Study population
The data came from the Screening Across the Lifespan Twin Study (SALT) [16, 17] which was conducted in 1998–2002 on all then living twin individuals born in 1958 or earlier (aged between 41 and 97 years) who were included in the Swedish Twin Registry [18]. SALT collected data on diseases, symptoms, lifestyle factors, and medication use through a computer-assisted telephone interview.
The selection of the study population is shown in Fig. 1. After linking to national health register data on dementia diagnoses using personal identification number, we excluded individuals with an onset of dementia before baseline or who had severe cognitive impairment at baseline (see below). This left us with 41,550 participants (full sample) available for the main analysis. A subsample of 10,487 individuals in SALT aged 65 years and older who received a cognitive assessment at baseline (cognitive sample) was used in a sensitivity analysis and to further adjust for baseline cognitive level. Given the significant role of the APOE ɛ4 allele as a risk factor of dementia, we further performed analyses adjusting for APOE ɛ4 carrier status in individuals with genotype data available. A subsample of 10,502 participants from the full sample (genotyped sample I) and a subsample of 3156 participants in the cognitive sample (genotyped sample II) were available for this analysis. Lastly, for the within-pair analysis, we had 11,031 DZ twin pairs and 4055 MZ twin pairs available in the full sample and 2176 DZ twin pairs and 766 MZ twin pairs in the cognitive sample (Fig. 1).
Assessment of the FI
Construction of the FI used in this study has been described previously [19]. Briefly, the FI is based on the Rockwood deficit accumulation model [7] and consists of 44 self-reported health-related items on general health status, diseases, signs and symptoms of disease, psychosocial health, and functional abilities. We included all available items in SALT that meet the standard inclusion criteria [7] and are also suitable for younger adults (< 65 years). For example, activities of daily living were only available for those aged 65 and older and were thus not included. The missing values across the FI items were imputed by chained equations as previously described [19]. A list of the included FI items and their coding is presented in Additional file 1: Table S1 [20,21,22]. The FI for each individual was calculated as the number of deficits present divided by the total number of deficits, yielding a continuous score ranging from 0 to a theoretical maximum of 1.
Ascertainment of all-cause dementia
Dementia diagnoses during the follow-up were retrieved from nationwide registers, namely, the National Patient Register (NPR), the Cause of Death Register (CDR), and the Prescribed Drug Register (PDR) [23]. Both the NPR (with nationwide coverage since 1987) and the CDR (with nationwide coverage since 1961) contain disease information based on the International Classification of Diseases (ICD) system. Dispensed dementia medication according to Anatomical Therapeutic Chemical (ATC) codes in the PDR was regarded as proxy for dementia diagnosis. ATC codes for anti-dementia drugs in the N06D group were considered. The primary dementia diagnosis or death was followed up from baseline until the last day of December in 2016, yielding an up to 19 years of follow-up. The ICD and ATC codes used to identify dementia are shown in Additional file 1: Tables S2 and S3. In addition, dementia information was available from ascertainments from STR studies of aging, including SALT for participants aged 65 or older (the cognitive sample), where diagnoses were set at consensus conferences based on DSM-III-R and DSM-IV criteria [20].
Covariates
In addition to sex and age at FI measurement, body mass index (BMI), tobacco use, years of education, living alone, and physical activity were considered as covariates because of their roles as common risk factors for dementia [3]. BMI, from self-reported data, was calculated as weight divided by height squared (kg/m2). Tobacco use was classified as non-user (reference category) or user if the participant was a current smoker or used smokeless tobacco regularly, had previously smoked, or used smokeless tobacco regularly. Living alone was a binary variable with “not living alone” as the reference category. For those who were born after 1926 (N = 37,218), physical activity was assessed based on the question “Of these 7 alternatives, which fits your annual exercise pattern?” The alternatives were 0 (almost never exercise), 1 (much less exercise than average), 2 (less exercise than average), 3 (average amount of exercise), 4 (more exercise than average), 5 (much more exercise than average), and 6 (maximum amount of exercise). For those who were born before 1926 (N = 4046), physical activity was assessed based on the question: “How much do you exercise?” The alternatives were 0 (almost no exercise), 1 (light exercise), 2 (regular median exercise), and 3 (hard physical exercise). The two physical activity variables were transformed to z scores (each unit representing one standard deviation from the mean) and combined into one variable for the analysis. Screening for cognitive function for those aged 65 and older (cognitive sample) is described in Additional file 1, Appendix S1: Supplementary methods.
