Moderating effects of educational inequality on education polygenic scores, attained education and dementia-risk relationships

Abstract

While education is among the most consistent predictors of cognitive health and dementia risk, it is unclear whether this reflects the benefits of schooling itself or other factors such as underlying genetic predispositions or family-based inequality. Further, the benefits of education to cognitive health may be dependent on larger structures of inequality. Using data from the IGEMS Consortium of over 4500 twins from seven cohorts in the USA, Sweden, Denmark, and Australia, we examine how genetic propensity for education and national/historical contexts of inequality shape dementia risk in later life. Dementia risk was measured with the Latent Dementia Index, a validated continuous indicator of likelihood of dementia, while key predictors included both attained education, a polygenic score for educational attainment (PGS-ED), and cohort- and country-specific Gini indices of educational inequality (GINI-ED). Ordinary least squares and within–between twin models were estimated to separate between-family from within-family effects. Results confirm that higher education is associated with lower likelihood of dementia; PGS-ED also predicts likelihood of dementia, although both associations are attenuated when comparing twins within families. Country-level educational inequality moderated these effects; education—and to a lesser extent PGS-ED—were stronger predictors of cognition in more egalitarian contexts. Our findings emphasize the interplay of both genes and social factors in potentially shaping dementia.

Highlights

• •Education's protective effect is strongest in low-inequality settings.
• •Educational inequality moderation is between-family; within-twin educational gains for reducing dementia risk fade with a BA or above.
• •The EAPGS predicts dementia risk; effect weakens with ecucational inequality.

1. Introduction

Education is one of the most widely studied protective factors for population health (Mackenbach, 2012; Mirowsky & Ross, 2005). Due to its high correlation with cognition and cognitive ability, education is also a known protective factor for age-related cognitive decline and Alzheimer's disease and related dementias (ADRD) (Katzman, 1993; Livingston et al., 2017; Mortimer & Graves, 1993; Sharp & Gatz, 2011). However, other studies suggest that preexisting cognitive ability (Kremen et al., 2025), genetic profile (Walters et al., 2025), and early-life health (Haas, 2006)—each influencing both schooling and dementia risk—may account for much of this association.

Given that education is generally thought to be a fundamental cause of population health and health inequalities in general (Phelan et al., 2010)—and given large historical improvements in educational attainment within countries and stark disparities between countries—there is substantial interest in exploring the pathways by which education can slow the pace of aging and potentially decrease cognitive decline and ADRD (Sugden et al., 2023).

However, there is growing evidence that the health benefits of educational attainment may be dissimilar across contexts; in fact, these effects may be contextually specific and vary widely across historical, geographic, and social contexts (Finkel et al., 2007, pp. P286–P294; Karlsson et al., 2015; Lipnicki et al., 2017; Norton et al., 2014; Ozyigit, 2021). Our inquiries into these specificities must become more nuanced in order to understand the conditions under which educational investment may more effectively reduce ADRD risk (Norton et al., 2014).

Despite extensive research on both educational attainment and genetic propensity for education, far less is known about how these influences operate jointly across macro-level contexts characterized by unequal opportunity structures. Prior work has shown that both attained education and polygenic scores for education (PGS-ED) vary in their predictive validity across national and historical settings, yet it remains unclear whether the cognitive consequences associated with these factors likewise differ across such environments. The present study addresses this gap by examining how both achieved and genetically influenced educational pathways relate to later-life dementia risk within cross-national contexts that differ markedly in educational inequality. By integrating attained education, PGS-ED, and contextual measures of inequality, we aim to clarify whether social and genetic determinants of cognitive aging are amplified or constrained by opportunity structures in childhood.

The present study focuses on two interrelated questions: After accounting for potential confounding from genetic and shared family background characteristics, (1) whether the impact of attained education on cognitive function and ADRD differs across countries with varying levels of educational inequality over time; and (2) whether the effect of genetic propensity for educational attainment is moderated by such inequality.

1.1. Education, cognition, and ADRD risk

Research has identified several pathways through which educational attainment relates to cognitive aging and dementia risk. These pathways can be grouped into three overarching mechanisms. First, direct pathways posit that education contributes to cognitive health by building reserve, shaping occupational complexity, fostering healthier lifestyles, and promoting lifelong cognitive engagement (Ross & Mirowsky, 1999; Ross & Wu, 1995). Higher education can lead to better and more high-paying jobs and less financial strain (Finkel et al., 2022), more cognitively complex occupations that stimulate brain activity and minimize stress (Winkleby et al., 1992), and healthier lifestyles (Margolis, 2013). Second, higher levels of education are thought to build both brain reserve and cognitive reserve, providing individuals with greater capacity to compensate for neuropathology and thereby delay the clinical onset of dementia (Stern, 2012; Stern et al., 2023). In addition, education is associated with a lifelong pattern of cognitive engagement, supporting the “use-it-or-lose-it” hypothesis whereby sustained intellectual stimulation helps maintain neural plasticity and slows cognitive decline (Coyle, 2003; Katzman, 2004).

Second, selection pathways emphasize that education is partly shaped by early-life factors that also influence dementia risk. Education is correlated with both cognition and cognitive performance and subsequently with reduced ADRD risk (Clouston, 2020, Cutler & Lleras-Muney, 2006). Several mechanisms for this relationship have been proposed and studied. First, educational attainment may be a marker for familial level factors that are advantageous for life-chances; these factors include better nutrition, richer cognitive stimulation, and a host of health resources that foster cognitive reserve and can help individuals to maintain a healthy life-style over the life-course (Haas, 2006). Some researchers have argued that these confounding factors are almost entirely responsible for the education/health gradient (Amin et al., 2015), while most have found that educational attainment remains an independent predictor (Banks & Mazzonna, 2012, Barcellos, 2025; Ericsson et al., 2023; Montez & Friedman, 2015, Fletcher, 2021, Nguyen, 2016). In addition, educational attainment exerts a moderate causal impact on later-life health and cognitive decline, even after accounting for cognitive ability (Clouston, 2020, Huh, 2024, Hayward, 2015). Finally, educational attainment may be confounded by prior causes (i.e. selection into education), in which lower childhood cognitive ability, brain injuries, or poor health constrains educational attainment, eventually raising ADRD risk (Haas, 2006; Kremen et al., 2019).

Third, genetic predispositions influence both educational attainment and cognitive functioning. The polygenic score for educational attainment (PGS-ED) captures genetic variants associated not only with schooling but also with cognitive ability and brain development. Recent evidence has suggested a genetic underpinning for educational attainment (Okbay et al., 2022) while other studies have found genetic correlations between genetic propensities for educational attainment and protective factors for health and mortality (Boardman et al., 2015). Specifically, single nucleotide polymorphism (SNP)-based linkage disequilibrium (LD) score regression analyses observed a genetic correlation of approximately −0.29 between education and dementia, suggesting that alleles linked to higher educational attainment also confer protection against ADRD (Okbay et al., 2022). However, these genetic underpinnings are not completely responsible for observed educational attainment/ADRD relationships (Ericsson et al., 2023; Walters et al., 2025). Finaly, although genetic correlations exist between education and dementia, these correlations are incomplete, suggesting that education and PGS-ED may exert both shared and independent influences on cognitive aging.

1.2. Social inequality as a moderator of genetic and educational effects

Evidence suggests ethno-regional differences in rates of dementia, even between high income areas such as North America and Western Europe (Rizzi et al., 2014), and even between wealthy Nordic countries such as Sweden and Denmark (Ozyigit, 2021). Further, low educational attainment is responsible for a substantial portion of population-attributable dementia risk (PAR), but there are large differences in the low education PAR between countries (Norton et al., 2014).

