Study population
The UK Biobank is a prospective study of more than 500,000 middle-aged and older adults (aged 37 to 73 years; target age range: 40 to 69 years), who were recruited between 2006 and 2010. The study rationale and design have been described elsewhere [23]. Briefly, individuals registered with the UK National Health Service (NHS) and living within a 25-mile (~40 km) radius of one of 22 assessment centres were invited to participate. Participants provided data on their sociodemographic characteristics, health behaviours and medical history and underwent physical examinations. Linked hospital inpatient records are available for most participants and primary care records are available for half of the participants. A third of the participants also completed an online follow-up mental health questionnaire (MHQ) between 2016 and 2017.
Mental disorders
We identified individuals with a lifetime history of depression, bipolar disorder or anxiety disorders using criteria that we have reported elsewhere [13,14,15]. Cases were ascertained from multiple data sources: the modified Composite International Diagnostic Interview Short Form (CIDI-SF), self-report questions on (hypo)mania and a question on psychiatric diagnoses (UK Biobank data field 20544) which were assessed as part of the MHQ; the nurse-led baseline interview in which participants reported medical diagnoses (field 20002); hospital inpatient records (ICD-10 codes); primary care records (Read v2 or CTV3 codes) and self-report questions on mood disorders from the baseline assessment (field 20126). Individuals with psychosis were excluded from all cases and individuals with bipolar disorder were excluded from anxiety disorder cases due to their increased risk of physical multimorbidity [24, 25]. The depression and bipolar disorder groups were mutually exclusive, but individuals could be included in both the anxiety disorder and the depression group. Individuals could have had other psychiatric comorbidities (e.g. substance use or eating disorders), however these were not the focus of this study.
A non-psychiatric comparison group included individuals who had no mental disorders: (i) had not reported “schizophrenia”, “depression”, “mania / bipolar disorder / manic depression”, “anxiety / panic attacks”, “obsessive compulsive disorder”, “anorexia/bulimia/other eating disorder”, “post-traumatic stress disorder” at the nurse-led interview; (ii) reported no psychiatric diagnoses on the MHQ; (iii) reported no current psychotropic medication use at baseline (field 20003) [26]; (iv) had no ICD-10 Chapter V code in their hospital inpatient record, except for organic causes or substance use; (v) had no diagnostic codes for mental disorders in their primary care record [27]; (vi) were not classified as individuals with probable mood disorder at the baseline assessment; (vii) had no Patient Health Questionnaire-9 (PHQ-9) or Generalised Anxiety Disorder Assessment (GAD-7) sum score of ≥ 5; (viii) did not report that they ever felt worried, tense or anxious for most of a month or longer (field 20421); and (ix) were not identified as cases based on the CIDI-SF and questions on (hypo)mania [13, 15].
Frailty phenotype
We derived the Fried frailty phenotype [1], adapted for the UK Biobank [28, 29]. Participants provided data on weight loss, exhaustion, physical activity and walking speed via touch-screen questionnaires at the baseline assessment (Additional file 1: Table S1). Hand-grip strength in whole kilogramme-force units was measured using a Jamar J00105 hydraulic hand dynamometer. We used the maximal grip strength of the participant’s self-reported dominant hand. If no data on handedness were available or the participant was ambidextrous, we used the highest value of both hands [30]. All variables were coded as zero or one and summed up. Participants with a total score of three or more were classified as frail, while participants with a total score of one or two and zero were classified as pre-frail and non-frail, respectively [1]. Participants with missing data for at least one criterion were excluded.
