Data

For this cohort study, secondary data were taken from UK Biobank (https://www.ukbiobank.ac.uk/), which achieved a 5.5% response rate.^{Reference Sudlow, Gallacher, Allen, Beral, Burton and Danesh16,Reference Allen, Sudlow, Downey, Peakman, Danesh and Elliott17} Over 502 000 community-dwelling individuals aged 37–73 years were recruited to UK Biobank during 2006–2010. Participants attended one of 22 assessment centres across England, Scotland and Wales. For this study, we limited the sample to those attending assessment centres in England. Baseline assessments were linked to Hospital Episode Statistics (HES) for England and death records provided by NHS Digital (both up to 31 December 2019). The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. UK Biobank was approved by NHS National Research Ethics Service North West (approval number 21/NW/0157). All adult participants provided written informed consent to participate in UK Biobank. We excluded participants who requested their data be withdrawn from UK Biobank (updated on 9 August 2021).

Cohort definition

Individuals with an SCMD were identified via linked clinical records and/or self-report, using the ‘first occurrence’ variables (see https://biobank.ndph.ox.ac.uk/ukb/label.cgi?id=1712 for full details). UK Biobank provides the date on which a diagnosis was recorded for the first time and the source (e.g. primary care, in-patient data or self-reported data). For each diagnosis group of interest (bipolar disorder (ICD-10 codes F30, F31), schizophrenia and other psychotic disorders (ICD-10 codes F20–F29), depression (ICD-10 codes F32–F39) and anxiety and related disorders (ICD-10 codes F40–F48)), the earliest date on which a diagnosis was recorded (from linked primary care or hospital in-patient data) was identified.

At the baseline assessment centre individuals could also self-report a lifetime diagnosis of ‘schizophrenia’; ‘mania/bipolar disorder/manic depression’; ‘depression’ or anxiety and related disorders: ‘anxiety/panic attacks’, ‘nervous breakdown’, ‘post-traumatic stress disorder’, ‘obsessive–compulsive disorder’ or ‘stress’. A subset of UK Biobank participants (those recruited in 2009–2010) also completed detailed questionnaires about lifetime depressive and mania symptoms at the baseline assessment, from which probable cases of major depression and bipolar disorder have been derived by clinicians.^{Reference Smith, Nicholl, Cullen, Martin, Ul-Haq and Evans18} For solely self-reported records, the date of diagnosis was recorded as the date the individual joined UK Biobank. An individual was classified as having the corresponding mental disorder diagnosis if they had either a clinical record and/or a self-reported record.

If a person had more than one SCMD diagnosis, we ranked them in the following order and classified the patient according to the highest ranked: schizophrenia, bipolar disorder, anxiety or depression. Participants who had an ICD-10 code under Chapter V (mental and behavioural disorders (e.g. eating disorders or as a result of psychoactive substance use)) not covered by the mental disorder categories above, were excluded from the sample (n = 31 923), and all those with no recorded mental or behavioural disorder were grouped into the control group. For the analysis, we grouped those with anxiety or depression (CMDs) together because of their significant comorbidity.^{Reference Merikangas and Kalaydjian19}

Outcome

The primary outcome of interest was emergency hospital admissions for an ACSC. ACSCs were defined according to the NHS England criteria, which includes 19 conditions divided into acute, chronic and vaccine-preventable.^{Reference Fleetcroft, Hardcastle, Steel, Price, Purdy and Lipp20} Acute conditions included cellulitis; dental conditions; ear, nose and throat infections; gangrene; gastroenteritis/dehydration; nutritional deficiency; pelvic inflammatory disease; perforated/bleeding ulcers and pyelonephritis. Chronic conditions included angina, asthma, chronic obstructive pulmonary disease, congestive heart failure, diabetes complications, convulsions/epilepsy, hypertension and iron deficiency anaemia. Vaccine-preventable conditions included influenza and pneumonia and ‘others’ such as tuberculosis and hepatitis B.

Hospital admissions for ACSCs were extracted from the HES for England supplied by NHS Digital via UK Biobank.²¹ HES reports data as episodes (period of care under a consultant), and there can be more than one episode during one hospital stay. Each episode can record multiple diagnoses (via ICD-10 codes), which can be used to identify ACSCs. An ACSC admission was defined as an ACSC condition recorded in the first emergency admission episode.

To identify ACSC admissions, we grouped together consecutive episodes of the same patient. UK Biobank does not report hospital codes, so these continuous periods in hospital can include transfers between different hospitals. To construct these ‘continuous in-patient spells’, we used information about the source and method of admission and the discharge destination, together with the start and end dates of the episodes to make sure the episodes were in the correct order. Supplementary Table 1 available at https://doi.org/10.1192/bjo.2023.602 contains detail on the data exclusions in HES to identify ACSC admissions.

Covariates

We included a range of potential variables that may influence the association between SCMD and ACSC admissions, grouped into sociodemographic, socioeconomic, health and biomarkers, health-related behaviours, social isolation and psychological factors. All data for the covariates were collected at the baseline assessment centre.

