Racial and Ethnic Differences in the Population-Attributable Fractions of Alzheimer Disease and Related Dementias

Abstract

Background and Objectives

Previous studies estimated that modifiable risk factors explain up to 40% of the dementia cases in the United States and that this population-attributable fraction (PAF) differs by race and ethnicity—estimates of future impact based on the risk factor prevalence in contemporary surveys. The aim of this study was to determine the race-specific and ethnicity-specific PAF of late-onset Alzheimer disease and related dementias (ADRDs) based on the risk factor prevalence and associations observed on the same individuals within a prospective cohort.

Methods

Data were from Multiethnic Cohort Study participants (African American, Japanese American, Latino, Native Hawaiian, and White) enrolled in Medicare Fee-for-Service. We estimated the PAF based on the prevalence of risk factors at cohort baseline and their mutually adjusted association with subsequent ADRD incidence. Risk factors included low educational attainment and midlife exposures to low neighborhood socioeconomic status, unmarried status, history of hypertension, stroke, diabetes or heart disease, smoking, physical inactivity, short or long sleep duration, obesity, and low-quality diet, as well as APOE ε4 for a subset.

Results

Among 91,881 participants (mean age 59.3 at baseline, 55.0% female participants), 16,507 incident ADRD cases were identified from Medicare claims (1999–2016, mean follow-up 9.3 years). The PAF for nongenetic factors combined was similar in men (24.0% [95% CI 21.3–26.6]) and women (22.8% [20.3–25.2]) but varied across Japanese American (14.2% [11.1–17.2]), White (21.9% [19.0–24.7]), African American (27.8% [22.3–33.0]), Native Hawaiian (29.3% [21.0–36.7]), and Latino (33.3% [27.5–38.5]) groups. The combined PAF was attenuated when accounting for competing risk of death, in both men (10.4%) and women (13.9%) and across racial and ethnic groups (4.7%–25.5%). The combined PAF was also different by age at diagnosis and ADRD subtypes, higher for younger (65–74 years: 43.2%) than older (75–84 years: 32.4%; ≥85 years: 11.3%) diagnoses and higher for vascular or unspecified ADRD than for AD or Lewy body dementia. An additional PAF of 11.8% (9.9–13.6) was associated with APOE ε4, which together with nongenetic risk factors accounted for 30.6% (25.8–35.1) of ADRD.

Discussion

Known risk factors explained about a third of the ADRD cases but with unequal distributions across racial and ethnic groups.

Introduction

Late-onset Alzheimer disease and related dementias (ADRDs) are the fifth leading cause of death among Americans aged 65 and older.¹ Age-specific incidence of ADRD has held steady or declined in the United States over the recent 2 decades,^2-4 likely owing to improved educational attainment and better medical care of cardiovascular conditions.^2,5 Nonetheless, the prevalence of ADRD is projected to double by 2050 because of population aging.¹ Approved treatments are limited to addressing symptoms or slowing the progression of early disease: the latter class of antiamyloid immunotherapy will likely remain access- and cost-prohibitive for broad uptake under the current Medicare coverage policy.⁶ Therefore, preventing or delaying ADRD onset through reducing exposure to modifiable risk factors is an urgent public health priority.

In estimating the preventable population-attributable fraction (PAF) of ADRD, studies have varied in the risk factors considered, and also most studies have used different sources of data for risk factor prevalence vs for magnitude of associations. For example, the 2020 Lancet Commission estimated that up to 40% of dementia cases worldwide may be explained by 12 potentially modifiable risk factors, including less education, hypertension, hearing impairment, smoking, obesity, depression, physical inactivity, diabetes, low social contact, excessive alcohol consumption, traumatic brain injury, and air pollution.⁷ They based their estimate on global risk factor prevalence data for the 2010s and risk factor associations with dementia (relative risks) derived from meta-analyses of published population-based cohort studies or trials.

Two recent US studies determined modifiable fractions and the most prominent risk factors in each racial and ethnic group,^8,9 to understand the reported disparities in ADRD.^10-13 Nianogo et al.⁸ considered 8 of the 12 risk factors suggested by the Lancet Commission using the prevalence from the 2018 US Behavioral Risk Factor Surveillance System and the relative risk estimates from meta-analyses of either ADRD or all-cause dementia, which yielded a PAF of 37% overall but varying estimates among Black (40%), American Indian and Alaska Native (39%), Hispanic (34%), White (29%), and Asian (16%) groups. Another study by Lee et al.⁹ included all 12 risk factors and relative risks from the global Lancet report,⁷ while using the prevalence from 4 cross-sectional surveys conducted in 2011–2018. They estimated the overall PAF at 41% and confirmed differences across Hispanic (47%), Black (46%), White (39%), and Asian (36%) groups.⁹ While emerging evidence points to structural and social determinants of health (SSDoH) as an important contributor to the large racial and ethnic ADRD disparities,¹⁴ PAF estimates made on global or multiethnic populations have not incorporated SSDoH data.

These recent studies underscore the potential for substantially reducing the future burden of ADRD and mitigating racial and ethnic disparities. Their estimation used risk factor prevalence from younger and more contemporary adults and risk factor association estimates from older and heterogeneous cohorts. This approach circumvented the limited availability of long-term cohort data that can provide a uniform assessment of risk factors and ADRD/dementia outcomes on a large number of US adults of multiple racial and ethnic backgrounds. However, this approach depends on a set of assumptions, such as homogeneous risk factor associations across populations, and involves calculating indirect measures, such as degree of risk factor co-occurrence that were then applied to obtain weighted PAFs from unadjusted relative risks.^7,8 In addition, the approach did not account for competing risk of death from other comorbidities that share risk factors with ADRD.⁸

In this study, we tested the hypothesis that the PAF of ADRD varies by race and ethnicity, directly based on observations of the Multiethnic Cohort Study (MEC) participants. In the MEC, most of the established or putative risk factors were prospectively assessed on a large population sample of African American, Japanese American, Latino, Native Hawaiian, and White individuals of middle age in Hawaii and southern California. We previously observed in the MEC a 2-fold difference in age-adjusted ADRD risk by race and ethnicity,^12,13 consistent with limited multiethnic studies to date.¹⁰ In the current analysis, we estimated the PAF based on the risk factor prevalence and mutually adjusted risk factor associations with ADRD observed on the same individuals, while accounting for competing risk of death. We also compared the PAF by sex, age at diagnosis, common ADRD subtypes, and the APOE genotype.

