APOE genotype and cognitive decline: educational context as a moderator of genetic risk

Abstract

Objectives

The apolipoprotein E (APOE) ε4 allele is linked to increased risk of Alzheimer’s disease and cognitive decline. While educational attainment appears to mitigate this risk, it is not a direct measure of the early-life environment. This study examined whether childhood exposure to well- versus poorly resourced educational contexts modified the association between APOE ε4 and cognitive decline, independent of parent and respondent educational attainment.

Methods

We used data from the Health and Retirement Study (2006–2018) linked with state-level data on public education systems. The sample included U.S.-born respondents who participated in the 2006–2012 genetic data collection, lived in one of the 48 contiguous states at age 10, and were born between 1914 and 1959 (n = 14,817 respondents, 76,806 person–period observations). To measure educational context, we created a standardized factor score representing state education resources, derived from indicators such as per-pupil spending, pupil–teacher ratios, and teacher salary. We used mixed effects models to estimate the interaction between APOE ε4 and state education resources, adjusting for parent and respondent education (measured in years of schooling).

Results

APOE ε4 carriers who grew up in states with low education resources experienced an accelerated rate of cognitive decline compared to carriers from states with high education resources. For noncarriers, the rate of cognitive decline was not statistically different across educational contexts.

Discussion

Ensuring access to high-quality education may serve as an effective social policy to slow cognitive decline among future cohorts of older adults at high genetic risk of dementia.

Substantial evidence links the ε4 allele in the apolipoprotein E (APOE) gene to increased risk of late-onset Alzheimer’s disease (AD) and cognitive decline via its hypothesized influence on Aβ deposition, tau tangle formation, and neuroinflammation, among other pathways (Fortea et al., 2024; Holtzman et al., 2012; Loy et al., 2014; Michaelson, 2014). Specifically, APOE ε4 carriers show a 3- to 15-fold increase in AD risk than noncarriers (Farrer et al., 1997; Holtzman et al., 2012). Currently, treatment for cognitive impairment and AD is limited; thus, identifying modifiable environmental factors that mitigate genetic susceptibility to cognitive decline and dementia is of high importance.

Educational attainment is one of the most important modifiable determinants of cognitive function and dementia (Crimmins et al., 2018; Hayward et al., 2021; Livingston et al., 2020; Walsemann, Kerr, et al., 2022). One explanation for this relationship is that education produces long-term structural changes in the brain—such as increased vascularization or enhanced neural connectivity—while another is that it fosters more flexible and resilient patterns of cognitive functioning (Stern et al., 2019). The former suggests that education builds underlying brain capacity; the latter emphasizes the brain’s ability to compensate for damage or age-related decline. Together, these mechanisms are thought to contribute to cognitive reserve, which helps individuals better withstand accumulating brain pathology as they age (Barulli & Stern, 2013; Mungas et al., 2018; Nelson et al., 2021; Stern, 2012, 2021). Given its theoretical role in developing cognitive reserve, researchers have identified educational attainment as an environmental factor that may moderate the relationship between APOE ε4 and cognitive decline. In a recent meta-analysis, 71%–79% of published studies reported that the completion of more years of schooling mitigated the association between APOE ε4 and cognitive decline and dementia onset (Aravena et al., 2024).

Although educational attainment reflects, in part, early childhood environments, it is measured at the individual level and is not, by definition, an environmental context. As such, it is likely confounded by genetics and other social processes (Boardman et al., 2013; Conley & Fletcher, 2017), including parental education, childhood economic resources, individual cognitive skills and ability, and access to high quality schooling. Moreover, attainment reflects education acquired not only in childhood but also in late adolescence and adulthood. Although brain plasticity occurs throughout life, childhood is a sensitive period for brain development (Lindenberger & Lövdén, 2019). Thus, early educational environments may play a large role in shaping genetic risk of cognitive decline and dementia independent of educational attainment.

