Age of Diabetes Diagnosis and Lifetime Risk of Dementia: The Atherosclerosis Risk in Communities (ARIC) Study

Recent studies suggest that risk factors for dementia are strongly age dependent (1–3). Diabetes is a potent midlife risk factor for dementia. Studies in late life show more modest risk associations between diabetes and dementia (2–5); however, existing work has primarily focused on relative risk. Estimates of absolute risk can complement relative risk and provide a more comprehensive understanding of the association between diabetes and dementia across the entire life course.

Several studies have characterized the lifetime risk of dementia (6–8). However, differences by the age of diabetes diagnosis are poorly characterized. Estimating lifetime dementia risk is complicated by the competing mortality risk. Standard approaches, such as the Kaplan-Meier estimator, fail to properly account for competing events and overestimate lifetime risk. Bias may be especially pronounced for late-life disease (e.g., dementia) when the competing mortality risk is very high. Competing risk models are important for accurately assessing differences in lifetime risk by age of diabetes diagnosis and age of dementia onset.

Our objective was to compare the lifetime risk of dementia by the age of diabetes diagnosis. To accomplish this aim, we applied cause-specific competing risk models to more than three decades of prospective data spanning midlife to old age.

The Atherosclerosis Risk in Communities (ARIC) study is an on-going prospective cohort study that recruited and monitored 15,792 adults since 1987–1989 (age 45–65 years) from four communities: Forsyth County, North Carolina; Jackson, Mississippi; suburbs of Minneapolis, Minnesota; and Washington County, Maryland. For this analysis, we used visit 2 (1990–1992) as the study baseline because it was the first visit with hemoglobin A1c (HbA_1c) and cognitive tests. Among 14,348 participants, we excluded 279 participants missing HbA_1c, 619 with missing information about covariate data, and 363 who had dementia, death, or loss to follow-up before age 60 as the origin. The final analytic sample included 13,087 participants.

Type 2 diabetes (diabetes) was defined based on self-reported physician diagnosis, diabetes medication use, or HbA_1c ≥6.5% by visit 2. Participants with diabetes were categorized as having middle age–onset diabetes (baseline age from 49 to <60 years) or older-onset diabetes (baseline age from 60 to 69 years) using the age by visit 2. A subset of participants also completed a questionnaire at baseline that included information on the age of diagnosis. We used this information to recategorize participants in the older-onset diabetes groups as having middle age–onset diabetes if the questionnaire-reported onset age was <60 years.

Dementia was ascertained using a thorough review process conducted by an expert committee. The review included cognitive function assessments, informant reports, hospitalization codes, and death records (9).

For individuals who attended in-person visits, the cognitive function assessments were performed through a three-test cognitive battery during visits 2 (1990–1992) and 4 (1996–1998) and through an expanded neuropsychological 10-test battery (10) that was administered starting from visit 5 (2011–2013). The Clinical Dementia Rating scale (11), the Functional Activities Questionnaire, and the Mini-Mental State Examination (12) were also administered. Initial diagnoses were given with low scores in these tests using a computer algorithm. Subsequently, the diagnoses were adjudicated by a team of clinicians and neuropsychologists.

For individuals who did not attend an in-person follow-up visit, the expert panel relied on the Telephone Interview for Cognitive Status-modified or Six-Item Cognitive Screener, which were given to the participant, as well as the Ascertain Dementia Eight-Item Informant Questionnaire, which was completed by an informant. International Classification of Diseases codes (Supplementary Table 2) from hospital discharge records or death certificates were also examined to ascertain dementia cases.

Death was ascertained via contacting participant proxies, obituaries, hospital records, death certificates, or vital statistics from the National Death Index.

Most covariates were ascertained at baseline visit 2 (1990–1992). When the covariates were not available, values from visit 1 (1987–1989) were carried forward.

Age at baseline was calculated using self-reported birth dates collected at visit 1. Sex, education level, and race were self-reported at visit 1. Education level was categorized into three categories: less than high school, high school graduate or equivalent, and some college or higher. Race was self-reported in four fixed categories (Asian, Black, American Indian/Alaskan Indian, or White). A small number of participants self-reported as being Asian or American Indian (n = 36). We combined these two racial groups and used race as a three-level covariate. We included ARIC field centers (Forsyth County, Jackson, Minneapolis Suburbs, and Washington County) as a covariate to account for potential geographical heterogeneity. During visit 2, the TaqMan assay (Applied Biosystems, Foster City, CA) was used to identify the apolipoprotein E (APOE) ε4 genotype based on the number of ε4 alleles (zero, one, or two alleles).