Inequality in educational access—whether across countries or birth cohorts—shapes both the distribution of education and the degree to which genetic potential for learning can be realized (Tucker-Drob & Bates, 2016). A recent study found that inequality of educational opportunity at age 10 was associated with poorer late-adulthood cognitive functioning, particularly among women (Leist et al., 2021). A central theme of recent gene–environment (G × E) research is that heritability and genetic influences on education vary across macro-level contexts (Branigan et al., 2013, Cesarini & Visscher, 2016, Fraemke et al., 2025, Harden, 2021, Herd 2019, Turkheimer et al., 2003, Turkheimer, 2000). A growing body of work suggests that structural inequalities condition the extent to which both social background and the genome shape educational outcomes. For instance, Adkins and Vaisey (Adkins & Vaisey, 2009) theorized that genetic influences are most visible under conditions of low inequality and high mobility, but are dampened where educational opportunities are tightly stratified. Similarly, evidence from twin and molecular genetic studies indicates that heritability of educational attainment varies substantially across countries (Branigan et al., 2013), with some of the strongest genetic effects observed in more egalitarian contexts (Tucker-Drob & Bates, 2016). A very recent study found that the magnitude of educational attainment polygenic associations with attainment was the same in East and West Germany before unification, but then became larger in East Germany after unification (Fraemke et al., 2025). Additional examples include the emergence of genetic effects with the decline of gender inequality (Herd et al., 2019) and strong SES by gene effects in more unequal countries (Tucker-Drob & Bates, 2016) in the US. In general, genetic influences on educational attainment often increase in more egalitarian systems, as barriers are reduced and individual differences in aptitude can more fully express. Conversely, in high-inequality settings, the social environment may overwhelm genetic potential, with attainment reflecting social class origins more than cognitive ability (Reiss et al., 2013).

Although most contextual genetic research focuses on variation in the heritability of educational attainment, similar principles imply that the cognitive consequences of genetics may also vary across socio-economic structures. If egalitarian educational systems allow genetic differences in cognitive and learning abilities to manifest more fully, genetically influenced advantages may accumulate across the life course, resulting in larger PGS-ED associations with late-life cognition in low-inequality settings. Conversely, where educational opportunities are tightly stratified, structural barriers may constrain the expression of genetic propensities, attenuating downstream cognitive effects. This reasoning motivates our expectation that national educational inequality may moderate not only attained education's association with dementia likelihood but also the association between PGS-ED and cognitive aging.

1.3. “Gene-Gini”: social inequality and GxE

Our study focuses specifically on country-level educational inequality, operationalized via the Gini coefficient for education (GINI-ED) to capture cross-national variation in opportunity structures (Danler & Pfaff, 2021). Large between-country differences in inequality provide the contextual heterogeneity needed to distinguish alternative models of G × E interplay (Seabrook & Avison, 2010). The “genes and Gini” perspective underscores that inequality is not merely a background condition but a force that can either magnify or obscure genetic influences (Selita & Kovas, 2019).

Educational inequality (GINI-ED) also moderates the downstream consequences of education. If the protective effect of education against ADRD arises mainly through cognitive stimulation and skill acquisition, then this effect may be particularly pronounced in high-inequality settings where fewer individuals achieve higher schooling and the benefits are concentrated among a select group (Sharp & Gatz, 2011). Conversely, if educational attainment largely proxies pre-existing advantages—including genetically influenced abilities—then the education/ADRD association may weaken under high-inequality conditions, where educational access is constrained by social class (Sharp & Gatz, 2011).

Importantly, inequality is not only observed cross-nationally, but also shifts over time within countries (Chancel et al., 2023). Opportunity structures differ widely across birth cohorts, reflecting both demographic and social policy changes across generations (Richards, 2022). For example, large expansions of secondary and post-secondary in Northern Europe in the 20th century created relatively egalitarian contexts for educational attainment, independent of parental social class, gender, or regional variations within countries (Heckman et al., 2018; Lynch, 2003). The post-WWI GI bill in the US offered improved opportunities for post-secondary education, although these opportunities were highly demarcated along racial and gender lines (Thomas, 2017).

These historical shifts altered the meaning of educational attainment across cohorts, with later-born cohorts both achieving higher average schooling and experiencing different returns to education in terms of wealth, occupational opportunities, health dividends, and ultimately, risk for dementia (Elder & Glen, 1975, Finkel et al., 2007, pp. P286–P294; Flynn, 1984; Glymour & Whitmer, 2019; Karlsson et al., 2015; Lipnicki et al., 2017; Lynch, 2006; Norton et al., 2014; Ozyigit, 2021, Rizzi et al., 2014). Situating gene–environment (GxE) interplay in these cohort-specific contexts is thus essential for understanding how social inequality conditions the influence of both educational attainment and the role of genetic propensity in educational attainment across the life course (Boardman, & DawFreese, 2013).

Our research design leverages country- and cohort-level variation in educational inequality (GINI-ED) to test these emerging perspectives. Educational inequality varied widely over time in the four countries studied here (see Fig. 1), although less so during the windows of observation for our study participants. Such differences allow us to examine whether polygenic scores for education (PGS-ED) exert stronger or weaker effects depending on the opportunity structure in place during childhood. They also allow us to test whether the returns to attained education—for later-life cognition and dementia risk—are amplified or suppressed in high-inequality settings. Cross-national comparisons are therefore critical, as the same level of educational inequality may be present in different cohorts in different historical time periods and allow researchers to explore the possibility that observed education/dementia associations in one context may reflect fundamentally different underlying processes than in another.

Fig. 1.

GINI-ED by Year and Country. Note: Colored dots indicate the year when twins reached age 10 in each country. Horizontal dashed lines demarcate low, medium, and high levels of inequality used in our analysis.

1.4. The present study

Leveraging the Interplay of Genes and Environment across Multiple Studies (IGEMS) consortium (Finkel et al., 2025; Pedersen et al. 2013, 2019b), we examine two questions in a harmonized dataset of twins from seven cohorts across four countries, spanning multiple birth cohorts. In general, IGEMS data provide: (1) harmonized measures of attained education using the International Standard Classification of Education (ISCED); (2) country-level indicators of educational inequality (GINI-ED) for relevant birth cohorts; (3) genotyping data for constructing PGS-ED and assessing Apolipoprotein E (APOE) genetic risks; and (4) assessment of cognitive and ADRD outcomes. By combining genetically informed designs with rich macro-level context (country and cohort), we aim to clarify how inequality conditions the effects of attained education and polygenic propensity for education on the likelihood of dementia.

A further strength of our approach lies in the use of within–between family models, which exploit the natural grouping of twins to sharpen causal inference. By decomposing variance into between-family and within-family components, these models help to address two critical issues. First, for achieved variables such as educational attainment, the within-family contrasts effectively control for shared family background, socioeconomic origin, and other environmental factors that might otherwise confound associations with later-life likelihood for dementia. This has implications for whether macro-level opportunity contexts might operate through family structures or independently for individuals within these structures. Second, for genetic predictors such as PGS-ED, the literature on within-family PGS models (Selzam et al., 2019) shows that this approach removes bias from factors such as population stratification (Wang et al., 2006), unmeasured linkage disequilibrium (Satten et al., 2001), assortative mating (Border et al., 2022), and indirect genetic effects from parents (i.e., genetic nurture) (Kong et al., 2018) as well as other stable family-level environmental influences correlated with genotype. This is especially important in cross-national research where population structures differ greatly.

In addition, decades of research has shown that twins are broadly representative of their source populations in terms of education, health, and cognitive outcomes (Christensen et al., 1995; Christensen et al., 2006; Eriksen et al., 2012; Petersen et al., 2011; Simmons et al., 1997); this broad trend of generalizability also extends to older twins with respect to cognitive functioning (Simmons et al., 1997).

2. Research hypotheses

This study investigates how educational attainment and genetic propensity for education relate to cognitive function and Alzheimer's disease and related dementias (ADRD) in later life, and how these relationships are shaped by country-level educational inequality.