We know of no gene-environment (G × E) association studies, however, that have examined educational context, such as per-pupil spending or resources, as a moderator of the APOE ε4–cognitive decline relationship. Educational context, shaped by state and local policies, structures life course opportunities for cognitive development and engagement (Walsemann et al., 2013). However, state and federal investment in public education varied considerably throughout much of the 20th century (Walsemann, Fisk, et al., 2023), creating disparities in school resources and learning environments across birth cohorts. Before the 1950s, for instance, school term length differed widely by state, with some providing far fewer days of instruction per year. By the late 1950s, however, school term lengths had become more standardized nationwide (Walsemann, Ureña, et al., 2022; Walsemann, Fisk, et al., 2023). Federal policies further reshaped educational opportunities, including the National Defense Education Act of 1958 and the Elementary and Secondary Education Act of 1965, which incentivized school desegregation through federal funding (Johnson, 2015; Walsemann, Fisk, et al., 2023; Walsemann, Hair, et al., 2023). While these policy shifts increased education funding, significant disparities persisted, leading to substantial state variation in students' access to high quality education.

There is growing attention to how educational contexts may shape cognitive function and decline. For example, among a nationally representative sample of midlife and older U.S. adults, growing up in states that invested more heavily in their public education system was associated with higher cognitive function in mid to late life, particularly among less educated individuals (Walsemann et al., 2024). Similarly, geographically localized studies find that more advantaged educational contexts—at the school, community, or state level—are associated with better cognitive function in older adulthood (Crowe et al., 2013; Moorman et al., 2019; Sisco et al., 2015; Soh et al., 2023).

Gene-environment association research

The diathesis-stress and social distinction models, commonly used in G × E association studies, provide two competing hypotheses for how educational contexts may alter genetic risk of dementia (Boardman et al., 2012, 2013). The diathesis-stress model posits that genetic risk of dementia (i.e., APOE ɛ4) would manifest when APOE ɛ4 carriers are exposed to disadvantaged educational contexts. Others have used the language “social trigger” model to describe how alleles associated with higher genetic risk may be activated following exposure to social disadvantage (Boardman et al., 2012). Results from G × E association studies examining educational attainment as an environmental exposure (where low education is considered a disadvantaged environment) provide preliminary support for this hypothesis (Cook & Fletcher, 2015; Fletcher et al., 2021; Frank et al., 2021; López et al., 2018; Ma et al., 2022; McArdle & Prescott, 2010; Reas et al., 2019; Wang et al., 2012). For example, a multi-center study found that, while APOE ɛ4 was inversely associated with memory retention and working memory among older Black and White adults, the association was null among those with at least a high school degree (Reas et al., 2019).

Conversely, the social distinction model posits that genetic factors would manifest within educational contexts that minimize noise—i.e., advantaged or high-quality educational contexts (Boardman et al., 2012). In these settings, a resource-rich environment maximizes cognitive benefits, creating conditions where genetic influence emerges and contributes to differences in dementia outcomes, including an increased risk associated with APOE ɛ4. While most existing evidence supports the diathesis-stress model, there are a few studies that have found suggestive evidence for the social distinction hypothesis (Aravena et al., 2024; Lee et al., 2022; Seeman et al., 2005; Xiang et al., 2021). One study found steeper declines in episodic memory among APOE ɛ4 carriers compared to noncarriers, but only among those who completed at least 8 years of schooling (Seeman et al., 2005). Using data from the Atherosclerosis Risk in Communities (ARIC) study, researchers found that the inverse association between educational attainment and incident dementia was smaller (and nonsignificant) among APOE ɛ4 carriers, suggesting that higher levels of education provide less protection for APOE ɛ4 carriers than noncarriers (Lee et al., 2022).

The purpose of our study is to test two G × E models–the diathesis-stress and social distinction models–to determine if exposure to more advantaged state educational contexts strengthens or weakens the relationship between APOE ɛ4 and cognitive decline. We estimate this G × E association in a nationally representative sample of U.S. older adults who experienced significantly different educational contexts across birth cohorts, independent from educational attainment (Walsemann, Fisk, et al., 2023). Based on prior findings, we hypothesize that more disadvantaged educational contexts will show stronger associations between APOE ɛ4 and cognitive decline, supporting the diathesis-stress models. Our study is one of the first to consider how early educational environments shape genetic risk of ­cognitive decline, independent from educational attainment.

Method

Study population

Data come from the Health and Retirement Study (HRS), a nationally representative, prospective study of U.S. adults over age 50. The HRS uses a multi-stage area probability sampling design to select age-eligible households—the primary sampling unit—from U.S. Metropolitan Statistical Areas (MSAs) and non-MSA counties, and includes oversamples of non-Hispanic Black and Hispanic adults. Starting in 1992, the HRS conducted core interviews with age-eligible respondents and their spouses approximately every 2 years. Using a steady-state design, the HRS enrolls new cohorts of 51–56-year-old adults every 6 years to ensure that the sample remains nationally representative of the U.S. population over age 50 (Sonnega et al., 2014). Beginning in 2006, the HRS collected salivary DNA from community-dwelling respondents who completed a face-to-face interview; APOE was genotyped for respondents who participated in the genetic data collection from 2006 to 2012 (Faul et al., 2021). HRS respondents who moved into nursing homes after providing genetic data are retained in the sample.

