Study population

Data used in this study were derived from the UK Biobank, a large population-based longitudinal study. Between 2006 and 2010, more than 500,000 individuals aged 37–73 years were recruited and underwent comprehensive assessments at 22 assessment centers across the United Kingdom. All enrolled participants provided informed and written consent. Of 502,412 participants in the baseline examination, we excluded 65,716 with a history of depression (n = 65,543) or dementia (n = 238) at recruitment, 268 who developed depression after the occurrence of dementia, and 196 who developed both dementia and depression on the same date. Overall, 436,232 participants were included in the current study (Supplementary Figure S1).

The UK Biobank received ethical approval from the North West Multi-Centre Research Ethics Committee (21/NW/0157), and our work was performed under the UK Biobank application number 67048 (PI: Weili Xu).

Data collection

At baseline, information on participants’ age, sex, race (white versus mixed, Asian or Asian British, black or black British, Chinese, or other ethnic groups), smoking status (never, previous, or current), alcohol consumption (never, previous, or current), and physical activity was self-reported through computerized touch-screen questionnaires. Physical activity was measured as total metabolic equivalents (MET) per week using the modified version of the International Physical Activity Questionnaire and classified as low (<600 MET-min/week), moderate (600 to <3000 MET-min/week), or high (≥3000 MET-min/week) [Reference Chudasama, Khunti and Zaccardi14]. Body weight and height were measured, with body mass index (BMI) calculated as weight (kg)/(height (m) ²). Hypertension was identified based on systolic blood pressure ≥ 140 mm Hg, diastolic blood pressure ≥ 90 mm Hg, self-reported history of hypertension, use of antihypertensive drugs, or medical records. Diabetes was defined as the presence of hemoglobin A1c ≥6.5%, fasting plasma glucose ≥126 mg/dL, self-reported history of diabetes, use of glucose-lowering medications, or medical records. Heart disease (including myocardial infarction, angina, atrial fibrillation, and heart failure) and stroke were ascertained through medical records and self-reported medical history.

Assessment of CR and generation of CR indicator

CR was assessed based on six factors including educational level, occupational complexity, mentally passive sedentary behavior, social connection, confiding in others, and leisure activity engagement, as defined in previous studies [Reference Wang, MacDonald, Dekhtyar and Fratiglioni15–Reference Yang, Wang and Dove18]. All information about these factors was self-reported at baseline.

Educational level was determined according to the years of regular schooling converted based on the International Standard Classification of Education scale, divided into (1) no educational qualifications (equal to 7 years), (2) Certificate of Secondary Education, Ordinary levels/General Certificate of Secondary Education (equal to 10 years), Advanced levels/Advanced Subsidiary levels or equivalent (equal to 13 years), (3) other professional qualifications (equal to 15 years), (4) National Vocational Qualification, Higher National Diploma, Higher National Certificate or equivalent (equal to 19 years), or (5) college/university degree (equal to 20 years) [Reference Okbay, Beauchamp and Fontana19].

Occupational complexity was assessed based on participants’ current (or, for retired people, longest-held) occupation and categorized according to the UK Standard Occupational Classification 2000 system, which was developed by the UK Office of National Statistics [20]. Occupation was further classified into one of the eight socioeconomic categories in the National Statistics Socio-economic Classification (SEC) [21], coded as ordinal variables ranging from 1 to 8, where lower values indicate higher occupational complexity and attainment (i.e., jobs requiring more thought and higher skill levels) [Reference Ko, Kim and Kim22]. Occupational complexity was categorized into five levels: (1) never worked and long-term unemployed (SEC 8) or routine occupations (SEC 7), (2) semi-routine occupations, small employers and own account workers, or lower supervisory and technical occupations (SEC 6–4), (3) intermediate occupations (SEC 3), (4) lower managerial and professional occupations (SEC 2), and (5) higher managerial and professional occupations (SEC 1).

Mentally passive sedentary behavior was assessed based on the time (in hours/day) that participants spent in watching television, categorized as (1) ≥4, (2) 3–3.9, (3) 2–2.9, or (4) <2.

Social connection was measured based on the frequency of participants visiting or being visited by friends or family, divided into (1) no friends/family outside household or about once a month or less, (2) about once a week, (3) 2–4 times a week, or (4) almost daily.

Confiding in others was determined based on the frequency of participants confiding in someone close to them, classified as (1) never or almost never, (2) about once a month or less, (3) 1–4 times a week, or (4) almost daily.

Leisure activity engagement was assessed according to the number of leisure activities (including sports club or gym, pub or social club, religious group, adult education class, and other group activity) participants engaged in at least once a week, classified as three levels: (1) low (none), (2) moderate (1 activity), or (3) high (2–5 activities).