Hypothesis 1

Educational Attainment × Inequality

We hypothesize that the protective association between educational attainment and likelihood of dementia will vary systematically by levels of educational inequality. In low-inequality settings, the returns of each additional level of education to cognitive outcomes may be smaller, reflecting a more homogeneous distribution of skills and resources across the population. In contrast, in high-inequality contexts, where education is more selective and concentrated among advantaged families, the benefits for dementia risk reduction may be amplified as educational attainment reflects not only acquired skills but also persistent lifetime advantages in health, wealth, and opportunity. As suggested by behavioral genetic studies, there may be a stronger association of education and dementia in more equal contexts where access to education is driven by ability rather than by privilege.

Hypothesis 2

Genetic Propensity for Educational Attainment × Inequality

We hypothesize that the association between an education polygenic score (PGS-ED) and likelihood of dementia will also be moderated by educational inequality (GINI-ED). In low-inequality contexts, reduced structural barriers may allow genetic propensities for education to manifest more fully, potentially strengthening PGS-ED associations with dementia risk. Conversely, in high-inequality settings, environmental constraints on educational attainment may attenuate the expression of genetic potential, weakening the link between PGS-ED and these outcomes. Alternatively, if the observed associations between PGS-ED and later-life cognition operate largely through access to elite educational and occupational pathways, we may observe stronger PGS-ED effects in high-inequality settings, where such advantages are concentrated.

3. Data, measures, and methods

3.a. Data for this study are drawn from the Interplay of Genes and Environment across Multiple Studies (IGEMS) consortium, which includes 21 twin studies of adult development and aging in five countries (Pedersen et al. 2013, 2019b). Our analyses pool data from four countries and seven IGEMS twin cohorts that contain measured genotypes. Swedish twin studies include: The Swedish Adoption/Twin Study of Aging (SATSA) (Finkel & Pedersen, 2004), The OCTO-Twin Study (Origins of Variance in the Old-Old) (McClearn et al., 1997), and Aging in Women and Men: A Longitudinal Study of Gender Differences in Health Behaviour and Health among Elderly (GENDER) (Gold et al., 2002). Danish twin studies include: The Longitudinal Study of Aging Danish Twins (LSADT) (Christensen et al., 1999) and The Middle-Aged Danish Twin Study (MADT) (Pedersen, Larsen, et al., 2019). US twin studies include The Vietnam Era Twin Study of Aging (VETSA) (Kremen et al., 2013). Australian twin studies include The Older Australian Twins Study (OATS) (Sachdev et al., 2009).

Our analytic sample includes 4505 individual twins with available data on: (1) polygenic scores for educational attainment (PGS-ED) and APOE genotype; (2) educational attainment (ISCED); (3) country-level education Gini coefficient at age 10 (GINI-ED); and (4) the latent dementia index (LDI). By country, the sample includes 1141 participants from Sweden, 1956 from Denmark, 888 from the United States, and 520 from Australia. We include only participants who were directly assessed, excluding those assessed by proxy. Participants were born between 1899 and 1954, with a mean age of 61.1 years at study intake. Forty-five percent were female and nearly all were White (99%). 41% were monozygotic twins (MZ), 36% were same-sex dizygotic twins (DZ), and 23% were opposite-sex DZ twins. Summary statistics are presented in Table 1.

Table 1.

Descriptive statistics.

Variable	N	Mean	SD	Median	Min	Max
Female	4505	0.45	0.5	0	0	1
Year of Birth	4505	1935.75	12.89	1939	1899	1954
Age at Intake	4505	61.06	10.58	58.33	35.9	92
Education
ISCED (twin)	4505	3.33	1.8	3	0	8
Parent ISCED	3886	2.45	1.54	2	0	8
Education Inequality GINI	4505	20.48	8.28	15.85	7.26	39.1
Genetic Variables
PGS-ED	4505	0.03	0.98	0.03	−4.1	3.36
Number of e2 alleles	4505	0.18	0.39	0	0	2
Number of e4 alleles	4505	0.32	0.51	0	0	2
Latent Dementia Indicator
LDI	4505	6.74	1.17	6.65	1.12	11.2

3.1. Measures

Our likelihood-of-dementia outcome is the Latent Dementia Index (LDI), derived within IGEMS studies by aligning cognitive, memory, and functional ability measures across studies and applying a latent variable model that effectively prioritizes dementia-relevant decline over general cognitive ability and includes functional abilities that may be affected by dementia (Beam et al., 2022). The LDI is a continuous score (0-12) that reflects an individual's estimated likelihood of dementia, validated against clinically diagnosed dementia where it was available in studies within IGEMS (See Fig. 2 for the distribution of LDI scores). Reliability of the LDI was high across studies (0.79–0.96), with strong correlations with clinical diagnoses (0.41–0.75) and mental status scores (0.48–0.80) (Beam et al., 2022). Sensitivity for registry dementia codes was greater than 88% and false positive rates were 8% or less (Beam et al., 2022). For those without dementia, scores from the last available contact were used to compute the LDI; for participants who developed dementia, scores from the last contact prior to onset were used. Higher values on the LDI indicate stronger cognitive performance, particularly on memory, higher functional ability, and lower likelihood of having dementia.

Fig. 2.

Distribution of the latent dementia index.

Educational Attainment (ISCED). Participant education was harmonized across studies and cohorts using the International Standard Classification of Education (ISCED-1997), with higher scores indicating greater educational attainment (UNESCO, 1997). The ISCED measure provides a standardized approach for describing educational programs and credentials that apply uniform and internationally agreed upon classifications which maximize commensurability across countries and over time. Levels correspond to: 0 = less than primary (no education or less than completed primary education), 1 = primary (complete primary education, typically 6 years), 2 = lower secondary education (grades 7-9), 3 = upper secondary education (grades 10-12), 4 = post-secondary non-tertiary education, 5 = short-cycle tertiary education, 6 = bachelor's or equivalent, 7 = master's or equivalent, 8 = doctoral or equivalent.

Educational Inequality (GINI-ED). Country-cohort level educational inequality was measured using an educational GINI (GINI-ED) from the World Income Inequality Database (Chancel et al., 2023). The Gini index, derived from the Lorenz curve, measures inequality as the ratio of the area between the Lorenz curve and the line of perfect equality to the total area under that line, ranging from 0 (perfect equality) to 100 (perfect inequality) (Gini, 1936). This measure is linked to IGEMS participants by using year at age 10 and their corresponding country of residence. In the sample, GINI-ED values ranged from 7.3 to 39.1, with higher values indicating greater inequality. We categorized the educational Gini into low, medium, and high levels rather than treat it as continuous as Selita and Kovas (Selita & Kovas, 2019) argue that inequality often alters opportunity structures in threshold-like ways, making categorical comparisons theoretically and empirically preferable to assuming a linear effect. We categorized GINI-ED into terciles using the full IGEMS sample distribution (10 times larger), not just our sample distribution. Using all data from these four IGEMS countries (not just our sample), our terciles are as follows: low (<13.2), medium (13.2–18.7), and high (≥18.8) educational inequality. These tercile cut-points are marked as horizontal lines on Fig. 1.

Polygenic Score for Educational Attainment (PGS-ED). The polygenic score (PGS) for educational attainment was computed based on summary statistics from the third iteration of the educational attainment (EA3) genome-wide association study using more than 1.1 million individuals (Lee et al., 2018). To generate the PGS-ED, allele counts for each variant were weighted by the effect sizes from the GWAS summary statistics. We processed the EA GWAS summary statistics using the SBayesR method (Lloyd-Jones et al., 2019) prioritizing HapMap3 variants and accounting for linkage disequilibrium (Zhao et al., 2024). The final EA polygenic score represents the sum of weighted alleles, using adjusted weights obtained from SBayesR, that reflect the cumulative genetic contribution to educational attainment. The PGS-ED scores in our sample were first standardized within each IGEMS study/genotyping array and then standardized across the 7 studies included in the analysis.