State-level data on public education systems come from two sources: (a) the 1919/20–1957/58 Biennial Surveys of Education (BSE) and (b) the 1959/60–1969/70 Statistics of State School Systems. These data were reported biennially for the 48 continental U.S. states and the District of Columbia. While state-level education data were largely complete, occasional missing values for select variables in certain states required imputation. For these cases, we used linear regression models incorporating data from adjacent years to estimate missing values. Additionally, because data were not collected in school years ending in odd numbers (e.g., 1920/21), we interpolated values using data from the two surrounding biennial years.

Our sample includes HRS respondents who participated in the 2006 to 2012 genetic data collection, were born between 1914 and 1959, and had at least one measure of cognitive function between 2006 and 2018 (N = 17,799). We further excluded respondents who were born outside of the United States (N = 2,210), could not be matched to a state where they lived at age 10 (N = 758), or who had missing race and ethnicity (N = 14) for a total sample of 14,817 respondents. We also excluded person–period observations that occurred before age 50 or with missing values of cognitive function. In total, there were 76,806 person–period observations. Item nonresponse was generally low, but 960 respondents were missing values on parental education. We used multiple imputation with chained equations to impute missing values on parental education. Our presented estimates are averaged across 20 data sets using Rubin’s Rules (Rubin, 2004).

Measures

Cognitive function

To assess cognitive function, the HRS administered the Telephone Instrument for Cognitive Status (TICS) at each interview. The TICS consists of immediate and delayed 10-word recall tests, a serial 7s subtraction test, and a counting backward test. We summed all TICS items, resulting in a total score ranging from 0 to 27, with higher scores indicating better cognitive function. TICS was administered either by phone, face-to-face, or, in 2018, web for a small subset (8%) of respondents. Per recommendations, we subtracted 1 point from the TICS score if the respondent completed their 2018 interview on the web to account for mode effects (Domingue et al., 2023).

APOE genotype

Two genetic variants (rs7412 and rs429358) contribute to 3 APOE alleles (ε2, ε3, and ε4), resulting in six potential APOE genotypes (ε2/ε2, ε2/ε3, ε2/ε4, ε3/ε3, ε3/ε4, and ε4/ε4). Genotyping was performed using TaqMan allelic discrimination SNP assays. When genotyping did not pass quality control (one or both SNPs failed), HRS imputed APOE status from previously genotyped array data using the 1000 Genomes Cosmopolitan Reference Panel (phase 3; n = 1,533). Imputations with posterior probabilities <0.8 for rs7412 (n = 20) or rs429358 (n = 76) were dropped per HRS recommendations (Faul et al., 2021). In total, 1,533 respondents had imputed APOE genotype. We classified respondents as APOE ε4 carriers or noncarriers because only 342 respondents were ε4 homozygotes.

State educational context

Following the methodology used by Walsemann and colleagues (Walsemann et al., 2024), we conducted a factor analysis to identify two factors that summarized the state educational context where respondents lived when they were aged 10. We used these factor scores, rather than individual indicators, because of the high correlation between available indicators of state educational context. Factors were estimated from seven annual indicators: pupil-teacher ratio, teacher salary, per-pupil spending, per-pupil revenue from local sources, per-pupil revenue from state sources, percent of revenue from local sources, and percent of revenue from state sources. Although we identified two factors that explained a high share (83.7%) of variation in the individual indicators, we included in this analysis only the second factor, “state-level education resources,” which captures the level of education resources a state had each year. This factor had large positive loadings of per-pupil expenditures, per-pupil revenue from local sources, and teacher salary. We excluded the first factor, “state-education funding source,” because prior studies found that this factor was not associated with cognitive function or decline (Walsemann et al., 2024), and preliminary analysis on our analytic sample also found a null association.

Educational attainment

We included respondent’s education measured as years of schooling completed (0 to 17 years).