Weight (kg) and height (m) were measured at visit 2. BMI was computed by dividing weight by height squared and then was categorized into four groups: normal or underweight (<25 kg/m²), overweight (25 to <30 kg/m²), obesity (30 to <40 kg/m²), and extreme obesity (≥40 kg/m²). We included self-reported smoking status from visit 2 as current, former, or never smokers. Alcohol consumption (current, former, and never users) was assessed at visit 2. Exercise physical activity was evaluated at visit 1 through the interview-administered Baecke questionnaire (13), including exercise types and frequency. Physical activity was then classified as recommended, intermediate, and poor, according to the recommendations of the American Heart Association and the Physical Activity Guidelines for Americans (14).

We defined hypertension as systolic blood pressure (BP) ≥140 mmHg, diastolic BP ≥90 mmHg, or the use of BP-lowering medication at visit 2. Plasma total cholesterol (mmol/L) and HDL cholesterol (mmol/L) were measured at visit 2. Prevalent stroke was identified based on self-report, hospitalization, or death related to stroke by visit 2. Prevalent atrial fibrillation was determined by self-report and electrocardiogram results by visit 2.

Baseline characteristics were compared using proportions and means (SD) for no diabetes, middle age–onset diabetes, and older-onset diabetes. We conducted survival analysis using age as the time scale with age 60 as the origin. Left truncation was considered to allow participants entering at different ages at or older than age 60. Follow-up time was measured from age at visit 2 (1990–1992) to age at first occurrence of dementia or death, loss to follow-up, or administrative censoring on 31 December 2020.

We used Kaplan-Meier survival analysis as the traditional way of estimating cumulative incidence as a comparison. The Kaplan-Meier estimator treats death as a type of censoring (equivalent as loss to follow-up) and assumes the censoring is independent from dementia events. Therefore, it estimates the probability of dementia conditional on no death occurring prior to dementia, which vastly overestimates risk because many people die before dementia onset. Cumulative incidence function (CIF), a nonparametric estimator, is a more appropriate method to compute crude risk when there are competing events (e.g., death) (15,16). The CIF estimator recognizes that once death occurs, the theoretical risk of subsequent dementia becomes irrelevant because death precludes subsequent risks to this participant. We interpreted the CIF at age 95 years as the lifetime risk of dementia because there were few participants who lived past this age. The equation is presented in Supplementary Method 1.

We used cause-specific Cox proportional hazard models for dementia and death, treating one cause as censoring when modeling the other. Covariates were sequentially adjusted in three sets: model A: demographic factors and APOE; model B: additionally adjusted for lifestyle factors based on covariates in model A; model C: additionally adjusted for clinical factors based on covariates in model B. Because the sequentially adjusted models gave consistent results (Supplementary Table 1), we then derived dementia-specific CIF based on the all covariates-adjusted model (model C) (Supplementary Method 2) (15–17). Proportionality was assessed using log-log plots.

Briefly, first, cause-specific baseline hazards for no diabetes were computed using the Breslow estimator (18) at each failure time. The estimation was based on means for both continuous and categorical variables. Second, cause-specific cumulative hazard functions for middle age–onset diabetes and older-onset diabetes were derived from model-based hazard ratios (HRs) together with baseline hazard. Third, survival function was estimated from both dementia- and death-specific cumulative hazard functions. Last, the covariates-adjusted CIF estimation was computed by integrating the product of survival function and cumulative hazard. We calculated 95% CIs using bootstrap bias-corrected and accelerated intervals with 1,000 replications.

