International Cross-Temporal Meta-Analysis of Assortative Mating for Educational Attainment

Abstract

Previous studies have found a high degree of assortative mating for educational attainment (r = .56). However, this can be confounded by cohort effects or country effects, where certain nations may have more pronounced assortative mating than others. In addition, method variance regarding how educational attainment is measured may also result in heterogeneity of effect sizes. Effect sizes were gathered from various datasets and from academic literature, resulting in a large collection of effect sizes (k = 1498, n = 9,159,098), spanning 84 different countries. Assortative mating for educational attainment was stronger than what previous literature suggested (r = .66, [.64, .68]), largely due to the fact that assortative mating for educational attainment is stronger when latent methods are used. The strongest predictors of assortative mating for education between countries were individualism (r = −.61, p < .001) and HDI (r = −.56, p < .001). Assortative mating over time was found to vary by region. Capitalist Europe experienced an increase in assortative mating for education, while Communist Europe experienced a decrease. The United States had a non-linear trend in assortative mating for educational attainment, as it decreased from 1875 to 1926, increased from 1926 to 1945, decreased from 1945 to 1958, increased from 1958 to 1977, and decreased from 1977 onwards.

Introduction

Phenotypic similarity between marital partners is the subject of scientific investigation for various reasons. The degree of similarity in phenotypes can bias heritability estimates downwards (Wolfram & Morris, 2023), as it will cause parents to be more genetically related than expected by chance, thereby inflating the correlations between non-monozygotic relatives. Also, phenotypic similarity increases the variance in a trait since offspring inherit genotypes that are correlated.

The impact of assortative mating on the human genome is examined by assessing the gametic phase disequilibrium (GPD) that occurs between phenotypically relevant loci. Studies measuring GPD between alleles, controlling for population stratification find that alleles for several traits, including height and educational attainment, are more positively correlated than what would be expected by chance in individuals of European ancestry (Yengo et al., 2018). A study based on evidence from a Japanese cohort replicated these findings (Yamamoto et al., 2022).

Increasing assortative mating has been the hypothesized cause of high social divisions in America according to Murray (2013), which can be observed by increased divergence in behaviour, beliefs, and achievement within White Americans. According to several researchers, assortative mating in the United States increased between the 60s and 80s (Schwartz & Mare, 2005), though Eika et al. (2019) found that this increase in observed assortative mating cannot explain the increase in income inequality. However, the latter paper does not take into account the fact that assortative mating increases genetic inequality as well as phenotypic inequality, thus it does not disprove that increased levels of assortative mating may have increased social inequality.

Evidence from a meta-analysis indicates that partners are more similar to each other in the vast majority of personal characteristics than would be expected by chance (Horwitz et al., 2022; McPherson et al., 2001). Some counterexamples include gender, chronotype, worrying, hearing difficulty, and irritability, which have a negative phenotypic similarity among partners. Educational attainment, in particular, was highly correlated among mates (r = .56), though this figure can be biased by moderators such as country of origin or reporting method. Some studies attempt to explicitly compare different countries, for example, Gonggrijp et al. (2023) found that educational attainment was more highly correlated between partners in Finland (0.51) than in the Netherlands (0.45). Theoretically, assortative mating for education can occur for two reasons: because of phenotypic assortment or social homogamy. Phenotypic assortment involves assortative mating occurring due to selection for the characteristic itself, while social homogamy results in assortative mating individuals who mate will typically inhabit similar environments. Gonggrijp et al. (2023) investigated the independent effects of both of these potential causes using twin models and came to the conclusion that both processes were involved, but that phenotypic assortment had a stronger effect.

Most studies in the literature suggest that assortative mating for educational attainment has increased within Western countries (Eika et al., 2019), however, some researchers have argued that this is not the case, notably (Heath et al., 1985) and (Clark & Cummins, 2022), who found stable levels of assortative mating within Norway and the UK respectively.

Three unresolved issues in the literature on assortative mating for educational attainment have been highlighted. The first is the issue of international differences – while these effects have been researched in various countries, the effect sizes used differ by study so they are difficult to compare. The second question is of secular differences - whether assortative mating for educational attainment has changed with the progress of time. The last is the mechanism by which this process occurs: whether it's social homogamy or phenotypic assortment. The meta-analysis can only resolve the first two questions, as we lack the twin data necessary to test whether phenotypic assortment or social homogamy drives assortative mating for educational attainment.