Covariates

We included birth cohort, race/ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic, and other race/ethnicity), age (in years), gender (male or female), parent’s education measured as the highest years of schooling of either parent (0 to 17 years), Census division of residence at age 10 (New England, Mid-Atlantic, South Atlantic, East South Central, West South Central, East North Central, West North Central, or Mountain/Pacific), interview mode (face-to-face or other), and whether APOE was imputed (yes/no). We included a measure of state income inequality, averaged over the three years when respondents were 6 to 8 years old, to capture time-varying state-level factors related to education resources and cognition. This measure was linked to HRS data at age 7 using respondents' state of residence at age 10, assuming they lived in the same state three years earlier. State income inequality was defined as the income share of the top 10% relative to the bottom 90%, based on IRS tax records compiled by Sommeiller and Price (2018) for the years 1917 to 1972. For our analyses, we used the subset of data spanning 1921 to 1966 to align with when respondents were 6 to 8 years old.

Statistical analysis

We estimated three-level linear mixed models to test if the association between APOE ε4 carrier status and level and decline in cognitive function varied by state-level educational context. Mixed models accounted for the non-independence of observations—i.e., observations nested within respondents who were also nested by state of residence at age 10. Age represents time and was centered at 70, the mean age of respondents across the observation period, and divided by 10 so that coefficients reflect change in cognitive function per decade.

Our models included an indicator of APOE ε4 carrier status, state-level education resources, and—to test the G × E association—an interaction of APOE ε4 and state-level education resources. We also tested for a G × E interaction between APOE ε4 and educational attainment and for a three-way interaction with state-level education resources; neither was statistically significant (results available upon request).

To account for observed nonlinearity, we modeled state-level education resources using both linear and quadratic terms and interacted these with APOE ε4 and age. All covariates were also interacted with linear age to assess their relationship with cognitive decline. Interacting age-squared with covariates did not improve model fit and caused convergence issues; however, excluding age-squared led to poorer model fit and yielded similar G × E associations.

We used a collapsed state-level random intercept. States with more than 25 respondents were included individually; those with fewer than 25 respondents were combined with an adjacent state in the same U.S. Census division. All models applied individual-level sampling weights that accounted for the HRS sampling design, nonresponse, respondent attrition, and genetic data collection. For respondents without a genetic data weight (N = 494), we used their most recent core interview weight.

To confirm that state education resources did not predict APOE ε4 status, we regressed APOE ε4 carrier status on state education resources and covariates using weighted logistic regression, which we report in Supplementary Table 1 (Boardman et al., 2013; Conley & Fletcher, 2017).

We then tested whether cognitive function was lower, or cognitive decline steeper, among APOE ε4 carriers who grew up in states with low versus high education resources, which would support the diathesis-stress hypothesis. To aid interpretation of the interaction, we also calculated predicted cognitive function scores at 5-year intervals by APOE ε4 status, setting state education resources at low (∼20th percentile) and high (∼80th percentile) levels. These predictions incorporated linear and quadratic terms and their interactions, holding all covariates at their sample means. Model coefficients and 95% confidence intervals (CI) are shown in Table 2.

Table 2.

Coefficients and 95% CI from mixed models predicting total cognitive function by sample, the Health and Retirement Study, 2006–2018 (weighted analysis).

Variable	At age 70		Linear rate of change
	b	95% CI	b	95% CI
APOE ε4	−0.65***	−0.80, −0.50	−0.37***	−0.56, −0.18
Education resources	0.26***	0.16, 0.36	0.07	−0.06, 0.20
Education resources-squared	−0.08**	−0.12, −0.03	−0.03	−0.09, 0.04
APOE ε4 × education resources	0.04	−0.07, 0.16	0.19*	0.04, 0.33
APOE ε4 × education resources-squared	−0.04	−0.12, 0.04	−0.10*	−0.20, −0.00
Constant	15.29***	15.12, 15.46	−2.00***	−2.23, −1.78

Note. Number of respondents = 14,817; number of person–period observations = 76,806. CI = confidence intervals. Linear rate of change is estimated using age which is centered at 70 and divided by 10 such that a one-unit change in the independent variable is equivalent to a 10-year change in the dependent variable. Quadratic rate of change is included as [age−70/10]² in all models to account for the functional form but was not interacted with any model covariates. The model adjusts for cohort, race/ethnicity, gender, interview mode, APOE imputation, parent’s and respondent’s education, average state-level income inequality at age 7, a Census division fixed effect at age 10, interactions of age with all covariates, and a collapsed state random intercept.