Methods

Study Population

The MEC was established in 1993–1996 with a population-based sample of more than 215,000 men and women, aged 45–75 years, living in Hawaii (Oahu, Big Island, Maui, Kauai, Lanai, and Molokai) or California (largely Los Angeles County and also Riverside, Orange, San Bernardino, San Diego, San Francisco, San Mateo, Contra Costa, and Alameda counties) who completed and returned a comprehensive questionnaire in mail.¹⁵ The primary sampling frames were driver's license files, supplemented by Hawaii voter's registration file to identify older Japanese American women who did not drive and California Health Care Financing Administration files to identify additional African American participants. The 5 racial and ethnic groups targeted for recruitment using ethnic-specific surname lists, which had been shown to be relatively effective, were African American, Japanese American, Latino, Native Hawaiian, and White. On the questionnaire, participants selected 1 or more of the following 9 categories as their racial or ethnic background: Black or African American, Chinese, Filipino, Hawaiian, Japanese, Korean, Mexican or other Hispanic, White or Caucasian, and other. Participants who selected multiple racial/ethnic categories were priority ranked for African American, followed by Native Hawaiian, Latino, Japanese American, White, and other. Participants were asked to indicate their sex as male or female: thus, the MEC was not designed to delineate persons of trans or nonbinary gender. We refer to male participants as men and female participants as women in this report, solely because of the limited space.

The cohort has been linked to the Centers for Medicare & Medicaid Services (CMS) claims data for 1999–2016¹⁶ for identifying chronic disease outcomes including ADRD based on the International Classification of Diseases codes included in the claims, in addition to linkage to Hawaii and California death files and the National Death Index for vital status. We restricted the current late-onset ADRD analysis to Medicare Fee-for-Service beneficiaries with available claims data (n = 123,196) and excluded participants who were not from the 5 targeted racial and ethnic groups (n = 7,511) or individuals whose claims started before age 64 years (n = 4,707). Following the best-practice recommendations to capture valid and incident cases of ADRD,¹⁷ we excluded individuals who were enrolled in Medicare for less than 2 years (n = 3,066), those with a disease claim within the first 2 years of the Medicare linkage, as of 1999 or later coverage start (n = 1,610), and those who reported AD on the MEC questionnaire (1998–2002; n = 222): we retained 106,080 eligible participants for analysis.

Standard Protocol Approvals, Registrations, and Patient Consents

The Institutional Review Boards at the University of Hawaii (CHS 9575) and the University of Southern California (HS-17-00714) approved the MEC study protocol and agreed that implicit consent was granted by participants who returned a completed baseline questionnaire. We adhere to the Data Use Agreement with CMS on our Limited Data Set use for research.

Exposures to Risk Factors

At cohort baseline, participants provided detailed information on the questionnaire regarding their education, marital status, disease history and medications, smoking history, physical activity, sleep duration, and body weight and height. Habitual food consumption was assessed using a quantitative food frequency questionnaire developed and validated specifically for a multiethnic population,¹⁸ from which the Healthy Eating Index (HEI-2015) score was calculated for all participants to determine their overall diet quality based on adherence to the Dietary Guidelines for Americans.¹⁹ An established composite measure of neighborhood socioeconomic status (nSES) was estimated for all participants based on their baseline residential address, geocoded and linked to the 1990 US Census block groups for Hawaii and California and the Census data on education, housing, employment, occupation, income, and poverty.^20,21 We modeled the nSES values as state-specific quintiles.

Based on the previous PAF studies^7,8,22 and the associations observed in the MEC, we selected for our PAF estimations 12 potentially modifiable risk factors assessed at cohort baseline that coincided with midlife for most participants (median age: 59 years, interquartile range: 53–66). As in the Lancet report, we included the following 7 exposures: less education (≤12th grade), history of hypertension, current smoking, obesity (body mass index ≥30 kg/m²), physical inactivity (<30 min/d of moderate or vigorous activity), history of diabetes, and being unmarried (separated, divorced, widowed, or never married) as a proxy for less social contact. We did not include alcohol consumption because it was not associated with ADRD risk in the MEC (data not shown) and did not consider hearing impairment, depression, traumatic brain injury, or air pollution because the data were not available consistently in MEC. In addition, based on the accumulating evidence, we included a history of stroke or heart disease, short or long sleep duration (≤5 or ≥9 h/d), and low-quality diet (HEI-2015 <60 points).^23-26 We also considered low nSES (defined as the lower 3 state-specific quintiles) as an important risk factor for ADRD beyond the effect of low educational attainment and a modifiable factor by policies at the local and national level.^27-29 We examined other putative risk factors, including sitting time and nonsteroidal anti-inflammatory drug use, as well as alcohol, but did not include them because they were not associated with ADRD risk. Details on how each risk factor was defined in the MEC in comparison with the recent 2 US studies can be found in eTable 1 (links.lww.com/WNL/D343).

Outcome Ascertainment

As detailed previously,¹² we identified ADRD cases during the Medicare linkage period based on the International Classification of Disease code in the claims data. We defined ADRD as AD or related dementia based on the definition by Medicare¹⁷ and Goodman et al.³⁰ and obtained common ADRD subtypes¹: degenerative dementia not otherwise specified (NOS), AD only, AD of mixed etiology, vascular dementia (VD) only, Lewy body dementia (LBD) only, and a small number (n = 43) of frontotemporal dementia (eMethods, links.lww.com/WNL/D343).

