Data and variables

We used population register data on all Finns born between 1900 and 1950 obtained from Statistics Finland. The cohort was followed, through record linkage using personal identification codes assigned to all permanent residents, for incident dementia at the age of 65 years and above in administrative health registers from 2001 through 2018. The baseline for follow-up thus varied from year 2000 to 2015 according to the year of birth. Information on sociodemographic characteristics was obtained from population censuses conducted in 1970, 1975, 1980 and 1985 and from the population register from 1987 to 2018. The sibling analysis was based on a 10% household sample from the 1950 Finnish population census that has been linked to subsequent census records and the population register. Children living in the same family at the age of 0 to 15 years in 1950 were identified as siblings by means of unique family identifiers. The family is defined based on the youngest generation in the household dwelling-unit, and thus siblings could only be identified if they did not have children of their own. We therefore limited the age range to 15 years, when childbearing was still rare. The sibling subsample thus included cohorts born between 1935 and 1950.

We identified dementia using the medication reimbursement register of the Social Insurance Institution of Finland and the hospital care register of the Finnish Institute for Health and Welfare. Thus our definition of dementia only included people on dementia medication or receiving hospital care. We collected the beginning (month and year) of entitlement to special state reimbursement of antidementia medication, and the dates of state-reimbursed purchases of antidementia medication using the World Health Organization Anatomical Therapeutic Chemical (ATC) code N06D. Specialised out-patient care and in-patient hospital admissions with a dementia diagnosis were collected with World Health Organization ICD-10^{Reference World Health Organization18} codes F00–03, F05.1 and G30. The earliest entry in any of these registers was set as the date of dementia incidence. These data sources present good sensitivity and high precision for dementia diagnosis,^{Reference Solomon, Ngandu, Soininen, Hallikainen, Kivipelto and Laatikainen19} and the age-specific incidence rates observed in our data (Supplementary Table 1 available at https://doi.org/10.1192/bjp.2021.217) are consistent with reports for screened community samples.^{Reference Wolters, Chibnik, Waziry, Anderson, Berr and Beiser20}

To restrict the study population to initially dementia-free individuals, we excluded those with pre-baseline reimbursements of antidementia medication (reimbursement became available in 1999) or hospital admissions with a dementia diagnosis indicated by ICD-10 codes in 1996–2000 and ICD-9^{Reference World Health Organization21} codes 290, 2912A, 2928C, 2941A, 3310, 3311 and 4378A in 1987–1995 (n = 24 543). Because dementia is among the main indications of institutional residence, we further restricted the study population to those living in private households at baseline (n = 49 384 excluded).

Data on clinical depression 15 to 30 years before baseline was collected from the hospital care register. In-patient care episodes with a depression diagnosis were identified using ICD-8^{Reference World Health Organization22} codes 2960, 2980, 3004 and 3011 in 1971–1986, ICD-9 codes 2961, 2968, 3004, 3009 and 3090 in 1987–1995, and ICD-10 codes F32, F33, F34.1 and F38.1 in 1996–2000. Individuals with at least one care episode with a depression diagnosis were classified as having a history of clinical depression.

Information on educational level and marital status was collected from the period 15 to 30 years before baseline. Education was indicated as the highest achieved qualification, categorised as tertiary (generally ≥13 years of education), secondary (10–12 years) and basic education or less (up to 9 years). Marital status was classified as married, divorced, widowed and never married. If education or marital status changed during the observation period, we considered the most recent status. We further included other medical conditions as covariates to indicate vascular risk factors (for example smoking, excessive alcohol use, diabetes) and cardiovascular diseases (CVDs) that are shown to be comorbid to depression^{Reference Penninx23} and to influence dementia risk.^{Reference Hernán, Alonso and Logroscino24,Reference Rehm, Hasan, Black, Shield and Schwarzinger25} These conditions were measured from 15 to 30 years before baseline from the hospital care register, and included alcohol-related diseases and accidental poisoning by alcohol, chronic obstructive pulmonary disease (COPD) or asthma, diabetes, coronary heart disease (CHD), and other non-stroke CVDs. Stroke was excluded because of the direct short-term effect on (post-stroke) dementia (for ICD-codes see Supplementary Table 2). Because depression history and the covariates could only be defined for Finnish residents, we excluded those living abroad over the period 15 to 30 years prior to baseline (n = 14 914).

Statistical analyses

We used Cox proportional hazards regression to estimate hazard ratios (HRs) and 95% CI for the association between a history of clinical depression and dementia between 1 January 2001 and 31 December 2018. Cohorts turning 65 years in year 2001 or later entered the analysis on the 1 January following their 65th birthday. Attained age in years was used as the underlying timescale in the analyses. Individuals were censored on the date of death, at the end of the year preceding emigration or at the end of 2018, whichever came first.

We first used conventional Cox models to estimate the association between a history of clinical depression and dementia, controlling for observed characteristics. In addition to underlying attained age, model 1 adjusted for gender and calendar year. Model 2 accounted additionally for education and marital status, and model 3 further adjusted for comorbid medical conditions. Second, we used Cox regression with sibling fixed-effects by assuming a separate baseline hazard for each childhood family (n = 23 626) to control for all time-invariant characteristics shared by siblings.^{Reference Allison26} We further controlled for the same observed characteristics as in the conventional Cox models. Third, modification by gender, educational level and marital status was analysed in the full cohort, with basic (model 1) and full adjustments (model 3). The significance of interaction was analysed using the likelihood ratio test. All analyses were performed with Stata 16.0.

Sensitivity analyses

As selection for health and survival may bias the estimates in studies of older people,^{Reference Hernán, Alonso and Logroscino24,Reference Euser, Schram, Hofman, Westendorp and Breteler27} we conducted age-stratified analyses to assess whether the associations changed with increasing age at study inclusion. We used three groups: cohorts born in 1935–1950 (aged 65 years at baseline), cohorts born in 1920–1934 (66–80 years) and cohorts born in 1900–1919 (81–100 years).

Ethics of research

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. All procedures involving human participants were approved by Statistics Finland Board of Ethics (permit no. TK-53-1490-18). Participant consent was not required as administrative register data can be used for scientific purposes under the Personal Data Act and the Statistics Act. Statistics Finland pseudonymised the data prior to providing it to researchers.