Non-random mating patterns within and across education and mental and somatic health
Studying the complete set of first-time Norwegian parents, we found positive partner correlations in GPA, EA, and all analysed mental and somatic health conditions, observed from 10 to 5 years before the birth of a couple’s first child. The initial similarity and later convergence were larger for mental than somatic health conditions. We also observed ubiquitous cross-trait correlations for mental health conditions, which in prospective analyses were approximately as large as the within-trait correlations. The pattern of correlations between relatives indicated deviations from direct assortment on several of the observed phenotypes. Although partner correlations could be partially explained as by-products of assortment related to education, this cannot be a primary explanation of partner correlations in mental health.
Mental health in early adulthood associated with partner selection
Our study expands on previous research by including the whole population, studying diagnosed health conditions, and contrasting the importance of mental versus somatic health conditions. To minimize the influence of convergence, we examined young adults before parenthood and typically before partnership formation. We demonstrate that partners resemble each other in mental health before they are likely to have met. As far as we are aware, partner resemblance in mental health assessed before couple formation has previously only been found for self-reported symptoms in a cohort study29. Our prospective analyses and use of proper diagnoses indicate that there is assortment on the liability to mental disorders, as questioned by Yengo11. The lack of correlations between partners’ polygenic indices in previous studies is likely due to limited discovery samples and small effects of each causal variant, giving the polygenic indices low predictive value for mental health conditions. An alternative explanation is that overrepresentation of healthy and well-educated individuals in cohort studies restricts the range and downwardly biases partner correlations. For example, we observed a partner correlation of 0.48 for EA, compared to 0.42 in a Norwegian cohort20. However, for mental health, our estimates of correlations between partners-to-be were slightly lower (median r = 0.13) than in a cohort study assessing global mental health among future partners (r = 0.16)29. Our study indicated that mental health conditions were more strongly related to partner selection than somatic health conditions common in young adulthood. This is not surprising, given that mental health is linked with marriage and fertility30 and could indicate desirability to potential partners.
Partner correlations in mental health were considerably higher at the end than at the start of the observational period. This highlights that studies on established couples can typically only inform on correlations, and that convergence needs to be addressed before interpreting correlations as indicative of assortment2,16. This increase does not necessarily reflect mutual influences or shared experiences; it could also be that partner selection is based on vulnerabilities to mental disorders that manifest as diagnosable conditions later in life (indirect assortment). We observed a change in resemblance from 10 to 5 years before childbirth until the years surrounding childbirth in the same couples—the increased resemblance may be even more pronounced among older couples.
Assortment across mental health conditions is ubiquitous
The cross-trait correlations for different mental health conditions in the two partners were almost as strong as within-trait correlations (median r = 0.14 vs 0.13). Hence, individuals tend to mate with partners who share similarly good or poor mental health, with the specific type of health condition being subordinate. Such results align with assortment on perceived attractiveness, itself influenced by both mental and educational traits. Thus, the partner correlations observed across different traits likely reflect indirect assortment. Our results differ from a study that found assortment primarily on symptoms of specific disorders8. However, that study used data on established couples, which, according to our results, have increased in within-trait correlations.
The positive manifold across mental health conditions in partners can in the next generation increase genetic correlations between traits; not because the same set of genes are associated with different traits, but because genetic liabilities to different traits co-occur in the same individuals4. This can contribute to the frequently observed “p-factor”31. In addition, cross-trait assortment can easily lead to bias in genetic studies, as unmeasured genetic variants can be related to measured variants as well as the outcome of interest. This can inflate estimates in genome-wide association studies and violate the exclusion criteria in Mendelian randomization studies19. In the presence of cross-trait assortment, the results of such studies should be interpreted with caution.