Frailty index
We also derived a frailty index, following the procedure previously used in the UK Biobank [31]. Health deficits included in this index met the following criteria: indicators of poor health, more prevalent in older individuals, neither rare nor universal, covering multiple areas of human functioning and available for ≥ 80% of participants. The index included 49 variables obtained via touch-screen questionnaires and nurse-led interviews at the baseline assessment, including cardiometabolic, cranial, immunological, musculoskeletal, respiratory and sensory traits, well-being, infirmity, cancer and pain (Additional file 1: Table S2). Categorical variables were dichotomised (no deficit = 0; deficit = 1), and ordinal variables were mapped onto a score between zero and one. The sum of deficits present was divided by the total number of possible deficits, resulting in frailty index values between zero and one, with higher values reflecting greater levels of frailty [32, 33]. Participants with missing data for ≥ 10 variables were excluded [31]. Participants with a frailty index value of ≤ 0.08 were classified as non-frail, while participants with values between 0.08–0.25 and ≥ 0.25 were classified as pre-frail and frail, respectively [34].
Ascertainment of mortality
The date of death was obtained through linkage with national death registries, NHS Digital (England and Wales) and the NHS Central Register (Scotland). The censoring date was 28 February 2021. The most recent death was recorded for 23 March 2021, although the data were incomplete for March 2021.
Covariates
Covariates were identified from previous studies and included age, sex, ethnicity (White, Asian, Black, Chinese, Mixed-race or other) and highest educational/professional qualification (four levels, reflecting similar years of education [35]: college/university degree; education to age 18 or above, but not reaching degree level (“A levels”/“AS levels” or equivalent, NVQ/HND/HNC or equivalent, other professional qualifications); education to age 16 qualifications (“GCSEs”/“O levels” or equivalent, “CSEs” or equivalent; no qualifications), Townsend deprivation index, which is a small-area level measure of socioeconomic status [36], cohabitation with spouse or partner (yes/no) [37], smoking status (never, former or current), alcohol intake frequency (never, special occasions only, one to three times a month, once or twice a week, three or four times a week, or daily or almost daily), systolic and diastolic blood pressure (mmHg), body mass index (BMI, kg/m2), cholesterol (mmol/L), multimorbidity count (zero, one, two, three, four, five or more) and assessment centre.
Statistical analyses
All statistical analyses and data visualisations were done in R (version 3.6.2).
Sample characteristics were summarised using means and standard deviations or counts and percentages. Differences in the frailty index between individuals with and without mental disorders were estimated using standardised mean differences ± 95% confidence intervals (CI) and ordinary least squares regression models. Group differences in the frailty phenotype (non-frail, pre-frail and frail) were estimated using ordinal logistic regression models. We fitted both unadjusted and fully adjusted models. Age-related differences in the frailty index were estimated using generalised additive models within the ‘mgcv’ package [38] in R.
We calculated person-years of follow-up and the median duration of follow-up of censored individuals. Unadjusted survival probabilities by frailty level and case status were estimated using the Kaplan-Meier (KM) method [39]. Hazard ratios (HRs) and 95% confidence intervals were estimated using Cox proportional hazards models [40] to examine associations between frailty and mortality by case status. Age in years was used as the underlying time axis, with age 40 as the start of follow-up. We fitted both unadjusted and fully adjusted models. Non-frail individuals without mental disorders were the reference group. We estimated the percentage risk difference between individuals with and without mental disorders at the pre-frailty and frailty levels using the formula: (HRdisorder – HRno disorder)/(HRno disorder − 1) × 100.
Adjusted P-values were calculated using the p.adjust function in R to account for multiple testing. P-values from the regression models were corrected for six tests (one parameter × two models × three disorders) and p-values from the Cox proportional hazards models for 30 tests (five parameters × two models × three disorders). Two methods were used: (1) Bonferroni and (2) Benjamini and Hochberg [41], two-tailed, with α = .05 and a 5% false discovery rate, respectively. We have opted for this approach because the Bonferroni correction may be too conservative and lead to a high number of false negatives.
Additional analyses
We repeated our main analyses of group differences in frailty and of all-cause mortality stratified by sex. As a sensitivity analysis, we repeated the analyses of all-cause mortality after excluding individuals with comorbid depression and anxiety disorders. Finally, we examined all-cause mortality in individuals with comorbid depression and anxiety disorders.