A composite CR indicator was constructed using latent class analysis (LCA) based on these six factors. LCA is a well-validated statistical approach that can identify hidden clusters by grouping multiple observed categorical variables (i.e., CR-related factors) into a latent variable (i.e., the CR indicator) with mutually exclusive latent classes. Three latent classes were identified after comprehensively considering statistics regarding model selection (with a relatively lower Bayesian information criterion value) and the uncertainty of posterior classification (with mean posterior probabilities in all latent classes >0.70), and they respectively represented a high (characterized by higher levels of education, occupational complexity, confiding in others, and leisure activity engagement as well as less mentally passive sedentary behavior), moderate (characterized by moderate levels of all CR-related factors), and low level (characterized by a higher level of social connection but less favorable levels of other CR-related factors) of CR according to the item-response probabilities (Supplementary Table S1). Similar calculations have been described previously [Reference Zhang, Chen and Pan23, Reference Paudel, Ahmadi, Phongsavan, Hamer and Stamatakis24], and the methodology details are available in Supplementary Material S1.

Ascertainment of depression, dementia, and death

Incident depression was identified through hospital admissions data (i.e., Hospital Episode Statistics-Admitted Patient Care in England, Scottish Morbidity Records-General/Acute Inpatient and Day Case Admissions in Scotland, and Patient Episode Database in Wales), primary care records, self-reported diagnoses of depression, and death registries. These events were recorded and coded based on the International Classification of Diseases, Version 10, and codes F32–F33 were used. Prevalent depression was detected at baseline using the hospital admissions data, primary care records, self-reported diagnoses, and the Patient Health Questionnaire-2 (PHQ-2) which assessed the frequency of depressed mood and anhedonia over the past 2 weeks, with cutoff ≥2 reflecting possible depression [Reference Levis, Sun and He25].

Dementia was ascertained based on algorithmic definitions developed by the UK Biobank outcome adjudication group, which combined multiple data sources including hospital admissions, self-reported diagnoses of dementia/Alzheimer’s disease/cognitive impairment, and/or death registries. Post-depression dementia was defined as having dementia after the occurrence of depression. The earliest recorded date of occurrences of depression and dementia were used and compared to ensure the chronological order of events. Data on deaths from all causes were extracted via linkage to national death registries.

Statistical analysis

Baseline characteristics of participants according to CR level were compared using a one-way analysis of variance for normally distributed continuous variables or chi-square test for categorical variables.

A multi-state Markov model was used to assess the influence of CR (reference group: low CR) on the risk of incident depression and subsequent dementia and/or death. Results are presented as transition-specific hazard ratios (HRs) and 95% confidence intervals (CIs) of depression and related outcomes. The proportional hazards assumption was checked using Schoenfeld residuals and no violations were detected. Follow-up time was calculated as the time from baseline to death or end of follow-up (January 31, 2022), whichever came first. In this study, five transition phases were considered: (1) baseline to depression, (2) depression to dementia, (3) baseline to death without depression, (4) depression to death without dementia, and (5) post-depression dementia to death (Figure 1). For participants whose depression/dementia diagnosis and death were recorded on the same date, the entry date of theoretically prior state was calculated as the entry date of the latter state minus the median interval time of corresponding stage (577 days for transition 4, n = 13; 304 days for transition 5, n = 21) [Reference Wu, Zhang and Qian26]. We only considered the first entry into a state, and no reversal of state was allowed.

Figure 1. Schematic representation of multistate model.

State-specific numbers of participants were reported in boxes, and numbers (percentages) of participants in transitions from baseline to depression, subsequently to dementia, and ultimately to death were reported on arrows.

To further explore the role of CR in the prevention of depression, we assessed the probability and duration of depression-free survival (i.e., an initial state without any transition to depression, post-depression dementia, or death) according to CR level. A combined outcome was defined as either incident depression or death. Cox proportional hazard regression was used to examine the longitudinal association between CR and the combined outcome. Follow-up time was calculated as the time from baseline until the earliest occurrence of depression, death, or end of follow-up. Laplace regression was used to estimate the absolute percentile difference in time until the occurrence of outcome according to CR level, so the results indicated depression-free survival times. Because nearly 10% of participants developed depression or died, we modeled and predicted differences in time (in years) by which the first 10% of participants would experience the outcome.

Given possible differences in prevalence and etiology of depression and dementia, as well as the complex relationship between the two diseases, at different ages, we also analyzed the aforementioned associations after stratifying by age group (middle age [<60 years, n = 244,368] versus older age [≥60 years, n = 191,864]). A multiplicative interaction was tested by incorporating the two factors (i.e., the CR indicator and age group) and their cross-product term in the same models. All analyses were adjusted for age, sex, race, smoking status, alcohol consumption, physical activity, BMI, hypertension, diabetes, heart disease, and stroke. Missing data were imputed using the fully conditional specification, with estimates pooled across five iterations.

Several supplementary analyses for the multi-state analysis were performed: (1) to minimize the influence of reverse causation, we excluded the cases of depression (n = 943) or dementia (n = 30) that occurred during the first year of follow-up, (2) we recalculated the entering date of the prior state using a 0.5 day of time interval for participants who entered different states on the same date in transitions 4 and 5 [Reference Han, Hu and Yu27], and (3) we further adjusted for the Townsend deprivation index, a variable reflecting neighborhood-level socioeconomic status. Statistical analyses were performed using Stata SE 15.0 (StataCorp, College Station, TX, USA) and R software (version 4.3.0). Results with a 2-sided P value <0.05 were considered statistically significant.