Population Attributable Fractions for Modifiable Risk Factors of Incident Dementia in Cognitively Normal and Mild Cognitively Impaired Older Adults: Data from Two Cohort Studies

Study participants

This study uses data from two cohort studies included in the Global Alzheimer’s Association Interactive Network (GAAIN) database [5, 11]. The GAAIN collects data of patients with dementia and healthy adults via data partners. Three of the 46 GAAIN datasets fulfilled the following criteria: 1) longitudinal information available on cognitive function, diagnosis of dementia and at least six out of twelve risk factors; and 2) include more than 100 adults above 65 years with normal cognition at baseline. Data from two of the three suitable datasets could be obtained, namely form the National Alzheimer’s Coordinating Centre (NACC) and Hellenic Longitudinal Investigation of Aging and Diet (HELIAD).

In 2005, the NACC commenced a cohort study with continuing recruitment of now 42,980 participants with normal cognition, MCI, or dementia in the United States of America [12]. All participants are annually followed, with a maximum of fifteen visits to date. Data were collected by 39 Alzheimer’s disease Centers form September 2005 to May 2020, each enrolled participants according to their own protocols. Participants were enrolled via clinician referral, self-referral, active recruitment by community organizations or on a voluntary basis [12]. Data were collected by study clinicians and health professionals linked to the participants or their co-participants, usually a close friend or family member of the participant. All participants of the NACC were expected to have a co-participant and only in very exceptional cases a participant did not have a co-participant [12]. Information was collected via home visits, office visits or via telephone calls. Milestone Forms were used to document the death or drop-out of a participant [12].

As determined by the University of Washington Human Subjects Division, the NACC database is exempt from Institutional Review Board (IRB) review and approval because it does not involve human subjects, as defined by federal and state regulations. However, all Alzheimer’s disease Centers have obtained informed consent from the participants and maintain their own separate IRB review and approval from their institution before submitting data to NACC [12].

Eligible for the current study were 16,147 participants of 65 years and older, who were free from dementia at the baseline and who had at least one follow up visit (Fig. 1).

OPEN IN VIEWER

Fig. 1

Flow diagrams of participants in the NACC and HELIAD study.

The HELIAD is a population-based study set up to investigate the prevalence and incidence of MCI and dementias in the Greek community dwelling population above 65 years, and to examine the associations between lifestyle risk factors and age-related neuropsychiatric diseases [13]. The HELIAD was conducted in the cities of Larissa and Marousi, Greece, with respective populations of 145,000 and 72,000 inhabitants. Data from 2,062 participants have been collected to date [13]. Of the 2,062 participants who completed the initial evaluation of the HELIAD, 1,141 participants of 65 years and older, who were free from dementia at baseline and who had at least one follow up visit were included for analysis after exclusions (Fig. 1). Information on clinical, lifestyle, and demographic variables was collected [14]. Participants were asked to bring a caregiver of informant, but were not always accompanied by an informant or caregiver [14]. Participants were followed up after approximately three years to observe the incidence of MCI or dementia. The assessments were conducted at day care centers for older adults, the participants’ homes, the University of Thessaly Medical Centre, or municipal public health clinics. The HELIAD study was approved by the local IRB.

Cognitive status

The subject’s cognitive status was determined in both studies at every visit. The evaluations at follow up were performed by neurologists and neuropsychologists who administered questionnaires, conducted neuropsychological assessments and reviewed all the information that was collected at expert meetings [13]. Subjects who were determined by the clinician(s) to have normal cognition and behavior were given the status cognitively normal. If the clinician(s) determined that the participant had a clinically significant impairment in cognition or behavior, the participant was either diagnosed with MCI or dementia. In both studies the dementia diagnoses were made by a multidisciplinary consensus team based on application of the widely accepted DSM-IV criteria [12, 15]. The diagnosis of MCI was made based on the presence of subjective memory complaints and objective cognitive impairment in at least one cognitive domain, without significant interference in daily functioning [12, 16]. At baseline, participants were classified as either cognitively normal (no presence of MCI or dementia) or cognitively impaired (presence of MCI) or dementia. Participants with dementia were excluded from the analyses. Participants with dementia at follow-up were classified as incident dementia.

