A Diabetes Genetic Risk Score Is Associated With All-Cause Dementia and Clinically Diagnosed Vascular Dementia in the Million Veteran Program

Diabetes and dementia are common polygenic diseases with a high health care burden. The type 2 diabetes prevalence in the U.S. as of 2018 was 10.5% (1) and Alzheimer disease (AD) and related dementias were diagnosed in >11.5% of adults older than the age of 65 (2). Not only does dementia have a devastating effect on quality of life during aging, the care of dementia patients also poses a severe impending health care crisis: the estimated lifetime costs of health care for a person with dementia are $184,000 greater than those of a person without dementia (3). Additionally, disease prevalence differs by race/ethnicity. For diabetes (1), the rate is 11.7% among non-Hispanic Black individuals, 12.5% among Hispanic individuals, and 7.5% among non-Hispanic White individuals. In the case of dementia (2), the numbers are 14.7% for non-Hispanic Black individuals, 12.9% for Hispanic individuals, and 11.3% for non-Hispanic White individuals.

Observational epidemiology studies have linked type 2 diabetes with multiple diagnoses of dementia, including late-onset AD and vascular dementia (VaD) (4,5). Compared to those without diabetes, the relative risk of dementia is increased 1.4- to 2.2-fold in individuals with diabetes. More targeted analyses have found associations of diabetes-related traits, such as insulin resistance and cardiovascular disease, with an increased incidence of dementia as well as evidence of a relationship between diabetes and nonamnestic mild cognitive impairment (6–8). Experiments focused on impaired glycemic control, hyperinsulinemia, and metabolic syndrome propose hypotheses for the relationship (9–11), such as disparate cerebrocortical activity in insulin-resistant individuals (10), and diminished blood-brain barrier integrity due to hyperglycemia (11,12). Nevertheless, studies of individuals who have avoided risk factors or controlled their diabetes offer hope for a decline in the incidence of dementia (13–15).

Genetics can be an effective tool for probing both disparities and biological mechanisms. With its well-known association established >25 years ago (16), APOE ε4 is the strongest common (>1% allele frequency) risk locus for AD, with risk increasing 2.84-fold with one allele and 8.07-fold with two alleles (16). The highly heterogeneous molecular architecture of the APOE region includes rs429358 and rs7412 that together encode ε4 alleles (deleterious) or ε2 alleles (protective) with respect to AD (16,17). Recent research has identified additional common AD-associated loci with smaller effect sizes in large-scale genome-wide association studies (GWAS) of European-descent cohorts (18,19) and novel loci unique to African Americans in a GWAS meta-analysis with an African genotyping panel (20). Regarding diabetes, the extensive studies of the genetics of type 2 diabetes and related glycemic traits have identified >300 independent variants with additive effects on type 2 diabetes risk (21,22). Several genetic studies assessing the causal relationship of type 2 diabetes and other glycemic traits with the risk of AD or reduced cognitive function have found little evidence that such a relationship exists (23–27). Yet, in one of those studies, by Thomassen et al. (24), the observational association between these two diseases was overwhelming, with hazard ratios of 1.13 (95% CI 1.06–1.21) for AD, and 1.98 (95% CI 1.83–2.14) for VaD.

To examine whether type 2 diabetes–associated genetic variants are associated with dementia and whether they differ by race/ethnicity or dementia diagnosis, we assessed the relationship between a genetic risk score (GRS) for type 2 diabetes and three classifications of clinical dementia diagnoses in the Million Veteran Program (MVP): all-cause dementia, VaD, and AD.

We conducted our study in the U.S. Department of Veterans Affairs MVP (28), the largest clinical biobank in the U.S. MVP links clinical data from the Veterans Affairs (VA) Healthcare System with genotype data on >650,000 individuals. MVP includes a racially/ethnically diverse population (29) with an overall diabetes prevalence of >35% (22). MVP participants went through a counseling process before they enrolled and provided consent to have their electronic health records reviewed. The VA Central Institutional Review Board (Washington, D.C.) gave approval for the study protocol in accordance with the principles of the Declaration of Helsinki (22).