We used restricted mean survival time (RMST) to estimate the survival years by calculating the area under the curve with restricted time points (19). The RMST provides a more robust estimation than a simple average of survival years because it accounts for the entire survival distribution and is not affected by extreme values or late-stage events. Dementia-free survival years were computed from the CIF curve, which was previously estimated from the all covariates-adjusted model (Supplementary Fig. 1). For survival years lived with dementia, we calculated from the survival curve based on the Cox proportional hazard model using time from dementia to death as the scale and adjusting for all covariates. To ensure the adequacy and robustness of person-time, we estimated years lived free from dementia and years lived with dementia using ages 70 and 80 years as the starting points. The RMST estimated survival years were performed by diabetes age of diagnosis.

As a sensitivity analysis, we separated underweight as another BMI category to account for potential increased mortality risk for underweight individuals due to other preexisting conditions. We also repeated our main analyses after excluding the small number of participants that self-reported being Asian or American Indian.

All analyses were done using Stata/SE 17.0 software (StataCorp, College Station, TX). P values <0.05 were considered statistically significant.

The ARIC data are not publicly available due to confidentiality issues. Investigators can access data from the ARIC study by submitting a manuscript proposal to aricpub@unc.edu.

Among a total of 13,087 participants, there were 2,982 individuals with dementia and 4,662 deaths without dementia during a median (percentile 25–percentile 75) follow-up of 24.1 (17.4–28.3) years. Participants with middle age–onset and older-onset diabetes were more likely to have education less than high school or to be Black compared with participants without diabetes (Table 1). Average BMI was highest in participants with middle age–onset diabetes (31.5 kg/m²), followed by participants with older-onset diabetes (30.8 kg/m²) and participants with no diabetes (27.4 kg/m²). Participants with diabetes were more likely to be physically inactive regardless of diagnosis age. Participants with diabetes had a higher level of total cholesterol, a lower level of HDL cholesterol, and a greater prevalence of hypertension, stroke, and atrial fibrillation.

The cumulative incidence of dementia computed from Kaplan-Meier estimation censoring deaths was much higher than from CIF estimation (Fig. 1). The Kaplan-Meier approach showed cumulative incidence estimates of dementia for participants without diabetes, with middle age–onset diabetes and with older-onset diabetes were 11.5% (95% CI 10.8–12.2), 25.2% (95% CI 22.1–28.5) and 14.3% (95% CI 10.7–18.8) at age 80, respectively; and 53.9% (95% CI 52.0–55.9), 74.2% (95% CI 66.9–80.9), and 60.7% (95% CI 53.0–68.5) at age 90, respectively. The cumulative incidence reached 80% by age 95 years in this method, which censors deaths prior to dementia. The mortality cumulative incidence calculated using Kaplan-Meier and CIF estimators is shown in Supplementary Fig. 2.

In contrast, using CIF estimation, the cumulative incidence of dementia before age 80 years was highest for individuals with middle age–onset diabetes, followed by similar dementia risks in those with older-onset diabetes and no diabetes. After age 80 years, the influence or mortality differences dominate the CIF dementia risk, and those with no diabetes had higher dementia CIF than older-onset diabetes, even surpassing middle age–onset diabetes at age 89. The CIF estimates of dementia for no diabetes, middle age–onset diabetes, and older-onset diabetes groups were 9.4% (95% CI 8.9–10.0), 16.1% (95% CI 14.1–18.3), and 9.1% (95% CI 6.6–12.0) by age 80, respectively. At age 95, the lifetime risk estimates were 45.6% (95% CI, 44.0–47.3), 36.0% (95% CI, 32.6–39.4), and 31.0% (95% CI, 26.5–35.6) for individuals with no diabetes, middle age–onset diabetes, and older-onset diabetes, respectively.

Table 2 shows HRs and predicted CIFs from cause-specific proportional hazard models adjusting for all covariates. Middle age–onset diabetes was significantly associated with dementia (HR 1.7, 95% CI 1.5–2.0), whereas older-onset diabetes was not (HR 0.9, 95% CI 0.8–1.1). In contrast, death without dementia was associated with both diabetes groups, with HRs of 1.9 (95% CI 1.8–2.1) for middle age–onset diabetes and 1.8 (95% CI 1.6–2.0) for older-onset diabetes. Accounting for competing risk, middle age–onset diabetes consistently had the highest composite CIFs. By age 90, the composite CIF estimates were 93%, 87%, and 77% for middle age–onset diabetes, older-onset diabetes and no diabetes, respectively. Diabetes, regardless of age of diagnosis, was associated with greater mortality CIFs compared with no diabetes. Middle age–onset diabetes was strongly associated with elevated dementia CIFs across all ages compared with older-onset diabetes. At ages 80 and 90, the predicted dementia CIFs for middle age–onset diabetes were 1.7 times (11.7% vs. 6.7%) and 1.5 times (32.3% vs. 22.2%) that of older-onset diabetes.