Statistical Analysis

We estimated the PAF, the proportion of ADRD that would be reduced if risk factors were removed; the PAF combines information on the prevalence of each risk factor and its relative risk (RR) in association with ADRD—that is, a higher PAF for a risk factor that is prevalent and/or has a strong association with the disease. We used the method by Laaksonen et al.^31,32 to estimate the PAF within the context of censored time-to-event cohort data, where the individual survival time is assumed to follow a parametric proportional hazards model with piecewise constant hazards.³³ This approach allows for the estimation of RRs over the entire follow-up by birth cohort categories (as opposed to the instantaneous time of the Cox model) and also support adjustment for competing risk of death. Maximum likelihood estimation was used to obtain the RR and the 95% CI.³¹

The hazards regression models, separately for men and women initially, were constructed as reported previously.^12,34 The follow-up began 2 years after January 1, 1999, for existing Medicare members at the time or 2 years after Medicare enrollment for newer members and ended at the earliest of the claim date for ADRD, the date of death, or the censor date (December 31, 2016). The models included the 12 selected risk factors and also the age at cohort entry when the risk factors were assessed, age at ADRD follow-up start on Medicare to adjust for the variation in enrollment age, race and ethnicity, and Medicare usage defined by the average number of inpatient claims (<1 or ≥1 per year) and outpatient claims (<1 or ≥1 per year). The PAFs ignoring censoring because of death will be biased as any modification of shared risk factors would also be expected to influence death from competing, non-ADRD causes.³² To estimate the extent of this bias, we additionally present PAFs using the risk factor associations with ADRD from models that accounted for the competing risk of death,¹² which adjusted for censoring because of death from competing causes.³²

Because the RRs and PAFs were similar between men and women, we performed subgroup comparisons in sex-combined data with adjustment for sex. We used general linear models to compare the PAFs across subgroups defined by race and ethnicity, age at diagnosis (65–74, 75–84, 85 years and older), or common subtypes of ADRD (AD-only, AD-mixed [with other subtypes], VD-only, LBD-only, and NOS). In addition, in the subset data with genotype information, we estimated the PAF for APOE by adding the number of ε4 risk alleles to the regression models above and also examined the modifiable PAF stratified by the ε4 carrier status.

All analyses were conducted using SAS, version 9.4 (SAS Institute Inc., Cary, NC).

Data Availability

The ADRD outcome data were obtained under a data use agreement with CMS that explicitly prohibits data sharing. The MEC baseline data are shared in dbGaP (accession number: phs002183.v1.p1).

Results

We excluded 14,199 of the 106,080 eligible participants for having any missing data for 1 or more of the 12 risk factors using a complete-case approach, which resulted in 91,881 for the main analysis: we also present the results for all participants using imputed data, as discussed below and included in a supplement. The participants with complete data were more likely to be younger (59.3 vs 60.6 years at cohort entry), male (45.0% vs 39.1%), and Japanese American (33.8% vs 20.6%) or White (28.0% vs 18.9%), compared with those with incomplete data (eTable 2, links.lww.com/WNL/D343). During an average follow-up of 9.3 years, 16,507 incident cases of ADRD were identified among 91,881 participants. ADRD cases were older than noncases at cohort entry and ADRD follow-up start (Table 1). Compared with noncases, ADRD cases were more likely to be African American, use more Medicare services over the follow-up, and be less educated. At cohort baseline, ADRD cases were also more likely to reside in lower nSES neighborhoods and report being unmarried, a history of cardiometabolic disease, lower physical activity, and short or long sleep duration, while less likely to report smoking, being obese, or consuming a low-quality diet.

Table 1.

Distribution of Risk Factors in ADRD Cases and Noncases in the Multiethnic Cohort Study

Risk factors	Men (n = 41,343)		Women (n = 50,538)
	Non-ADRD (n = 34,521)	ADRD (n = 6,822)	Non-ADRD (n = 40,853)	ADRD (n = 9,685)
Age at cohort baseline
Mean (SD)	58.2 (8.1)	66.1 (6.2)	57.5 (8.0)	66.3 (6.0)
Interquartile range	51–64	62–71	51–63	62–71
Age at ADRD follow-up start, mean (SD)	69.0 (4.4)	73.4 (5.0)	68.6 (4.2)	73.6 (4.9)
Age at ADRD claim, mean (SD)		82.0 (6.2)		82.5 (6.2)
Race and ethnicity, n (%)
African American	3,010 (7.7)	900 (11.4)	5,452 (11.5)	1,829 (15.4)
Japanese American	11,934 (30.6)	2,399 (30.5)	13,363 (28.2)	3,331 (28.1)
Latinoa	7,075 (18.1)	1,281 (16.3)	7,282 (15.4)	1,558 (13.1)
Native Hawaiian	2,515 (6.4)	434 (5.5)	3,279 (6.9)	539 (4.5)
White	9,987 (25.6)	1,808 (23.0)	11,477 (24.3)	2,428 (20.5)
Medicare use over follow-up, n (%)
In-patient service (≥1 per year)	1,073 (2.7)	475 (6.0)	1,168 (2.5)	598 (5.0)
Out-patient service (≥1 per year)	16,329 (41.8)	4,704 (59.7)	24,074 (50.9)	7,352 (62.0)
Education, n (%)
≥Vocational school/some college	22,749 (58.3)	3,753 (47.7)	25,098 (53.0)	4,640 (39.1)
9th–12th grade	8,771 (22.5)	2,303 (29.2)	12,435 (26.3)	3,912 (33.0)
≤8th grade	3,001 (7.7)	766 (9.7)	3,320 (7.0)	1,133 (9.6)
nSES at cohort baseline, n (%)
Quintile 4 and 5 (high)	17,450 (44.7)	3,277 (41.6)	19,529 (41.3)	4,177 (35.2)
Quintile 1–3 (low)	17,071 (43.7)	3,545 (45.0)	21,324 (45.1)	5,508 (46.5)
Marital status at baseline, n (%)
Married	27,286 (69.9)	5,393 (68.5)	25,868 (54.7)	5,341 (45.0)
Not marriedb	7,235 (18.5)	1,429 (18.1)	14,985 (31.7)	4,344 (36.6)
Disease history at baseline, n (%)
Hypertension	12,521 (32.1)	3,051 (38.7)	13,540 (28.6)	4,462 (37.6)
Stroke	644 (1.7)	264 (3.4)	573 (1.2)	306 (2.6)
Diabetes	3,206 (8.2)	967 (12.3)	3,158 (6.7)	1,294 (10.9)
Heart disease	2,818 (7.2)	811 (10.3)	1,946 (4.1)	843 (7.1)
Current smoking at baseline, n (%)	5,685 (14.6)	904 (11.5)	5,542 (11.7)	1,027 (8.7)
Physical inactivity (<30 minc/d) at baseline, n (%)	10,505 (26.9)	2,456 (31.2)	15,927 (33.7)	4,173 (35.2)
Sleep duration at baseline, n (%)
6–8 h/d	29,104 (74.6)	5,533 (70.3)	34,015 (71.9)	7,692 (64.9)
≤5 or ≥9 h/d	5,417 (13.9)	1,289 (16.4)	6,838 (14.4)	1,993 (16.8)
Obesity (BMI ≥30 kg/m2) at baseline, n (%)	5,659 (14.5)	970 (12.3)	7,977 (16.9)	1,840 (15.5)
Low-quality diet (HEI-2015 <60) at baseline, n (%)	10,638 (27.3)	1,616 (20.5)	8,241 (17.4)	1,436 (12.1)
No. of APOE ε4 alleles, n (%)
0	4,690 (12.0)	855 (10.9)	4,593 (9.7)	789 (6.7)
1	1,437 (3.7)	394 (5.0)	1,470 (3.1)	410 (3.5)
2	90 (0.2)	47 (0.6)	128 (0.3)	49 (0.4)
Not available	28,304 (72.5)	5,526 (70.2)	34,662 (73.2)	8,437 (71.2)

