The Second Person in the Room: Late Diagnosis and Relationship Outcomes

The question

Autism and ADHD are now routinely diagnosed in adults in their thirties, forties and fifties. At those ages most people are partnered, and have been for some time. A late diagnosis therefore does not arrive into a life. It arrives into a relationship that was built, negotiated and defended for years without it.

Does the relationship change? In what direction, over what period, and for which partner first?

As far as I can find, nobody has measured it.¹

Where the question came from

Adults are being diagnosed with autism and ADHD in numbers that have no historical precedent. Referrals to adult services have climbed steeply across the English-speaking world, driven by public awareness, social media, and the belated recognition that a generation of people - disproportionately women, disproportionately those who learned to camouflage well - were simply missed as children.² The average age of an adult autism diagnosis sits somewhere in the thirty-one to forty-nine range.³

That age range is the point. It is not a number about diagnosis; it is a number about life stage, and it is the reason this is a question about couples rather than individuals.

The event itself has a recognisable shape. It is worth treating as a shock in the technical sense - a discrete, dateable event that arrives from outside the relationship and changes the terms on which the two people understand each other. Something happens in the months that follow. Both partners must reinterpret a shared history. One of them acquires a new framework for understanding their own behaviour, usually at speed and often obsessively, while the other receives that framework in fragments. Roles that were stable become negotiable. Old arguments are reopened with new evidence. The explanation that was supposed to resolve everything initially makes everything harder.

I described that pattern at length in the relationship chapter of my book. I described it because it is what clinicians and couples consistently report. What I could not do, when writing it, was cite a single study that had followed couples through it and measured what actually happened. I looked. The chapter rests on clinical observation and qualitative accounts, and I was uncomfortable about that then. I am more uncomfortable now.

Why divorce statistics cannot answer it

The obvious move is to reach for divorce rates. It is the wrong move, and I want to be precise about why, because the error is everywhere in the popular literature on this topic.

Divorce is dissolution conditional on marriage, and the neurotypes do not enter marriage at equal rates. In a study of 306 adults grouped by ADHD and autism traits, the autism-traits group contained fewer married people and more people who had never married, while the elevated divorce rate appeared in the ADHD-traits group instead.⁴ Those are different denominators. If autistic adults marry later and more selectively, the married subset is already survivor-selected, and a low divorce rate among them could mean unusually good matches or simply that the fragile relationships never became marriages and so were never counted.

Worse, divorce collapses a continuous state into a binary event. It cannot see the marriage that persists while both people are miserable, which is precisely the cell this question is about. Clinical estimates of marital “maladjustment” in ADHD-affected couples run near sixty per cent, far above any divorce figure - and that gap between not divorced and well is where the entire phenomenon lives.⁵

Then the practical objections. Marriage rates are falling and de facto partnership is not registered anywhere. A relationship can end by death, by drift, by two people continuing to share an address. And divorce is a slow legal artefact that may be recorded years after the relationship it describes actually finished, which makes it close to useless for dating an effect against an event.

None of this means the question is unanswerable. It means the outcome measure has to be built rather than borrowed.

What the evidence actually says

Here is the honest inventory. There are two literatures, both real, and they do not touch.

The first is about neurodiverse couples. It is small but no longer trivial. Yew, Hooley and Stokes surveyed 95 autistic adults and 65 non-autistic partners and found that partner responsiveness - feeling understood and cared for - predicted relationship satisfaction for both, which shifts the focus off the autistic partner as the sole variable.⁶ Khaw and Vernon compared 106 autistic adults grouped by their partner’s neurotype and found no statistically significant differences in satisfaction between autistic/autistic, autistic/neurotypical and autistic/neurodivergent dyads, though the specific enabling and challenging factors varied.⁷ Smith and colleagues explored what actually sustains these relationships qualitatively.⁸ Attwood and Aston, drawing on clinical practice and survey work, suggest satisfaction may be higher in autistic-autistic couples, though they had few such couples respond.⁹

Notice what none of these do. Not one of them treats diagnostic timing as a variable. Participants are recruited as diagnosed autistic adults and their partners. Whether the diagnosis arrived at seven or forty-seven, before the relationship began or fifteen years into it, is not asked, not reported, and not analysed.

The second is about late diagnosis. It is larger and growing quickly. A 2026 systematic review of twenty-five studies on identity reconstruction following adult diagnosis of autism or ADHD found a consistent pattern of relief mixed with grief, frustration and disorientation, and concluded that post-diagnostic support must address identity.¹⁰ Older adults diagnosed in later life describe existential upheaval, and the authors argue that diagnosis should be treated as a starting point rather than a solution.¹¹ A study of nineteen autistic adults and four support persons found that informal support - which in practice usually means a partner - improved outcomes while creating strain in the very relationships providing it.¹²

Notice what none of these do. The relationship appears as context, as a source of support, occasionally as collateral damage. It is never the outcome. Nobody administers a dyadic measure. Nobody follows the couple.