Given the known difference in diabetes and dementia prevalence by race/ethnicity and the difference in the effect of variants by ancestry (1,2), we analyzed participants from MVP stratified by race/ethnicity group using the harmonized ancestry and race/ethnicity classification (HARE) method established in this population (30). HARE was developed through a machine learning algorithm combining self-reported race with inferred ancestry from genotype data to produce four HARE groups: non-Hispanic White (EUR), non-Hispanic Black (AFR), Hispanic (HIS), and non-Hispanic Asian (ASN). All participants in our study were at least 65 years of age. We focused our research on EUR, AFR, and HIS because the sample size for ASN was too small (<3,000 participants with phenotype and genotype data).

The processing of genotype data within MVP has been described elsewhere (31). Briefly, our work was based on release 4 of the MVP genotype data (N = 658,582 total participants), which was imputed to the African Genome Resources reference panel by the Sanger Institute (1000 Genomes Project plus 1,500 unrelated pan-African samples). The genotyping array used was a custom array (MVP 1.0) designed specifically for the MVP population based on the Thermo Fisher Scientific Axiom Genotyping Platform. We removed individuals who were more closely related than second cousins (i.e., a coefficient of kinship >0.088) to arrive at a sample of completely independent participants. From the imputed genotypes, we built a weighted GRS for type 2 diabetes using 331 variants and their effect sizes from published GWAS. Specifically, we used the loci from the Diamante Consortium (21) and the European effect sizes from Mahajan et al. (32) to build the GRS for all race/ethnicity groups. A more recent GWAS of type 2 diabetes (22) included MVP participants, so we chose not to use its effect sizes due to concerns of model overfitting. Similarly, we did not use the effect sizes from the Diamante Consortium because they reflected a meta-analysis with MVP (21). Thus, to avoid overfitting, the GRS included loci identified in a multiancestry GWAS of type 2 diabetes weighted by effect sizes from a GWAS in European-ancestry individuals only. We standardized the GRS values to a mean of 0 and a SD of 1.

APOE isoform calls for the MVP participants (e.g., ε2/ε2, ε3/ε3, ε3/ε4, etc.) were determined from the genotypes of the two single nucleotide polymorphisms (SNPs) whose haplotypes define the different isoforms: rs429358 and rs7412. Best-guess genotypes for these SNPs were generated from the imputed data. Both SNPs were well imputed in MVP (r² = 1 and r² > 0.87 for all HARE groups, respectively). From the APOE isoform genotypes, we derived the additively coded APOE ε4 dosage; that is, individuals were coded as having zero, one, or two ε4 alleles.

To pursue one potential biological mechanism, we constructed an HbA_1c GRS consisting of 91 variants from the transancestry results identified in Chen et al. (33). Additionally, we constructed a GRS of 23 variants from the glycemic class (cluster G) when identified in the “hard” clustering structure from the same analysis (33). Both of these GRS values were standardized to a mean of 0 and SD of 1.

Our outcomes were three clinical diagnoses of dementia (Supplementary Fig. 1): AD, VaD, and all-cause dementia based on International Classification of Diseases, Ninth Revision (ICD-9), (34) or ICD-10 diagnosis codes (Supplementary Table 1). In the VA electronic health record, these codes are logged by presiding physicians in the course of routine clinical care as administrative documentation of their professional assessment of a patient’s condition. A case was defined as a participant with at least two ICD-9 or ICD-10 codes corresponding to one or more of the outcomes. For AD, case participants had at least two codes related to AD (Supplementary Table 1); and for VaD, case participants had at least two ICD-9 or ICD-10 codes related to VaD (Supplementary Table 1). Note that individuals meeting the case definition of AD and VaD were included as case participants for both; that is, AD and VaD were not mutually exclusive. Control subjects were those participants without a mention of even one dementia ICD-9 or ICD-10 code in VA clinical records.

We compared individuals across HARE groups using χ² tests for categorical variables and t tests for continuous variables. We evaluated associations of the GRS with dementia outcomes using logistic regression stratified by EUR, AFR, and HIS races/ethnicities. We considered P < 0.017 (0.05/3 outcomes) to be statistically significant, accounting for multiple testing, and also reported nominally significant results using a threshold of P < 0.05. We adjusted the models for biological sex, age, and 10 genetic principal components to account for potential confounding due to global ancestry. Since rs429358, one of the variants in the APOE isoform, is also in the GRS, we conducted a sensitivity test by removing this variant from the GRS to construct a new GRS called GRS-nonAPOE. GRS-nonAPOE was evaluated using the same models as the GRS.