Data

Various datasets were consulted to calculate effect sizes, including all PISA waves, the PIRLS 2001 wave, the General Social Survey (GSS), the European Social Survey (ESS), the World Values Survey (WVS), the US Census, the Current Population Survey (CPS), and the China Family Panel Study (CFPS). The US Census and CPS were obtained from the supplemental material provided by economists who were studying assortative mating (Eika et al., 2019). In addition, academic literature that assessed assortative mating for education was gathered from another meta-analysis (Horwitz et al., 2022). Studies that assessed samples that were not representative of their host country (e.g., individual provinces or states) were excluded from the meta-analysis. Within the literature, there was weak evidence for positive-results bias according to the regression test (p = .07). Adjusted for the observed publication bias, the correlation between partners in educational attainment would be 0.53 instead of 0.56. A funnel plot of these effect sizes has been plotted in Figure 1.

Figure 1.

Funnel plot of effect sizes compiled from the literature.

Studies on assortative mating for educational attainment often use crude categories (e.g., years of education, educational categories) that may not accurately represent how educated a person is. Education could be defined as an underlying phenotype that is associated with educational attainment or a social signal that is transmitted from a human to other humans; regardless of which definition is used, education cannot be reduced to one metric. Therefore, latent associations between partners’ educational attainment were calculated when possible and were used as the reference group for the meta-analysis. Latent associations were computed by modeling educational attainment as a latent variable, and years of education and degree category as indicators of that latent variable. Since years of education and educational category do not perfectly correlate (r = .85 in the General Social Survey, r = .77 in the European Social Survey), estimates from latent models are more robust than those using basic Pearson correlation coefficients, as shown in Table 1.

Table 1.

Correlation Between the Educational Attainment of Spouses in the General Social Survey Within Waves by Method.

Method	Correlation
Latent	0.64
Years of education	0.61
Categories (ordinal)	0.56

We also tested whether this was the case for European countries using the ESS (n = 500k). The polychoric correlation, which estimates the true correlation between two latent normally distributed variables based on an observed ordinal variable, was used in place of the method that uses latent associations. Using the polychoric correlation instead of the Pearson correlation increased the magnitude of assortative mating for educational attainment within Europe from 0.52 to 0.62 – similar to the increase observed within America using the latent correlation (0.56 → 0.64). In addition, the rank order of the correlations between countries was roughly the same regardless of the method used (r = .91, n = 33), so using poor measurements of educational attainment should still result in the same rank order differences between countries. Admittedly, the polychoric correlation is only adjusting for the measurement error that occurs due to the use of a category variable, not the error that occurs due to the fact it's an imperfect indicator of educational attainment. In practice, they appear to have the same effect.

Given the large amount of data collected, it was tested whether certain variables correlated with assortative mating between nations. Differences in cultural values between nations were hypothesized to relate to assortative mating as culture may influence mate choices. For example, cultures with more hierarchical beliefs may select more strongly for indicators of social status. Besides this, income inequality between countries was hypothesized to correlate with assortative mating for educational attainment because the phenotypic assortment may increase the genetic and environmental inequality within the society. As assortative mating for education should theoretically lead to increased phenotypic inequality, it was tested whether cognitive variation, measured with PISA scores, correlates with assortative mating for educational attainment. To avoid confounders from biasing these estimates, HDI and regional classifications were used as covariates in the analysis between regions.

Data on country populations and HDI were collected from the United Nations (United Nations, 2023). Data on cultural values such as individualism, power distance, indulgence, masculinity, uncertainty avoidance, and long-term orientation were sourced from Hofstede (2023). Gini coefficients were obtained from the CIA (2023). Information regarding the methodological choices used when analyzing each dataset is available in Table 2, and a table that details these national variables in Table 3.

Table 2.

Methodological Decisions Made in the Calculation of Effect Sizes. If two Different Metrics are Named, Then it Means That a Latent variable was Constructed Based on Those Metrics.

Dataset	Weights	Measure of Educational Attainment	Method
PISA (International)	yes	5–8 categories	pearson correlation
PIRLS (International)	yes	7 categories	pearson correlation
WVS (International)	yes	ISCED categories	pearson correlation
GSS (USA)	no	years of education / 5 categories	latent
ESS (Europe, self-reports)	yes	12 categories	pearson correlation
ESS (Europe, child reports)	yes	5–12 categories	pearson correlation
Census (USA)	yes	years of education / 6 categories	latent
CPS (USA)	yes	years of education / 6 categories	latent
CFPS (China)	yes	years of education / 6 categories	latent

The latent models used in all cases were based on using years of education and degree as indicators of a latent variable, not polychoric correlations.

Table 3.

Timeranges, Number of Countries, and Sources of National Variables Tested Within This Study.