Abbreviations: ADRD = Alzheimer disease and related dementia; BMI = body mass index; HEI-2015 = Healthy Eating Index-2015; nSES = neighborhood socioeconomic status.

^aMost of the MEC Latino individuals were of Mexican descent.

^bSeparated, divorced, widowed, or never married.

^cModerate or vigorous physical activity.

Table 2 shows the PAF for 12 nongenetic risk factors at cohort baseline that were significantly associated with ADRD risk in the mutually adjusted models. The PAF for individual risk factors ranged up to 3.6% (less education) in men and 3.9% (low nSES) in women. Similarly high PAFs were observed for not being married, having a history of hypertension or diabetes, and reporting lower physical activity in both sexes. The joint PAF (95% CI) for all 12 risk factors was similar in men (24.0%) and women (22.8%) (p for sex difference >0.5), for a PAF of 23.4% (95% CI 21.6–25.2) in the merged data. The joint PAF was reduced when accounting for competing risk of death because of likely different risk profiles for ADRD between MEC participants who were observed and those who had died, more so in men (to 10.4%) than in women (to 13.9%). The competing risk-adjusted PAF remained significantly positive for several individual risk factors (less education, not being married, diabetes, and physical inactivity), while it was negative for heart disease in men, indicating that preventing or delaying heart disease at midlife would likely lower the competing risk of death from it and contribute to higher ADRD incidence at older ages.³¹

Table 2.

PAF of ADRD for Modifiable Risk Factors in the Multiethnic Cohort Study

	Prevalence, %	ADRD, n	Competing event,a n	RR (95% CI)b	PAF% (95% CI)	PAF% (95% CI)c
Men (n = 41,343)
Less education (≤12th grade)	35.9	3,069	4,970	1.14 (1.08–1.20)	3.6 (2.1 to 5.1)	2.2 (0.4 to 3.9)
Low nSES (quintile 1–3)	49.9	3,545	6,029	1.10 (1.04–1.16)	3.4 (1.5 to 5.2)	1.6 (−0.6 to 3.7)
Not married	21.0	1,429	2,582	1.20 (1.13–1.27)	2.7 (1.8 to 3.6)	1.3 (0.3 to 2.4)
History of hypertension	37.7	3,051	5,245	1.11 (1.05–1.16)	3.0 (1.5 to 4.4)	0.0 (−1.7 to 1.7)
History of stroke	2.2	264	444	1.43 (1.27–1.63)	0.7 (0.5 to 1.0)	0.3 (0.0 to 0.6)
History of diabetes	10.1	967	1,804	1.43 (1.33–1.53)	3.1 (2.5 to 3.7)	0.8 (0.0 to 1.5)
History of heart disease	8.8	811	1,677	1.06 (0.98–1.14)	0.4 (−0.2 to 1.0)	−1.3 (−2.0 to −0.7)
Current smoking	15.9	904	2,346	1.34 (1.25–1.44)	3.2 (2.4 to 4.1)	−0.3 (−1.2 to 0.6)
Physical inactivity (<30 min/d)	31.3	2,456	4,119	1.15 (1.09–1.21)	3.2 (2.0 to 4.4)	1.7 (0.2 to 3.1)
Short/long sleep (≤5 or ≥9 h/d)	16.2	1,289	2,195	1.07 (1.00–1.13)	0.8 (0.0 to 1.6)	0.3 (−0.6 to 1.2)
Obesity (BMI ≥30)	16.0	970	2,008	1.10 (1.03–1.19)	1.1 (0.3 to 1.9)	−0.1 (−1.0 to 0.8)
Low-quality diet (HEI-2015 <60)	29.6	1,616	3,079	1.02 (0.96–1.08)	0.4 (−0.7 to 1.5)	−0.4 (−1.7 to 0.9)
All combined					24.0 (21.3 to 26.6)d	10.4 (7.1 to 13.5)e
Women (n = 50,538)
Less education (≤12th grade)	41.2	5,045	4,789	1.08 (1.04–1.13)	2.6 (1.2 to 4.0)	1.9 (0.3 to 3.5)
Low nSES (quintile 1–3)	53.1	5,508	5,586	1.11 (1.06–1.16)	3.9 (2.2 to 5.5)	2.4 (0.5 to 4.3)
Not married	38.2	4,344	4,327	1.11 (1.07–1.16)	3.1 (1.9 to 4.3)	2.1 (0.7 to 3.5)
History of hypertension	35.6	4,462	4,560	1.09 (1.05–1.14)	2.5 (1.3 to 3.7)	0.0 (−1.4 to 1.4)
History of stroke	1.7	306	313	1.46 (1.30–1.64)	0.6 (0.4 to 0.8)	0.4 (0.1 to 0.6)
History of diabetes	8.8	1,294	1,495	1.48 (1.39–1.57)	3.0 (2.5 to 3.5)	1.2 (0.7 to 1.8)
History of heart disease	5.5	843	977	1.10 (1.03–1.19)	0.5 (0.1 to 0.9)	−0.2 (−0.6 to 0.3)
Current smoking	13.0	1,027	1,811	1.27 (1.19–1.36)	2.2 (1.5 to 2.8)	−0.1 (−0.8 to 0.5)
Physical inactivity (<30 min/d)	39.8	4,173	4,122	1.09 (1.04–1.13)	2.4 (1.2 to 3.6)	1.6 (0.2 to 3.0)
Short/long sleep (≤5 or ≥9 h/d)	17.5	1,993	1,973	1.12 (1.06–1.18)	1.5 (0.8 to 2.2)	1.2 (0.5 to 2.0)
Obesity (BMI ≥30)	19.4	1,840	2,271	1.06 (1.01–1.12)	0.9 (0.1 to 1.6)	−0.0 (−0.9 to 0.8)
Low-quality diet (HEI-2015 <60)	19.1	1,436	1,732	1.09 (1.03–1.16)	1.1 (0.4 to 1.8)	0.6 (−0.2 to 1.4)
All combined					22.8 (20.3 to 25.2)	13.9 (11.0 to 16.8)