Individuals with better grades or higher education were less likely to have partners with mental and somatic health conditions. This suggests a trade-off between different attractive traits in partners, indicating competition for healthy partners rather than matching on similarity. Nevertheless, the remarkably high correlation for substance use disorder could indicate genuinely different lifestyle preferences. Modelling of variations in preferences may be vital to fully understand cross-trait assortment3,32. There was little assortment across different somatic conditions or across mental and somatic conditions. Still, most correlations were positive, and mostly so among those involving various types of pain, possibly reflecting the mental aspect of pain. Regardless of genetic consequences, the widespread cross-trait assortment could enhance negative consequences for children, as they may be influenced by both low education and poor mental health in their parents33,34.
Partner correlations are generally inconsistent with direct assortment
When accounting for assortative mating to avoid bias, studies make assumptions about the mechanisms involved. Typically, they assume direct assortment on the observed variables35,36. Our results challenge this notion. The siblings-in-law correlations exceeded those expected under direct assortment, suggesting that direct assortment is not a sufficient explanation for partner resemblance and that studies relying on this assumption can be biased. Deviations from direct assortment have previously been reported for EA20,24,25,37. We extended this observation to GPA and a range of health conditions. Our results align with another study that observed deviations from direct assortment in 29 of 51 traits27, mainly different traits than those studied here.
Although the phenotypic model could be falsified, the underlying mechanisms remain elusive. Both indirect assortment and social stratification38 could increase in-law correlations disproportionately and explain our observations. In any case, partner resemblance is not solely due to assortment based on the observed phenotypes. Whether parts of the partner correlations in mental health are due to causal influences on partner choice remains to be determined. Identifying the traits that actively determine assortment is an important question for future studies. It might be more strongly related to general vulnerability to psychopathology31 than to specific disorders. Due to the strong cross-trait assortment, such causal effects may be more plausible at the level of general mental health, rather than for specific diagnoses. A previous study indicated that partner similarity in many traits was driven by assortment on a few key traits38, but it did not include mental disorders. Future studies may explore whether partner resemblance across many traits can be more parsimoniously explained by assortment on one or a small number of dimensions.
Indirect assortment need not be based on symmetric assortment on a manifest phenotype. Measurement error can be indistinguishable from indirect assortment on an unknown trait. Assortment may then be said to be direct for the true values of a trait, but indirect for an imperfect indicator. As measurement error is widespread and relatively easy to estimate, accounting for measurement error could improve future studies on assortment. Indirect assortment could also be related to impression management, whereby partner selection could take place on successful misrepresentations of one’s characteristics. This should, however, not influence sibling correlations. Finally, correlations in trait preferences among siblings can increase correlations between distant affines, such as co-siblings-in-law32. Hence, models of preferences may be needed to fully understand similarities in wider family networks.
Assortment leads to correlations between all genetic and environmental influences in one partner and those in the other. When parental traits leave a mark on their children through vertical transmission, this assortment leads to an intertwining of genetics and environment in the children. This can substantially increase gene-environment correlations in the child generation, which again increases the genetic similarity between partners9 (formula S1.8). If there is indirect assortment, the partner similarity in assorted factors will be larger than indicated by the observed variables, and the intergenerational consequences can be underestimated. Intergenerational studies therefore need to carefully model indirect assortment. Regardless of mechanism and possible genetic consequences of assortative mating18, the potential social consequences of partnership composition could remain.
Assortment on educational attainment partially explains health similarity
Given the known correlation between education and health status, one should expect a partner with higher education to, on average, also enjoy better health. Indeed, when we adjusted for both partners’ GPA or EA, partner correlations within and across mental disorders were reduced. Hence, similarities in mental health could to some degree be by-products of on education or its precursors. Nevertheless, correlations within and across mental disorders remained significant, indicating that these were not solely by-products of assortment based on education. Hence, mental health is related to partner selection independently of observed education. Partner correlations within and across different somatic health conditions were close to zero both before and after these adjustments.