Variable	Time Range	Number of Countries	Source
Hofstede Individualism	1967–1973	75	Hofstede, 2023
Hofstede Power Distance	1967–1973	75	Hofstede, 2023
Hofstede Uncertainty Avoidance	1967–1973	75	Hofstede, 2023
Hofstede Masculinity	1967–1973	93	Hofstede, 2023
Hofstede Long Term Orientation	1967–1973	93	Hofstede, 2023
Hofstede Indulgence	1967–1973	93	Hofstede, 2023
Gini Coefficient	2023	178	CIA, 2023
HDI	2000–2013	188	United Nations, 2023
Standard Deviation of PISA Scores	2000–2018	70	Self-computed
Looseness	2000	68	Uz, 2014
Postmaterialism	2010–2014	59	Nový et al., 2017

In the European Social Survey, General Social Survey, CPS, US Census, and World Values Survey, effect sizes were also separated by cohort to detect secular trends in assortative mating. The same process was done for the parents of the respondents. Cohort assignments were based on the year of birth and the year of data collection. If reports of educational attainment were obtained from respondents, then their cohort is assumed to be the year of data collection minus their age. If the reports were taken from the children, then the cohort was assumed to be the year of data collection subtracted by the age of the child subtracted by 30. This is because the average age at which mothers bear children is slightly under 30 (Wikipedia, 2023; United Nations, 2022), and as fathers tend to be slightly older than mothers, 30 would be a rough estimate of the age of the average parent at the time of the birth of their children. There are two exceptions to this - the PISA dataset, where parents were assumed to be 45 years old, and the PIRLS dataset, where parents were assumed to be 40 years old. This is because the PISA test-takers were 15 years old, and the PIRLS test-takers were 10 years old. If the age of the respondents is missing, it is assumed to be 40, the average age of respondents within this meta-analysis.

Methodology

Various moderators were tracked when collecting effect sizes, including:

• Source of the report: typically, educational attainment is reported by the children of the parents or the parents themselves. Reports from parents were used as the reference group. Not adjusting for this bias would lead to an inflated meta-analytic correlation because reports of educational attainment from the children of parents tend to be more highly correlated in comparison to reports that are taken from the parents themselves.

• Country: the country in which this data was collected.

• Year: if the study was sampling the assortative mating of the participants, the sampling year or the study year was used. If the assortative mating of the parents of the respondents was used, the birth date of the participants was used instead.

• Source: the source of this data. All academic sources were grouped into one category, as were international datasets (World Values, PISA, PIRLS). The reference group are the national datasets (GSS, ESS, Census, CFPS).

• Cohort: the year in which these individuals were born.

• Method: whether years of education, degree categories, or latent methods were used to calculate the correlation in education between spouses. Reference group is the latent method.

• Country: it was tested whether countries differed in educational attainment.

The descriptive statistics of the moderators have been provided in Table 4. A few of the assortative mating correlations were above 1 due to modeling error; there are only two effect sizes with assortative mating coefficients above 1 and they had sample sizes of 16 and 23, respectively. The maximum and minimum ages are not whole numbers because they were estimated from ranges.

Table 4.

Descriptive Statistics of the Study.

Variable	Minimum	Maximum	Mean (SD)
Assortative mating for education	−0.0054	1.07	0.61 (0.11)
Sample size	11	417746	5403 (14955)
Uses sample weights	0	1	0.85 (0.36)
Year of data collection or study	1920	2022	2002 (21)
Age	24.5	69.5	40.5 (7.4)

Given that assortative mating for educational attainment was found to vary between cohorts, the meta-analysis was restricted to individuals born on or after 1950. To avoid differences in assortative mating patterns within the youth between nations from biasing the estimates, assortative mating correlations that were generated within the age category of 15–34 was removed from the between-country comparisons. The age categories of 30–60, 30–36, 36–40, 41–45, 46–50, 51–55, 25–55, 35–44, 45–54, 55–65, 65–74, 30–37, 32–57, 40, 40–70, and 16–99 were deemed sufficiently representative and included in the analysis. To avoid unreliable estimates of individual countries causing the international relationships to be less accurate, countries with a sample size of under 2500 or that only had one effect size available were removed from the meta-analysis.

This search for effect sizes concluded with the discovery of 2559 effect sizes covering 13.826.069 individuals born from 1861 to 1998 across 106 different countries. The exclusion criteria resulted in a reduction of the amount of effect sizes in the meta-analysis to 1498, the number of individuals sampled to 9.159.098, and the number of countries to 84. Fortunately, most countries had samples of above 10,000 individuals, which makes the estimations of assortative mating for educational attainment within countries highly precise, as shown in Figure 2.

Figure 2.

Sample size by country.