Abbreviations: ADRD = Alzheimer disease and related dementia; BMI = body mass index; HEI-2015 = Healthy Eating Index-2015; nSES = neighborhood socioeconomic status; PAF = population-attributable fraction.

^aDeaths from other causes than ADRD.

^bThe models included age at cohort entry, age at start of Medicare surveillance, sex, race and ethnicity, Medicare usage, and all 12 risk factors.

^cAccounting for competing risk of death.

^dp for difference by sex = 0.6299.

^ep for difference by sex = 0.1510.

In the comparison across the 5 racial and ethnic subgroups, the joint PAF varied 2-fold (p for heterogeneity <0.0001) (Table 3), highest in the Latino group (33.3%) followed by the Native Hawaiian (29.3%) and African American (27.8%) groups, while lower in the White group (21.9%) and lowest in the Japanese American group (14.2%). When accounting for competing risk, the corresponding PAFs were 25.5%, 13.7%, 13.5%, 10.9%, and 4.7%, showing a 5-fold difference by race and ethnicity. Individual risk factors with the largest PAF in Latino (diabetes 6.8%), African American (low nSES, 5.5%), White (low nSES, 4.0%), Native Hawaiian (low nSES, 5.2%), and Japanese American (physical inactivity 3.6%) groups generally coincided with the most prevalent risk factors in each group (eTable 3, links.lww.com/WNL/D343).

Table 3.

PAF of ADRD for Modifiable Risk Factors by Race and Ethnicity in the Multiethnic Cohort Study

Risk factors	African American	Japanese American	Latino	Native Hawaiian	White
	PAF% (95% CI)a	PAF% (95% CI)a	PAF% (95% CI)a	PAF% (95% CI)a	PAF% (95% CI)a
Total no. 91,881	11,191	31,027	17,196	6,767	25,700
ADRD case no. 16,507	2,729	5,730	2,839	973	4,236
Age at diagnosis	82.1 ± 6.3	83.5 ± 5.8	81.1 ± 6.1	80.1 ± 6.3	82.1 ± 6.2
Less education (≤12th grade)	3.4 (1.4 to 5.5)	1.3 (−0.5 to 3.0)	6.1 (1.7 to 10.3)	4.1 (−1.4 to 9.2)	1.8 (0.5 to 3.1)
Low nSES (quintile 1–3)	5.5 (0.8 to 10.0)	1.0 (−0.6 to 2.5)	5.7 (0.5 to 10.6)	5.2 (0.0 to 10.0)	4.0 (2.2 to 5.8)
Not married	5.1 (2.2 to 7.9)	1.0 (0.0 to 2.0)	4.9 (2.7 to 7.0)	3.4 (0.3 to 6.4)	2.8 (1.3 to 4.3)
History of hypertension	4.4 (1.4 to 7.3)	1.4 (−0.2 to 3.1)	3.7 (1.5 to 5.9)	3.4 (−1.2 to 7.8)	2.7 (1.2 to 4.1)
History of stroke	1.2 (0.8 to 1.6)	0.6 (0.3 to 0.9)	0.9 (0.5 to 1.3)	0.7 (0.0 to 1.3)	0.3 (0.0 to 0.6)
History of diabetes	4.0 (3.0 to 4.9)	2.3 (1.7 to 3.0)	6.8 (5.6 to 8.1)	2.9 (1.1 to 4.8)	1.3 (0.8 to 1.8)
History of heart disease	0.2 (−0.6 to 1.1)	0.2 (−0.3 to 0.7)	0.8 (−0.2 to 1.7)	0.3 (−1.0 to 1.6)	0.9 (0.2 to 1.5)
Current smoking	3.3 (1.9 to 4.6)	1.5 (0.8 to 2.2)	2.2 (1.1 to 3.4)	3.8 (1.0 to 6.5)	3.3 (2.3 to 4.3)
Physical inactivity (<30 min/d)	−0.1 (−2.5 to 2.3)	3.6 (2.2 to 4.9)	2.1 (−0.8 to 4.9)	4.3 (1.1 to 7.3)	2.7 (1.3 to 4.0)
Short/long sleep (≤5 or ≥9 h/d)	0.6 (−0.8 to 2.0)	1.1 (0.4 to 1.9)	1.8 (0.4 to 3.3)	2.0 (−0.5 to 4.5)	0.8 (−0.2 to 1.7)
Obesity (BMI ≥30)	1.0 (−0.8 to 2.7)	0.2 (−0.4 to 0.7)	0.7 (−1.0 to 2.3)	1.2 (−2.5 to 4.7)	1.5 (0.4 to 2.5)
Low-quality diet (HEI-2015 <60)	0.6 (−0.6 to 1.7)	0.5 (−0.6 to 1.6)	1.3 (−0.4 to 3.0)	1.2 (−2.3 to 4.5)	1.4 (0.3 to 2.5)
Joint PAF	27.8 (22.3 to 33.0)c	14.2 (11.1 to 17.2)	33.3 (27.5 to 38.5)	29.3 (21.0 to 36.7)	21.9 (19.0 to 24.7)
Joint PAF adjusted for competing riskb	13.5 (6.6 to 19.9)d	4.7 (1.1 to 8.2)	25.5 (18.9 to 31.5)	13.7 (3.6 to 22.8)	10.9 (7.5 to 14.1)

Abbreviations: ADRD = Alzheimer disease and related dementia; BMI = body mass index; HEI-2015 = Healthy Eating Index-2015; nSES = neighborhood socioeconomic status; PAF = population-attributable fraction.

