Open Dialogue missed its primary endpoint. It still cut inpatient admissions.

Open Dialogue is a model of mental health care built around network meetings: the service user, members of their social network, and usually two practitioners who stay with that network for the duration of care, with an emphasis on continuity, therapeutic relationships and collaboration [s1]. It originated in Finnish practice and spread through NHS services on the strength of non-randomised evidence and enthusiasm. Until now it had never been tested in a randomised controlled trial [s2].

The result, published in The Lancet Psychiatry on 26 August, is the kind that gets read two different ways depending on which line you stop at. Open Dialogue did not beat treatment as usual on the outcome the trial was designed around. It did better on most of the others.

The design

ODDESSI was a multicentre, parallel two-arm, cluster-randomised superiority trial run in mental health services across five NHS trusts in London and the south of England [s1]. Clusters were defined at the level of primary care practices within service catchment areas, randomised 1:1, stratified by catchment area and balanced on average GP list size and the 2015 Index of Multiple Deprivation [s1]. Participants were adults aged 18 or older presenting in crisis to mental health services and registered with a practice inside a trial cluster [s1].

Of 185 general practices screened, 105 were excluded and 80 formed 32 clusters, 16 assigned to each arm; one trust withdrew with two clusters, leaving five trusts and 30 clusters [s1]. Between 25 June 2019 and 9 December 2021, 494 participants consented — 223 to treatment as usual, 271 to Open Dialogue — with a mean age of 38.1 years (SD 13.4); 266 (54%) were of female gender, 221 (45%) male gender, and 341 (69%) White British [s1].

The primary outcome was time in days to first relapse after initial recovery from the index crisis, censored at the end of two years of follow-up [s1]. The chief investigator, senior statistician and the assessors of that outcome were masked [s1]. The protocol, published in 2021, had prespecified time to relapse as primary and named shared decision-making and social network quality and size as the mediators the trial would investigate [s2].

What it found

There was no significant difference in time to relapse: a marginal hazard ratio of 0.95 (95% CI, 0.67–1.32) [s1]. Of those consented, 174 (78%) in the treatment-as-usual arm and 225 (83%) in the Open Dialogue arm recovered and had data allowing relapse to be determined [s1].

The secondary outcomes went the other way. Open Dialogue was associated with significantly lower probabilities of psychiatric inpatient admission and of re-referral to crisis care or to secondary mental health services, and with improvements in self-rated recovery, health-related quality of life and satisfaction with services [s1]. Social network quality and size — one of the two mediators the protocol had flagged [s2] — did not differ significantly between groups [s1]. There were 386 serious adverse events, 281 in the treatment-as-usual arm and 105 in the Open Dialogue arm; 376 (97%) were deemed unrelated to the intervention [s1].

How to read a missed primary endpoint

The honest summary is that the trial's own hierarchy says Open Dialogue did not work as hypothesised. Time to relapse was the question asked in advance, and the answer was null. Reading the secondary outcomes as if they carried the same weight is the error that produces a decade of irreproducible practice.

But the secondaries are not decoration either, and two features make them harder to dismiss than usual. Inpatient admission and re-referral were extracted from electronic health records rather than self-reported, so they are less vulnerable to the unblinded participants' expectations. And the pattern is internally coherent: fewer admissions, fewer re-referrals, better self-rated recovery and higher satisfaction all point the same way, at a service model whose stated aim is continuity rather than symptom suppression.

That coherence is also the caution. In a trial where participants and clinicians cannot be masked to a service-delivery model, satisfaction and self-rated recovery are exactly the measures most likely to move for reasons other than clinical benefit. The bed-use finding is the one that would survive that objection, and it is the one worth replicating.

The authors' own conclusion is measured: Open Dialogue did not reduce time to first relapse, but it reduced acute inpatient bed use, improved service-user-reported outcomes and experience, and raised no significant safety concerns; further investigation is required to determine whether it can improve the effectiveness and acceptability of crisis and continuing care [s1]. The trial was funded by the National Institute for Health Research and is complete [s1].

What happens next is a commissioning question as much as a scientific one. A model that does not change relapse timing but reduces admissions has a cost case that is separable from its clinical one — and the health-economic analysis the protocol specified [s2] is where that argument will be settled.

Sources

• [s1] "Open Dialogue versus treatment as usual for adults presenting in crisis to mental health services in England (the ODDESSI Trial): a multisite cluster-randomised trial," The Lancet Psychiatry, 26 August 2026. https://doi.org/10.1016/S2215-0366(26)00229-4
• [s2] "Open Dialogue compared to treatment as usual for adults experiencing a mental health crisis: Protocol for the ODDESSI multi-site cluster randomised controlled trial," Contemporary Clinical Trials, 24 December 2021. https://doi.org/10.1016/j.cct.2021.106664