Given that the influence of each moderator could vary by country, four different meta-analyses were conducted to test the best way to compute effect sizes. Countries have different educational systems; thus, the categorization may have different effects depending on the country. Therefore, models where the effect of categorization was allowed to vary by country were considered. In addition, the effect of data sources on the assortative mating of each country may vary, so some models allowed for the effect of individual sources to vary. Based on these two choices, four different models were considered based on whether they considered ones of these interactions or not. Besides there four models, two others were considered: one in which random effects meta-analyses are conducted within individual countries without adjusting for moderators, and a fifth model that averaged all five of these estimates of the four methods. Then, the national differences in assortative mating that were estimated with these models were contrasted with variables that are the most strongly related to assortative mating for education across countries. It was determined that the meta-analysis that did not consider interactions had more predictive validity than the others, as demonstrated in Table 5, so it was used to estimate differences between countries in assortative mating. The modeling statistics of each meta-analytic regression can be found in Table 6.

Table 5.

Correlation Between Assortative Mating for Education and 4 Different Variables. in the Case of Regions, the Effect Size is the Amount of Variation in Assortative Mating for Education That can be Explained by Regional Differences.

Method	Gini Coefficient	Hofstede IDV	Region
No interactions	0.38***	−0.47***	0.30***
Country x Categorization Method	0.29**	−0.37***	0.28***
Country x Data Source	0.19	−0.27*	0.27**
Both interactions	0.2	−0.32*	0.26**
Within-country meta-analysis	0.35**	−0.48***	0.33***
Composite	0.29**	−0.40**	0.29***

*** → p < .001, ** → p < .01, * → p < .05. Hofstede IDV - individualism.

Table 6.

Modeling Statistics of Each Mixed Effects Meta-Analytic Model.

Parameter	Model 1	Model 2	Model 3	Model 4
Child-Reporting	Yes	Yes	Yes	Yes
Country	Yes	Yes	Yes	Yes
Format	Yes	Yes	Yes	Yes
Source	Yes	Yes	Yes	Yes
Cohort	Yes	Yes	Yes	Yes
Categorization Format x Country	No	Yes	Yes	Yes
Data Source x Country	No	No	No	Yes
I^2	90.82%	89.22%	85.56%	85.57%
R^2	76.85%	80.57%	86.15%	86.14%

Previous literature has attempted to assess the relationship between assortative mating for educational attainment and income inequality (Eika et al., 2019). Given that this study managed to estimate assortative mating for educational attainment in a large number of countries, it is possible to test for whether income inequality and assortative mating for educational attainment are interconnected. While the Gini coefficient and assortative mating for educational attainment correlate positively (r = .37, p < .001), it is unclear whether this relationship survives applying controls. To determine this, Bayesian Model Averaging was employed, a technique that accounts for model uncertainty by considering a large number of candidate models, selecting the most likely models, and then determining the independent effect of each predictor based on these models (Hinne et al., 2019). From this analysis, the posterior inclusion probability (PIP), the probability that each predictor would be in the hypothesized ideal model; and the expected value (EV), the expected effect of each predictor are calculated.

Some prior studies, including (Eika et al., 2019), have reported that assortative mating for educational attainment has increased in the United States. However, many of these papers use categories to measure educational attainment, which may not be invariant over time in their measurement of educability. To ascertain if the method of measuring educational attainment has an impact on the secular trend, trends in educational attainment were calculated for three different methods of measuring educational attainment: categories, years of education, and latent educational attainment, a latent variable constructed from years of education and categories. In order to make the analysis more comparable to the one from Eika et al., only data from the Census and CPS was used in this particular analysis. According to a mixed-effects meta-analytic model which predicted the correlation of educational attainment between spouses within these datasets, the increase in assortative mating over time only holds when categories are used as a measurement of educational attainment (correlation between cohort and strength of assortative mating = .59, p < .001). When years of education (r = -.31, p < .001) or latent methods (r = -.22, p = .021) are used, the relationship between assortative mating for education and cohort reverses. The moderator table of the mixed-effects analysis is available in Table 7, and a figure which plots the interaction is displayed in Figure 3.

Table 7.

Results of a Mixed Effects Model Which Tests the Impact of the Moderators on the Effect Size.

Moderator	Estimate
Intercept	0.52 (0.0048)***
Method: Latent	0.16 (0.0067)***
Method: Years of Education	0.11 (0.0068)***
Cohort	0.0010 (0.000070)***
Latent x Cohort	−0.0012 (0.000098)***
Years of Education x Cohort	−0.0012 (0.000098)***
I^2	99%
R^2	80.14%

The cohort variable was adjusted to have a minimum of 0. Parameter estimates are unstandardized.

Figure 3.

Cohort trends in assortative mating for education in American partners by method used to measure educational attainment. Linear trends were calculated using a mixed-effects meta-analytic model which used cohort as a moderator. 95% CIs are shaded on the outside.