Figure 2 shows that mean dementia-free survival years were shortest for individuals with middle age–onset diabetes, even after adjustment for all covariates. Starting from age 70, they experienced 4.1 fewer years (13.4 [95% CI 12.9–13.9] vs. 9.3 [95% CI 8.5–10.1]) and 1.4 fewer years (10.7 [95% CI 9.5–11.8] vs. 9.3 [95% CI 8.5–10.1]) compared with no diabetes and older-onset diabetes. Starting from age 80, individuals without diabetes and with older-onset diabetes had 2.0- and 1.4-times longer dementia-free survival years than middle age–onset diabetes (5.4 [95% CI 5.0–5.8] and 3.7 [95% CI 3.0–4.6] vs. 2.7 [95% CI 2.1–3.3] years). As for survival years lived with dementia, people without diabetes had the longest duration, whereas those with middle age–onset diabetes had similar years lived with dementia as those with older-onset diabetes.

Owing to the small number of underweight individuals (n = 71 [0.5%]), the results were essentially unchanged after separating these individuals as another BMI category. The results remained the same as well when individuals reported being Asian or American Indian (n = 36 [0.3%]) were excluded.

In this prospective community-based cohort study, participants with middle age–onset diabetes had a higher lifetime risk (at age 95) of dementia than those with older-onset diabetes (36% vs. 31%). People with middle age–onset diabetes developed dementia 4 years earlier than those with older-onset diabetes and 1 year earlier than those with no diabetes. These findings suggest that preventing or delaying diabetes may be important for reducing the risk of dementia throughout the life course.

Our study extends the existing research on the lifetime risk of dementia. The Framingham Heart Studies (FHS) and the Rotterdam Study have presented overall lifetime risk by sex (6,7,20), and a recent study in the Whitehall II study (2) showed relative risks of dementia by age of diabetes onset. However, previous studies have not examined the absolute lifetime risk of dementia by age of diabetes diagnosis. Our research bridges these two areas, highlighting how the risk of dementia evolves with the age at diabetes diagnosis in both relative risk and absolute risk scales.

The FHS and Rotterdam Study both reported lifetime risk estimates of dementia that were ∼15–20% lower than our study. The higher risk estimates in our study may be explained by dementia ascertainment. The FHS and Rotterdam study mostly ascertained dementia through study visits, where individuals with cognitive impairment were less likely to show up for visit-based evaluations (21). The ARIC study has more thorough dementia ascertainment strategies, including not only visit-based evaluations but also telephone calls, informant interviews, hospital discharge codes, and death records (Supplementary Table 2).

Our findings are consistent with research indicating that the dementia risk is significantly higher when diabetes is diagnosed at an earlier age (2–4). In the Whitehall II study, the risk of dementia was two-times higher in individuals with diabetes diagnosed before the age of 60 compared with those without diabetes (HR 2.12, 95% CI 1.50–3.00), and the risk of dementia increased by 24% for every 5-years younger onset of diabetes (HR 1.24, 95% CI 1.06–1.46). The Swedish Twin Registry study also showed that only diabetes <65 years (odds ratio 2.41, 95% CI 1.05–5.51) was associated with a higher risk of dementia (3). Duration of diabetes and age at diabetes onset were correlated but independently confer risk for dementia (2). Cumulative, long-term exposure to hyperglycemia is likely an important pathway by which diabetes affects neurocognitive outcomes (22). Prolonged exposure to factors such as inflammation, oxidative stress, and mitochondrial dysfunction among individuals with diabetes appear to play a critical role in the deterioration of neuronal structure and function (23–25).

In the past two decades, the incidence and prevalence of diabetes in the U.S. have been increasing concomitantly with rates of overweight and obesity (26,27). However, prevention programs, such as the Centers for Disease Control and Prevention-led National Diabetes Prevention Program (28) are underused (29). Improving diabetes prevention has the potential to convey long-term benefits on reducing absolute dementia risk and extending healthy survival years.