^aThe models included age at cohort entry, age at start of Medicare surveillance, sex, race and ethnicity, Medicare usage, and all 12 risk factors.

^bAccounting for competing risk of death.

^cp for difference across racial and ethnic groups <0.0001.

^dp for difference across racial and ethnic groups <0.0001.

Comparing by age at diagnosis (Table 4), the PAFs for individual risk factors were generally higher among cases diagnosed at younger vs older ages, with a joint PAF of 43.2% for diagnoses at age 65–74 years compared with 32.4% for 75–84 years or 11.3% for 85 years and older. These findings suggest that the modifiable risk factors contribute less to ADRD in older elderly than in younger elderly. Accounting for competing risk of death did not meaningfully affect the PAF of ADRD diagnoses before age 75 (40.7%) but reduced the preventable proportion of diagnoses at 75–84 years (21.6%) and diagnoses at older ages (−14.0%). These results indicate that removing the modifiable risk factors at midlife would prevent earlier deaths from other competing causes and lead to an increase in ADRD at age 85 years and older.

Table 4.

PAF of ADRD for Modifiable Risk Factors by Age at Diagnosis in the Multiethnic Cohort Study

Risk factors	Age at ADRD diagnosis
	65–74 y (1,958 cases)		75–84 y (8,250 cases)		85 y and older (6,299 cases)
	RR (95% CI)a	PAF% (95% CI)	RR (95% CI)a	PAF% (95% CI)	RR (95% CI)a	PAF% (95% CI)
Less education (≤12th grade)	1.16 (1.05–1.28)	5.4 (1.6 to 9.1)	1.13 (1.08–1.18)	4.8 (2.9 to 6.7)	1.05 (1.00–1.11)	1.7 (−0.0 to 3.4)
Low nSES (quintile 1–3)	1.09 (0.98–1.20)	4.4 (−1.0 to 9.5)	1.12 (1.07–1.18)	5.4 (3.2 to 7.5)	1.09 (1.04–1.15)	3.0 (1.2 to 4.7)
Not married	1.30 (1.18–1.44)	8.7 (5.4 to 11.9)	1.13 (1.08–1.19)	3.6 (2.1 to 5.0)	1.04 (0.98–1.10)	0.8 (−0.5 to 2.0)
History of hypertension	1.15 (1.05–1.27)	5.4 (1.7 to 8.9)	1.13 (1.08–1.18)	4.6 (2.8 to 6.4)	1.05 (1.00–1.11)	1.5 (−0.0 to 3.0)
History of stroke	1.42 (1.08–1.86)	0.8 (0.1 to 1.6)	1.55 (1.38–1.73)	1.4 (1.0 to 1.8)	1.35 (1.17–1.56)	0.7 (0.4 to 1.0)
History of diabetes	1.93 (1.70–2.18)	8.5 (6.6 to 10.3)	1.56 (1.46–1.65)	5.3 (4.5 to 6.1)	1.18 (1.09–1.29)	1.4 (0.7 to 2.0)
History of heart disease	1.24 (1.06–1.45)	1.8 (0.4 to 3.2)	1.13 (1.05–1.21)	1.2 (0.5 to 1.9)	1.01 (0.92–1.10)	0.1 (−0.6 to 0.7)
Current smoking	1.36 (1.22–1.53)	5.8 (3.5 to 8.0)	1.32 (1.23–1.41)	3.2 (2.4 to 4.0)	1.18 (1.07–1.30)	1.1 (0.4 to 1.8)
Physical inactivity (<30 min/d)	1.18 (1.08–1.30)	6.3 (2.7 to 9.8)	1.15 (1.09–1.20)	4.7 (3.1 to 6.2)	1.03 (0.98–1.09)	0.7 (−0.6 to 2.1)
Short/long sleep (≤5 or ≥9 h/d)	1.14 (1.02–1.28)	2.6 (0.4 to 4.8)	1.10 (1.04–1.16)	1.6 (0.7 to 2.6)	1.07 (1.01–1.14)	0.9 (0.1 to 1.7)
Obesity (BMI ≥30)	1.11 (1.00–1.24)	2.7 (−0.1 to 5.4)	1.05 (0.98–1.11)	0.7 (−0.3 to 1.6)	1.06 (0.97–1.14)	0.5 (−0.3 to 1.3)
Low-quality diet (HEI-2015 <60)	1.18 (1.06–1.30)	4.4 (1.5 to 7.2)	1.07 (1.02–1.14)	1.2 (0.2 to 2.2)	0.94 (0.87–1.01)	−0.7 (−1.4 to 0.1)
Joint PAF		43.2 (38.0 to 48.0)c		32.4 (29.6 to 35.0)		11.3 (8.1 to 14.4)
Joint PAF adjusted for competing riskb		40.7 (35.3 to 45.6)d		21.6 (18.4 to 24.7)		−14.0 (−18.2 to −9.9)

Abbreviations: ADRD = Alzheimer disease and related dementia; BMI = body mass index; HEI-2015 = Healthy Eating Index-2015; nSES = neighborhood socioeconomic status; PAF = population-attributable fraction.

^aThe models included age at cohort entry, age at start of Medicare surveillance, sex, race and ethnicity, Medicare usage, and all 12 risk factors.

^bAccounting for competing risk of death.

^cp for difference from overall PAF (23.4%) <0.0001 for all age categories.

^dp for difference from overall PAF (12.5%) <0.0001 for all age categories.