The cause of the statistical artefact is that the relationship between years of education and educational attainment categories has increased with the passage of time, so the categories become better indicators of educational attainment. Figures 4 and 5 plot the relationship between the correlation between years of education and educational attainment categories in women and men respectively. The presence of two clusters after 1950 in both figures is due to the fact that the two datasets used have different correlations between years of education and educational attainment categories.

Figure 4.

Cohort trends in the correlation between years of education and educational attainment categories within women.

Figure 5.

Cohort trends in the correlation between years of education and educational attainment categories within men.

Because of this artefact, in examining secular trends in assortative mating for educational attainment, only datasets that used a large number of categories, latent methods, or educational attainment were selected. Only the ESS (12 categories), GSS (latent), and American Census/CPS (latent) datasets met this criterion. To increase the power of this analysis, effect sizes were calculated for individual cohort years, and trends within major regions were tested first. Next, the relationship between cohort and assortative mating in each country was assessed; if the p-value of the linear or non-linear relationship was below 0.01, the country's secular trend in assortative mating was plotted next to its regional neighbours. In this analysis, corrections were made for using child reports within regions. If a country had a non-linear relationship between time and the strength of assortative mating, a spline was used to plot the relationship between assortative mating and time, otherwise linear analysis was used.

Secular trends in assortative mating were calculated using the entire dataset of effect sizes, though these are potentially biased by changes in the validity of the measurement of education across time, and should be treated with skepticism - particularly the increases. Countries that had less than 11 observations or had meta-analytic models that failed to converge were removed from this analysis. The same methodology that was used to plot the trajectories earlier was used to plot these trajectories as well, with the exception that the effect of the data source was controlled for in this analysis. The only countries that reached significance in the extended data that did not reach significance in the high quality are the Philippines and Serbia. These analyses are provided in the Appendix.

Results

The mean correlation between partners in educational attainment was 0.66 (95% CI: [0.64, 0.68]) when weighed by the square root of the population sizes (0.64 otherwise). There was a large amount of heterogeneity in effect sizes (I^2 = 93%), though it is to be expected when the quantity of data collected is taken into account (n = 9.2 M, k = 1498). The details regarding the parameter estimates for the model that was used to calculate the international differences are presented in the supplement. National differences in assortative mating for education have been posted in Figure 6.

Figure 6.

Assortative mating for educational attainment by country.

There were substantial international differences in assortative mating (F = 5.5, p < .001), ranging from .52 in Nordic countries to .76 in Sub-Saharan Africa, as displayed in Table 8.

Table 8.

Regional Differences in Assortative Mating for Educational Attainment.

Region	Correlation	Number of Countries	95% CI
Sub-Saharan Africa	0.76	4	[.70, .82]
East Asia	0.69	14	[.65, .73]
Latin America and the Caribbean	0.67	11	[.62, .72]
Western Asia and Northern Africa	0.66	17	[.63, .70]
Balkan	0.65	5	[.61, .69]
Southern Europe	0.64	5	[.58, .71]
Northern America	0.62	2	[.55, .68]
Continental Western Europe	0.60	7	[.57, .63]
Eastern Europe	0.59	14	[.55, .63]
Anglo countries outside North America	0.55	4	[.52, .58]
Nordics	0.52	4	[.51, .54]

Two rounds of Bayesian model averaging were run - one that included every single predictor that had a statistically significant correlation with assortative mating, and one that only included predictors with a posterior inclusion probability (PIP) above 10%. This was done because the model averaging that included only four predictors had less missing values, so it had enhanced power. Bayesian model averaging suggested that the strongest predictors of assortative mating were individualism (PIP = 81.9%) and power distance index (PIP = 77.2%). Gini coefficient was not a predictor of assortative mating for educational attainment (PIP = 0%) in the second round of Bayesian model averaging. Controlling for Hofstede's individualism measurement alone makes the relationship between assortative mating for educational attainment and the Gini coefficient statistically insignificant (b = 0.0019, p = .15). Statistics related to the modeling are available in Table 9 and the correlation matrix of the national variables is available in Table 10.

Table 9.

Expected Value of Each Parameter in Each Round of Bayesian Model Averaging.

Parameter	Round 1	Round 2
Individualism	−0.46 (88.2%)	−0.33 (81.9%)
Power Distance Index	0.91 (26.6%)	0.31 (77.2%)
Gini Coefficient	0.11 (34.7%)	0.0 (0%)
PISA Standard Deviation	−0.0038 (1.6%)	Not included
Looseness	0.0 (0%)	Not included
HDI	−0.081 (24.6%)	0.0 (0%)
Number of Countries	42	60
Adjusted R2 (most likely model)	0.5	0.44

Posterior inclusion probability is in parenthesis. Not shown: regional controls.