On average, polygenic scores capture the aggregate effects of thousands of single-nucleotide polymorphisms (SNPs) (Boyle et al., 2017; Rietveld et al., 2013; Visscher et al., 2017), including the 1271 statistically significant SNPs identified in the genome wide-association study for EA3. In independent samples (AddHealth and HRS), EA3 predicted between 11 and 13% of the variance in attained education (Lee et al., 2018).

While the PGS-ED is a strong predictor of educational attainment, it may also predict dementia risk because it captures genetic variants associated not only with schooling but also with correlated traits such as general cognitive ability, brain development, and health behaviors. Its link to dementia could therefore reflect direct biological effects on neural structure and function, as well as gene–environment correlations (rGE). Finally, the PGS-ED may reflect inequality in the largely European samples used to train the GWAS, which could attenuate the country-level inequality effects.

Covariates. Age at LDI assessment and sex are included as covariates in all models, along with the numbers of APO-E ε2 and ε4 alleles as controls for the genetic risk of dementia (Liu et al., 2013) and as controls for any potential genetic confounds with PGS-ED (Okbay et al., 2022).

3.2. Statistical analyses

We examined how educational attainment and genetic factors interact with country-cohort level educational inequality to predict dementia risk. Analyses were conducted using linear regression models estimated in Stata 19 with clustered standard errors adjusting for twin-pairs. Specifically, we estimated (a) marginal effects of ISCED on LDI at low, medium, and high levels of educational inequality, and (b) marginal effects of PGS-ED on LDI at low, medium, and high levels of inequality.

Model 1

Hypothesis 1. In the first model, we assessed whether the association between educational attainment (ISCED) and dementia risk (LDI) differed across levels (L = low, M = medium, H = high) of country-level educational inequality (GINI-ED). We regressed LDI on ISCED, PGS-ED, and the interaction between ISCED and terciles of GINI-ED for the ith person in the jth country. We first specified a population-average model that ignores the twin structure of our data (see Equation (1.1) below).

$${\text{LDI}}_{ij}={\mathrm{\beta }}_0+{\mathrm{\beta }}_E{\text{ISCED}}_{ij}+{\mathrm{\theta }}_M{\text{GINI}}_{Mj}+{\mathrm{\theta }}_H{\text{GINI}}_{Hj}+{\mathrm{\delta }}_M\left({\text{ISCED}}_{ij}\times {\text{GINI}}_{Mj}\right)+{\delta }_H\left({\text{ISCED}}_{ij}\times {\text{GINI}}_{Hj}\right)+{\mathrm{\gamma }}^{\prime }{Z}_{ij}+{ϵ}_{ij}$$ Equation 1.1

This model controls for the covariates $Z_{ij}$ outlined above and estimates both the overall/main effect of ISCED and its interaction with low- (reference category), medium-, and high-levels of the GINI coefficient. Standard errors are clustered at the twin-pair level to account for non-independence within twin pairs.

We next specified a within-between twin model that specifically accounted for the twin structure of our data and allow for family background controls, which would include contributions from such unmeasured factors as childhood socio-economic status (i.e., parental income, education, and occupational status) (Sjölander et al., 2012).

$$\overline{{\text{ISCED}}_f}=\frac{1}{n_f}\sum _{i=1}^{n_f}\mathrm{r}ISCE{D}_{fi}\text{,}$$ Equation 1.2

$$\widetilde{{\text{ISCED}}_{fi}}=\text{ISCE}{\mathrm{D}}_{\text{fi}}-\overline{{\text{ISCED}}_f}\text{.}$$ Equation 1.3

For this specification, we first computed an average $\overline{{\text{ISCED}}_f}$ for family/twin pairs (see Equation (1.2)) and then computed an individual ($i$) twin deviation $\widetilde{{\text{ISCED}}_{fi}}$ from each $\overline{{\text{ISCED}}_f}$ mean for the family/twin pairs ($f$) as in Equation (1.3). This allows for the simultaneous estimation of both a between-twin/family pair effect ($\overline{{\text{ISCED}}_f}$) that we obtain from Equation (1), as well as a within twin/family effect ($\widetilde{{\text{ISCED}}_{fi}}$). This is illustrated in Equation (1.4) with the regular normality assumptions in Equation (1.5).

$$\underset{\text{Outcome}}{\underbrace{\text{LD}{\mathrm{I}}_{\text{fi}}}}={\beta }_0+\underset{\text{Between}‐\text{family}\text{effect}}{\underbrace{{\beta }_B\overline{{\text{ISCED}}_f}}}+\underset{\text{Within}‐\text{family}\text{effect}}{\underbrace{{\beta }_W\widetilde{{\text{ISCED}}_{fi}}}}+{\mathrm{\gamma }}^{\prime }{\mathrm{Z}}_{\text{fi}}+{\epsilon }_{fi}\text{.}$$ Equation 1.4

$$u_f\sim \mathcal{N}\left(0,{\mathrm{\sigma }}_u^2\right),{\mathrm{\epsilon }}_{\text{fi}}\sim \mathcal{N}\left(0,{\mathrm{\sigma }}^2\right)\text{.}$$ Equation 1.5

In these models, ${\beta }_B$ is the between-family (between-twin-pair) gradient and indicates how LDI differs across families that, on average, have higher vs. lower ISCED. It captures both the effects of attained education plus shared family background. ${\beta }_W$ is the within-family (within-twin-pair) gradient and indicates how LDI differs between co-twins who differ in ISCED, holding all shared family factors constant. ${\mathrm{Z}}_{\text{fi}}$​ is a vector of covariates, including: age, sex, APO-E allele counts, and the PGS-ED.

This model can then be extended to include interactions with the categorical GINI, as shown in Equation (1.6).

$$\text{LD}{\mathrm{I}}_{\text{fi}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_B\overline{{\text{ISCED}}_f}+{\mathrm{\beta }}_W\widetilde{{\text{ISCED}}_{fi}}+{\mathrm{\theta }}_M\text{GIN}{\mathrm{I}}_{\mathrm{M}}+{\mathrm{\theta }}_H\text{GIN}{\mathrm{I}}_{\mathrm{H}}+{\mathrm{\delta }}_{\text{BM}}\left(\overline{{\text{ISCED}}_{\mathrm{f}}}·\text{GIN}{\mathrm{I}}_{\mathrm{M}}\right)+{\mathrm{\delta }}_{\text{BH}}\left(\overline{{\text{ISCED}}_{\mathrm{f}}}·\text{GIN}{\mathrm{I}}_{\mathrm{H}}\right)+{\mathrm{\delta }}_{\text{WM}}\left(\widetilde{{\text{ISCED}}_{\text{fi}}}·\text{GIN}{\mathrm{I}}_{\mathrm{M}}\right)+{\mathrm{\delta }}_{\text{WH}}\left(\widetilde{{\text{ISCED}}_{\text{fi}}}·\text{GIN}{\mathrm{I}}_{\mathrm{H}}\right)+{\mathrm{\gamma }}^{\prime }{\mathrm{Z}}_{\text{fi}}+{\mathrm{\epsilon }}_{\text{fi}}$$ Equation 1.6

Model 2

Hypothesis 2. In a similar model, we tested whether associations between PGS-ED and LDI varied across levels of inequality by including an interaction between PGS-ED and GINI-ED terciles (see Equation (2.1)).

$$LD{I}_{ij}={\mathrm{\beta }}_0+{\mathrm{\beta }}_P{PGSED}_{ij}+{\mathrm{\theta }}_{M}GINI\_M_j+{\mathrm{\theta }}_{H}GINI\_H_j+{\mathrm{\delta }}_{PM}\left({PGSED}_{ij}\times GINI\_M_j\right)+{\mathrm{\delta }}_{PH}\left({PGSED}_{ij}\times GINI\_H_j\right)+{\mathrm{\gamma }}^{\prime }{Z}_{ij}+{\mathrm{\epsilon }}_{ij}$$ Equation 2.1

In this model, $Z_{ij}$ is a vector of covariates and ${\mathrm{\epsilon }}_{ij}$ is the individual-level error term.