^*** p < .001.

^** p < .01.

^* p < .05.

^‡ p < .10.

Figure 1 displays the resulting trajectories for carriers and noncarriers in states with low and high education resources, allowing us to visualize how differences in cognitive function by APOE ε4 carrier status unfold with age in each type of educational context. Predictions begin at age 60, when differences by carrier status typically emerge, and extend to age 80, beyond which sample sizes become too small to support reliable estimates.

Figure 1.

Predicted cognitive function from age 60 to 80 by APOE ε4 carrier status in states with low (top panel) and high (bottom panel) education resources.

Note. Predicted TICS scores were generated from mixed-effects regression models by varying age, APOE ε4 carrier status, and state education resources, holding all other variables at their sample means. Predictions are shown from age 60 to 80, separately for low and high resourced states (defined as the 20th and 80th percentiles of the education resource distribution). Shaded areas represent 95% CI. Noncarriers are shown with dashed lines and teal shading; APOE ε4 carriers are shown with solid lines and purple shading.

To further interpret the interaction, we compared predicted TICS scores between carriers and noncarriers in states with low and high education resources at 5-year age intervals between 60 and 80. Table 3 presents these predicted values—also shown in Figure 1—along with statistical tests of whether cognitive function differs across educational contexts, separately for carriers and noncarriers at each age interval. We also tested whether the amount of cognitive decline from age 60 to 80 differed significantly by level of education resources for carriers and noncarriers.

Table 3.

Predicted cognitive function by age, APOE ε4 carrier status, and level of state education resources.

Age	Low education resources		High education resources		Difference
	ŷ	95% CI	ŷ	95% CI	ŷlow− ŷhigh	95% CI
Carriers
60	17.2	16.9, 17.4	17.3	17.0, 17.5	−0.1	−0.4, 0.2
65	16.3	16.1, 16.5	16.6	16.4, 16.8	−0.3*	−0.1, −0.5
70	15.4	15.2, 15.6	15.9	15.7, 16.0	−0.5*	−0.3, − 0.8
75	14.3	14.0, 14.6	15.0	14.7, 15.3	−0.7*	−0.4, −1.0
80	13.2	12.7, 13.6	14.1	13.6, 14.5	−0.9*	−0.5, −1.3
Δ 60 to 80	4.1*	3.6, 4.5	3.2*	2.8, 3.6	0.8*	0.2, 1.3
Noncarriers
60	17.3	17.1, 17.5	17.6	17.5, 17.8	−0.3*	−0.6, −0.1
65	16.7	16.6, 16.9	17.1	17.0, 17.3	−0.4*	−0.6, −0.2
70	16.1	15.9, 16.2	16.5	16.4, 16.7	−0.4*	−0.6, −0.3
75	15.3	15.1, 15.5	15.8	15.6, 16.0	−0.5*	−0.7, −0.3
80	14.4	14.1, 14.8	15.0	14.7, 15.3	−0.6*	−0.9, −0.2
Δ 60 to 80	2.9	2.6, 3.2	2.7	2.4, 2.9	0.2	−0.2, 0.7

Note. CI = confidence intervals. Predicted values calculated using margins command in Stata 18. APOE ε4 carrier status is set at 0 (noncarrier) or 1 (carrier). State education resources are varied to be low (approximately 20th percentile) or high (approximately 80th percentile). Age is varied at 5-year intervals between 60 and 80. All other covariates are set at the mean.

^* p < .05.

We conducted sensitivity analyses to determine the robustness of our findings, which we report in the sensitivity analysis section and in Supplementary Material (Supplementary Tables 2–4). First, because TICS is only available for respondents able to complete it, our sample likely skews cognitively healthier, potentially underestimating cognitive decline between a respondent’s last observation and death. To address this, we assigned a TICS score of 6, the threshold for dementia, for waves between a respondent’s last observation and death, then re-­estimated the models (Supplementary Figure 1). Second, we excluded respondents with ε2/ε4 genotype, as the ε2 and ε4 alleles may have opposite effects on cognitive decline, potentially obscuring the G × E associations (Supplementary Figure 2) (Narasimhan et al., 2024). We also estimated two additional models excluding (a) respondents with less than four years of schooling due to their limited exposure to educational contexts (Supplementary Figure 3) and (b) those who completed the 2018 TICS observation via web (Supplementary Figure 4). Additionally, we excluded respondent’s educational attainment from our models, treating it as a mediator rather than a covariate, to avoid an overly conservative test of G × E associations (Supplementary Figure 5). Finally, we tested G × E associations using relative per-pupil expenditures (state expenditures/average state expenditures in a year) instead of the state education resources factor (Supplementary Figure 6).

