Epigenetic g predicts cognitive function in a diverse, nationally representative sample of older adults Authors: Jessica D. Faul, Stacey Collins, Trey Smith, Eric T. Klopack, Colter Mitchell, Eileen M. Crimmins, Mateo P. Farina Published: 2026 ABSTRACT BackgroundAlzheimer’s disease and related dementias (ADRD) are major public health concerns. DNA methylation (DNAm)–based biomarkers such as GrimAge and PhenoAge predict aging and health risk, but were not designed to optimize prediction of cognitive function or decline. Epigenetic g—a DNAm-derived index of general cognitive ability—is a promising marker of cognitive function that has not been assessed in a racially and socioeconomically diverse population.MethodsWe used data from the 2016 Venous Blood Study of the Health and Retirement Study (HRS), a nationally representative cohort of U.S. adults aged ≥ 51 years (N = 3575 with high-quality DNAm). Epigenetic g scores were computed using CpG weights from a BayesR+ model of general cognitive ability developed in Generation Scotland. Cognitive function was measured with a modified version of the Telephone Interview for Cognitive Status (TICS) at each interview wave. Linear regression estimated associations with cognitive scores; mixed-effect growth curve models estimated the association with cognitive change. Models were adjusted sequentially for demographics, education, parental education, APOE ε4 status, and blood-based neurodegeneration markers (NfL, GFAP, Aβ42/40, pTau181).ResultsHigher epigenetic g was associated with better baseline cognition (β = 2.55, 95% CI 1.80–3.30)) and cognition at the time DNAm was measured (β = 2.30, 95% CI 1.47–3.13) after demographic adjustment. Associations were attenuated but remained significant with education and parental education (β = 1.23–1.89). In growth curve models, Epigenetic g was associated with higher cognitive performance but did not have a statistically significant association with decline over a 6-year period. Results were robust to adjustment for APOE ε4 and neurodegeneration biomarkers.ConclusionsEpigenetic g is a scalable, blood-based marker of cognitive function, and potentially or cognitive reserve, that adds predictive value beyond demographics, socioeconomic indicators, APOE, and neuropathology. Its validation in a diverse, nationally representative U.S. cohort underscores its potential for early risk profiling and for research on social determinants of cognitive aging in cross-national samples.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13148-026-02142-9. Background Alzheimer’s disease and related dementias (ADRD) are a growing public health concern that affect millions worldwide. To identify at-risk groups and individuals, researchers have collected and analyzed biomarkers of cognitive impairment. While several studies have incorporated biomarkers from multiple biological systems (i.e., cardiovascular, metabolic, endocrine, etc.), one of the most promising areas of research has been the identification of epigenetic patterns – namely DNA methylation (DNAm) – as predictors of age-related outcomes, including cognitive function and impairment. The promise of the DNAm-based biomarkers has led to a proliferation of health measures predicting mortality (i.e. GrimAge), multisystem dysregulation (i.e. PhenoAge), and system specific health outcomes (i.e. inflammation, immune functioning, etc.). Many of these markers are associated with cognitive functioning and impairment in older adulthood [1, 2], providing researchers opportunities to examine individual and group-based differences in risk along with examining the association with social and environmental exposures across the life course (e.g. educational attainment) [3, 4]. However, as these markers were not developed to predict cognitive functioning, DNA methylation sites that are associated with cognitive function or decline may be excluded from these predictive algorithms. As a result, existing measures may not be optimized to predict cognitive functioning or health risks in later life. Thus, developing a DNAm-based biomarker tailored to predict cognitive functioning or decline holds substantial promise for advancing the study of cognitive aging. Such a measure could enhance our ability to not only detect individuals or groups who are at greater risk of cognitive impairment years or decades before the onset of systems, but also uncover molecular pathways that underlie social disparities in brain health. Prior work has constructed epigenetic based algorithms of cognitive functioning. One notable approach is the development of “epigenetic g,” an aggregate measure derived from DNA methylation sites associated with measures of general cognitive ability (g) in adulthood, conceptually paralleling the psychometric construct of g in cognitive psychology [5]. The concept of g is thought to represent mental capacity, broadly, which influences performance on a wide array of cognitive tasks like problem-solving, reasoning, and learning [6]. However, it also likely reflects inputs like education and cognitive stimuli across the life course. Using the Generation Scotland Study, researchers developed an epigenetic measure of cognition based on several domains of cognitive performance that was replicated and predictive of cognitive functioning in two external samples of adults from the Lothian Birth Cohort Study. While this existing epigenetic g measure represents an important advancement, this measure was derived in a relatively homogeneous, midlife cohort with limited racial and ethnic diversity, raising questions about its generalizability to populations in other contexts [7]. DNA methylation and cognitive