is not confirmed and is not printed here as fact.
(link)
⏺️ Sweden’s grant for organisations working against involuntary loneliness closes on 15 October, ten days from now, with 98 million kronor to distribute for 2027. Socialstyrelsen opened the round on 1 September. Eligible applicants are non-profit associations, non-profit foundations not run by the state or a municipality, and faith communities and congregations. The total is not on the grant page itself but in the Anvisningar 2027, which also carries the condition that the money is paid only if the Riksdag grants funds for the purpose. Decisions are planned for the first quarter of 2027, and the page notes they may be affected by high application pressure. The previous round drew 801 applications, a demand figure that does not seem to have been reported anywhere. The legal basis is förordningen (2019:474). This is the organisations stream; Socialstyrelsen runs a separate municipalities stream for projects on involuntary loneliness among older people, on its own cycle, and the two are easily confused.
(link)
⏺️ Eighteen days after the Scottish Government said full details of the 2026-27 Community Cohesion Fund would be published “shortly”, they have not been. The fund rises from £300,000 to £500,000 for 2026-27, announced on 17 September and committed in the Programme for Government published on 1 September. No application route, eligibility criteria, opening date or closing date has appeared on gov.scot or from the delivery partner, the STV Children’s Appeal, whose page still describes the £300,000 round and carries only the 2025-26 awards. The announced Ministerial Taskforce on Community Cohesion still has no published membership and no start date. The counts do not agree. The September release says 38 organisations benefited in 2025-26, while gov.scot’s own February release says 32 projects shared £300,000, as does STV plc’s release of the same day. The STV Children’s Appeal award table enumerates 38 named organisations. We used 38, because it is the figure the awarding partner’s published table actually lists and the one the government itself settled on seven months later; the likeliest explanation is that 32 was a launch-day count that grew as awards were finalised, but we could not verify that.
(link)
Two absences in this section are tooling failures and not findings, and I would rather say so than let them read as news. GOV.UK’s Find a Grant service ignored the search term and returned its unfiltered default list of 116 grants, so we cannot say whether a new English competition on loneliness was listed this week. And the Horizon Europe Funding and Tenders portal has now resisted four attempts by three different routes, most recently a 180-second timeout on the SEDIA search API, so the CL2 2027 budgets for HERITAGE-04, TRANSFO-01 and TRANSFO-09 remain secondary-sourced and no euro figure for them appears here. Nothing new opened on Horizon Europe or EU4Health within the window.
Corrections, arguments and anything I have missed are welcome in the thread. They go into the next issue and into the brief the AI works from, which is where most of the improvements so far have come from. If you know the social prescribing literature and think the isolation item has been handled better somewhere than it was in the national evaluation, I would particularly like to hear it.
Hans
Discuss this issue in the General space (members only).