We found that individuals without diabetes had the highest lifetime risk of dementia: 45.6% for no diabetes, 31.0% for older-onset diabetes, and 36.0% for middle age–onset diabetes. This finding is almost certainly due to the competing risk of mortality. Old age is the strongest risk factor for dementia, and individuals without diabetes live substantially longer than those with diabetes. These results suggest that effective diabetes prevention interventions may extend survival and paradoxically increase the lifetime dementia risk for some adults. Thus, it may be important to pair prevention with additional resources and interventions aimed at promoting brain health in older age. Novel disease-modifying therapies have recently been approved and may be especially important. Other nonpharmacological strategies, including healthy lifestyle behaviors (diet, exercise, and cognitive activity) (30–32), hearing-aid adoption (33), and social engagement (34,35), may also be beneficial.

Strengths of our study include the rigorous consideration of competing risk of death, using CIF estimation and cause-specific proportional hazard models. Kaplan-Meier estimation substantially overestimates risk compared with CIF estimation. Using nonparametric risk estimation from cause-specific proportional hazard models, compared with Fine-Grey methods, is useful when the CIFs by exposure groups are nonproportional. The estimation of survival years based on competing risk and Cox proportional hazard models are also helpful for interpretation. Our study also benefited from the community-based, large-scale, long-duration cohort with rigorous information on risk factors and comprehensive ascertainment and adjudication of dementia.

Our study has certain limitations. First, we broadly categorized age of diabetes diagnosis because exact age was not systematically captured in our study. Second, people who entered the study with age ≥60 (range 60–69) might be healthier and have later dementia incidence compared with those <60 at baseline. However, the dementia rate is low from age 60 to 69, so any bias is likely to be minimal. Third, we did not have information to differentiate dementia subtypes. Last, owing to the observational nature of our study, we cannot rule out residual confounding.

In conclusion, younger age of diabetes diagnosis was strongly associated with earlier dementia onset. Compared to those without diabetes, adults with middle-age onset diabetes had a higher risk of dementia by age 80 years, but a lower lifetime risk because of increased mortality. Preventing or delaying diabetes is an important approach for reducing dementia risk throughout the life course. However, the risk of dementia is very high at oldest ages, regardless of diabetes status. Thus, there is a need for additional strategies that promote brain health in older age for all individuals.

Acknowledgments. The authors thank the staff and participants of the ARIC study for their important contributions.

E.S. is an editor of Diabetes Care but was not involved in any of the decisions regarding review of this manuscript or its acceptance.

Funding. The ARIC study has been funded in whole or in part with funds from the National Heart, Lung, and Blood Institute (NHLBI), National Institutes of Health (NIH), Department of Health and Human Services, under contract nos. 75N92022D00001, 75N92022D00002, 75N92022D00003, 75N92022D00004, and 75N92022D00005. The ARIC Neurocognitive Study is supported by U01HL096812, U01HL096814, U01HL096899, U01HL096902, and U01HL096917 from the NIH (NHLBI, National Institute of Neurological Disorders and Stroke, National Institute on Aging, and National Institute on Deafness and Other Communication Disorders). P.L.L. was supported by NIH/NHLBI grant K24 HL159246. R.F.G. was supported by the National Institute of Neurological Disorders and the Stroke Intramural Research Program. E.S. was supported by NIH/NHLBI grants K24 HL152440 and R01 HL158022, NIH/National Institute of Diabetes and Digestive and Kidney Diseases grants R01 DK089174 and R01 DK128837, and NIH/National Institute on Aging grant RF1 AG074044. M.F. was supported by NIH/NIDDK grant K01 DK138273-01.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. J.H. conducted the analyses and wrote the initial manuscript. J.H., E.S., M.F., and J.C. designed the study, researched the data, and contributed to the discussion. J.R.P., P.L.L., A.R.S., L.E.W., T.M.H., J.C.S., R.F.G., and T.H.M. provided substantial contributions to the interpretation of data, made critical revisions to the manuscript, and approved the final manuscript. J.H. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Handling Editors. The journal editor responsible for overseeing the review of the manuscript was Frank Hu.