ability are sensitive to social and environmental exposures throughout life that vary significantly across national and demographic contexts, including cumulative disadvantage, educational quality, and structural racism—factors that also shape cognitive health disparities. Other studies, with diverse samples quite different from the Scottish training sample in terms of age, race/ethnicity and nationality, have used this measure of epigenetic g to evaluate social differences in children, adolescents and adults [8, 9]. However, it is unknown whether the structure and correlates of epigenetic g replicate in diverse populations of older adults and to what extent it is associated with cognitive function and decline in later life. Using the Health and Retirement Study (HRS)—a nationally representative panel of U.S. adults over age 50—we apply an epigenetic measure of cognition in a sample that reflects the social and economic diversity of cognitive aging in the United States. This approach extends prior work on longitudinal disease prediction and risk profiling for dementia by integrating molecular data and life course perspectives on cognitive health and dementia risk. Specifically, we seek to examine the relationship between epigenetic g and educational attainment, parents’ education, cognitive function, and other measures of dementia risk to determine whether a DNA methylation-based score of epigenetic g provides unique insights into cognitive function and decline beyond existing genetic and biological measures. Methods Sample To examine the association of epigenetic g on general cognitive functioning, we used data from the Health and Retirement Study (HRS), a nationally representative, longitudinal study of U.S. adults 51 years and older. The HRS collects extensive demographic and cognitive functioning information, alongside a rich set of biomarker and genetic data [10–12]. The HRS is sponsored by the National Institute on Aging (NIA U01AG009740) and is conducted by the University of Michigan. In 2016, the HRS conducted the Venous Blood Study (VBS). To be eligible for the VBS, respondents had to reside in the community, complete the 2016 HRS core interview without a proxy, and consent to and complete the blood draw [13]. Blood samples were collected in respondents’ homes by trained phlebotomists and processed at the CLIA-certified Advanced Research and Diagnostic Laboratory at the University of Minnesota. The sample with DNAm measurement and assays of blood-based neurodegenerative markers was selected from 2016 HRS respondents born in 1959 and earlier who participated in the 2016 VBS. From this group, a probability sample was selected from HRS respondents who were 65 and older and thus eligible for the 2016 Harmonized Cognitive Assessment Protocol (HCAP) or who are eligible for a future HCAP (a random half of HRS sample members not yet 65 years old). This subsample fully represents the entire HRS sample when sample weights are used. DNA methylation profiling was performed using the Illumina Infinium MethylationEPIC BeadChip V1 (850 K) array using DNA extracted from the buffy coat collected from blood samples in 2016. DNA samples were randomized across analytic plates by age, cohort, sex, education, and race/ethnicity along with 39 pairs of blinded duplicates. Data preprocessing and quality control were performed using the minfi package in R. A total of 3.4% of the methylation probes were removed from the final data because their detection P-value fell below the threshold of 0.01 (n = 29,431 out of 866,091). Analysis for failed samples was done after removal of detection P-value failed probes. A total of 58 samples were removed after applying a 5% cutoff. Sex-mismatched samples and any controls (cell lines, blinded duplicates) were also not included in the analyses. The final analytic dataset includes 97.9% of the originally plated samples (n = 4,018). More information is available elsewhere [14]. For this analysis, sporadic missing beta methylation values were assigned a value of 0 after standardization. The analytic sample includes individuals with complete DNAm data, cognitive assessments from the 2016 core interview, salivary DNA, and relevant demographic information (N = 3575). Measures Cognitive functioning Cognitive functioning was assessed using the modified version of the Telephone Interview for Cognitive Status (TICS). The total score ranges from 0 to 27, with higher scores indicating better cognitive functioning. It consists of immediate word recall (0–10), delayed word recall (0–10), backwards counting from 20 (0–2), and serial 7s subtraction (0–5). We used cognitive functioning at baseline for each individual, usually when participants were between 51 and 56 years of age, and cognitive functioning in 2016 when the DNA methylation data was collected, on average 13.7 years later. In order to minimize the effects of item-level non-response among self-respondents, we used the imputed cognition data released by HRS [15]. Dementia incidence We classified dementia using the Langa–Weir algorithm [16], which incorporates both self-respondent and proxy information. For self-respondents, we used the modified 27-point TICS, with scores between 0 and 6 identifying those with probable dementia. For proxy respondents, we used the 11-point proxy scale, which includes interviewer assessments (0–2), proxy-rated memory (0–4), and instrumental activities of daily living (0–5). Proxy respondents scoring 6–11 were classified as having probable dementia. This classification criteria have been validated against clinical diagnoses in the Aging, Demographics, and Memory Study (ADAMS), a sub study of the HRS with in-person