Table 10.

Correlation Matrix of the various Variables in the Dataset.

	IDV	PDI	IVR	LTO	MAS	UAI	GINI	AME	PSD	HDI	LOS
IDV
PDI	−0.60***
IVR	0.157	−0.306*
LTO	0.084	0.046	−0.449***
MAS	−0.178	0.251*	−0.067	−0.032
UAI	0.078	0.098	0.079	0.016	−0.062
GINI	−0.509***	0.323**	0.196	−0.401***	0.039	0.047
AME	−0.606***	0.526***	−0.183	−0.195	0.105	0.213	0.373***
PSD	0.542***	−0.384**	0.187	0.196	0.23	0.16	−0.347**	−0.278*
HDI	0.656***	−0.577***	0.255*	0.376***	−0.089	−0.038	−0.318***	−0.556***	0.632***
LOS	0.512***	−0.382**	0.371**	0.237	0.061	−0.098	−0.117	−0.321*	0.501***	0.61***
PSM	−0.118	−0.317	0.354	0.073	−0.372	−0.318	0.321	−0.211	−0.054	0.352*	0.396

IDV - Individualism, PDI - Power Distance Index, IVR - Indulgence, LTO - Long-term orientation, MAS - Masculinity, UAI - uncertainty avoidance, GINI - Gini coefficient, AME - assortative mating for educational attainment, PSD - PISA score standard deviation, HDI - Human Development Index, LOS - Looseness, PML - Postmaterialism. * → p < .05, ** → p < .01, *** → p < .001

The two strongest predictors of assortative mating for educational attainment between countries were Hofstede's individualism measurement and HDI. The relationship between these two variables and assortative mating for education between countries has been plotted in Figures 7 and 8.

Figure 7.

Relationship between individualism and assortative mating for educational attainment.

Figure 8.

Relationship between HDI and assortative mating for educational attainment.

As the United States has a large amount of data, extended analysis of the country's secular trends in assortative mating was done. A model which used restricted cubic splines explained much more variance in effect sizes (R^2 = .63) than a linear model (R^2 = .049), so it was used instead. According to the spline model that used sample sizes as weights and cohort as the independent variable, assortative mating for education decreased from 1875 to 1926, increased from 1926 to 1945, decreased from 1945 to 1958, increased from 1958 to 1977, and decreased from 1977 onwards. The time trend has been plotted in Figure 9.

Figure 9.

Secular trends in assortative mating in the USA. Controlled for the effect of child-reporting and format. 95% CI is shaded in grey.

As not as much data was available for individual European countries, they were first categorized based on whether they were Capitalist or Communist countries during the Cold War. It was found that assortative mating for educational attainment decreased in Communist countries within the cohorts that were raised under Communism, but increased in the cohorts who were getting married after the Berlin Wall fell. Secular trends within the United States, Capitalist Europe, and Communist Europe have been plotted in Figure 10. In the cases of both Communist (F = 4.2, p < .001) and Capitalist (F = 8.2, p < .001) Europe, the nonlinear trend was a better fit for the data than the linear fit.

Figure 10.

Secular trends in assortative mating for educational attainment within major regions. 95% CI intervals have been shaded in.

Countries that were identified to have secular trends in assortative mating based on a p-value threshold of 0.01 were plotted in Figures 11, 12, and 13 next to their regional neighbours. Within North-east Europe, all countries have decreased in assortative mating, except for Czechia, which increased in assortative mating within recent cohorts. The same is true for South-east Europe, with all countries examined experiencing a decline. Within North-west Europe, all identified countries increased in assortative mating for educational attainment, except for Sweden, which experienced a decrease.

Figure 11.

Trends in assortative mating within north-east Europe.

Figure 12.

Trends in assortative mating within south-east Europe.

Figure 13.

Trends in assortative mating within North-west Europe.

Discussion

This study replicates the prior literature, suggesting that assortative mating for education exists and is large in magnitude. However, the correlation found in this study (.66) is reasonably higher than the .55 correlation found in another meta-analysis (Horwitz et al., 2022). This discrepancy is primarily because this meta-analysis used latent methods of measuring educational attainment as the reference group, which have higher correlations because they measure educational attainment more accurately.

The Gini coefficient correlated with assortative mating for educational attainment (r = .37, p < .001); however, it was not a robust predictor when other variables such as region and culture were considered (PIP = 0%), according to Bayesian model averaging. This does not imply that assortative mating doesn't contribute to income inequality, as the analysis was not powerful (n = 60). Within our dataset, the variables that correlated the strongest with assortative mating between countries were Hofstede's individualism dimension (r = -.61, p < .001) and HDI (r = -.56, p < .001), though the causal nature of these relationships cannot be determined at the moment.