As a further sensitivity test, we excluded individuals who met case definitions for both clinically diagnosed AD and VaD, creating mutually exclusive cases for those two outcomes. Additionally, we analyzed the interaction of the APOE ε4 dosage with GRS-nonAPOE using a model in which APOE ε4 dosage was included as a linear (additive) term along with an APOE ε4 × GRS-nonAPOE interaction term. If we found that the interaction term did not meet the significance threshold (0.017), we removed the interaction to assess the model with the two genetic main effects: APOE ε4 dosage and GRS-nonAPOE.

We evaluated the association of each individual variant making up the GRS with each clinical dementia diagnosis to determine whether specific type 2 diabetes–associated loci were associated with dementia. For these association tests, our multiple testing significance threshold was 1.51e−4 (0.05/331 variants). For variants that achieved nominal significance (P < 0.05) for association with any dementia outcome, we examined functional annotation information from publicly available resources (35).

Finally, we examined associations of GRS related to HbA_1c with each clinical dementia diagnosis, stratified by race/ethnicity group, to evaluate the potential role of glycemia in associations of diabetes with dementia outcomes. For this analysis, we applied the same significance thresholds as in the primary diabetes GRS analyses of P < 0.017 (0.05/3 outcomes) for statistical significance and P < 0.05 for nominal significance.

There were 334,672 MVP participants (Table 1) who met our inclusion criteria as case participants (n = 24,043 [7.18%]) for at least one of the clinical dementia diagnoses or control participants (n = 310,629). The average age for AFR and HIS was younger than that of EUR. The study sample was mostly male (96.7%); 81.7% of the sample were EUR, 12.9% were AFR, and 5.4% were HIS (Table 1). Clinically diagnosed VaD and AD prevalence differed by race/ethnicity, but all-cause dementia did not. Diabetes prevalence differed by race/ethnicity: 33.53% for AFR, 22.78% for EUR, and 34.44% for HIS. The APOE ε4 carrier status differed by race/ethnicity, with the highest rate of AD risk-raising carriage in AFR (P < 0.001). The GRS and GRS-nonAPOE differed by race/ethnicity as well, with AFR being lower than the other two HARE groups (Supplementary Fig. 2).

Both the GRS and GRS-nonAPOE were significantly associated with diabetes status for all races/ethnicities (odds ratio [OR] > 1.28, P < 3.18e−118). In EUR, we found associations of the GRS (Fig. 1) with all-cause-dementia (OR 1.06 per 1-SD increase in the GRS, P = 1.60e−07), and clinically diagnosed VaD (OR 1.12 per 1-SD increase in the GRS, P = 5.2e−05) but not clinically diagnosed AD (OR 1.02, P = 0.43). The GRS was not associated with any of the dementia outcomes in AFR or HIS (all P > 0.05). When removing rs429358 to construct GRS-nonAPOE, we found that the effect size estimates in EUR increased (Fig. 1) and the P values decreased for all-cause dementia (OR 1.08, P = 2.6e−12) and clinically diagnosed VaD (OR 1.14, P = 7.2e−07) and that there was a nominal association with clinically diagnosed AD (OR 1.06, P = 0.018). For AFR, the association of GRS-nonAPOE with all-cause dementia (OR 1.07, P = 0.042) was nominally significant, and clinically diagnosed VaD (OR 1.15, P = 0.016) crossed the significance threshold. There were no significant findings for HIS. The effect sizes across all races/ethnicities were consistently risk-raising for dementia with GRS-nonAPOE even while not meeting the statistical significance threshold. When we excluded individuals meeting the case definitions for both VaD and AD, the effect size for GRS-nonAPOE in association with clinically diagnosed VaD increased, while the association with clinically diagnosed AD decreased in EUR (Supplementary Table 2). In AFR, the effect sizes were stable but now only nominally significant.

APOE ε4 dosages were available for almost 97% of the study population. In the model of the interaction between the GRS-nonAPOE and the APOE ε4 dosage (Supplementary Table 3), we found a significant interaction for clinically diagnosed AD in EUR (β = 0.144, P = 1.02e−03). None of the interaction terms were significant for any other outcome and race/ethnicity combination. When we removed the interaction term from the model, we found that APOE ε4 dosage was associated (OR > 1.5 per allele, P < 1.6e−06) with all diagnoses of dementia for all race/ethnicity populations in a model that included GRS-nonAPOE. Additionally, we found that GRS-nonAPOE was associated with all-cause dementia (OR 1.09, P = 2.3e−13), clinically diagnosed VaD (OR 1.15, P = 3.70e−07), and clinically diagnosed AD (OR 1.07, P = 0.012) in EUR in a model that included the APOE ε4 dosage. Clinically diagnosed VaD was associated (OR 1.18, P = 0.01) with GRS-nonAPOE in AFR. The effect size for the APOE ε4 dosage term in the model increased from all-cause dementia to VaD to AD (Fig. 2A). Conversely, the effect size for GRS-nonAPOE decreased from VaD to AD in this model (Fig. 2B).