neuropsychological assessments and expert adjudication of cognitive status, has been widely used in population research, and aligns with Medicare-based prevalence estimates [16, 17]. Using the Langa-Weir classification, incident dementia was defined as dementia in 2018, 2020 or 2022 among respondents who were cognitively normal or cognitively impaired but not demented (CIND) in all survey waves up through 2016. Epigenetic g Epigenetic g was created using the following three steps. First, we used effect size weights for CpG sites derived from a BayesR+ epigenome-wide prediction model of general cognitive ability—corrected for age, sex, BMI and epigenetic smoking—developed in the Generation Scotland cohort [5]. Second, for each participant we standardized DNA methylation beta values (mean = 0, SD = 1) at those CpG sites. Third, we multiplied the standardized values by the published weights and summed across all sites to yield a continuous epigenetic predictor of general cognitive ability (g). All steps were performed using R Statistical Software (v4.5.1; R Core Team 2025) [18]. Epigenetic g was standardized for comparability to other markers. Neurodegenerative markers The neurodegeneration biomarkers included in this study were measured from plasma and serum using blood was collected from consenting HRS participants as part of the 2016 Venous Blood Study. Prior pilot work established the stability and validity of measuring blood-based biomarkers of neuropathology using this collection protocol [19]. Neurofilament light chain (NfL), glial fibrillary acidic protein (GFAP), amyloid-β 42:40 ratio, and Phosphorylated Tau at threonine 181 (pTau181) were all measured on the same HD-X analyzer (Quanterix Inc.). Aβ-40, Aβ-42, NfL, and GFAP were measured in plasma using the Quanterix Neurology 4-Plex E kit. Phosphorylated tau (pTau-181) was measured in serum using the Quanterix Advantage V2.0 kit. For analysis, the Aβ42/20 Ratio was rescaled by multiplying by 100. Furthermore, we standardized these variables for interpretability. For NfL, GFAP and pTau181, we top coded at the 99th percentile and log-transformed due to substantial skewness and then standardized within the analytic sample. Coefficients for these biomarkers can be interpreted per one-standard-deviation difference in the transformed measure. APOE genotyping and polygenic risk scores APOE genotyping was performed on salivary DNA samples collected between 2006 and 2012. TaqMan allelic discrimination SNP assays, C___3084793 (rs429358) and C____904973 (rs7412) using a 7900HT Sequence Detection System and quantified using SDS v2.1software (Applied Biosystems, Foster City, CA) were performed. Genotypes were determined using TaqMan Genotyper software. For participants for whom direct genotyping was not available, APOE status was imputed from previously genotyped array data using the Illumina HumanOmni2.5 array. Dosage data for rs429358 and rs7412 was used to assign the “best guess genotypes” to infer the APOE isoform. The salivary DNA samples were genotyped and analyzed at the Center for Inherited Disease Research (CIDR) Genetic Resources Core Facility (GRCF) and Fragment Analysis Facility (FAF) at Johns Hopkins University (https://cidr.jhmi.edu). A complete description is available from the HRS [20]. GrimAge age acceleration GrimAge was developed based on the 7 DNAm surrogates of plasma proteins and smoking pack years in a two-stage procedure [21]. GrimAge age acceleration was calculated using the residual that results from regressing GrimAge on chronological age. This results in a measure of age acceleration without the effect of chronological age. For GrimAge, sex is also regressed out of the measure because both age and sex were used in the identification process. GrimAge age acceleration was standardized to more directly compare with other biomarkers. Sociodemographic characteristics Age was based on birth year. Race and ethnicity were self-reported and categorized into Non-Hispanic White/Other, non-Hispanic Black, and Hispanic. Sex was reported as male or female. Respondent’s parental education was based on father’s and mother’s highest level of education, which was self-reported. From this information, we created a dichotomous variable indicating whether either parent had at least a high school education. Own education was based on the highest completed year of education (0–17 years). We created three categories: 0–11 years, 12 years, and > 12 years. Statistical analysis To examine the association of epigenetic g and cognitive functioning as well as the sensitivity of this association after the inclusion of other key social and biomarker predictors of cognitive health, we used a series of linear regression models predicting cognitive functioning at baseline and in 2016. We estimated linear mixed-effects growth-curve models of cognitive score across four waves (2016–2022), with repeated observations nested within respondents. Models included a random intercept and random slope for time at the respondent level. Because HRS does not provide longitudinal weights, we used the 2016 VBS baseline respondent weight for this subsample (vbsi16wgtra) as a base-year weight for prospective analysis of the 2016 VBS subsample and estimated weighted models using probability weights with robust standard errors. Age was centered at age 50 to facilitate interpretation of the intercept. As a supplementary analysis, Cox proportional hazard models were used to examine the association between epigenetic g and dementia incidence over a 6-year follow-up period. In the first model, we evaluated the association between epigenetic g with each cognitive functioning outcome adjusted