Prior work demonstrates that Black adults were more likely to grow up in states that had lower levels of education resources, and the association between education resources and cognitive function varies by race (Walsemann et al., 2024). Our main analysis pools respondents and includes race/ethnicity as a covariate due to limited power for G × E interactions in non-Hispanic Black and Hispanic subsamples. However, to explore potential differences, we estimated supplemental models for non-Hispanic White and Black respondents (Supplementary Table 5 and Supplementary Figures 7 and 8), recognizing the exploratory nature and the constraints of statistical power, although the Hispanic sample was too small to support a separate analysis.

Results

Sample characteristics

Table 1 presents sample characteristics by APOE ε4 carrier status. APOE ε4 carriers had lower scores on cognitive function at first and last observation relative to noncarriers, but were slightly younger at both observations. Educational attainment (parent’s and own) did not vary by carrier status; however, APOE ε4 carriers grew up in states with slightly lower education resources. This association, however, is not significant after adjusting for race, birth cohort, and division of residence in a logistic regression model predicting APOE ε4 carrier status (Supplementary Table 1). Additional bivariate associations are reported in Table 1.

Table 1.

Sample characteristics by APOE ε4 carrier status, the Health and Retirement Study, 2006–2018 (weighted estimates).

Characteristics	Noncarrier Mean (SE) or N (%)	APOE ε4 carrier Mean (SE) or N (%)
Cognitive function (range 0 to 27)
First observed	16.3 (0.05)	15.8* (0.08)
Last observed	15.0 (0.05)	14.2* (0.09)
State education resources, z-score	0.36 (0.01)	0.27* (0.02)
Age, years
First observed	62.4 (0.11)	61.9* (0.17)
Last observed	71.3 (0.11)	70.6* (0.16)
Respondent’s education, years	13.3 (0.03)	13.2 (0.05)
Parent’s education, years	11.2 (0.04)	11.2 (0.07)
Female	6,703 (53%)	2,376 (53%)
Race/Ethnicity
Non-Hispanic White	8,269 (85%)	2,926 (80%)*
Non-Hispanic Black	1,610 (9%)	1,017 (15%)*
Hispanic	590 (4%)	159 (3%)*
Other	173 (2%)	73 (2%)*
Birth cohort
AHEAD	766 (6%)	234 (4%)*
CODA	1,343 (9%)	468 (8%)*
HRS	3,298 (19%)	1,302 (20%)*
War babies	1,455 (18%)	593 (20%)*
Early baby boomers	1,922 (26%)	826 (27%)*
Mid-baby boomers	1,858 (21%)	752 (21%)*
Census Division at age 10
East North Central	2,233 (23%)	812 (21%)*
East South Central	850 (7%)	371 (8%)*
Mid-Atlantic	1,641 (17%)	597 (14%)*
Mountain/Pacific	1,492 (15%)	554 (14%)*
New England	505 (6%)	198 (6%)*
South Atlantic	1,773 (14%)	778 (16%)*
West North Central	1,070 (10%)	408 (11%)*
West South Central	1,078 (7%)	457 (8%)*
State income inequality (z-score)	0.21 (0.01)	0.22 (0.01)
Sample size, unweighted	10,642	4,175

Note. AHEAD = b. <1924; CODA = b. 1924–1930; HRS = b. 1931–1941; War babies = b. 1942–1947; Early baby boomers = b. 1948–1953; Mid baby boomers = b. 1954–1959. Statistical differences within-sample by APOE carrier status determined from t-tests for continuous variables and corrected F-tests for categorical variables.

^* p < .05, two-tailed test.

Linear mixed models

In Table 2, we show select coefficients from the mixed-effect models used to calculate our predicted cognitive function trajectories for APOE ε4 carriers and noncarriers. Greater state-level education resources were associated with better cognitive function at age 70 (b = 0.26 [95% CI = 0.16, 0.36], p = .001) and the quadratic term was significant suggesting diminishing returns to additional resources (−0.08 [−0.12, −0.03], p < .01); however, there was no significant interaction between APOE ε4 carrier status and education resources or education ­resources-squared. This suggests that growing up in states with fewer education resources was associated with poorer cognitive function at age 70 for both APOE ε4 carriers and noncarriers alike.