Sociodemographic factors included age (years), gender (male, female), ethnicity (White British, White Irish, other White background, South Asian, Black, mixed or other), urban/rural residence (based on home postcode population density) and assessment centre attended. Socioeconomic factors included education level (1: university or college degree; 2: A-levels or equivalent; 3: O-levels, General Certificate of Secondary Education (GCSE), vocational Certificate of Secondary Education (CSE) or equivalent; 4: other (e.g. National Vocational Qualifications or other professional qualifications) or 5: none of the above), deprivation at the output area level (assessed with the Townsend index,^{Reference Townsend22} converted to a Z-score (number of s.d.s from the mean value) where higher levels reflect higher levels of deprivation), employment status (paid employment or self-employment, retired, looking after home and/or family, unable to work because of sickness or disability, unemployment or other), housing tenure (owner-occupier or renter/other) and household income (before tax, self-reported: <£18 000, £18 000–£30 999, £31 000–£51 999, £52 000–£100 000 or >£100 000).

Health measures included multimorbidity (a count of the number of self-reported chronic physical health conditions (0, 1, 2, 3 or ≥4), based on a previously published approach,^{Reference Jani, Hanlon, Nicholl, McQueenie, Gallacher and Lee23} excluding mental health conditions) and body mass index (BMI) category (underweight, normal weight, overweight, obese). We included three biomarkers indicative of inflammation (C-reactive protein (CRP), logged because of its skewed distribution), metabolic function (waist circumference) and cardiovascular function (pulse rate). Indicators of health behaviours included smoking (never, previous, current) and alcohol consumption (daily or almost daily, 3–4 times a week, once or twice a week, 1–3 times per month, special occasions, former drinker or never). Physical activity (walking, moderate and vigorous) in a typical week was also recorded with self-reported items from the International Physical Activity Questionnaire Short Form,^{Reference Booth24} from which a single measure of total physical activity in metabolic equivalent of task hours per week was derived; this was converted into quintiles.^{Reference Cullen, Smith, Deary, Pell, Keyes and Evans25}

A number of measures were used to capture social isolation: living arrangements (with spouse/partner, with other people, live alone), social contact (visit friends/family less than weekly versus at least once a week) and social participation (one or more activity, e.g. sports club, at least once a week versus no activities). Finally, psychological factors included loneliness (whether participants often feel lonely, yes or no), current depressive symptoms (measured with an adapted Patient Health Questionnaire-4)^{Reference Kroenke, Spitzer, Williams and Löwe26} and sleeplessness (never/rarely, sometimes, usually).

Statistical analysis

First, descriptive statistics for the sample were calculated, including the number of hospital admissions by SCMD diagnosis. We then ran several survival models to assess the relationship between SCMD (schizophrenia, bipolar disorder, depression or anxiety, and those with no disorder as the reference group) and ACSC admissions. We ran models in the following order to examine the associations using different groups of covariates, in particular those over and above sociodemographic and socioeconomic factors that are most often included in previous studies:^{Reference Davydow, Ribe, Pedersen, Fenger-Grøn, Cerimele and Vedsted6}

1. (a) unadjusted;

2. (b) model 1 plus age, gender, ethnicity, urban/rural and assessment centre location (sociodemographic factors);

3. (c) model 2 plus education, deprivation, employment status, housing tenure and income (socioeconomic factors);

4. (d) model 3 plus multimorbidity, BMI, pulse rate, waist circumference and CRP (health and biomarker variables);

5. (e) model 3 plus smoking, alcohol consumption and physical activity (health-related behaviours);

6. (f) model 3 plus living arrangements, social participation and social contact (social isolation factors);

7. (g) model 3 plus depressive symptoms, sleeplessness and loneliness (psychological factors);

8. (h) all variables.

The observation period for each person started on the date of the initial baseline UK Biobank assessment centre attendance or when they were diagnosed with an SCMD, if the diagnosis was later than the assessment. During the observation period, a patient could have none, one or more than one hospital admission, and we included participants who had ACSC admissions before joining UK Biobank, because of the older age of participants. Models were censored at the earliest date of ACSC admission, date of death or the end of follow-up on 31 December 2019. We use two model specifications, one modelling the time to first admission within the study period (Cox proportional hazard model) and one that considered all admissions (Prentice–Williams–Peterson total time (PWP-TT) model).^{Reference Westbury, Syddall, Simmonds, Cooper and Sayer27,Reference Amorim and Cai28} The time to first admission model does not use all data (it ignores second and later admissions), and can show associations between covariates and admissions that do not hold once all admissions are considered.^{Reference Westbury, Syddall, Simmonds, Cooper and Sayer27} The PWP-TT analyses ordered multiple events via stratification, based on the prior number of events during the follow-up period.^{Reference Amorim and Cai28,Reference Prentice, Williams and Peterson29} It therefore takes into account that having a prior admission affects the risk of future admissions, and that the effect of covariates may differ in subsequent events.^{Reference Amorim and Cai28} Further details on these models and how they can be implemented in Stata software can be found elsewhere.^{Reference Westbury, Syddall, Simmonds, Cooper and Sayer27} To maximise the use of available data, participants with missing data for any variable were excluded from the analysis by using pairwise deletion (models contain a different number of individuals and therefore should not be directly compared). The extent of missing data varied from 1.5% in model 2 to 35.5% in model 8 (mainly because of the high proportion of missing data relating to physical activity). Violations of the proportional hazard assumption were examined graphically by plotting scaled Schoenfeld residuals. Statistical analysis was performed with software Stata/MP version 17 for Windows.