for age, sex and race/ethnicity. Subsequent models additionally adjust for own education and having at least one parent with 12 years or more of education (Model 2), APOE ε4 carrier status (Model 3), blood-based neurodegeneration markers (Model 4), and GrimAge age acceleration (Model 5). To facilitate interpretation and comparison, values for epigenetic g, the blood-based neurodegeneration markers and GrimAge acceleration were standardized. Sampling weights were used to account for the sampling design and differential non-response. This approach yields estimates generalizable to the US population over 55 years of age. Linear regressions (Proc Survey Reg) and hazard models (Proc Survey Phreg) were completed using SAS software, version 9.4 (SAS Institute Inc., 2023) [22]. Linear mixed-effect growth curve models were run using STATA version 17 [23]. Results Sample characteristics Table 1 shows the descriptive information for the analytical sample for cognitive functioning at baseline and 2016 (n = 3575). The average age of the samples was 69.6 years old, and age ranged from 56 to 93 at the time of DNAm assessment in 2016. Most of the sample was non-Hispanic White (71.4%), followed by non-Hispanic Black at 15.5%, and Hispanic at 13.1%. 15.2% of the sample had less than a high school diploma, 52.5% had a high school diploma or GED, and 32.3% had more than a high school diploma. 59.9% of the sample had at least one parent with a high school diploma. At baseline, scores on the modified TICS averaged 16.8 (SD = 4; range = 1–27) and at the time of DNAm assessment the average was 15.1 (SD = 4.3; range = 0–27). In 2016, 6.5% of respondents were classified as having probable dementia. Table 1Descriptive statistics of the cogntiive function sample and incident dementia sample—HRS 2016Cognitive function linear modelsGrowth modelIncident dementian = 3575n = 3183n = 2762Mean / %SDMinMaxMean / %SDMinMaxMean / %SDMinMaxAge in 2016 (yrs)69.68.9569370.09.0569369.58.75693Female57.6%57.658.5%Race/ethnicityNon hispanic white71.4%72.5%74.3%Non hispanic black15.5%14.8%13.9%Hispanic13.1%12.7%11.8%EducationLess than high school15.2%14.9%11.7%HighSchool/GED52.5%53.3%53.5%More than high school32.3%31.8%34.8%ParenteducationNo parents > 12th grade40.1%40.1%37.5%1 or moreParents > 12th grade59.9%59.9%62.5%Cognitive test scoreBaseline16.84.012716.93.932717.33.67.027201615.14.302715.14.302715.93.87.027Cognitive status 2016Dementia6.5%6.3%0%No dementia93.5%93.7%100%APOE e4074.8%75.6%122.7%21.7%22.5%2.6%NeurodegenerationNfL25.3126.421.70590.5023.6523.452.90456.34GFAP108.9474.825.70789.10104.8370.705.70789.10AB42/20 Ratio*6.562.241.1058.446.572.311.1058.44pTau-1812.122.15.0343.002.052.13.0343.00Incident dementia in 6 yearsNo onset (n = 2541)92%Dementia onset (n = 221)8%Years beween baseline enrollment and 2016Years13.76.1021145.9421145.9421 Table 1Descriptive statistics of the cogntiive function sample and incident dementia sample—HRS 2016Cognitive function linear modelsGrowth modelIncident dementian = 3575n = 3183n = 2762Mean / %SDMinMaxMean / %SDMinMaxMean / %SDMinMaxAge in 2016 (yrs)69.68.9569370.09.0569369.58.75693Female57.6%57.658.5%Race/ethnicityNon hispanic white71.4%72.5%74.3%Non hispanic black15.5%14.8%13.9%Hispanic13.1%12.7%11.8%EducationLess than high school15.2%14.9%11.7%HighSchool/GED52.5%53.3%53.5%More than high school32.3%31.8%34.8%ParenteducationNo parents > 12th grade40.1%40.1%37.5%1 or moreParents > 12th grade59.9%59.9%62.5%Cognitive test scoreBaseline16.84.012716.93.932717.33.67.027201615.14.302715.14.302715.93.87.027Cognitive status 2016Dementia6.5%6.3%0%No dementia93.5%93.7%100%APOE e4074.8%75.6%122.7%21.7%22.5%2.6%NeurodegenerationNfL25.3126.421.70590.5023.6523.452.90456.34GFAP108.9474.825.70789.10104.8370.705.70789.10AB42/20 Ratio*6.562.241.1058.446.572.311.1058.44pTau-1812.122.15.0343.002.052.13.0343.00Incident dementia in 6 yearsNo onset (n = 2541)92%Dementia onset (n = 221)8%Years beween baseline enrollment and 2016Years13.76.1021145.9421145.9421 Descriptive statistics of the cogntiive function sample and incident dementia sample—HRS 2016 High School/GED Parent education No parents > 12th grade 1 or more Parents > 12th grade We also show the descriptive information for the analytic sample used to evaluate cognitive decline (also shown in Table 1; n = 3183). Most demographic characteristics were similar. For this sample, neurofilament light (NfL) levels averaged 25.31 pg/mL (SD = 26.42; range = 1.7–590.50). Glial fibrillary acidic protein (GFAP) averaged 108.94 pg/mL (SD = 74.82; range = 5.7–789.1). The rescaled Aβ42/40 ratio had a mean of 6.56 (SD = 2.24; range = 1.1–58.4). Phosphorylated tau-181 (pTau-181) averaged 2.12 pg/mL (SD = 2.15; range = 0.03–43). APOE ε4 carriage was distributed as follows: 74.8% of participants carried no ε4 alleles, 22.7% carried one ε4 allele, and 2.5% carried two ε4 alleles. Over the 6 years of follow-up, 8% of the sample developed dementia. The mean epigenetic g was 0.02 with a standard deviation of 28.30, indicating quite wide dispersion. In our sample, epigenetic g predicted 3% of cognitive functioning at baseline. We evaluated the correlations between each neurodegenerative marker and epigenetic g—none of the correlations exceeded 0.06 indicating very low correlation levels between epigenetic g and the blood-based neurodegenerative measures (shown in Supplemental Table 1). Epigenetic g was also not highly correlated with GrimAge age acceleration (0.11). Association of Epigenetic G with Cognitive Functioning at Baseline, Cognitive Functioning at the time of DNAm assessment (2016), and Dementia Incidence (2016–2022). Table 2 shows the associations of epigenetic g with cognitive