No substantial net increase in assortative mating has taken place in the United States. Previous findings of increases (Eika et al., 2019) can be attributed to categories becoming increasingly valid predictors of educational attainment. This lack of a change mirrors prior literature (Gihleb & Lang, 2020) which has also noted that increases in assortative mating for education found in the literature are sensitive to changes in statistical methods. These findings challenge the argument of Murray (2013), who contended that increasing educational assortment in the United States caused increases in class divides in the United States. While it could still be possible that class divides are increasing within the United States, this cannot be ascribed to assortative mating for education due to the relative stability of this metric across time. This concords with other researchers who studied trends in assortative mating in other regions such as England (Clark & Cummins, 2022) and Norway (Heath et al., 1985) and observed stable levels of assortative mating. Any reports of changes in assortative mating for education across time that do not test whether the method of measuring educational attainment affects the results should be viewed with skepticism, as these effects are sensitive to changes in methodology.

While there has been no net increase in assortative mating within the United States, the strength of the relationship fluctuated depending on the time period. Notably, assortative mating for education reached a low point in the cohort born during the roaring 20 s, a high point in the cohort born during World War 2, another low point in the cohort born during the baby boom in the 50 s, and a high point in the cohort born in the 70 s. It is not clear how exactly these increases and decreases came to be, but they seem to overlap with shifts in political life in the United States - assortative mating peaks in cohorts born during times of strife, and bottoms in cohorts born in times of economic and social harmony. Admittedly, this is just a post-hoc theory, and the only other piece of evidence that supports it is an observed negative correlation between socioeconomic development and assortative mating for educational attainment (r = -.56) between countries.

Within the United States, years of education and educational attainment categories have become increasingly correlated over time. This appears to be a novel finding, as no other researchers have been observed reporting it. It is possible that it could be due to political standardization, where schools become more similar to each other due to the influence of the federal government. Alternatively, it could be due to fewer people dropping out of school, leading to individuals within educational degree categories sharing similar years of educational attainment.

In Capitalist Europe, assortative mating across time has slightly increased, while Communist Europe's assortative mating for education has decreased. There are a few countries that do not follow this pattern, such as Sweden which experienced a decrease, and Czechia which experienced an increase. Considering the negative correlation between assortative mating for educational attainment and individualism, it's possible that global westernization may be driving decreases in assortative mating for educational attainment in certain regions. However, more work needs to be made to determine whether these time trends in Europe are sensitive to changes in statistical methodology, like the American ones.

Appendix

Table A1.

Effect of Cohort on Assortative Mating for Educational Attainment Within High Quality Data (ESS, GSS, American Census Data).

Country	p-value (linear)	beta (linear)	p-value (nonlinear)	k
Finland	0.01172431033	−0.001166614524	0.004139275347	157
Germany	9.43E-20	0.003981064145	0.009254411561	155
France	0.01468205929	0.00122063242	0.0319219396	153
United Kingdom	0.003326699626	0.001629328552	0.01808101978	151
Belgium	0.6468491915	2.28E-04	0.01544502029	153
Bulgaria	0.01385316547	−0.001457106316	0.06905056557	147
Greece	0.3759063881	0.0007853369282	0.104682678	133
Norway	0.009462176589	0.001391171721	0.1359822462	153
Portugal	0.1392677044	0.0009270586415	0.03571590125	152
Sweden	1.64E-02	−0.001214442398	0.08087666612	153
Switzerland	3.42E-06	0.002555759726	0.9749746634	152
Denmark	0.9065970304	−6.70E-05	0.004777721984	152
Spain	0.08710275309	0.0008580800912	0.04771315063	151
Hungary	0.04142307944	0.001188532855	0.6553184287	150
Ireland	0.03512083653	−0.0009473627051	0.7376653006	149
Israel	2.91E-03	−0.001414319379	2.95E-10	152
Slovakia	0.06991610487	−0.001446143388	0.165874255	138
Netherlands	0.09629737915	0.0008637899081	0.004399440584	149
Cyprus	0.848343735	−0.0001625200391	0.4424815373	139
Czechia	0.5538415055	0.0003691889616	9.31E-01	145
Estonia	4.52E-08	−0.002802586936	0.06105931693	151
Slovenia	1.08E-05	−0.00266207109	0.5470588769	147
Ukraine	0.07587501688	−0.001531889145	0.03774348529	138
Lithuania	5.95E-12	−0.003675714304	6.01E-05	147
Poland	4.95E-10	−0.003254113589	0.9061710666	146
Croatia	9.30E-05	−0.003231311912	0.1345117189	138
Russia	4.83E-04	−0.002160020673	0.7028545184	140
Austria	0.2674673242	0.000853187009	7.77E-01	142
Italy	0.0007523106039	−0.00202902428	0.3018654489	139
Iceland	0.1869010351	0.00126674154	0.02084616258	136
Serbia	0.7841338888	−0.000306076491	7.59E-05	127
Kosovo	0.6590061822	−0.0008390190479	7.59E-01	108
Latvia	0.8625301412	0.0003824861085	0.1951726059	88
Albania	0.05511173355	−0.002807313113	0.1372181683	116
Montenegro	0.942263601	−0.0001313890367	0.6687929334	109
United States	6.43E-04	0.0001439032956	1.42E-31	873