Genotype data
APOE ɛ4 carrier status was additionally adjusted for in the genotyped sample I (N = 10,502) and genotyped sample II (N = 3156). The APOE ɛ4 genotypes were either directly genotyped or determined from Illumina OmniExpress imputed to 1000 Genomes Project [24] using a pipeline with high accuracy [25]. Individuals carrying the genotypes ɛ2/ɛ2, ɛ2/ɛ3, or ɛ3/ɛ3 were categorized as non-carriers (reference category); those carrying the genotype ɛ2/ɛ4 or ɛ3/ɛ4 were categorized as heterozygous; and those carrying the genotype ɛ4/ɛ4 were categorized as homozygous.
Statistical analysis
Two main analytical approaches were used in this study: a cohort analysis to obtain population-level estimates of the association between frailty and dementia and a within-pair analysis to control for familial (i.e., genetic and shared environmental) effects on the association. All analyses were performed in the same manner in the full and cognitive samples.
We first performed Cox proportional hazard models with time since the FI measurement as the underlying timescale to estimate the hazard ratios (HRs) for a 10% increase (i.e., 0.1 increment) in the FI on dementia. Individuals who died during the follow-up were censored at the date of death. Age at FI assessment, sex, BMI, tobacco use, years of education, living alone, physical activity, and cognitive function score (in the cognitive sample, described in Additional file 1, Appendix S1: Supplementary methods) were first tested for their association with dementia in univariate Cox models. Following the rule of parsimony, those variables that were statistically significant or had an effect on the FI estimate were included as covariates in the multivariate Cox model. The proportional hazard assumption was tested using an interaction term between the covariates and time in the model. APOE ɛ4 carrier status was additionally adjusted for in the multivariate Cox models in the genotyped samples I and II that consisted of individuals who had genotype data available (Fig. 1). Cluster-robust standard errors were used to correct for the correlation within twin pairs. Next, taking into account the competing risk of death, a competing risk model based on the Fine and Gray method using subdistribution HRs (SHRs) [26] was performed in the full and cognitive samples. The functional form of the relationship between the FI and dementia was assessed using the log-likelihood test between the quadratic and the linear model and plotting the log-HRs for a quadratic, cubic spline-transformed and linear FI. Lastly, to facilitate clinical interpretations, we assessed the relationship between categorized FI and the incidence of dementia by categorizing the FI into three levels: non-frail (FI ≤ 0.08), pre-frail (0.08 < FI ≤ 0.25), and frail (FI > 0.25), according to pre-established cut-offs [27]. The categorized FI (non-frail as reference) was then tested for its association with dementia in a Cox model. Kaplan-Meier curves were used to assess the probability of being dementia-free during the follow-up by the aforementioned frailty categories in the full and cognitive samples.
The within-pair analysis was conducted in DZ and MZ twin pairs that were complete, i.e., both members of the pair had relevant information on FI and dementia (Fig. 1). A between-within (BW) model, a random-effects model incorporating a BW decomposition [22], was applied in a survival analysis framework to conduct the within-pair analysis. In this analysis, we tested the time-constant within-pair effect of the FI on dementia. To explore whether the associations differed by sex, the within-pair analysis was additionally performed separately in men and women (like sexed twin pairs). To facilitate conclusions about the independent, potentially causal role of frailty on dementia, we used a Wald test to formally test whether the within-pair estimate in MZ pairs differed from the population-level estimate in the Cox model for the MZ twins. We fitted the two models in parallel to allow for the statistical test of the two regression coefficients.
To assess whether the risk carried by increased FI varies over age at FI measurement and whether the association is independent of familial effects throughout the age range, from adulthood into old age, we modeled the association between FI and dementia including a statistical interaction between FI and age and modeled this interaction as a natural cubic spline function. We fitted the model to the standard association (population-level estimate), and to the within-pair association, henceforth referred to as the “standard interaction model” and the “within-pair interaction model,” respectively. To further test for genetic effects in the association between frailty and dementia, we formally assessed the difference in the MZ and DZ estimates across the age range in the within-pair interaction model by deriving the ratio of the HR between DZ and MZ (HRDZ/HRMZ) as a function of age. Details of the models are presented in Additional file 1, Appendix S1: Supplementary methods.
Lastly, as our FI included a number of conditions that are known risk factors for dementia, similar to the study by Song et al. [28], we conducted a sensitivity analysis by dividing the FI into those items that are traditional risk factors for dementia (FI-TRF) and those that are not (FI-NTRF) and analyzed them separately using Cox regression. Hypertension, high cholesterol or triglycerides, cerebral hemorrhage or stroke, TIA attacks, irregular cardiac rhythm/atrial fibrillation, diabetes, kidney disease, migraine, depression, and hearing were included in the FI-TRF due to their established roles as dementia risks [3, 29,30,31] and the remaining 34 items in the FI-NTRF (Additional file 1: Table S1). A two-sided P-value < 0.05 was considered statistically significant. All analyses were performed using STATA 15.1 and R version 3.6.1.