We repeated the analyses using multiple imputation for missing data on any of the 12 risk factors with 5 iterations, assuming missingness completely at random, and conditional on age, sex, and race and ethnicity. PAFs for subgroups by sex, race and ethnicity, and diagnosis age based on the imputed data (total n = 106,080) were similar to those from the complete-case approach (eTable 4, links.lww.com/WNL/D343).

Some differences in the PAF were also suggested by common ADRD subtypes (eTable 5, links.lww.com/WNL/D343). The joint PAF was higher for VD-only (40.9%) and NOS degenerative dementia (37.1%) compared with AD-only (17.9%), AD of mixed etiology (17.0%), or LBD (14.3%). The PAF for education or nSES was higher for AD-only dementia and NOS than others, while PAFs for history of cardiometabolic conditions were greater for VD-only and NOS, reflecting differences in the strength of the risk factor associations (eTable 6).

For a subset of 14,952 participants, genotype data were available from previous genome-wide association studies (GWASs) in the MEC (eMethods, links.lww.com/WNL/D343), from which we extracted APOE genotype information. The demographic characteristics, the distribution of risk factors, and age at diagnosis (overall or race-specific) in this GWAS subset were representative of the larger analysis data.¹² In this subset (Figure, eTable 7), the joint PAF for the 12 nongenetic factors (20.4%) was higher than the PAF for carrying the APOE ε4 allele (11.8%): this trend was similarly observed in African American (22.7% vs 11.5%), Latino (36.6% vs 11.2%), and White (21.7% vs 12.4%) groups, but not in the Japanese American (9.2% vs 10.3%) or Native Hawaiian (11.9% vs 14.2%) group. The nongenetic risk factors and APOE ε4 together accounted for 30.6% of ADRD cases with a range of 18.9% (Japanese American) to 44.9% (Latino) across the racial and ethnic groups. As expected, APOE ε4 accounted for a greater proportion of AD-only (21.5%) or AD of mixed etiology (25.8%) cases than of other subtypes (eTable 8). When the joint modifiable PAF of ADRD was examined stratified by the APOE ε4 carrier status (eTable 9), we found that the preventable proportion of cases was greater among noncarriers (25.0%) than carriers (12.7%) (p for difference = 0.047).

Figure. PAF of ADRD for Modifiable Risk Factors and Genetic Risk Because of APOE ε4 by Race and Ethnicity in the Multiethnic Cohort Data Subset With Genotype Information (n = 14,952).

The hazards regression models for PAF estimation included 12 nongenetic risk factors and the number of the APOE ε4 allele and adjusted for age at cohort entry, age at start of Medicare follow-up for ADRD, sex, race and ethnicity, and Medicare usage. ADRD = Alzheimer disease and related dementia; PAF = population-attributable fraction.

Discussion

In this large, multiethnic cohort of Hawaii and California populations, we determined that modifiable risk factors assessed at midlife accounted together for 23% of the late-onset ADRD cases that developed during follow-up. This preventable proportion of ADRD was similar by sex but varied substantially across the 5 racial and ethnic groups examined (14%–33%). As expected, because late-onset ADRD shares most of its risk factors with other metabolic diseases, accounting for competing risk of death reduced the PAF of ADRD from the above-mentioned levels to 10% in men and 14% in women and to 5%–26% across the racial and ethnic groups. We also observed that the preventable fraction was greater for cases diagnosed at younger (65–74 years, 43%) vs older ages (75–84 years, 32%; 85 years and older, 11%) and for VD (41%) and NOS ADRD (37%) vs AD or LBD (14%–18%). When the APOE genotype was considered additionally, the number of ε4 risk allele accounted for half the proportion of cases (12%) as those explained by nongenetic risk factors and for relatively consistent proportions (10%–14%) across the racial and ethnic groups.

Previous studies reported a wide range of PAF estimates for ADRD or dementia, between 10% and 66%,^35-37 likely reflecting the variation in the number, definition, assessment method, and prevalence of the risk factors examined, as well as the variation in the strength of the risk factor-ADRD/dementia associations, across heterogeneous study populations. As noted in the Introduction, the latest and most comprehensive PAF estimates both for dementia worldwide⁷ and for ADRD⁸ or dementia⁹ in the United States have been around 40%. Our estimate was lower in part because of the discrepancy in the risk factors examined and the prevalence of commonly examined risk factors. We did not include 5 of the 12 risk factors suggested by the Lancet Commission (hearing impairment, depression, traumatic brain injury, air pollution, and excessive alcohol use).⁹ We instead included low nSES, a history of stroke or heart disease, short or long sleep duration, and a low-quality diet, which were independently associated with ADRD in mutually adjusted models in the MEC. In addition, some of the 7 commonly examined risk factors showed a lower prevalence in MEC as of the 1990s than in the cross-sectional surveys of more contemporary adult populations in the 2010s as included in the other studies, which likely affected the PAF: for example, obesity was prevalent in 18% of the MEC vs 36% in Nianogo et al.⁸ and 44% in Lee et al.⁹; physical inactivity in 32% vs 24% and 63%; diabetes in 10% vs 11% and 29%; and hypertension in 37% vs 37% and 42%.

In addition, differences in the methodological approach likely contributed to the inconsistent PAFs. In the MEC, we could directly estimate the PAFs based on the prevalence of risk factors and the strength of their association with incident ADRD in a cohort of same individuals, while adjusting each risk factor-ADRD association for confounding by other risk factors. By contrast, the other studies^7-9 adopted the approach by Norton et al.³⁸ to overcome the lack of prospective data to simultaneously adjust for the frequently co-occurring risk factors and presumably to project future impact. Specifically, the unadjusted combined PAF was weighted by communality (each risk factor's variance shared with other factors) and uniqueness (defined as 1 minus the communality) for each risk factor. There were some differences in the communality estimates between the 2 US studies, despite both using nationally representative surveys from similar periods.^8,9