Risk factors

Data were available for nine of the twelve risk factors in NACC and for ten risk factors in HELIAD. Supplementary Table 1 provides an overview of the operational definitions of the risk factors. The data on the risk factors were collected with questionnaires. Participants were asked if a physician had given them the diagnosis hypertension, diabetes, hearing loss, TBI or depression. The participants were also screened for neuropsychiatric conditions including depression [14]. Depression was operationalized in HELIAD as diagnosis at baseline and in NACC as diagnosis at baseline or self-reported history of depression. The diagnosis was made according to DSM-IV criteria following a structured Geriatric Mental State interview. Low education was defined as having had less than twelve years of education or less than four years of high school. Excessive alcohol use was defined as alcoholism in the NACC dataset. All variables were dichotomized. The cut-off points were chosen based on risk-factor definitions used in Livingston et al., Mukadam et al. [5, 10] and the available data.

Analysis

Competing risk analyses were conducted to examine the associations between each of the risk factors and time to incident dementia. These models were based on a sub-distribution hazard risk function with incident dementia as the outcome of interest and death as the competing risk. Time was measured as the time in years from baseline to date of dementia diagnosis, death or drop-out, whichever occurred first. Analyses were conducted separately for three subgroups: participants with normal cognition at baseline, participants diagnosed with MCI at baseline and all participants combined. Our intention was to include the same risk factors in the calculation of the PAFs as previously reported in the literature [5, 10], and to adjust for confounders and other (modifiable) risk factors for dementia. We adjusted for several additional lifestyle factors, reasoning that in a population without dementia these factors are more likely confounders (common causes) than colliders (common effects). However, as some of these variables were not available in either of the cohorts, the analytical models differed across the cohorts. In NACC, PAFs were calculated for low education, hearing loss, obesity, hypertension, diabetes, depression, smoking, excessive alcohol use, and TBI with adjustment for the confounders age, sex, race, and APOE ɛ4. In HELIAD, PAFs were calculated for low education, hearing loss, physical inactivity, obesity, hypertension, diabetes, depression, social isolation, smoking, and TBI with adjustment for the confounders age, sex, sleep disorder, alcohol use, cognitive activity and adherence to a Mediterranean diet pattern [17, 18]. Confounding variables were measured as described in Supplementary Table 2.

The hazard ratios (HRs) for each risk factor for dementia from the competing risk analyses were used to calculate PAFs using formula P(HR - 1) / (1 + P(HR - 1)) [19]. P is the prevalence of the risk factor in the datasets of this study. Negative PAFs were omitted in the calculation of the overall PAF such that this value represents the potential effect of risk factors with positive PAFs. See the Supplementary Material for more information.

While the results based on competing risk analysis are presented as our main analyses (to account for study discontinuation), we also computed odds ratios based on logistic regression to facilitate direct comparison of our results to Livingston et al. [5]. To account for potential selection bias in the NACC cohort (which is not population-based), sensitivity analyses were done in which odds ratios were derived from logistic regression models weighted for sex and education using the U.S. Census data of 2006 as reference [20].

Overall, 3.2% (NACC) and 1.3% (HELIAD) of the data were missing and 24.8% (NACC) and 9.7% (HELIAD) of the participants had missing data on at least one variable. Comparison of participants with missing data with participants with complete data showed no substantial differences in characteristics, risk factors and dementia risk between the subgroups (Supplementary Tables 3and 4). Missing data were thus imputed using multiple imputation by chained equations (MICE) [21], which, in order to account for the uncertainty in the imputed values, created fifteen imputed data sets. The pooled results of these data sets are presented. To examine the influence of imputation, sensitivity analyses were done using participants with complete data only. The analyses were performed in IBM SPSS Statistics 26 and Stata 16.

NACC

The participants of the NACC had a mean (SD) age of 75.2 (6.9) years old and 59.0% were female (Table 1). Most participants (94.7%) were highly educated with an average of 15.8 (5.7) years of education. Of the 16,147 participants, 2,630 developed MCI (25.1%) and 3,333 developed dementia (20.7%) during a median follow-up of 4.9 (3.5) years. The characteristics of participants, for both the CN and MCI subgroups, are presented in Table 2.