Table A2.

Effect of Cohort on Assortative Mating for Educational Attainment Within all Data.

Country	p-value (linear)	beta	p-value (nonlinear)	k
Albania	0.0551117	−0.0028073	0.1349293	118
Argentina	0.3225574	0.0028773	0.6686990	13
Australia	0.1513922	−0.0015621	0.1533108	14
Austria	0.2674673	0.0008532	0.7776580	146
Bulgaria	0.0213235	−0.0013429	0.0718323	157
Brazil	0.6931445	0.0007195	0.5528437	14
Canada	0.2149549	0.0008381	0.6907315	14
Switzerland	0.0000034	0.0025558	0.9761761	156
Chile	0.0222216	−0.0031667	0.3735015	20
Czechia	0.4470964	0.0004459	0.8805645	160
Germany	0.0000000	0.0039961	0.0284144	183
Denmark	0.7955402	−0.0001511	0.0034119	164
Finland	0.0117243	−0.0011666	0.0034383	162
France	0.0146821	0.0012206	0.0320749	157
United Kingdom	0.0023018	0.0016688	0.0182073	162
Greece	0.7823696	0.0002040	0.1095327	148
Hong Kong SAR China	0.9008538	−0.0001543	0.2529578	28
Indonesia	0.1747785	−0.0013363	0.5449937	12
Iceland	0.4399884	0.0007623	0.0170889	147
Italy	0.0018056	−0.0018363	0.2710422	163
Japan	0.8573197	0.0002546	0.8404893	14
South Korea	0.4729154	−0.0010129	0.2461616	47
Latvia	0.8625301	0.0003825	0.1984509	92
Mexico	0.7921020	0.0004085	0.3859934	17
Netherlands	0.2771179	0.0005199	0.0033732	166
New Zealand	0.8701914	−0.0002325	0.3668005	22
Peru	0.0377035	−0.0021074	0.0529312	13
Poland	0.0000000	−0.0033980	0.8560302	156
Portugal	0.1852235	0.0007973	0.0233159	169
Romania	0.9432117	−0.0000749	0.1661069	13
Russia	0.0000266	−0.0023432	0.7801816	155
Sweden	0.0164172	−0.0012144	0.0520571	158
Thailand	0.5506352	−0.0009865	0.1426428	14
United States	0.0009029	0.0001411	0.0000000	893
Macao SAR China	0.0863587	−0.0024504	0.3627252	26
Slovakia	0.0608624	−0.0013314	0.2549681	151
Tunisia	0.9102216	−0.0002799	0.9154642	12
Turkey	0.0740254	−0.0028574	0.0352886	20
Uruguay	0.3810789	0.0023574	0.6031953	13
Colombia	0.5812903	−0.0008770	0.5018129	19
Croatia	0.0000003	−0.0038648	0.1125689	153
Lithuania	0.0000000	−0.0036757	0.0000583	151
Estonia	0.0000000	−0.0028026	0.0604803	153
Montenegro	0.9422636	−0.0001314	0.6685913	111
Singapore	0.2035659	−0.0014222	0.8148455	13
Slovenia	0.0000108	−0.0026621	0.5406471	150
Vietnam	0.1756928	0.0036321	0.8455048	12
Jordan	0.6345097	−0.0006426	0.2393881	11
Kazakhstan	0.0339610	−0.0030282	0.9022720	11
Lebanon	0.2667970	−0.0021274	0.0613826	11
Morocco	0.2533677	0.0042222	0.1667919	11
Malaysia	0.8623567	−0.0003924	0.2526559	11
Philippines	0.0043194	−0.0055862	0.6569749	11
Serbia	0.2704560	−0.0009760	0.0001890	138
Ukraine	0.0493563	−0.0014276	0.0295572	149
Cyprus	0.7491394	−0.0002461	0.3582815	150
Iran	0.5601345	0.0013046	0.0697877	11
China	0.0256207	0.0029619	0.0839647	13