Nonetheless, as in the previous US studies,^8,9 we similarly found that the modifiable proportion of ADRD cases varied by race and ethnicity. The PAF was greater for Latino (33.3%), Native Hawaiian (29.3%), and African American (27.8%) groups than for White (21.9%) and Asian American (14.2%) groups. We also similarly observed that the top contributing risk factors varied by race and ethnicity. In the MEC overall, low educational attainment was the leading risk factor in men (3.6%), followed closely by low nSES (3.4%), and ranked high in women (2.6%) after low nSES (3.9%), not being married (3.1%) as a proxy for low social contact, and a history of diabetes (3.0%). The importance of low education and nSES was most prominent for ADRD cases of Native Hawaiian descent as the first and second contributor, respectively. Low education and nSES also ranked high for ADRD cases of African American, Latino, and White race and ethnicity. Conversely, physical inactivity and diabetes were most contributory factors in the Asian American group, as observed previously.⁸ Notably, we also found that the modifiable PAF is lower for ADRD cases diagnosed among the oldest-old over the age of 85. A large proportion of ADRD diagnoses in this age group is reported to be limbic-predominant age-related TDP-43 encephalopathy dementia.³⁹ While it is unclear what is driving the lower PAF in the oldest-old cases in our data, our finding suggests that the etiology and effective intervention strategies may differ in this group.

Strengths of this study include the population-based prospective design with a long-term follow-up, a large number of participants and cases from diverse racial and ethnic backgrounds, and the availability of the APOE genotype data for a subset. In the genotyped subset, we observed that, compared with potentially modifiable risk factors, this predominant genetic risk associated with APOE ε4 was lower and more consistent across racial and ethnic groups. The comprehensive information on demographics and risk factors enabled us to adjust for various potential confounders and evaluate independent associations. In addition, the large size allowed for PAF comparisons by age at diagnosis, common ADRD subtypes, and the APOE genotype, which provided novel findings for further investigations. The regular and near-complete updating of the cohort for vital status allowed for competing risk adjustment and more direct estimation of modifiable case fractions.

Limitations of our study include that the ADRD case definition was based on Medicare claims data and likely involved some misclassification. However, studies have shown that Medicare claims are largely concordant with clinical diagnoses of dementia,¹⁷ including in Los Angeles County,⁴⁰ where the MEC California recruitment was based,^17,41 and that the accuracy has improved over time.^41-43 Medicare claims likely reflect the limitation of a large proportion of clinical dementia diagnoses that are entirely symptom-based, not Alzheimer pathology biomarker-based.⁴⁴ Another limitation is that our PAF estimates did not account for hearing impairment, depression, or history of traumatic brain injury because these data were not available in MEC at baseline, or exposures to air pollution that are more meaningful and variable for California residents, and not those in Hawaii, in MEC. This likely has led to the slightly lower PAF in our study, although our estimate included some other risk factors that were independently associated with ADRD in our population, such as low residential SES, poor quality diet, and less optimal hours of sleep. Finally, our PAF estimates, like the ones before, did not directly assess the likely critical contribution of SSDoH, other than individual education and neighborhood-level SES, to the racial and ethnic ADRD disparities, which warrants systematic collections of individual-level SSDoH data.^14,45 One of such SSDoH is access to quality health care and relevant social and structural resources, for example, to prevent and manage cardiometabolic risk factors for ADRD that are known barriers for minoritized populations.

In conclusion, our findings confirm that education and midlife exposures to risk factors account for substantial but varying proportions of late-onset ADRD cases across racial and ethnic populations. In particular, our findings call for interventions in the racial and ethnic groups with a large proportion of preventable cases and individually modifiable risk factor profiles, while they suggest the importance of discovering additional risk factors in other racial and ethnic groups.

Glossary

AD

Alzheimer disease

ADRD

Alzheimer disease and related dementia

CMS

Centers for Medicare & Medicaid Services

GWAS

genome-wide association study

HEI

Healthy Eating Index

LBD

Lewy body dementia

MEC

Multiethnic Cohort Study

NOS

degenerative dementia not otherwise specified

nSES

neighborhood socioeconomic status

PAF

population-attributable fraction

RR

relative risk

SSDoH

structural and social determinants of health

VD

vascular dementia

Appendix. Authors

Name	Location	Contribution
Song-Yi Park, PhD	Population Sciences in the Pacific Program, University of Hawaii Cancer Center, University of Hawaii at Manoa, Honolulu	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data
Veronica Wendy Setiawan, PhD	Department of Population and Public Health Sciences, Keck School of Medicine and Norris Comprehensive Cancer Center, University of Southern California, Los Angeles	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Eileen M. Crimmins, PhD	Leonard Davis School of Gerontology, Andrus Gerontology Center, University of Southern California, Los Angeles	Drafting/revision of the manuscript for content, including medical writing for content
Lon R. White, MD, MPH	Pacific Health Research and Education Institute, Honolulu, HI	Drafting/revision of the manuscript for content, including medical writing for content
Anna H. Wu, PhD	Department of Population and Public Health Sciences, Keck School of Medicine and Norris Comprehensive Cancer Center, University of Southern California, Los Angeles	Drafting/revision of the manuscript for content, including medical writing for content
Iona Cheng, PhD	Department of Epidemiology and Biostatistics, University of California, San Francisco	Drafting/revision of the manuscript for content, including medical writing for content
Burcu F. Darst, PhD	Public Health Sciences Division, Fred Hutchinson Cancer Center, Seattle, WA	Drafting/revision of the manuscript for content, including medical writing for content
Christopher A. Haiman, ScD	Department of Population and Public Health Sciences, Keck School of Medicine and Norris Comprehensive Cancer Center, University of Southern California, Los Angeles	Drafting/revision of the manuscript for content, including medical writing for content
Lynne R. Wilkens, DrPH	Population Sciences in the Pacific Program, University of Hawaii Cancer Center, University of Hawaii at Manoa, Honolulu	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Loїc Le Marchand, MD, PhD	Population Sciences in the Pacific Program, University of Hawaii Cancer Center, University of Hawaii at Manoa, Honolulu	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Unhee Lim, PhD	Population Sciences in the Pacific Program, University of Hawaii Cancer Center, University of Hawaii at Manoa, Honolulu	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data

Study Funding

This work was supported by grants from the National Cancer Institute and a supplement from the National Institute on Aging at the NIH to the Multiethnic Cohort Study (U01 CA164973; CA164973 08S1).

Disclosure

The authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.