The closest thing I have found to my question is Holmes, an exploratory phenomenological study of later-in-life diagnosis and marital satisfaction.¹³ It exists, it is qualitative, it is small, and it is one study. That is the state of the field.

So: two mature-ish literatures, one measuring couples without asking about timing, one measuring timing without asking about couples. The question sits exactly in the seam.

The part where I argue against myself

I would rather state the weaknesses in my own idea than have a supervisor state them for me. There are three, and the first is severe enough to sink a badly designed version of this study.

Reverse causation. Adults do not seek assessment at random. They seek it because something forced the question, and relationship difficulty is one of the most common triggers. A comparative analysis of 160 adults referred for ADHD and autism assessment found that those on the ADHD pathway were typically responding to acute crisis and seeking legitimisation, while those on the autism pathway were seeking frameworks for self-knowledge and for mutual comprehension within relationships and institutions.¹⁴ Read that second clause carefully. People pursue autism assessment partly because a relationship has stopped making sense.

Which means a naive finding that relationships deteriorate after late diagnosis may be measuring nothing but the crisis that produced the diagnosis-seeking in the first place. The arrow runs backwards. Any design that recruits people after diagnosis and asks them to look back is measuring a population selected on the outcome, and will produce a confident, publishable, worthless result.

Attrition that correlates with the outcome. This is the quiet killer in longitudinal couples research. The couples who separate are exactly the couples who stop answering surveys. If separation is coded as missing data rather than as data, the study will report that relationships are broadly stable after diagnosis while the failures walk silently out of the sample. Any design has to treat dissolution as a measured outcome with a defined exit protocol, not as dropout.

The construct problem. “Did the relationship survive” is not the same as “did it succeed”, and defaulting to the first imports an assumption that lifelong cohabitation is the only good ending. A partnership that converts into a stable co-parenting friendship, or that ends cleanly with both people better off than they were, counts as a failure by the survival measure and a success by almost any other. If a study of neurodivergent relationships is going to impose a definition of success, it should at minimum notice that it is doing so, and ask the participants what they were aiming at.

Access to participants, which may be the one that kills it. Adult autism research runs on small samples for structural reasons, and this design needs a harder sample than most. Recruitment leans on online convenience sampling, and the resulting cohorts are demonstrably unrepresentative: a review of 36 autism studies recruiting through social media found samples skewing markedly female and markedly more educated than the autistic population they claim to describe.¹⁵ The pipeline leaks well before that, too - women face greater barriers to diagnosis in the first place, shrinking the pool before recruitment even begins.¹⁶ Sitting behind both is a trust problem the field has earned. A UK study of 1,516 researchers and community members found researchers believed they were engaged with the autistic community while autistic people and their families did not agree, against a background of funding directed away from the questions the community says matter.¹⁷ Asking late-diagnosed adults to hand over the state of their marriage, twice a year, for two years or more, is a substantial request to make of a population with good historical reasons for declining.

Then the design doubles it. Dyadic enrolment requires both people to screen, consent and complete independently, and the funnel is brutal: in one large online dyadic study, of 3,826 individuals assessed for eligibility only 1,076 enrolled, with a third of the losses caused simply by a partner who never started the screening.¹⁸

And here is the part that should worry me most. Partner referral is not random with respect to the thing I want to measure. Research on dyadic recruitment has found that participants who successfully brought their partner into a study were significantly more satisfied in their relationships than those who did not.¹⁹²⁰ That is selection directly on the outcome variable. The couples who enrol as couples are the couples doing better - which means a study of relationship deterioration recruits, by construction, the people least likely to show it. It compounds the attrition problem above rather than sitting beside it: the sample is skewed toward the happy at intake and further purged of the unhappy at every follow-up.

I do not think this makes the study impossible, but I think it makes a single-site, single-researcher, individually-recruited version of it impossible, and any proposal that does not budget for this is a proposal that will under-recruit and quietly conclude nothing.

There is a fifth possibility that I should name even though it is not a design flaw: the effect may not exist. It is entirely possible that once you control for the crisis that prompted the referral, late diagnosis makes no measurable difference to a relationship’s trajectory at all - that the destabilisation clinicians describe is real at the level of experience but washes out at the level of outcomes, or is simply the ordinary base rate of relationship difficulty in the fifth decade of life. That would be a genuinely useful finding, and a study designed only to confirm the shock would be unable to produce it.

The design that solves the hard problem

The reverse-causation objection looks fatal until you notice something about how adult assessment actually works.

Not everyone who is assessed is diagnosed. In one UK specialist service, 89% of 422 referrals received an autism diagnosis - which means eleven per cent did not.³ Rates vary considerably by service and by pathway, and in more general adult services the negative proportion is substantially larger.

That group is the control that this question needs. People who present for assessment and do not receive a diagnosis arrived by the same route, under the same pressures, in the same state of relationship strain, with the same expectations - and then the shock did not land. They control for the crisis that prompted the referral, because they had it too. If relationships deteriorate at the same rate in the assessment-negative group, the diagnosis is a marker rather than a cause. If they diverge, that divergence is the effect, cleanly identified.

