Our multinational cohort study utilized individual-level data from two well-characterized cohorts with the same biennial longitudinal design and comparable survey protocols: the HRS and the SHARE. The HRS is a nationally representative longitudinal survey of adults aged 50 years and older in the United States (US) [15, 16]. The SHARE is a longitudinal household survey focusing on retirement and health among the elderly in continental Europe [17]. More information on the sample design and procedures can be found in their cohort profiles [15, 18]. This study adheres to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.

We used data from the HRS and the SHARE from 2012 to 2018. Baseline data was collected in 2012. The follow-up assessments were conducted until 2018. For HRS and SHARE, we included 30,313 and 39,300 participants aged 60 years or older at baseline. After excluding participants with dementia, Alzheimer’s disease, cognitive impairment, or those who lacked the aforementioned information, we included 9926 participants with normal cognitive function in the HRS and 6699 participants with normal cognitive function in the SHARE. Then, we excluded participants who lacked relevant information on physical and psychological multimorbidity and covariates, and finally included 8543 and 5939 participants in HRS and SHARE, respectively (Fig. 1).

Study flowchart. HRS, Health and Retirement Study; SHARE, Survey of Health, Ageing and Retirement in Europe

The HRS has received approval from the University of Michigan Institutional Review Board (IRB Protocol: HUM00061128). The SHARE was reviewed and approved by the Ethics Committee of the University of Mannheim and the Ethics Council of the Max Planck Society (IRB: No 723/2009). All participants provided informed consent.

Psychological disorder was identified using the 8-item Center for Epidemiologic Research Depression (CES-D) scale in the HRS. This scale measured the frequency of feelings on eight dichotomous items in the past week, including “depressed,” “everything was an effort,” “happy,” “life was enjoyable,” “sad,” and “unable to get going” [6, 19, 20]. We reverse-coded the items for “happy” and “life was enjoyable” and then summed all the items. The total scores ranged from 0 to 8, with a cutoff value greater than 3 indicating a psychological disorder. In SHARE, the EURO-D, which comprises 12 items (depressive symptoms, pessimism, death wish, guilt, irritability, crying, fatigue, sleep problems, loss of interest and appetite, reduced ability to concentrate, and capacity to enjoy things over the last month), was used to assess psychological disorder [21]. The cutoff for a clinically relevant psychological disorder is ≥ 4 [21].

Physical disorders included seven self-reported physician-diagnosed conditions: hypertension, diabetes, cancer, lung disease, heart disease, stroke, and arthritis in the HRS and the SHARE [6, 22]. Reported presence of any of the aforementioned seven chronic physical conditions was considered as a physical disorder [6]. Participants were categorized into four groups based on the presence of physical and psychological disorders: none, only physical disorder, only psychological disorder, and physical and psychological multimorbidity.

Referring to previous studies, dementia was determined by a combination of self-reported physician diagnosis of dementia or Alzheimer’s disease, or total scores below the cutoff value of 7 on the HRS cognitive scale [23, 24]. This cognitive scale included immediate and delayed 10-noun free recall to assess memory, serial sevens subtraction to evaluate working memory, and counting backwards to measure the speed of mental processing [22, 25, 26]. The total scores ranged from 0 to 27, with a score of 6 or less indicating dementia, while a score of more than 11 indicates normal cognitive function [27, 28]. In the SHARE, cognitive function was assessed using episodic memory and verbal fluency tasks [29]. A memory score of 1.5 standard deviations below the age-specific mean had been considered an indicator of cognitive impairment. If respondents failed to name at least 15 correct words in verbal fluency tasks, they had a verbal fluency problem. Normal cognitive function was defined as having no issues with both episodic memory and verbal fluency [29]. We included participants who reported no dementia and had normal cognitive function at baseline.

The baseline covariates included age (< 70 years, 70 ~ 79 years, ≥ 80 years), gender (female, male), educational level (less than high school, high school or associate degree, college degree or above), marital status (married, unmarried), total wealth income (the lowest quartile, quartile 2, quartile 3, the highest quartile), self-reported body mass index (BMI; underweight, normal, overweight, obesity), physical activity (no, yes), drinking habits (no, yes), and smoking status (no, yes). BMI was classified based on the standard from WHO [30]. The total wealth income (including housing, vehicles, and saving accounts) minus other debts at the household level has been assessed (including secondary residence, if any) in local currencies (HRS: dollars, SHARE: euros), and further divided into four groups based on quartile range [6]. The physical activity was assessed based on the frequency of participating in light, vigorous, or moderate physical activity [6]. Options from any of the activities, including “everyday,” “more than once a week,” and “once a week,” were recoded as “yes,” while other options, including “one to three times a month” and “hardly ever or never,” were recoded “no.” Drinking habits and smoking status were both self-reported based on questions about ever drinking any alcohol or ever smoking.

The characteristics of participants were compared using the χ2 test for categorical variables. We considered mortality as the competing event and utilized univariate and multivariable competing risk models to estimate the crude hazard ratios (cHRs) and adjusted hazard ratios (aHRs), along with their corresponding 95% confidence intervals (95% CIs). For all analyses, we fitted three models: model 1 was a univariate model; model 2 adjusted for age, gender, educational level, marital status, and total wealth income; and model 3 was a full model adding self-reported BMI, physical activity, drinking, and smoking. We conducted subgroup analysis based on age, gender, educational level, marital status, income, BMI, physical activity, drinking habits, and smoking status to assess the robustness of the results. Then, we conducted DerSimonian-Laird random-effects meta-analyses [31] to calculate the pooled HRs and 95% CIs.

In addition, we conducted three sensitivity analyses to test the robustness of the results: (1) due to discrete follow-up data from the limitations of study design, we utilized robust Poisson regression models instead of competing risk models to observe the risk ratios (RRs) of dementia; (2) after interpolating the missing covariates using the random forest method, we examined the association of physical and psychological multimorbidity with dementia among 9926 in the HRS and 6699 participants in the SHARE, respectively; (3) taking into account the impact of the number of physical and psychological disorders, we reclassified physical and psychological multimorbidity as eight groups (none, only one physical disorder, only two physical disorders, only three or more physical disorders, only psychological disorder, psychological disorder with one physical disorder, psychological disorder with two physical disorders, psychological disorder with three or more physical disorders) and calculate HRs. All analyses were conducted using R software, version 4.2.1 for Windows. The package “cmprsk” was used to fit the competing risk model (Fine-Gray subdistribution hazard model). Two-sided P values less than 0.05 were considered statistically significant.