While no significant interaction was found at age 70 between APOE ε4 carrier status and education resources, the interaction of APOE ε4 carrier status, education resources, and age was significant. Growing up in states with fewer education resources (lower z-scores) was associated with accelerated cognitive decline for APOE ε4 carriers (0.19 [0.04, 0.33], p < 0.05). The interaction between APOE ε4 carrier status and education resources-squared was also significantly associated with the rate of cognitive decline (−0.10 [−0.20, −0.00], p < 0.05). This result indicates that growing up in states with low education resources was particularly harmful to APOE ε4 carriers, while states with high education resources offered diminishing returns. For noncarriers, state-level education resources was not associated with the rate of cognitive decline.

Predicted values and 95% CI derived from the regression coefficients of linear mixed models are reported in Table 3 and plotted in Figure 1. At age 60, there was no significant difference in cognitive function among APOE ε4 carriers who grew up in states with low versus high education resources. However, by age 65, significant differences emerged (Figure 1), with carriers from states with low education resources showing lower predicted cognitive scores than their peers from states with high education resources (ŷ_low65 − ŷ_high65: −0.3 [−0.1, −0.5], p<.05). Over the two-decade span, APOE ε4 carriers who grew up in low-resourced states experienced an additional 0.8-point decline in predicted cognitive function ([0.2, 1.3], p < 0.05) compared to those from high-resourced states.

In contrast, APOE ε4 noncarriers who grew up in states with low education resources consistently had lower predicted cognitive function scores than those from high-resourced states across the plotted age rage, beginning at age 60 (ŷ_low60 − ŷ_high60: −0.3 [−0.1, −0.6], p < 0.05) and continuing through age 80 (ŷ_low80 − ŷ_high80: −0.6, [−0.2, −0.9], p<.05) (Figure 1). However, unlike carriers, noncarriers in low-resourced states did not experience significantly greater cognitive decline between ages 60 and 80 compared to their peers in high-resourced states (0.2, [−0.2, 0.7], p > .10).

Sensitivity analyses

Findings from sensitivity analyses consistently show that APOE ε4 carriers who grew up in states with low education resources had significantly accelerated cognitive decline compared with carriers who grew up in states with high education resources. The level of state education resources was not associated with the rate of cognitive decline for noncarriers. The only exception to these findings occurred when we used per-pupil expenditures, which were marginally associated with the rate of ­cognitive decline.

Race-stratified results indicated that relationships between state education resources, APOE ε4 carrier status, age, and cognitive decline held for White adults but were not significant for Black adults. The findings for Black adults, however, should be treated with caution given the small sample size, which we discuss in the limitations section.

Discussion

Our study tested two G × E models—the diathesis-stress and social distinction models—by examining whether state education resources at age 10 moderated the relationship between APOE ε4 and cognitive decline. Prior research has focused on educational attainment, an individual-level attribute, in shaping genetic risk (Aravena et al., 2024; Cook & Fletcher, 2015; Fletcher et al., 2021; Frank et al., 2021; Lee et al., 2022; López et al., 2018; Ma et al., 2022; McArdle & Prescott, 2010; Reas et al., 2019; Seeman et al., 2005; Wang et al., 2012; Xiang et al., 2021), but few studies have considered the broader educational environments in which individuals were embedded during childhood—a sensitive period for brain development (Lindenberger & Lövdén, 2019). Supporting the diathesis-stress hypothesis, our findings indicate that APOE ε4 carriers who grew up in states with low levels of education resources experienced faster cognitive decline than those from states with high levels of education resources. This G × E association remained after adjusting for both the respondent’s and their parent’s educational attainment, making this a conservative test of the G × E association.

Other studies have also documented an association between educational quality and context on cognitive function and dementia risk, but they have not considered how this exposure might modify the expression of genetic risk (Crowe et al., 2013; Moorman et al., 2019; Sisco et al., 2015; Soh et al., 2023; Walsemann et al., 2024; Walsemann, Hair, et al., 2023). Importantly, the associations reported in this literature are relatively modest. For instance, using the HRS core sample, Walsemann and colleagues (Walsemann et al., 2024) found that White high school graduates who grew up in states with low-resourced education systems scored 0.3 points lower on cognitive function than their counterparts in high-resourced states—a gap that did not widen with age because education resources were not associated with cognitive decline.