We then apply the within-between twin methodology for the polygenic score for educational attainment to obtain Equation (2.2). Since polygenic scores are not influenced by family background, these within-between estimates take on a different meaning. Largely, these are causal genetic effects that are unbiased by population artifacts including population stratification, assortative mating, and indirect parental genetic effects.

$$LD{I}_{ij}={\beta }_0+{\beta }_W\left({PGSED}_{ij}-{\overline{PGSED}}_j\right)+{\beta }_B{\overline{PGSED}}_j+{\theta }_{M}GINI\_M_j+{\theta }_{H}GINI\_H_j+{\delta }_{WM}\left[\left({PGSED}_{ij}-{\overline{PGSED}}_j\right)\times GINI\_M_j\right]+{\delta }_{WH}\left[\left({PGSED}_{ij}-{\overline{PGSED}}_j\right)\times GINI\_H_j\right]+{\delta }_{BM}\left({\overline{PGSED}}_j\times GINI\_M_j\right)+{\delta }_{BH}\left({\overline{PGSED}}_j\times GINI\_H_j\right)+{\gamma }^{\prime }{\mathrm{Z}}_{\text{ij}}+{\epsilon }_{ij}$$ Equation 2.2

In within–between twin models using a PGS, the within-twin effect is identified only from twins who differ in their genetic score. Because monozygotic twins (MZ) have (essentially) identical genotypes, they contribute no within-pair variation, so the within estimate is driven entirely by dizygotic twins (DZ). In this design, the within-family coefficient captures the association between having a higher PGS-ED than one's DZ co-twin and differences in late-life likelihood for dementia, while removing much of the confounding that affects between-family PGS associations. The “between” effect, by contrast, is estimated from variation in mean PGS-ED across twin pairs and is still susceptible to multiple sources of bias.

4. Results

4.1. Preliminary results

Although not an explicit hypothesis—as the literature is already rife with examples—we first replicated prior analyses (Amin et al., 2015; Ericsson, Pedersen, et al., 2023; Walters et al., 2025) and specified unmoderated relationships between education and LDI using the within/between twin approach (full results available upon request). The between-twin estimate for ISCED was significantly associated with lower dementia risk (B = 0.181, SE = 0.012, p < 0.001), although the within-twin estimate (which accounted for within-family influences while not fully accounting for genetic influences) was attenuated, but still statistically significant (B = 0.117, SE = 0.021, p < 0.001). Next, specifying an additional interaction term as the cross-product of the between- and within-twin variables revealed that the within-pair effect varied by average twin pair education. Specifically, within-pair educational differences had a strong protective influence in twin pairs with lower average education (B = 0.396, SE = 0.106, p < 0.001), but this effect diminished as family mean education increased, as indicated by a negative interaction between family mean and within-pair education (B = −0.068, SE = 0.024, p < 0.01). Based on this interaction, the protective within-pair effect becomes negligible once family mean education is approximately ISCED of six (BA or higher). Thus, the protective benefit of additional education against dementia is concentrated in lower-educated families, with diminishing returns at higher levels of family education. We next turn to our hypothesized relationships with a focus on educational inequality interactions.

As a preliminary step, we examined whether within-pair education differences varied according to the twin pair's average level of education. This descriptive analysis helps clarify whether within-family contrasts differ across the wider distribution of educational attainment, a question relevant to broader debates about the causal interpretation of education effects. Because this interaction is not central to our focal research questions concerning educational inequality, we report it briefly for context but do not draw substantive conclusions from it.

4.2. Model 1 results

In Model 1—a population averaged model which does not include within–between twin decomposition—we observed clear differences in the magnitude of the education–cognition association across levels of educational inequality (see Table 2 and Fig. 3). We present population-average models alongside within–between twin models because most research on education, PGS-ED, and cognitive aging—particularly in cross-national contexts—relies on such estimates. Including these models facilitates comparability with prior work and highlights how quasi-causal twin estimates refine interpretation. In low-inequality contexts, each additional ISCED level is associated with an average 0.239-point increase in latent dementia index (LDI) scores, corresponding to a standardized effect of 0.369 standard deviations. This effect is attenuated in medium-inequality settings, where the marginal coefficient is 0.167 (0.258 SD), and in high-inequality settings, where it is 0.163 (0.251 SD). Thus, while a higher level of education is consistently associated with better cognitive performance, the magnitude of this relationship is smaller in more unequal educational contexts.

Table 2.

Regression Results for ISCED x Educational Inequality (population-average model).

VARIABLES	LDI
Age	0.017∗∗∗
	(0.003)
Female	0.432∗∗∗
	(0.042)
APO-E: Number of e2 alleles	0.067
	(0.047)
APO-E: Number of e4 alleles	−0.158∗∗∗
	(0.037)
PGS-ED	0.045∗∗
	(0.019)
ISCED	0.239∗∗∗
	(0.031)
GINI-ED: Medium Inequality	0.395∗∗∗
	(0.133)
GINI-ED: High Inequality	0.492∗∗∗
	(0.120)
Medium Education Inequality#ISCED	−0.072∗∗
	(0.034)
High Education Inequality#ISCED	−0.077∗∗
	(0.035)
Constant	4.443∗∗∗
	(0.246)
Observations	4505
R-squared	0.103
Model	OLS
Y Mean	6.740

Marginal Effects of ISCED on LDI by Educational Inequality.

Low inequality: 0.239 (SE = 0.031, p < 0.001).

Medium inequality: 0.167 (SE = 0.015, p < 0.001).

High inequality: 0.163 (SE = 0.018, p < 0.001).

Clustered standard errors (twin-pair level) in parentheses.

∗∗∗p < 0.01, ∗∗p < 0.05, ∗p < 0.1.

Marginal effects calculated using post-estimation linear combination tests (lincom). These values represent the estimated change in LDI associated with a one-unit increase in ISCED at each inequality level.

Fig. 3.

Average Marginal Effects of ISCED on LDI by Education Inequality (population average).

The education polygenic score (PGS-ED) is also a significant predictor of LDI, but with a substantially smaller effect size than attained education. Across all contexts, a one-unit increase in PGS-ED is associated with a 0.045-point increase in LDI scores. Given the standard deviations of the variables, this translates to a standardized effect of 0.038 SD—suggesting that, although genetic propensity for education (albeit imperfectly measured) contributes directly to later-life dementia risk, the effect is modest relative to the influence of achieved education.

We additionally estimated an expanded model that simultaneously included the ISCED × GINI-ED and PGS-ED × GINI-ED interaction terms. Results from this fully specified model (not reported) were consistent with those reported above. The ISCED × GINI-ED interactions remain statistically significant and directionally consistent with the main model, indicating that the protective association of attained education for dementia risk is weaker in settings with higher educational inequality. In contrast, the PGS-ED × GINI-ED interaction terms are negative but not statistically significant, suggesting that inequality does not meaningfully moderate the association between genetic propensity for education and dementia risk. The main effects of ISCED and PGS-ED remain positive and statistically significant in this specification. These results reinforce that country- and cohort-level educational inequality moderates the attained education pathway more strongly than the genetic pathway. Clustering standard errors at the twin-pair level did not meaningfully change effect estimates or inference; all central conclusions remain the same.