In the association tests for the individual GRS variants with the three clinical diagnoses of dementia and all three races/ethnicities using a significance threshold of 0.05/331 = 1.5e−04, we found an association of all-cause dementia, clinically diagnosed VaD, and clinically diagnosed AD with rs429358, one of the variants that defines the APOE isoform genotype (all OR < 0.65, all P < 3.84e−06) (Supplementary Table 4). This variant is located in the fourth exon of APOE and promotes the expression of ε4 alleles that increase AD risk. The minor allele (C) is associated with decreasing risk in type 2 diabetes and increasing risk for AD (Fig. 3).

While none of the other individual variants met our significance threshold (Supplementary Table 4), we found a total of 178 associations with 127 unique variants (∼38%) that were nominally significantly associated (P < 0.05) with at least one dementia subtype in at least one race/ethnicity (Supplementary Table 4). In a few cases, the frequency of the nominally significant associations was significantly greater than what would be expected by chance. In EUR, 30 of 330 variants were nominally significantly associated with all-cause dementia (binomial P = 0.0013), and in HIS, 25 of 330 variants were nominally significantly associated with clinically diagnosed AD (binomial P = 0.027) (Supplementary Table 5).

Additionally, in an assessment of direction of effect of the 330 type 2 diabetes–associated variants (excluding rs429358), we observed evidence of concordance with dementia diagnoses at a higher frequency than what would be expected by chance (Supplementary Table 6). In EUR, 214 of 330 variants had the same direction of effect with all-cause dementia (binomial P = 3.77e−08), 184 of 330 with clinically diagnosed AD (binomial P = 0.0208), and 196 of 330 with clinically diagnosed VaD (binomial P = 0.0004). In AFR, 189 of 330 variants had the same direction of effect with all-cause dementia (binomial P = 0.0048) and 182 of 330 with clinically diagnosed AD (binomial P = 0.0346). In HIS, 181 of 330 had the same direction of effect with clinically diagnosed VaD (binomial P = 0.0439). When limiting the analysis to variants with a nominally significant association with a dementia trait, there was no statistically significant evidence (all P > 0.07) that the direction of effect was the same for diabetes and dementia (Supplementary Table 6).

In a functional annotation assessment, 71 of the nominally associated variants were near protein-coding genes, most frequently APOE, KCNQ1, and BCAM. Commonly associated pathways for these variants were gene expression, metabolism, the neuronal system, signal transduction, immune function, and developmental biology. In a comparison of our nominally significant variants with the summary statistics of a recent GWAS of AD (https://www.niagads.org/datasets/ng00075) (18), 16 of these variants met nominal significance, and 3 of them (rs12419690, rs1871045, and rs429358) met our significance threshold of 0.00015.

Note that there are seven variants missing from Fig. 3 because their effect sizes were too large for readability of the chart. Two of these variants, rs76263492 and rs80102379, are described in more detail in the nominally significant list (Supplementary Table 4).

To clarify the role that glycemia plays in the development of dementia, as opposed to diabetes, we conducted association tests of all-cause dementia, clinically diagnosed VaD, and clinically diagnosed AD with a GRS for HbA_1c and a sub-GRS corresponding to glycemia (cluster G) derived from a recent multiancestry GWAS meta-analysis of HbA_1c by Chen et al. (33) (Supplementary Material). In these association tests, we found some suggestive relationships in the EUR and AFR races/ethnicities (Supplementary Fig. 3). Two associations met our significance threshold in EUR testing (Supplementary Fig. 3A) of the HbA_1c GRS with dementia: cluster G with all-cause dementia (OR 1.03 per 1-SD increase in the GRS, P = 0.002) and cluster G with clinically diagnosed VaD (OR 1.06, P = 0.002). In general, genetically upregulated HbA_1c and cluster G were associated with an increased risk of dementia in EUR, even when the significance threshold was not met. In AFR (Supplementary Fig. 3B), no associations met our significance threshold, but cluster G had a large effect estimate with clinically diagnosed AD and nominal significance (OR 1.13, P = 0.019).