Our findings suggest that prior studies may have underestimated the role of education resources by not accounting for variation in genetic risk. Using the HRS genetic subsample, we find that the modest associations reported in earlier research may primarily reflect the experiences of noncarriers, who make up more than two-thirds of the sample. Among noncarriers, we observe small but consistent differences in cognitive function by state education resources, but no evidence of differential cognitive decline across context. In contrast, APOE ε4 carriers who grew up in states with low education resources experience markedly steeper cognitive decline than those who grew up in states with high education resources. Between ages 60 and 80, this difference amounts to an additional 0.8 points of decline, roughly equivalent to the cognitive aging expected over five years. These results suggest that education resources may play a stronger role in shaping cognitive trajectories for individuals at elevated genetic risk.

A key contribution of our study is the use of a direct measure of educational context—state-level education resources—rather than relying on individual educational attainment, which can conflate cognitive ability, family background, and access to opportunity (Walsemann et al., 2013). Because our measure captures conditions at age 10, we cannot determine whether our findings reflect a sensitive window of heightened brain plasticity or broader cumulative exposure to under-resourced educational environments. Distinguishing between these processes—sensitive period, accumulation, or both (Glymour & Manly, 2008)—is essential for understanding how educational environments interact with genetic predisposition for dementia. Future studies should incorporate data on educational exposures across childhood to examine how timing, duration, and intensity of contextual disadvantage shape cognitive aging. These insights could inform upstream interventions aimed at improving early educational conditions for populations at ­elevated genetic risk of dementia.

Limitations

Our study includes several limitations. First, older adults who were unable to complete the cognitive assessment on their own were excluded from our sample (Langa et al., 2023). Consequently, our analytic sample is likely more cognitively healthy than the general U.S. older adult population, which may result in lower bound estimates of the G × E association. Second, our sample included only 342 respondents with ε4/ε4 genotypes, limiting our ability to differentiate between homozygotes and heterozygotes in our tests of the G × E association, although patterns for homozygotes were similar to those for ε4 carriers. Third, our ability to examine whether G × E associations vary by race and ethnicity is limited by the small number of non-White respondents with genetic data in the HRS. In supplemental race-stratified analyses, we found a significant G × E interaction for non-Hispanic White adults but not for Black adults. However, in simulations where we assumed state education resources had the same effect on cognitive decline for Black adults as in the pooled sample, we found that a statistically significant G × E association was detected in only 55% of simulations. As the HRS continues to add new cohorts with genetic data (with half of the new respondents identifying as Black and Hispanic), the sample size will increase, providing for a better-powered test of this G × E association.

Finally, our measure of education resources is assessed at the state level, which likely obscures within-state heterogeneity in education resources. This may be particularly salient for Black adults who often attended schools with fewer resources than the estimated state average, particularly if they grew up in the U.S. South (Walsemann et al., 2024; Walsemann, Ureña, et al., 2022). However, using a state-level measure may reduce, albeit not eliminate, selection effects that can bias the G × E association estimates, such as parents choosing to move to neighborhoods with better resourced schools.

Conclusion

Our findings suggest that state investment in public education can positively impact cognitive function for APOE ε4 carriers. Therefore, ensuring access to high-quality education may serve as an effective social policy to maintain cognitive performance and slow cognitive decline among future cohorts of older adults at high genetic risk of dementia.

Supplementary material

Supplementary material is available online at The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences.

Data availability

This study uses restricted HRS data, which are available by application. The historical state education data are available at OpenICPSR (https://doi.org/10.3886/E233063V1) and in MiCDA’s Geographic Linkages Repository (GLR).

Author contributions

K.M.W., J.B., and P.H. contributed to the conception and study design. H.J. performed the statistical analyses, created the graphics, and drafted the methods and results section. K.M.W. wrote the first version of the manuscript. All authors interpreted the data and provided critical revisions to the manuscript.

Funding

This work was supported by the National Institute on Aging at the National Institutes of Health (R01AG067536, K02AG075237). HRS is sponsored by the National Institute on Aging (U01AG009740) and is conducted by the University of Michigan. The funding bodies had no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the manuscript; and in the decision to submit the manuscript for publication.

Conflicts of interest

None declared.