When accounting for shared family background using a random-intercept mixed model using twin pair as a level 2 cluster, the main association between attained education and dementia risk remains positive and statistically significant, indicating a robust within-family effect. The moderating influence of educational inequality was attenuated in this specification, as expected given that inequality is a contextual exposure that does not vary within twin pairs. The interaction between attained education and medium educational inequality remained negative and statistically significant, whereas the interaction with high educational inequality remained negative but was no longer statistically significant. A three-level random-effects specification (individuals nested within twin pairs nested within countries) and a country fixed-effects model produced more pronounced shifts in the GINI-ED coefficients, as expected given only four countries and very limited within-country variation in educational inequality.

In the within–between twin specifications, both average (between-twin) education and within-pair differences in education are modeled simultaneously, allowing us to distinguish family-level associations from within-family contrasts (see Table 3 and Fig. 4). The between-twin ISCED effect is strong and statistically significant. For the twin-pair mean, each one-unit increase in ISCED is associated with a 0.284-point increase in LDI (SE = 0.038, p < 0.001). This effect is moderated by country-level educational inequality. In medium-inequality settings, the between-twin ISCED effect declines by −0.090 (SE = 0.042, p < 0.01), producing an estimated slope of 0.193. In high-inequality contexts, the reduction is −0.119 (SE = 0.042, p < 0.001), yielding a slope of 0.165. These patterns closely parallel the population-average model (review Table 2 and Fig. 3), showing that family-level education differences remain strongly linked to dementia risk and that the magnitude of this link declines with higher educational inequality.

Table 3.

Regression Results for ISCED x Educational Inequality (within/between model).

VARIABLES	LDI
Age	0.018∗∗∗
	(0.003)
Female	0.435∗∗∗
	(0.042)
APO-E: Number of e2 alleles	0.068
	(0.047)
APO-E: Number of e4 alleles	−0.157∗∗∗
	(0.038)
PGS-ED	0.040∗∗
	(0.019)
ISCED twin mean	0.284∗∗∗
	(0.038)
ISCED within twin difference	0.095∗∗
	(0.038)
GINI-ED: Medium Inequality	0.451∗∗∗
	(0.155)
GINI-ED: High Inequality	0.632∗∗∗
	(0.138)
Medium Education Inequality#ISCED within twin	−0.001
	(0.043)
High Education Inequality#ISCED within twin	0.068
	(0.048)
Medium Education Inequality#ISCED between twin	−0.090∗∗
	(0.042)
High Education Inequality#ISCED between twin	−0.119∗∗∗
	(0.042)
Constant	4.267∗∗∗
	(0.266)
Observations	4505
R-squared	0.106
Model	OLS
Y Mean	6.740

Marginal Effects of between-twin ISCED on LDI by Educational Inequality.

Low inequality: 0.284 (SE = 0.038, p < 0.001).

Medium inequality: 0.193 (SE = 0.019, p < 0.001).

High inequality: 0.165 (SE = 0.020, p < 0.001).

Clustered standard errors (twin-pair level) in parentheses.

∗∗∗p < 0.01, ∗∗p < 0.05, ∗p < 0.1.

Fig. 4.

Average Marginal Effects of ISCED on LDI by Education Inequality (within/between model).

The within-twin ISCED effect is smaller. A one-unit increase in ISCED relative to a co-twin is associated with a 0.095-point increase in LDI (SE = 0.038, p < 0.05), and this association is statistically significant. Interaction terms with medium (−0.001, SE = 0.043) and high (+0.068, SE = 0.048) inequality are not statistically significant, indicating that inequality does not systematically moderate the within-twin education effect. Although not reported here, models with MZ twins only showed very similar patterns.

Relative to the population-average specification, the within–between twin model produces two key differences. First, the between-twin ISCED coefficients are nearly identical to the population-level results, reinforcing that much of the education–cognition association operates at the family level and is sensitive to broader country-level inequality. Second, the within-twin ISCED effect is considerably smaller and shows no evidence of moderation by inequality, in contrast to the significant interactions observed at the between-family level. This divergence suggests that the apparent moderation by educational inequality in population-average models is primarily driven by between-family differences in opportunity structure and shared background, rather than by causal effects of within-family differences in education.

4.3. Model 2 results

Model 2 examines the association between the education polygenic score (PGS-ED), attained education (ISCED), and national educational inequality (GINI-ED) in predicting late-life cognition (LDI). The main effect of PGS-ED is positive and statistically significant (see Table 4 and Fig. 5): a one-unit increase in PGS-ED is associated with a 0.118-point increase in LDI (SE = 0.042, p < 0.01). Given the LDI standard deviation of 1.17, this corresponds to a standardized effect of ∼0.10 SD. The effect is smaller than that of attained education but indicates that genetic propensity toward educational attainment carries direct implications for cognitive function.

Table 4.

Regression Results for PGS-ED x Educational Inequality (population average results).

VARIABLES	LDI	LDI
Age	0.018∗∗∗	0.017∗∗∗
	(0.003)	(0.003)
Female	0.434∗∗∗	0.433∗∗∗
	(0.042)	(0.042)
APO-E: Number of e2 alleles	0.068	0.068
	(0.047)	(0.047)
APO-E: Number of e4 alleles	−0.158∗∗∗	−0.158∗∗∗
	(0.037)	(0.037)
ISCED	0.118∗∗∗	0.095∗∗
	(0.046)	(0.047)
PGS-ED	0.176∗∗∗	0.232∗∗∗
	(0.011)	(0.032)
GINI-ED: Medium Inequality	0.184∗∗	0.379∗∗∗
	(0.087)	(0.135)
GINI-ED: High Inequality	0.262∗∗∗	0.468∗∗∗
	(0.070)	(0.121)
Medium Education Inequality#ISCED		−0.065∗
		(0.035)
High Education Inequality#ISCED		−0.067∗
		(0.036)
Medium Education Inequality#PGS-ED	−0.072	−0.045
	(0.052)	(0.053)
High Education Inequality#PGS-ED	−0.099∗	−0.072
	(0.055)	(0.057)
Constant	4.576∗∗∗	4.455∗∗∗
	(0.229)	(0.247)
Observations	4505	4505
R-squared	0.102	0.103
Model	OLS	OLS
Y Mean	6.740	6.740

Clustered standard errors (twin-pair level) in parentheses.

∗∗∗p < 0.01, ∗∗p < 0.05, ∗p < 0.1.

Fig. 5.

Average Marginal Effects of PGS-ED on LDI by Education Inequality (population average model).

Attained education (ISCED) remains a strong predictor, with each additional ISCED level linked to a 0.176-point increase in LDI (SE = 0.011, p < 0.001), equivalent to ∼0.15 SD. This standardized effect is more than twice as large as the PGS-ED effect, underscoring the primacy of achieved education as a determinant of dementia risk.

Critically, the PGS-ED × GINI-ED interactions reveal that the strength of the PGS-ED effect diminishes as inequality rises. In medium-inequality settings, the interaction is negative but not statistically significant (−0.072, SE = 0.050). In high-inequality settings, the interaction is negative and marginally significant (−0.099, SE = 0.055, p < 0.10). This implies that the PGS-ED effect on LDI is attenuated across contexts: from ∼0.10 SD in low-inequality contexts, to ∼0.04 SD in medium-inequality, and to near zero (∼0.01 SD) in high-inequality settings. This pattern is consistent with our hypothesis, although the interaction reaches only marginal statistical significance, suggesting more tentative support.

In the within–between twin specification (see Table 5 and Fig. 6), PGS-ED effects on cognition are decomposed into between-twin (pair-average genetic propensity) and within-twin (differences in genetic propensity between co-twins) components. The between-twin effect of PGS-ED is significant: each one-unit increase in the pair-average PGS-ED is associated with a 0.121-point increase in LDI (SE = 0.049, p < 0.05). The within-twin effect of PGS-ED is smaller (0.096) and imprecisely estimated (SE = 0.102, ns), suggesting no detectable impact of differences in PGS-ED within twin pairs.

Table 5.