functioning at baseline (Panel A). These linear models passed the assumptions for linearity, homoscedasticity, normality of residuals, and independent errors. With adjustments for chronological age, sex/gender, and race/ethnicity (Model 1), the association between epigenetic g and cognition remained positive (β = 2.55; 95% CI = 1.80, 3.3). Adding educational attainment and parental education (Model 2) attenuated the coefficient but the association persisted (β = 1.89; 95% CI = 1.29, 2.50). Table 2Association between epigentic g and cognitive performance at baseline and in 2016 (n = 3575)A. Baseline CognitionModel 1Model 2estimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueIntercept21.1019.6622.53 < 1E-3020.4619.2021.73 < 1E-30Epigenetic g2.551.803.303.029E-111.891.292.501.062E-09Age (Baseline)− 0.07− 0.10− 0.059.771E-09− 0.05− 0.07− 0.028.013E-05Female0.840.551.121.007E-081.100.861.34 < 1E-30Male (ref)––––––––Hispanic− 2.58− 3.14− 2.02 < 1E-30− 1.59− 2.01− 1.189.259E-14Non-hispanic black− 2.27− 2.72− 1.81 < 1E-30− 2.05− 2.40− 1.69 < 1E-30Non-hispanic white/Other (ref)––––––––Less than high school− 3.09− 3.52− 2.66 < 1E-30High School/GED− 1.24− 1.50− 0.99 < 1E-30More than high school (ref)––––No parent education ≥ 12− 0.36− 0.64− 0.098.767E-03Parent education ≥ 12 (ref)––––B. 2016 CognitionModel 1Model 2EstimateLower 95% CIUpper 95% CIp valueEstimateLower 95% CIUpper 95% CIp valueIntercept26.9325.7428.12 < 1E-3025.5024.4526.55 < 1E-30Epigenetic g2.301.473.136.006E-081.230.551.913.776E-04Age (2016)− 0.17− 0.18− 0.15 < 1E-30− 0.13− 0.14− 0.11 < 1E-30Female0.880.561.193.881E-080.980.741.236.661E-15Male (ref)––––––––Hispanic− 2.33− 2.88− 1.782.220E-16− 1.06− 1.50− 0.622.744E-06Non-hispanic black− 2.36− 2.87− 1.86 < 1E-30− 1.82− 2.19− 1.45 < 1E-30Non-hispanic white/Other (ref)––––––––Less than high school− 3.95− 4.38− 3.52 < 1E-30High school/GED− 1.68− 1.95− 1.40 < 1E-30More than high school (ref)––––No parent education ≥ 12− 0.70− 1.00− 0.412.264E-06Parent education ≥ 12 (ref)–––– Table 2Association between epigentic g and cognitive performance at baseline and in 2016 (n = 3575)A. Baseline CognitionModel 1Model 2estimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueIntercept21.1019.6622.53 < 1E-3020.4619.2021.73 < 1E-30Epigenetic g2.551.803.303.029E-111.891.292.501.062E-09Age (Baseline)− 0.07− 0.10− 0.059.771E-09− 0.05− 0.07− 0.028.013E-05Female0.840.551.121.007E-081.100.861.34 < 1E-30Male (ref)––––––––Hispanic− 2.58− 3.14− 2.02 < 1E-30− 1.59− 2.01− 1.189.259E-14Non-hispanic black− 2.27− 2.72− 1.81 < 1E-30− 2.05− 2.40− 1.69 < 1E-30Non-hispanic white/Other (ref)––––––––Less than high school− 3.09− 3.52− 2.66 < 1E-30High School/GED− 1.24− 1.50− 0.99 < 1E-30More than high school (ref)––––No parent education ≥ 12− 0.36− 0.64− 0.098.767E-03Parent education ≥ 12 (ref)––––B. 2016 CognitionModel 1Model 2EstimateLower 95% CIUpper 95% CIp valueEstimateLower 95% CIUpper 95% CIp valueIntercept26.9325.7428.12 < 1E-3025.5024.4526.55 < 1E-30Epigenetic g2.301.473.136.006E-081.230.551.913.776E-04Age (2016)− 0.17− 0.18− 0.15 < 1E-30− 0.13− 0.14− 0.11 < 1E-30Female0.880.561.193.881E-080.980.741.236.661E-15Male (ref)––––––––Hispanic− 2.33− 2.88− 1.782.220E-16− 1.06− 1.50− 0.622.744E-06Non-hispanic black− 2.36− 2.87− 1.86 < 1E-30− 1.82− 2.19− 1.45 < 1E-30Non-hispanic white/Other (ref)––––––––Less than high school− 3.95− 4.38− 3.52 < 1E-30High school/GED− 1.68− 1.95− 1.40 < 1E-30More than high school (ref)––––No parent education ≥ 12− 0.70− 1.00− 0.412.264E-06Parent education ≥ 12 (ref)–––– Association between epigentic g and cognitive performance at baseline and in 2016 (n = 3575) Table 2 also shows the results for cognitive functioning in 2016 (Panel B). In Model 1, after adjusting for chronological age, sex/gender, and race/ethnicity, epigenetic g is positively associated with cognitive functioning (β = 2.30; 95% CI = 1.47, 3.13). In Model 2, after adjusting for own education and parent’s education that were associated with cognitive function, the association of epigenetic g and cognitive functioning in 2016 remained statistically significant (β = 1.23; 95% CI = 0.55, 1.91). In addition, we examined how much of the effect of education on cognition in 2016 was mediated by epigenetic g. To do this, we needed to include education as a binary or continuous measure. We found that epigenetic g mediates 2.57% of the impact of education on cognitive score using a binary education measure (less than High School education / High School degree or greater), while epigenetic g mediates 2.25% of the impact of education on cognitive score using a continuous measure of education (in school years). Table 3 presents results from sequential mixed-effects growth curve models of cognitive functioning with a random intercept and random slope for age. In Model 1, higher epigenetic g was associated with better cognitive functioning at age 50 (b = 0.601, 95% CI 0.04, 1.16, p = 0.036). Cognitive functioning also followed a nonlinear age pattern, with a positive linear age term (b = 0.172, 95% CI 0.12, 0.22, p < 0.001) and a negative quadratic age term (b = − 0.007, 95% CI − 0.008, − 0.006, p < 0.001), consistent with curvilinear change across age. However, there was little evidence that epigenetic g moderated age-related cognitive trajectories: neither the interaction between epigenetic g and age (b = − 0.036, 95% CI − 0.088, 0.016, p = 0.172) nor the interaction between epigenetic g and age squared (b = 0.001, 95% CI 0.000, 0.002, p = 0.149) reached conventional levels of statistical significance. Table 3Mixed-effects growth curve models predicting cognitive performance and decline: 2016–2022 (n = 3183)Model 1Model 2Model 3Model 4estimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueConstant16.72216.14617.2980E + 0016.92616.34617.5070E + 0016.42815.7517.1050E + 0016.53115.85617.2070E + 00zEpigenetic g0.6010.041.1623.58E-020.5960.0311.1623.87E-020.555− 0.0091.1195.38E-020.52− 0.0431.0837.03E-02Age (centered at 50)0.1720.120.2235.02E-110.1720.1210.2234E-110.1990.1480.253.18E-140.20.1490.2511.84E-14Age squared− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00zEpigenetic g × Age− 0.036− 0.0880.0161.72E-01− 0.034− 0.0860.0181.97E-01− 0.029− 0.0810.0232.71E-01− 0.031− 0.0820.0212.46E-01zEpigenetic g × Age squared0.00100.0021.49E-010.00100.0021.83E-010.001− 0.0010.0022.79E-010.00100.0022.5E-01Female1.0920.8631.320E + 001.0810.8541.3090E + 001.0740.8381.3090E + 000.7990.5521.0462.3E-10Hispanic− 1.018− 1.422− 0.6138.35E-07− 1.027− 1.431− 0.6246.03E-07− 1.126− 1.524− 0.7282.94E-08− 1.197− 1.592− 0.8032.72E-09Non-hispanic black− 1.772− 2.115− 1.4290E + 00− 1.614− 1.959− 1.2690E + 00− 1.716− 2.058− 1.3730E + 00− 1.595− 1.94− 1.250E + 00Less than high school− 3.818− 4.234− 3.4020E + 00− 3.836− 4.252− 3.4190E + 00− 3.721− 4.137− 3.3050E + 00− 3.556− 3.975− 3.1380E + 00High school/GED− 1.457− 1.7− 1.2140E + 00− 1.461− 1.703− 1.2190E + 00− 1.394− 1.634− 1.1540E + 00− 1.275− 1.517− 1.0330E + 00No parent education > 12− 0.65− 0.912− 0.3881.13E-06− 0.653− 0.913− 0.3928.97E-07− 0.629− 0.887− 0.3721.67E-06− 0.606− 0.862− 0.3493.6E-06APOE e4 = 1− 0.886− 1.166− 0.6065.82E-10− 0.835− 1.112− 0.5583.37E-09− 0.818− 1.094− 0.5426.18E-09APOE e4 = 2− 0.886− 1.69− 0.0823.07E-02− 0.777− 1.5710.0175.52E-02− 0.787− 1.5840.015.29E-02z ln(NfL)− 0.401− 0.556− 0.2454.56E− 07− 0.307− 0.464− 0.151.24E-04z ln(GFAP)− 0.106− 0.2610.051.82E− 01− 0.167− 0.323− 0.0123.53E-02Aβ42/40 ratio*− 2.777− 6.891.3351.86E-01− 3.408− 7.610.7941.12E-01z ln(pTau181)− 0.091− 0.2210.041.72E-01− 0.107− 0.2360.0211.02E-01zDNAmGrimAge acceleration− 0.384− 0.509-0.261.58E-09N10,69310,69310,69310,693 Table 3Mixed-effects growth curve models predicting cognitive performance and decline: 2016–2022 (n = 3183)Model 1Model 2Model 3Model 4estimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueestimateLower 95% CIUpper 95% CIp valueConstant16.72216.14617.2980E + 0016.92616.34617.5070E + 0016.42815.7517.1050E + 0016.53115.85617.2070E + 00zEpigenetic g0.6010.041.1623.58E-020.5960.0311.1623.87E-020.555− 0.0091.1195.38E-020.52− 0.0431.0837.03E-02Age (centered at 50)0.1720.120.2235.02E-110.1720.1210.2234E-110.1990.1480.253.18E-140.20.1490.2511.84E-14Age squared− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00− 0.007− 0.008− 0.0060E + 00zEpigenetic g × Age− 0.036− 0.0880.0161.72E-01− 0.034− 0.0860.0181.97E-01− 0.029− 0.0810.0232.71E-01− 0.031− 0.0820.0212.46E-01zEpigenetic g × Age squared0.00100.0021.49E-010.00100.0021.83E-010.001− 0.0010.0022.79E-010.00100.0022.5E-01Female1.0920.8631.320E + 001.0810.8541.3090E + 001.0740.8381.3090E + 000.7990.5521.0462.3E-10Hispanic− 1.018− 1.422− 0.6138.35E-07− 1.027− 1.431− 0.6246.03E-07− 1.126− 1.524− 0.7282.94E-08− 1.197− 1.592− 0.8032.72E-09Non-hispanic black− 1.772− 2.115− 1.4290E + 00− 1.614− 1.959− 1.2690E + 00− 1.716− 2.058− 1.3730E + 00− 1.595− 1.94− 1.250E + 00Less than high school− 3.818− 4.234− 3.4020E + 00− 3.836− 4.252− 3.4190E + 00− 3.721− 4.137− 3.3050E + 00− 3.556− 3.975− 3.1380E + 00High school/GED− 1.457− 1.7− 1.2140E + 00− 1.461− 1.703− 1.2190E + 00− 1.394− 1.634− 1.1540E + 00− 1.275− 1.517− 1.0330E + 00No parent education > 12− 0.65− 0.912− 0.3881.13E-06− 0.653− 0.913− 0.3928.97E-07− 0.629− 0.887− 0.3721.67E-06− 0.606− 0.862− 0.3493.6E-06APOE e4 = 1− 0.886− 1.166− 0.6065.82E-10− 0.835− 1.112− 0.5583.37E-09− 0.818− 1.094− 0.5426.18E-09APOE e4 = 2− 0.886− 1.69− 0.0823.07E-02− 0.777− 1.5710.0175.52E-02− 0.787− 1.5840.015.29E-02z ln(NfL)− 0.401− 0.556− 0.2454.56E− 07− 0.307− 0.464− 0.151.24E-04z ln(GFAP)− 0.106− 0.2610.051.82E− 01− 0.167− 0.323− 0.0123.53E-02Aβ42/40 ratio*− 2.777− 6.891.3351.86E-01− 3.408− 7.610.7941.12E-01z ln(pTau181)− 0.091− 0.2210.041.72E-01− 0.107− 0.2360.0211.02E-01zDNAmGrimAge acceleration− 0.384− 0.509-0.261.58E-09N10,69310,69310,69310,693 Mixed-effects growth curve models predicting cognitive performance and decline: 2016–2022 (n = 3183) Adding APOE-ε4 status in Model 2 only minimally changed the association between epigenetic g and cognition, and the coefficient remained positive and statistically significant (b = 0.596, 95% CI 0.031, 1.162, p = 0.039). With further adjustment for plasma biomarkers in Model 3, the epigenetic g coefficient was modestly attenuated and became borderline statistically significant (b = 0.555, 95% CI − 0.009, 1.119, p = 0.054). In the fully adjusted Model 4, which additionally included DNAm GrimAge Acceleration, the epigenetic g coefficient was further attenuated and no longer statistically significant (b = 0.520, 95% CI − 0.043, 1.083, p = 0.070). Taken together, these findings suggest that the positive association between epigenetic g and cognitive functioning was reduced after accounting for APOE genotype, neurodegenerative biomarkers, and epigenetic aging. The largest attenuation occurred after inclusion of DNAm GrimAge Acceleration, consistent with partial overlap between epigenetic g and broader methylation-based aging processes, although the coefficient for epigenetic g remained positive across all models. Table S2 presents results from sequential Cox proportional hazards models predicting 6-year incident dementia. The proportional hazards assumption was checked in SAS without weights, and there was no evidence that the proportional hazards assumption was violated for epigenetic g. In Model 1, adjusted for age, sex/gender, and race/ethnicity, higher epigenetic g was associated