In the total sample, low education, diabetes, excessive alcohol use and depression were significantly associated with a higher risk of dementia (Table 3). In contrast, obesity and smoking were significantly associated with a lower risk of dementia.

In the CN subgroup, low education and depression were significantly associated with a higher risk of dementia (Table 3).

In the MCI subgroup, obesity was significantly associated with a lower risk of dementia (Table 3). In the analysis of the NACC datasets before imputation, the HRs were in the same range as the analyses in the imputed dataset (Supplementary Table 5).

PAFs in the total sample added up to an overall weighted PAF of 19.4% (12.6–28.2) for the nine risk factors (Table 4; Supplementary Figure 1). In the CN subgroup, the PAFs added up to 15.9% (5.1–34.5). In the MCI subgroup, the PAFs added up to 3.3% (0–13.7). The PAFs of the NACC dataset before imputation are presented in Supplementary Table 6.

HELIAD

The participants enrolled in the HELIAD were on average (SD) 72.9 (5.0) years old and had 8.2 (5.0) years of education (Table 1). Of the 1,141 participants, 58.0% were female, 131 developed MCI (11.2%), and 68 participants developed dementia (5.9%) during follow-up of on average 3.10 (0.86) years.

In the total sample, physical inactivity was significantly associated with a higher risk of dementia (Table 5).

In the MCI subgroup, social isolation was significantly associated with a lower risk of dementia (Table 5). In the analysis of the HELIAD datasets before imputation, the HRs were in the same range as the analyses in the imputed dataset (Supplementary Table 7).

PAFs in the total sample added up to an overall weighted PAF of 65.5% (0–65.4) for the ten risk factors (Table 6; Supplementary Figure 2). In the CN subgroup, the PAFs added up to 65.8% (0–66.1). In the MCI subgroup, the PAFs added up to 63.4% (0–66.2). The PAFs of the HELIAD dataset before imputation are presented in Supplementary Table 8.

Strengths and limitations

A strength of this study is the use of two different cohorts, which provides insight into the robustness of the findings across study populations. Another strength is the use of competing risk analyses, which correct for time until dementia diagnosis, discontinuation, death, or most recent visit, and thereby minimize potential bias (e.g., attrition bias). To compare our results to previous research [5, 10], we also ran logistic regression models but did not observe substantially different results (Supplementary Table 9). Livingston et al. [5, 10] calculated PAFs based on pooled data from systematic reviews. Hence, the PAFs were calculated separately for each risk factor and communality across risk factors had to be accounted for using an extra step that required additional data. In this study all risk factors and confounders were included in the same model so communality was accounted for implicitly. Moreover, we adjusted our models for several potentially confounding lifestyle factors, such as sleep, cognitive activity, alcohol use and dietary habits, as well as age, sex, APOE ɛ4, and race. Nevertheless, it remains possible that other factors confound the identified associations, which limits the interpretation of the corresponding PAFs. Sleep disorder and cognitive activity may be signs or symptoms of MCI or dementia so sensitivity analyses in which sleep and cognitive activity were not included as confounders were performed. The analyses showed minimal differences in results (Supplementary Table 10). A fourth strength is the large sample of participants in NACC, which was maximized by multiple imputation of missing values. Unfortunately, the sample size of HELIAD was smaller than the NACC database, but this dataset was very interesting because of the large differences in socio-economic and lifestyle environment. Another strength of HELIAD is that it is population-based, resulting in a representative study population with a relatively high external validity in the Greek population. However, there were insufficient data available on APOE ɛ4 in the HELIAD dataset to adjust for it in the analyses. A final strength of our study is that we included nine (NACC) and ten (HELIAD) of the twelve modifiable risk factors identified in the landmark study by Livingston et al. [5]. Unfortunately, insufficient information was available about physical or social activity (NACC) and air pollution (both studies). Hence, we were unable to include all relevant risk factors, which may have contributed to underestimation of the overall PAFs.