Regression Results for PGSED x Educational Inequality (within/between model).

VARIABLES	LDI
Age	0.018∗∗∗
	(0.003)
Female	0.433∗∗∗
	(0.042)
APO-E: Number of e2 alleles	0.067
	(0.047)
APO-E: Number of e4 alleles	−0.158∗∗∗
	(0.038)
ISCED	0.176∗∗∗
	(0.011)
PGSED twin mean	0.121∗∗
	(0.049)
PGSED within twin difference	0.096
	(0.102)
GINI-ED: Medium Inequality	0.184∗∗
	(0.087)
GINI-ED: High Inequality	0.263∗∗∗
	(0.071)
Medium Education Inequality#PGS-ED within twin	−0.084
	(0.114)
High Education Inequality#PGS-ED within twin	−0.035
	(0.119)
Medium Education Inequality#PGS-ED between twin	−0.070
	(0.056)
High Education Inequality#PGS-ED between twin	−0.107∗
	(0.060)
Constant	4.574∗∗∗
	(0.229)
Observations	4505
R-squared	0.102
Model	OLS
Y Mean	6.740

Clustered standard errors (twin-pair level) in parentheses.

∗∗∗p < 0.01, ∗∗p < 0.05, ∗p < 0.1.

Fig. 6.

Average Marginal Effects of PGS-ED on LDI by Education Inequality (within-between model).

Interactions with PGS-ED reveal moderation primarily in the between-twin component. The between-twin PGS-ED × high-inequality interaction is negative and marginally significant (−0.107, SE = 0.060, p < 0.10), indicating that the positive effect of pair-average PGS-ED on LDI is weakened in high-inequality settings. By contrast, the within-twin PGS-ED × GINI-ED interactions are negative but non-significant (−0.084 in medium-inequality and −0.035 in high-inequality contexts), suggesting that no consistent moderation of within-pair PGS-ED differences is detectable.

Relative to the population-average model, which estimated a significant PGS-ED effect (0.118) that was attenuated in high-inequality settings (interaction = −0.099, p < 0.10), the within–between results show that this moderation is concentrated in the between-twin component. That is, the attenuation of PGS-ED effects by inequality appears to operate through between-family differences, consistent with population-level artifacts such as assortative mating, gene–environment correlation, or contextual/environmental opportunities tied to genetic propensity for education. Once these factors are purged in the within-twin contrasts, the moderation disappears, leaving no evidence that genetic differences within DZ twin pairs are differentially expressed across inequality contexts.

5. Discussion and conclusions

5.1. Model 1 discussion

Our first set of analyses examined the population-averaged (between-family) effects of education on dementia likelihood, operationalized by the latent dementia index (LDI), and how these effects varied across contexts of educational inequality. Consistent with prior work, higher education was strongly associated with lower likelihood of dementia, but the magnitude of this association depended on country-level inequality (Leist et al., 2021). The marginal benefit of education was largest in low-inequality settings and attenuated as inequality increased. This pattern suggests that more egalitarian social environments maximize the cognitive return to education, whereas in more unequal societies, the advantages conferred by schooling are less evenly distributed.

When we turned to within–between twin models, however, a more nuanced picture emerged. The moderation of education by inequality was robust in population-level models, but within-twin contrasts—where early-life family background (childhood SES, e.g.) is held constant—showed no significant evidence of moderation. Regardless, a main effect of educational attainment (ISCED) persists, although it is substantially lower once the shared family environment is accounted for. The attenuation of attained education when controlling for family background, is also consistent with literature noting significant confounding by parental education (Lin, 2020).

This observed divergence in the within/between models highlights several possible mechanisms. First, between-family estimates are influenced by the fact that education is correlated with parental resources, early-life nutrition, and cognitive stimulation (Hackman et al., 2010). In low-inequality contexts, these advantages may overlap substantially with educational attainment, inflating the apparent protective effect of schooling. In contrast, the within-twin models control for shared family background, including shared childhood resources and environments which leaves only the most direct effects on education which are both smaller and not moderated by educational inequality.

A second explanation to explain this divergence is that educational attainment may partially reflect genetic propensities for cognitive ability, personality, perseverance, and other relevant individual factors (Kremen et al., 2025). In high-inequality settings, access to schooling is more strongly stratified by social origin, so the observed between-family education effects may partly reflect rGE—genetic differences aligned with differences in family environment, and, as Sharp and Gatz (2011) suggest, this stratification dampens the education–dementia association relative to contexts where ability drives access. Within-twin designs purge some of these rGE factors, since MZ and DZ twins share genetic background to differing degrees, and differences in attainment within pairs are less correlated with genome-wide differences. In this case, the attenuation of the within-twin educational attainment by inequality interaction may reflect the removal of some of the genetic confounders that bias the between-family effects (Walters et al., 2025).

Third, twin differences in education can emerge from other factors not directly related to SES, including: illnesses, accidents, personality differences, and/or school and teacher quality. These may be weaker pathways and have more divergent implications for dementia risk than SES which is captured in the between-family estimates. Because nonshared factors like illness or teacher quality both shape educational attainment and relate to dementia risk, failing to control for them can cause within-twin estimates to underestimate education's effect.

Fourth, in more unequal environments, the range in difference between lower and higher educational attainment may be large at the population level, but much smaller within families and as such, the nominal benefit of one twin achieving slightly higher education may not be as pronounced, nor show much of an effect. This dampening of variation in twin studies is widely recognized, but not often accounted for in most studies (Boardman, & DawFreese, 2013). At the same time, both twins may have been given similar opportunities, so any observed differences in education may result in smaller, less consequential differences in dementia risk. Using ISCED, which marks larger credentialing differences, does minimize this possibility, however—much less so than using years of education or even incorporating differences in quality of educational exposures, for example.

Finally, measurement error and attenuation bias may play a role. Education is measured categorically (ISCED) and may not reflect the full range of both credentials and years and quality of education. As such, small within-pair differences may be noisier, less statistically powerful, and may not capture the same contrasts as the broader between-family distribution. At the same time, approximately 42% of MZ- and 50% of DZ-twin pairs differed on their ISCED. Similarly, because the variance in education within twin pairs is much smaller (σ² = 1.237) than the between-family variance (σ² = 2.034), the within-twin estimates are especially vulnerable to attenuation. With variance nearly half as large, any given amount of measurement error will exert roughly double the proportional impact, biasing the within-twin coefficient toward zero and reducing power to detect an effect.

As a whole, our findings suggest that the observed moderation of education effects by educational inequality largely reflect family-level differences, even if educational inequality may condition family-level inequalities and the environment and resources in which family structures are embedded. Once these shared contexts are held constant, inequality does not significantly alter the marginal benefit of education within families. This does not rule out causal effects of education—it underscores that much of the inequality moderation operates through family-level pathways and possibly gene–environment correlation, rather than through differential returns to individual educational differences within families.

5.2. Model 2 discussion

Our models estimate that a larger PGS-ED predicts lower dementia risk, but the relationship weakens in higher-inequality countries—as suggested by the negative PGS-ED x GINI-ED interaction. This finding is generally consistent with the GxE literature showing that genetic influences are more fully expressed in more equitable social environments and are attenuated when environmental influences become stronger in highly unequal environments and conditions of deprivation (Turkheimer et al., 2003).

While the population-average models show that the effect of PGS-ED on LDI is attenuated, with marginal statistical significance, in high-inequality settings, the within-family PGS-ED models suggest a different interpretation. By comparing twins, these models strip out population-level artifacts—including stratification, assortative mating, and indirect parental genetic effects—that can inflate associations between polygenic scores and outcomes. Once these sources of bias are purged, the interaction between PGS-ED and educational inequality is substantially weakened and no longer statistically significant. This contrast suggests that much of the observed moderation of PGS-ED by inequality at the population level reflects contextual and demographic processes tied to family background and social stratification, rather than a direct causal interaction between genetic propensity and national inequality. These results are somewhat different from some of the population-averaged models in the literature that show variation in genetic effects across inequality contexts. We do acknowledge that these models are very taxing on statistical power and heavily dependent on DZ twins, and thus, our confidence intervals are very large. In addition, the between-twin variance in polygenic scores (0.702) is more than 2 ½ times as large as the within-twin variance (0.270).