We found a relationship of a diabetes GRS with clinically diagnosed VaD and all-cause dementia in EUR, particularly when rs429358 of the APOE isoform was excluded from the analysis. The GRS, however, did not meet our significance threshold with clinically diagnosed AD in any of the races/ethnicities but did exhibit nominal significance in EUR after removing rs429358. The GRS-dementia relationship was statistically significant for clinically diagnosed VaD in AFR after removing rs429358. When rs429358 was tested by itself, this APOE variant was significantly associated with clinically diagnosed AD and VaD and all-cause dementia in all three populations.

We observed a stronger association of the GRS with clinically diagnosed VaD than with all-cause dementia, consistent with prior epidemiologic studies (4). It is notable that we were able to detect this association despite considerably lower statistical power due to the number of case subjects (3,145 for clinically diagnosed VaD vs. 19,627 for all-cause dementia) and the imprecision of a clinical diagnosis in electronic health records. Similarly, the GRS association with clinically diagnosed VaD was significant in AFR even with only 821 case participants in our study population. Consistent with the primary analysis demonstrating a stronger association between GRS and clinically diagnosed VaD than with clinically diagnosed AD, the association of GRS-nonAPOE with clinically diagnosed VaD was stable in the sensitivity analysis excluding individuals meeting both VaD and AD case definitions in EUR. Previous studies have not observed a relationship between the genetics of diabetes and a risk of AD or impaired cognitive function (23–27). Our study offers new insight to these prior works, most likely due to the opportunity to analyze multiple diagnoses of dementia and the detection of a pleiotropic variant: rs429358.

Our study with its large AFR sample size (n = 43,089) adds to the literature demonstrating the value of studying genetics, diabetes, and dementia in African Americans. Previous studies with an African ancestry focus have offered fresh discoveries: a GWAS meta-analysis with an African genotyping panel identifying novel loci in African Americans (20), a local ancestry analysis (36) deciphering racial differences in the risk conferred by APOE ε4 in Europeans compared with African Americans, and a third study in which 27.1% of the participants were Black, demonstrating comparable associations of diabetes and APOE ε4 genotype with dementia risk (37). With the higher prevalence of diabetes in African Americans (1) and the disparity in cardiovascular disease in non-Hispanic Black individuals compared with the non-Hispanic White population (35), the significant association we found in AFR with clinically diagnosed VaD (and the larger effect size than in EUR) provides novel insight into the relationship between diabetes and dementia in African Americans that strongly warrants further investigation.

The fact that the GRS without the rs429358 variant had a greater effect size estimate for the risk of dementia suggests that the role of diabetes in the development of dementia is different from the well-known path of the APOE gene. Furthermore, our findings in the association of the cluster G GRS with clinically diagnosed VaD and all-cause dementia in EUR are consistent with this proposition. Previous work in this area has demonstrated the complexity of the relationship between these two diseases: Hersi et al. (38) and Corder et al. (17) showing a protective role of APOE ε2 against AD, Li et al. (39) demonstrating an increased incidence of dementia in patients with diabetes with the APOE ε4 genotype, and Irie et al. (40) finding multiplicative synergy of diabetes and APOE ε4 versus Shinohara et al. (41) discovering a significant interaction with diabetes increasing the risk of cognitive decline in APOE ε3 and APOE ε2 carriers but not APOE ε4 carriers. These seemingly conflicting studies in combination with our current study indicate a need to continue to disentangle the joint contributions of diabetes and APOE genotype leading to AD versus VaD.

Our study has important limitations. First, using effect sizes from a European-ancestry GWAS for the GRS does not account for genetic differences that might be present in an African-ancestry population. Second, our sample was predominantly men, precluding evaluation of sex interactions on dementia risk and potentially limiting generalizability of our results. Third, the cross-sectional analysis precludes causal inference. Fourth, associations with specific dementia subtypes may be impacted by imprecise diagnosis code use in the electronic health record. However, the sensitivity analysis of mutually exclusive definitions of clinically diagnosed VaD and AD demonstrated consistent results with the primary analyses.