with a lower hazard of incident dementia, although the association did not reach conventional levels of statistical significance (HR = 0.999, 95% CI 0.997, 1.000, p = 0.066). After further adjustment for educational attainment and parental education in Model 2, the association was further attenuated (HR = 0.999, 95% CI 0.997, 1.001, p = 0.201). In Model 3, which additionally adjusted for APOE ε4 status, the epigenetic g coefficient was further reduced (HR = 0.999, 95% CI 0.998, 1.001, p = 0.308). With the addition of neurodegenerative biomarkers in Model 4, the association remained small and non-significant (HR = 0.999, 95% CI 0.998, 1.001, p = 0.408). In the fully adjusted Model 5, which additionally included DNAm GrimAge Acceleration, epigenetic g remained unrelated to incident dementia risk (HR = 0.999, 95% CI 0.998, 1.001, p = 0.445). Discussion In a nationally representative cohort of 3,575 older U.S. adults, we tested a DNA methylation–derived index of general cognition to assess its utility as a readily obtainable blood-based marker of cognitive performance. Overall, we show a robust association between epigenetic g and cognitive functioning at baseline and cognitive functioning at the time DNAm was assessed. There was some suggestion that epigenetic g was associated with cognitive decline over a 6-year follow-up period but this relationship was not statistically significant. Importantly, epigenetic g remained strongly associated with cognitive function after the inclusion of own education, parental education, APOE ε4 status, blood-based neurodegeneration markers, and GrimAge age acceleration. These findings highlight epigenetic g as a promising, scalable blood-based biomarker for identifying cognitive reserve in older adults beyond traditional demographic, genetic, and neuropathological markers that are increasingly being collected in community and population-based studies. Epigenetic g was first validated in the Lothian Birth Cohort, which consisted of adults born in 1921 and 1936 in Scotland. In that study, epigenetic g predicted general cognitive ability independent of genetic predictors, demonstrating the utility of DNA methylation markers as potential indicators of cognitive health risk. Our study extended this work to a racially and ethnically diverse, nationally representative sample of U.S. older adults born between 1923 and 1960. We find that epigenetic g predicted 3% of cognitive functioning at baseline. It was highly predictive of cognitive performance over and above educational attainment and indicators of parental education. By demonstrating comparable prediction in a racially and socioeconomically diverse U.S. cohort, our study expands the generalizability of epigenetic g beyond the primarily Scottish samples in which it was first validated, broadening its potential application to more diverse contexts. Additionally, this work advances the application of epigenetic g as a life course biomarker of cognitive functioning. Prior work has shown a strong association of epigenetic g to cognitive performance at early ages. Specifically, using the Texas Twin Study, a sample of approximately 1000 children aged 8–19 years, researchers found that lower epigenetic g scores were associated with lower scores on tests of reasoning, comprehension, reading, and math [9]. Our work shows that association between epigenetic g and cognitive performance can also be found in adults in midlife and older adulthood in the US. As such, epigenetic g may be an important biomarker of cognitive performance and/or cognitive health risks in older adulthood that can be used to investigate aging processes across the life course and evaluate individual and population level differences in cognitive health risk. Future research should investigate whether epigenetic g is indicative of cognitive development in early life that is carried into older adulthood, associated with cognitive change in later life, or some combination of the two processes that determine cognitive health risk in older adults. Further work also remains to be done in determining exactly which aspects of cognitive function are captured by epigenetic g, and teasing apart the connections between specific CpG sites and gene expression in the brain. Finally, it will be important to clarify why epigenetic g is associated with cognitive functioning across the life course. This association could reflect stable individual differences established early in life, but it may also capture the biological embedding of socially patterned exposures and resources—such as educational opportunities, occupation, and material hardship —that are strongly linked to cognitive health but not fully operationalized in this analysis. Future research should test which life course social and economic conditions are most strongly associated with epigenetic g, whether these associations vary across contexts. In our analyses, epigenetic g remained strongly associated with cognition even after adjusting for key social determinants such as educational attainment and parental education. Future analysis should consider how epigenetic g could help clarify how early-life disadvantage, educational experience, and cumulative stress associated with low SES may contribute to disparities in cognitive aging in older adulthood. A recent study in children (n = 407, mean age 12 years) that evaluated change among commonly used DNA methylation markers (i.e. PhenoAgeAccel, GrimAgeAccel, DunedinPACE) over a time period of approximately 2.4 years found that epigenetic g changed the most over that time [24]. Future work should evaluate change in epigenetic g across adulthood to better examine whether epigenetic