Alternatively, the differences in magnitude of our estimates across levels of inequality largely match, so it might be too early to draw large conclusions. It might be very difficult to achieve the level of genotyping of twins across hundreds of years of cohorts and dozens of countries required to re-examine our results and determine if they do in fact contrast the existing literature.

Some of the baseline effect of PGS-ED operates through attained education (ISCED) although not much is explained by genetic correlations with the APO-E control variables (i.e., Ɛ2 and Ɛ4 allele counts, not shown). The effect of PGS-ED on LDI was virtually unchanged, and slightly larger, when APOE was included as a control (results not shown). Rather than indicating shared genetic pathways, this pattern suggests independence between the two predictors and implies that adjusting for APOE removes unrelated variance, yielding a cleaner estimate of the PGS-ED effect.

The PGS-ED aggregates many variants tied to neurodevelopment, synaptic plasticity, and brain maintenance that are unrelated to APO-E, pointing to direct biological pathways (Lee et al., 2018; Okbay et al., 2022). Further, many of the variants in PGS-ED affect baseline cognitive ability, executive function, and lifelong learning capacity, which build cognitive reserve and shape health behaviors (Okbay et al., 2022). Finally, gene-environment correlation (rGE) may play a role. With respect to active rGE, individuals with higher PGS-ED may actively create or seek out stimulating contexts and socially enriching environments, reinforcing lifelong cognitive benefits that are not captured by years of schooling alone. For example, individuals with higher PGS-ED may be more likely to select into cognitively demanding occupations which foster cognitive resilience throughout adulthood. Finally, evocative rGE may play a role as PGS-ED shapes how individuals are perceived and treated by others (Boardman, & HarrisFinch, 2024) and could explain why PGS-ED continues to be associated with the likelihood of dementia well after educational attainment has ceased.

5.3. Implications for research and policy

Our findings highlight the importance of incorporating macro-level contexts—such as historical variation in educational opportunity—into studies of dementia risk and gene–environment interplay. For future research, three implications emerge. First, cross-national harmonization efforts such as IGEMS provide essential infrastructure for testing whether genetic and educational predictors operate similarly across diverse institutional settings. Second, studies employing polygenic scores should more explicitly consider how contextual moderators shape the expression and consequences of genetic propensities across the life course. Third, our results illustrate the value of distinguishing within-family from between-family pathways when interpreting population-average associations, particularly in international samples where stratification mechanisms differ.

Policy implications also follow from these patterns. Stronger associations of education and PGS-ED with cognitive aging in egalitarian contexts suggest that policies reducing educational inequality may enhance the long-term cognitive benefits of schooling. Efforts to equalize educational access may therefore amplify both the social and biological pathways through which education protects against age-related cognitive impairment.

5.4. Limitations

Several limitations of this study should be acknowledged, including mortality selection. The mean age of our sample is 61 years old and according to life-tables in the United States, approximately 14% of the population will have died by this age. Further, mortality rates for those with less than a high school education are approximately 2-4 times more likely to die than those with a bachelor's degree (Hummer & Hernandez, 2013, Paglino et al., 2025). Survivorship bias remains a concern because analyses require both co-twins to survive and participate, which can selectively exclude pairs with lower education, poorer health, or higher genetic risk. This conditioning on joint survival may attenuate observed education/dementia risk associations and reduce the variance in within-family differences. This could disproportionately affect our within-family estimates and bias these effects downward, affecting our substantive conclusions that tend to wash out many of the observed between-family effects.

Second, although the Latent Dementia Index offers a validated cross-sectional indicator of dementia likelihood, our analyses do not incorporate repeated measures of cognitive change. As a result, we cannot examine intra-individual trajectories of decline or test whether inequality conditions rates of cognitive decline over time. The absence of longitudinal assessments may attenuate associations that unfold gradually across the life course.

Third, our cohorts exhibit a wide-range in age at intake (36-92) and these variations can complicate making direct comparisons. For example, the same “age group” in one country may experience a distinct historical experience in another, raising the possibility that observed effects partly reflect cohort-linked exposures rather than pure age or inequality differences. In addition, not all IGEMS cohorts were included as not all were genotyped nor contained the measures necessary for computation of the LDI. This does limit generalizability of our sample, but this is common in cross-national research where cohort representation is nationally based. In addition, although we directly specify cohort effects by the inclusion of GINI-Ed and control for age at LDI assessment, period effects may remain as confounders. The IGEMS data consortium provides enormous cohort depth and breadth, but no extant data would allow for a fully specified age-period-cohort model with the variables we are examining.

Fourth, our classification of educational inequality into low, medium, and high categories is necessarily constrained by the distribution across the four countries in our twin cohorts. While our empirically derived cut-points provide a useful heuristic for testing interactions, they do not reflect thresholds with social or biological meaning. Indeed, no empirically or theoretically grounded thresholds are available, so categorization necessarily relies on sample-derived distinctions.

Fifth, our study sample is largely White and European, which is both a strength and a limitation: polygenic scores for educational attainment show their highest predictive validity in these groups, but their portability to more admixed or non-European populations is limited, meaning our findings may not generalize broadly (Ding et al., 2023).

Despite these constraints, the present study makes a unique and important contribution. IGEMS harmonizes multiple genotyped twin cohorts spanning multiple countries and historical birth cohorts. By leveraging within-family twin designs, we distinguish between family-level confounding and individual-level effects, shedding light on the pathways through which education, genetics, and inequality shape dementia risk. Our results confirm the important role of education and genetic propensities in lowering the likelihood of dementia, but also shed further light on the contextual conditions under which these effects vary.

CRediT authorship contribution statement

Brian Karl Finch: Writing – review & editing, Writing – original draft, Supervision, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Sneha Nimmagadda: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Data curation. Deborah Finkel: Writing – review & editing, Supervision, Methodology, Funding acquisition, Data curation, Conceptualization. Margaret Gatz: Writing – review & editing, Project administration, Funding acquisition, Data curation. Chandra A. Reynolds: Writing – review & editing, Methodology, Funding acquisition, Data curation, Conceptualization. Marianne Nygaard: Writing – review & editing, Resources, Data curation. Vibeke Catts: Writing – review & editing, Resources, Data curation. Anbu Thalamuthu: Writing – review & editing, Data curation. Perminder Sachdev: Writing – review & editing, Data curation. Malin Ericsson: Writing – review & editing, Validation, Methodology, Data curation, Conceptualization. Ida Karlsson: Writing – review & editing, Validation, Methodology, Data curation. Valgeir Thorvaldsson: Writing – review & editing, Data curation. Linda Hassing: Writing – review & editing, Data curation.

Ethical statement

This study used data from the Interplay of Genes and Environment across Multiple Studies (IGEMS) consortium, which harmonizes information from multiple national twin registries. All data were accessed under restricted-use agreements (both institutional data usage agreements and individual data use agreements) and analyzed in de-identified form. Ethical approval for each participating registry was obtained from the appropriate institutional review boards and national ethics committees, and all participants provided informed consent at the time of original data collection. The present analyses were reviewed and approved by the Institutional Review Board at the University of Southern California (UP-16-00315). All research procedures complied with the ethical standards of the institutional and national research committees.

Declarations of competing interest

This manuscript was partially supported by the Center for Economic and Social Research at the University of Southern California as well as the following grants from the National Institutes of Health: R01AG059329, R01 AG081248, and R01 AG089666.

Data availability

The authors do not have permission to share data.