Despite the limitations, our study provides evidence for common genetic pathways leading to diabetes, VaD, and all-cause dementia. Moreover, we have found different effect sizes in the relationships between genetic diabetes risk and clinical dementia diagnosis suggesting distinct biology mechanisms in both EUR and AFR populations. Future work will include Mendelian randomization analysis in this MVP population to determine causal associations of diabetes with dementia and differentially methylated DNA analysis based on diabetes case status in a separate African American cohort.

Acknowledgments. The authors would like to thank Dr. Maggie Stanislawski, assistant professor at the University of Colorado, for her R code examples that made the analysis steps more straightforward. The authors would like to thank Dr. Qin Hui, Department of Epidemiology, Rollins School of Public Health, for his assistance in relatedness calculations and ongoing data steward work.

Funding. E.M.L. is supported by National Institutes of Health National Institute of Diabetes and Digestive and Kidney Diseases award P30 DK116073, and by funds from the Boettcher Foundation Webb-Waring Biomedical Research Program. Phenotype development in the MVP was supported by U.S. Department of Veterans Affairs award BLR&D 1 I01BX004192 (M.W.L., primary investigator). L.S.P. is supported in part by Veterans Administration awards CSP no. 2008, I01 CX001899, I01 CX001737, and I01 BX005831, National Institutes of Health National Institute of Diabetes and Digestive and Kidney Diseases awards R01 DK127083 and U01 DK098246, National Institute of Allergy and Infectious Diseases awards R03 AI133172 and R21 AI156161, National Center for Advancing Translational Sciences award UL1 TR002378, and Cystic Fibrosis Foundation award PHILLI12A0. R.L.H. is supported by the MVP022 award no. I01 CX001727, VISN-22 Veterans Administration Center of Excellence for Stress and Mental Health (CESAMH), and National Institute of Aging RO1 grants AG050595 (The Vietnam Era Twin Study of Aging [VETSA] Longitudinal Twin Study of Cognition and Aging VETSA 4), AG05064 (Effects of Androgen Deprivation Therapy on Preclinical Symptoms of Alzheimer’s Disease), and AG065385 (Novel Antagonists of the N-terminal Domain of the CRF Receptor Type 1 for Alzheimer’s Disease). S.R. is supported by U.S. Department of Veterans Affairs award IK2-CX001907, by National Institute of Diabetes and Digestive and Kidney Diseases award P30DK116073, by funds from the Boettcher Foundation Webb-Waring Biomedical Research Program, and has previously received research grant funding from the American Heart Association. This research is based on data from the MVP, Office of Research and Development, Veterans Health Administration, and was supported by awards MVP003, MVP009, MVP015, and MVP022.

This publication does not represent the views of the U.S. Department of Veteran Affairs or the U.S. Government.

Duality of Interest. Within the past several years, L.S.P. has served on Scientific Advisory Boards for Janssen and the Profil Institute for Clinical Research, and has or had research support from Merck, Amylin, Eli Lilly, Novo Nordisk, Sanofi, PhaseBio, Roche, AbbVie, Vascular Pharmaceuticals, Janssen, Glaxo SmithKline, and Pfizer. In the past, L.S.P. was a speaker for Novartis and Merck, but not for the last 5 years. L.S.P. is also a cofounder, officer and board member, and stockholder of DIASYST, Inc., which is developing software aimed to help improve diabetes management. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. E.M.Li. designed the study, performed the literature review, conducted the analysis, and wrote the manuscript. M.W.L. contributed to the dementia phenotype definitions, provided guidance on the APOE haplotypes, and edited the manuscript. R.Z. implemented the dementia phenotypes. B.R.C. implemented the diabetes phenotype. E.M.La. provided guidance on statistical methods and edited the document. J.E.H. provided guidance on applicable epidemiology concepts and study design. J.A.L. designed the dementia phenotype groupings. M.V. provided epidemiology and study design guidance. L.S.P. designed the diabetes complications study. L.A.L. provided epidemiology and study structure guidance. R.L.H. provided dementia subject matter expertise and edited the manuscript. S.R. offered diabetes subject matter expertise, study design and analysis guidance, and edited the manuscript. E.M.Li. and S.R. are the guarantors of this work and, as such, had full access to all the data in this study and take responsibility for the integrity of the data and accuracy of the data analysis.

Prior Presentation. Preliminary results of this study were presented as a poster at the Alzheimer's Association International Conference, Denver, CO, and online, 26–30 July 2021.