It must be noted that EA was not measured prospectively; at the young age of approximately 20 years, many individuals are yet to obtain their highest education. Individuals can select partners based on the traits that exist at this age and that lead to later EA, in which case the adjustment is defendable. However, it is also possible that the adjustment for EA is an overadjustment because one’s own or the partner’s health could influence education. Using GPA as an alternative indicator of educational potential reduces this issue because it is typically achieved before partners meet. However, each partner’s mental health could have influenced their own GPA, meaning that the true assortment on mental disorders could in principle be slightly larger than indicated by our study. GPA was somewhat less strongly linked to the partner’s health than EA was. This could suggest that traits that influence EA are more important for mate choice than traits influencing GPA. Interestingly, siblings were more similar in GPA than EA, but this was reversed in partners. Cognitive abilities and conscientiousness influence both GPA and EA, but there could be differences in ambitions, achieved status, or social background. Roughly half of the variance in EA was not shared with GPA, indicating that there are important differences between the two.
The current study indicates, as do also previous studies20,25, that the strong partner resemblance in EA is due to even stronger resemblance in an unobserved factor. This was also the case for GPA. Assortment for EA and GPA was itself indirect; therefore, unidentified factors must exist that contribute to partner similarity in education as well as in health. This aligns with a previous study that chain-linked in-laws and inferred far greater partner similarity in latent (unidentified) advantages than in the observed level of education25. Our inclusion of GPA early in life is novel, however, it did not capture these latent factors any better than EA. Future research could try to identify traits that account for the sorting process and understand how they relate to partner similarity across observed traits. This may include social status more broadly defined or health in childhood.
Limitations
This study has some limitations that one should consider when interpreting the findings. First, the medical records are proxies for actual health conditions, as not all individuals with health issues seek medical care. This prevented the study of conditions below the threshold of medical attention. This issue is reduced as the tetrachoric correlations model these thresholds. Also, our use of primary care data captures a larger proportion of cases than specialist care data alone39, which has been used in previous studies2. This further mitigates potential biases. We could only study somatic conditions that were common among parents-to-be in young adulthood. The results are not necessarily representative for other somatic health conditions; in particular, assortment on rare health conditions is unknown. This also prevented the study of health conditions with a higher average age of onset, such as cardiometabolic conditions and cancers. However, conditions that develop after couple formation cannot directly influence its composition. Second, our focus on parents of children born in Norway between 2016 and 2020 could limit the generalizability to other populations or time periods. Third, we cannot rule out that some partners had already influenced each other at the start of the observational period in early adulthood. Nevertheless, the prospective nature of our study is a major advancement over previous studies, and the comparison with cross-sectional data emphasizes the impact of this analytic decision. The gap of 5 years between the end of health observation and the birth of the first child exceeds the median duration of relationships, suggesting that most couples were unacquainted during the health observation period. Fourth, we used tetrachoric correlations, based on the assumption of an underlying normally distributed liability. Whereas this could be reasonable for mental health conditions, some somatic health conditions are binary in their nature, such as fractures. This could lead to an over-estimation of partner correlations. However, this would, if anything, make the difference between mental and somatic health conditions larger. In addition, this did not affect the tests of direct assortment (Supplemental Scripts S1–S2), and results were consistent in logistic regression.
In conclusion, this study provides evidence for assortative mating patterns across GPA, EA, and 20 health conditions, up to 10 years before partners had their first child in data without participation bias. Among the health conditions, mental health conditions were particularly strongly related to partner selection. We observed vast cross-trait assortment for mental health conditions, indicating that individuals match on overall mental health, rather than on specific health conditions. The link with education might indicate trade-offs for overall attractiveness. This questions assumptions in genetic designs and could have consequences for the distribution of risk factors among children. In general, partner resemblance could not be explained with direct assortment, however, GPA or EA could only to a moderate degree account for partner similarity in mental health. The use of prospective data ensured that partner resemblance was not merely due to convergence, and the comparison with cross-sectional data indicates that studies without prospective data do not precisely reflect assortative mating. Indirect assortment appears the best explanation for partner similarity, raising important questions on mate choice and complicating modelling of partner similarity.
link
