This study complies with all relevant ethical regulations. All participants provided written informed consent. Ethical approval for the most recent wave of data collection for this data was obtained from the South East Coast—Brighton and Sussex MREC (15/LO/1446). Information about ethical approvals obtained for previous waves of data collection can be accessed online: https://cls.ucl.ac.uk/wp-content/uploads/2017/07/BCS70-Ethical-review-and-Consent-2019.pdf.
Data
This study used data from the 1970 British Cohort Study (BCS70), a nationally representative longitudinal cohort study of 17,198 individuals born in England, Scotland and Wales during a single week in 1970 (ref. 25). Owing to attrition and missing data over follow-up, the number of participants with valid data for analyses was lower than the original 17,198 participants enrolled at birth. The main analytic sample comprised 9,280 participants. Characteristics of the analytic sample and differences to the sample with missing data are described in detail in ‘Results’. Follow-up data have been collected over 46 years, capturing a broad range of demographic, health and lifestyle information. Further details on data collection methods and participation rates at each wave have been published previously.
MeasuresADHD traits
ADHD traits were derived using a validated 14-item measure based on items collected at age 10 (ref. 26) (9 items corresponding to hyperactivity and 5 to inattention). These items were part of standard behavior questionnaires completed by parents and teachers in BCS70 (refs. 27,28). Cotton and Baker (2019) applied a data mining framework to map these items from BCS70 onto the DSM-5 criteria for ADHD, identifying 14 items that corresponded closely to these criteria. A zero-inflated item response theory mixture model was used to derive a continuous dimensional score of ADHD traits, alongside a binary indicator of high ADHD traits based on DSM-5-aligned thresholds (N = 469, 5.05% of analytic sample met criteria). This approach demonstrated strong model fit and discriminative validity, with the derived scale showing expected associations with known correlates of ADHD, including male sex, social disadvantage, and later educational and behavioral outcomes26. Psychometric evaluation has shown strong reliability and validity of this measure, with a Kaiser–Meyer–Olkin value of 0.90 and Cronbach’s α = 0.83. It showed high correlations with a mapped Strengths and Difficulties Questionnaire hyperactivity subscale (r = 0.74, P < 0.001) and an ADHD proxy derived from the same cohort (r = 0.82, P < 0.001), supporting construct validity. The binary variable showed comparable prevalence and composition (in terms of gender and subtype) to meta-analytic estimates of ADHD in the general population, supporting criterion validity.
The dimensional measure was used in the main analyses to capture the full population range of ADHD traits and to improve power. A detailed user guide describing how measures can be derived is available from the Centre for Longitudinal Studies (CLS). Items included are presented in Supplementary Material 1.
Psychological distress
Psychological distress was assessed at ages 26, 30, 34, 42 and 46 using the Malaise Inventory Scale29. This questionnaire has shown good internal consistency and convergent validity29. At some time points, the full 24-item questionnaire was administered, but at other time points, the 9-item short form was used. To maintain consistency in assessment over time, the nine items from the short form were used for all time points. At each time point, a sum score indicating the number of items cohort members endorsed was derived. In addition, an ‘accumulation’ score was also created. To do this, binary variables were derived for each time point, indicating whether cohort members met or exceeded the threshold for clinically relevant distress (≥4). A cumulative distress measure was created by calculating the proportion of time points at which individuals reported high distress, based on their available data. A proportion score was used instead of a sum score (that is, number of time points with clinically relevant distress) to maximize sample retention among cohort members with missing data. Cohort members with data available for fewer than three time points were excluded from this calculation.
Societal exclusion factors
Societal exclusion has previously been conceptualized as a multidimensional construct30,31 encompassing key domains captured within BCS70 at age 34 (ref. 32): health, relational, political, economic and service exclusion. Measures based on this framework has been validated in BCS70 using data and items from the age 34 sweep32, with confirmatory factor analysis supporting its construct validity across domains.
Health exclusion refers to poor health and comprised 4 items: self-assessed health (5-point scale, ranging from excellent to very poor), whether health limits daily activities (3-point scale, including yes, no but health problems, and no and no health problems), life satisfaction (measures on a 0–10 scale) and self-efficacy (a combined score including three items: whether the cohort member gets what they want out of life, has control over their life and can run life as they want). The item indicating risk of depression used in ref. 32 was removed owing to overlap between this measure and the study outcome.