g may capture meaningful increases in risk in later life associated with life course exposures. Further research should examine the association of epigenetic g on cognitive decline. While we did not find statistically significant evidence, the coefficient suggests that higher epigenetic g might be associated with faster cognitive decline with age. If this is found in other studies, with larger samples, it may indicate that epigenetic g has a similar association with cognitive change that has been identified between higher education, cognition, and rate of cognitive decline [25–27]. This has been found to be true especially when cognitive performance is based on tests that depend on verbal recall and memory tasks, as is the case when using the modified TICS as a measure of cognitive function [25, 27, 28]. This may suggest that epigenetic g, like education, contributes to or is a proxy for cognitive reserve. The cognitive reserve hypothesis suggests that protective socio-behavioral factors, like education, contribute to increased cognitive function and performance but that cognitive decline occurs more rapidly in individuals with greater levels of cognitive reserve once symptoms present [29, 30]. It is likely that the suggestive result that higher epigenetic g is associated with reduced incidence of dementia is largely because it is associated with higher cognitive performance earlier in life and less because it is an indicator of subsequent cognitive decline. DNAm surrogates, such as epigenetic g, allow for the investigation of exposures even if those specific exposures were not directly measured in the cohort. In our case, the HRS does not have a direct measure of cognitive “g” in its broadest sense. However, there are many examples of DNAm surrogates predicting diseases and outcomes better than the original (measured) biomarkers as has been shown with traditional cardiometabolic risk factors and cardiovascular disease [31], smoking and lung disease [32], and a DNAm surrogate for C-reactive protein predicting brain health [33]. Just as DNAm clocks, themselves a type of DNAm surrogate, are designed to predict specific outcomes like chronological age, mortality, and changes in blood-based markers and physical measures, epigenetic g predicts cognition and cognitive phenotypes well because that is what it was designed to do. If predicting the likelihood of cognitive decline or dementia incidence is an end goal, it is possible that we could refine a DNAm surrogate to optimize prediction of those outcomes. One can imagine a DNAm surrogate of neurodegenerative markers, brain imaging, and other cognitive disease phenotypes outperforming epigenetic g in the prediction of dementia risk. While our study leverages a large, racially and ethnically diverse, population-based cohort with integrated social, genetic, and blood-based neurodegeneration biomarkers and longitudinal follow-up, we acknowledge some limitations. First, the cognitive performance tests in the HRS include a limited set of cognitive domains. Epigenetic g has been shown to be related to performance across a wide spectrum of cognitive domains in prior work [9]. Future work with better measurements across multiple cognitive domains should evaluate the association in older adults. Second, our ability to investigate cognitive change and incident dementia was limited by a relatively short follow-up period (6 years) and small sample size (incident dementia). Furthermore, dementia incidence was defined using survey-based cognitive performance measures and not a clinical assessment and diagnosis. However, this method has been used elsewhere, was validated using an expert panel with a more extensive neuropsychological battery of tests, and produces similar results to those found in Medicare claims [16, 17, 34–36]. Future work with larger samples and/or longer time of follow-up should investigate whether epigenetic g can be used as an early detection biomarker for dementia. Conclusion In this study, we evaluated the performance of a DNA methylation–based predictor of cognitive function (epigenetic g) in a racially and socioeconomically diverse cohort of older adults. Epigenetic g showed a consistent association with cognitive performance and decline even after accounting for other cognitive risk factors, such as educational attainment, neurodegenerative markers, and genetic risk. This finding provides additional evidence that this DNAm signature is not limited to the original sample in which it was developed and tested. These findings contribute to the growing literature on the generalizability and portability of DNAm-derived predictors across populations. Although our results support the use of epigenetic g as a cross-study and cross-cohort marker of cognitive function, additional work is needed to establish robustness across cognitive batteries, to assess calibration and predictive performance for clinically meaningful outcomes, to determine how longitudinal changes in DNAm relate to cognitive change and neurodegenerative processes, and to evaluate how these epigenetic biomarkers of risk may perform differently across population subgroups (i.e. race/ethnic, sex, SES etc.). Because DNA methylation reflects both biological and environmental influences, future studies should test which behavioral, social, and environmental determinants shape epigenetic g and whether modifying these determinants alters DNAm-based cognitive signatures. Together, our results position epigenetic g as a promising epigenetic tool for harmonizing cognitive risk phenotypes and studying cognitive aging in across country contexts. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1