The second structural gift is the waiting list. Median waits for adult neurodevelopmental assessment ran to 252 days in one Scottish audit, with wide variation and long tails.²¹ From a clinical standpoint that is a scandal. From a design standpoint it is an eight-month window in which a cohort can be recruited, baselined and observed before anyone knows the answer - which is the only way to get genuine pre-shock measurement rather than retrospective reconstruction through the lens of a diagnosis already received.

What I would actually do

• Recruit at referral, not at diagnosis. Enrol partnered adults at the point of joining an adult assessment waiting list, before the outcome is known. Neither the participant nor the researcher can know which arm they will fall into, which removes the selection problem at the root.
• Enrol both partners, and analyse them dyadically. Separate responses, not a joint one. Where the two accounts diverge is not noise; the size and direction of that divergence is one of the more interesting things this study could measure.
• Baseline the relationship trajectory before the result arrives. Standard dyadic satisfaction instruments, plus individual wellbeing for both partners, plus a structured account of when the difficulties began relative to the referral.
• Record the referral trigger as a stratifying variable. Self-driven recognition, a child’s diagnosis, workplace breakdown, or relationship crisis. If the effect appears only in those referred because of the relationship, that is the reverse-causation story, and the design should be able to see it.
• Follow the assessment-negative group with equal rigour. They are not attrition. They are the comparison arm, and they will be tempting to lose.
• Recruit through services, not social media. The waiting list is not only a pre-shock observation window, it is the only route to a sample that is not self-selected from the more articulate, more online, more educated end of the population. It requires multi-site NHS / NDIS or equivalent partnerships from the outset, which is a supervisor-and-institution problem rather than a student one.
• Treat single-partner enrolment as a permitted state. Requiring both partners to consent before either counts imports the satisfaction bias described above. Better to enrol the referred adult regardless, invite the partner separately, and treat partner participation as a measured variable in its own right - because whether the partner agrees to take part is itself informative about the relationship.
• Budget for a feasibility phase and publish it. Recruitment rate and dyad-completion rate at one site, reported honestly, before committing to a powered design. If the dyadic version proves unrecruitable, the fallback is a longitudinal qualitative design on a smaller cohort, which answers less but answers something.
• Follow for at least twenty-four months post-result. The clinical accounts describe a destabilisation that resolves or fails to resolve over a year or more. A six-month follow-up would catch the disruption and miss the outcome.
• Treat separation as an outcome, with a paid exit interview. Consent to this at enrolment. A couple who separate at month fourteen are the single most informative data point in the study and must not be allowed to vanish.
• Ask both partners to define success at baseline, and compare against it at follow-up. This costs almost nothing and is the only way to avoid smuggling in an assumption about what these relationships were for.

My honest prediction is that the raw comparison will show deterioration after late diagnosis, that a substantial portion of it will survive the assessment-negative control, and that the surviving effect will be smaller than the clinical literature implies and concentrated in the first twelve months. I also expect the two partners’ trajectories to diverge sharply and asynchronously, and that the divergence, rather than the average, will turn out to be what matters.

I could be wrong on all counts. That is the point of doing the research.

Why this one is on the list

Because there is currently no post-diagnostic support for adults worth the name, and the little that exists is aimed at the individual.

A clinician hands over a diagnosis that reframes forty years of a person’s life, and then the appointment ends. The person goes home to a partner who was not in the room, has read nothing, and is about to spend the next year being told that things they experienced as unkindness were neurology. There is no leaflet for that. There is no session for that. There is no evidence base telling either of them what is normal, how long it lasts, or whether what they are going through is the start of a breakdown or the middle of an adjustment. That void is a large part of why I wrote Loving Variant Minds. My working assumption - as someone coming into psychology rather than out of it - is that the discipline’s primary instrument is education rather than cure: most of what a psychologist usefully does is explain to people what is happening to them and hand them a strategy for it. A book was the only way I could see to put something into the space between the newly diagnosed person and the partner trying to understand them. But a book assembled from clinical accounts and lived experience is a stopgap, not an evidence base. I could describe the pattern. I could not tell anyone how common it is, how long it runs, or which couples come through it - because the profession cannot tell them either.

If the trajectory turns out to be predictable - if there is a reliable window of destabilisation with a typical shape and duration - then it can be prepared for. That is not an expensive intervention. It is a conversation at the point of diagnosis, with the partner present, saying: the next twelve months will probably be harder than the last twelve, here is why, here is what tends to help, and here is what it means if it has not settled by then. Nobody can have that conversation now, because nobody has done the study that would tell them what to say.

The alternative finding is just as useful. If late diagnosis makes no measurable difference once the precipitating crisis is accounted for, then a great deal of anxiety currently circulating among newly diagnosed adults and their partners is misplaced, and saying so with evidence would be worth doing on its own.

Either way, the cohort exists, it is large, it is growing, and it is sitting on waiting lists right now.

---