Relational exclusion refers to lack of emotional support and distrust in others. This was made up of 3 items: relationship status (currently in a relationship, or not), trust in people (4-point scale, ranging from ‘a lot’ to ‘not at all’) and emotional support (number of people cohort member reported being able to turn to for support).
Political exclusion refers to minimal or no political engagement, and comprised 4 items: whether cohort member voted in the last general election (yes or no), whether interested in politics (4-point scale, ranging from ‘very interested’ to ‘not at all interested’), belief in influence on decisions affecting local area (4-point scale, ranging from ‘definitely agree’ to ‘definitely disagree’) and active participation (self-report of contact with governments, attending public meetings, participation in public demonstration or signing petitions in the last year).
For parsimony and owing to high correlations between income and employment, resource and labor market exclusion were combined into one domain indicating economic exclusion. Economic exclusion refers to financial instability and unemployment, and was comprised of 5 items: income poverty (reported income below 60% of median income), subjective poverty (5-point scale, ranging from ‘living comfortably’ to ‘finding it very difficult’), saving opportunities (self-report of whether cohort member saves any amount of their income), credit market exclusion (self-report of whether cohort member borrowed money from pawnbroker, money lender, friends or family in the last year) and economic activity (employed or unemployed).
Services exclusion refers to limited access to high-quality public services. This was based on cohort members’ ratings of public services in their area, including social and leisure facilities, health services, education services, police services and public transport services. These items were all scored on a five-point scale, ranging from ‘very good’ to ‘very poor’.
In line with the validated model in previous research32, each item was dichotomized to indicate high (1) versus low (0) exclusion. Thresholds used for dichotomization are available elsewhere32. Next, a summary score was created for each domain of societal exclusion by summing the relevant indicators, with higher scores reflecting greater levels of exclusion.
Covariates
Covariates included sex assigned at birth, ethnicity and social class at age 10. Sex was classified as male or female. Ethnicity was categorized as white or minoritized ethnicity. Owing to the small sample size of minoritized ethnic groups in this sample (<3%), a more detailed breakdown was not possible. Social class at age 10 was determined based on the father’s occupation, or the mother’s occupation if this was unavailable. Classification followed the Registrar General’s Social Class (RGSC) framework, which includes unskilled, partly skilled, manual, non-manual, managerial and technical, and professional categories33.
Statistical analyses
The sample with missing data was compared with the complete-case sample to assess differences in key variables and covariates, using t-tests and χ2 tests as appropriate.
First, linear regression models were used to test whether ADHD traits were associated with cumulative distress from ages 26 to 46 (proportion of time points in which clinically relevant distress was reported). Next, growth curve models were fit to the data to model distress scores from ages 26 to 46. Linear and quadratic models were fitted and compared based on standard fit indices (CFI and TLI ≥ 0.95, RMSEA ≤ 0.06, lower χ2 values indicating better fit25), with the best fitting model used in subsequent analyses. Next, growth mixture models were fitted to identify distinct trajectories of distress. Models with a 2-, 3-, 4-, 5- and 6-class solution were fitted and compared using a range of standard indices, including AIC and BIC (lower scores indicating better fit), entropy (values closer to 1 indicating better classification, and a threshold of ≥0.80 used to indicate acceptable entropy), percentages in each class (to avoid small and unstable groups) and Lo–Mendell–Rubin adjusted likelihood ratio test (to test whether there is significant improvement in fit compared with a model with one less class). Once the optimal number of trajectories was identified, class membership was extracted to use in subsequent analyses. Multinomial logistic regressions were run to test associations between childhood ADHD traits and distress trajectory class membership. The class with the largest sample size was used as the reference category.
Predicted probabilities of having clinically relevant distress at age 46 were estimated using the margins command following adjusted logistic regression in Stata, comparing individuals with and without high ADHD traits, using the binary indicator.
Finally, path models were run to test (1) direct effects of ADHD traits on distress at age 46, and (2) indirect effects through domains of adult societal exclusion: health, relational, political, economic and services. Covariances were included between societal exclusion domains to account for known correlations. Model fit was assessed using standard fit indices (χ2, CFI, TLI and RMSEA)34.
Models were run unadjusted (model 1) and adjusted for key covariates (sex, ethnicity and social class at age 10) (model 2). Missing data were dealt with in trajectory models and path models using full information maximum likelihood. Analyses were conducted in Stata v18 and MPlus v8.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.