Limitations of our study include possible selection bias (NACC) and healthy survivor bias (NACC and HELIAD) and reporting bias (NACC and HELIAD). The NACC study population was recruited via clinician referral, self-referral, and by community organizations [12]. Therefore, selection bias may be present, and results may not generalize to the overall US population. For example, the enrolled volunteers with normal cognition, appeared to have an above average level of education [12], which may lead to false-negative diagnoses [29]. The sensitivity analyses in which models were weighted by sex and education, showed similar results to the unweighted models (Supplementary Table 9), which makes it less likely that deviations in the distribution of sex and education from the general population explained some of the unexpected findings. However, selection bias may still be present based on other factors that we were unable to weigh for, e.g., age and higher prevalence of risk factors. Selection bias may also have been introduced by excluding participants without follow-up data on dementia. However, comparison of the prevalence of risk factors in the analysis sample with the total sample at baseline showed no marked differences. As described earlier in the discussion, healthy survivor bias likely explains the negative associations for, for example, smoking and obesity in survival analysis. Nevertheless, the used method in this study, competing risk analysis, aims to correctly estimate marginal probability of an event in the presence of competing events, such as death from cardiovascular disease, which reduces the chance of healthy survivor bias [30]. As most risk factors were measured using self-report, the negative associations may also be explained by reporting bias. However, the negative associations were also found for obesity, which was based on objective measurement of height and weight, which makes reporting bias less likely. As the risk factors with negative associations were excluded from calculating the PAFs, the PAFs likely underestimate the true risk associated with the modifiable factors. Reporting bias may have been an issue particularly in the MCI group, as their declined cognitive functions may have hindered reliable self-report of risk factors. However, the influence of reporting bias was minimized by having each participant be accompanied by an informant or caregiver who could confirm or complement the participant during their assessments [12, 14]. Another potential limitation is reverse causation. However, sensitivity analyses excluding participants who developed dementia in the first two years of follow-up showed minimal differences in the findings (Supplementary Table 11). Furthermore, because this study uses observational data, caution has to be taken when causally interpreting our results. Finally, the relatively small sample size and short follow-up in HELIAD may have resulted in underestimation of some risks and subsequently PAFs, due to a relatively low number of dementia cases at follow-up. A limitation in calculating PAFs based on individual studies is that these studies often have insufficient power which increases the risk of spurious associations [31]. The limited statistical power, particularly in the subgroups by cognitive status, may also affect the current findings. However, a limitation of PAFs calculated from data pooled in meta-analyses is that the specific characteristics of the individual studies are lost, and the results may not accurately reflect the situation in the region where the individual studies were conducted. Regional PAFs are important to inform local policy. When interpreting the PAFs, policy makers should consider the limitations of these metrics. The accuracy of the PAFs highly depends on the representativeness of the population-based sample. For conditions with long lead times, better metrics are needed that are capable of maintaining their relevance for both risk and protective factors.

Future work

Due to the contrasting findings between HELIAD and NACC, further research is needed to examine whether elimination of risk factors may have the same potential to prevent dementia in MCI as in CN subgroups. Given our results it is likely that similar prevention programs would result in, highly different prevention effects on dementia incidence in ethnically, genetically, socially, or culturally different populations. Additionally, future research should differentiate between different subtypes of dementia, such as Alzheimer’s disease and vascular dementia. In this study, any cause dementia was taken as outcome, but risk factors and their associated PAFs likely differ per dementia subtype. Furthermore, the PAFs of other risk factors, such as air pollution to dementia, should be included in follow-up studies in different cohorts. Future studies could also investigate interactions between risk factors. Such studies require ‘big data’, both rich in terms of participants as well as in measured variables. Hence, we encourage to widely utilize available dementia data sets that can be accessed using platforms like GAAIN [11].

Conclusion

In conclusion, the overall PAF in HELIAD of the known risk factors for dementia was higher than expected, while the overall PAF in the NACC was lower than expected based on previously reported PAFs in other studies. This was explained by differences between cohorts in risk factor prevalence and the strength of their relation to dementia. These differences suggest that translation of the valuable global meta-analysis data on modifiable risk factors should be carefully carried out per country to tailor preventive public health strategies and optimize their effectiveness [6]. Finally, the high overall PAF in the MCI subgroup of the HELIAD data suggests that preventive interventions continue to be relevant in people with MCI.