Divorce rates remain high across the industrialized world (Perelli-Harris et al., Reference Perelli-Harris, Berrington, Sánchez Gassen, Galezewska and Holland2017; Smock & Schwartz, Reference Smock and Schwartz2020), and marital dissolution predicts poorer mental health, economic decline, and educational disadvantages for adults and their children (Kailaheimo-Lönnqvist et al., Reference Kailaheimo-Lönnqvist, Jalovaara and Myrskylä2025; Kalmijn, Reference Kalmijn2017, Reference Kalmijn2023; Leopold, Reference Leopold2018; Metsä-Simola & Martikainen, Reference Metsä-Simola and Martikainen2013). Explaining why some marriages dissolve while others endure has occupied researchers in sociology, psychology, economics, and behavioral genetics for decades. Within behavioral genetics, studies have investigated why divorce runs in families (D’Onofrio et al., Reference D’Onofrio, Turkheimer, Emery, Harden, Slutske, Heath, Madden and Martin2007; McGue & Lykken, Reference McGue and Lykken1992; Salvatore et al., Reference Salvatore, Larsson Lönn, Sundquist, Sundquist and Kendler2018), how it is related to getting married in the first place (Jerskey et al., Reference Jerskey, Panizzon, Jacobson, Neale, Grant, Schultz, Eisen, Tsuang and Lyons2010), how parental divorce affects children (D’Onofrio et al., Reference D’Onofrio, Turkheimer, Emery, Slutske, Heath, Madden and Martin2005, Reference D’Onofrio, Turkheimer, Emery, Slutske, Heath, Madden and Martin2006; Van Winkle & Baier, Reference van Winkle and Baier2025), and what traits are genetically correlated with divorce (Jocklin et al., Reference Jocklin, Mcgue and Lykken1996; Jørgensen et al., Reference Jørgensen, Cheesman, Andreassen and Lyngstad2025). Additionally, twin studies have provided heritability estimates varying from 15% (D’Onofrio et al., Reference D’Onofrio, Turkheimer, Emery, Slutske, Heath, Madden and Martin2005) to 59% (Jocklin et al., Reference Jocklin, Mcgue and Lykken1996) depending on the specific sample and operationalization.
It has been a productive research field, but there is a problem: Much of this empirical work implicitly treats divorce as an individual-level outcome despite the obvious fact that the end of a marriage is dyadic: No individual can divorce without their spouse becoming divorced too. Thus, whether an individual becomes divorced depends not only on their own liability, but also on the liability of the spouse and on factors specific to the couple. Treating divorce as an individual-level outcome may therefore yield results that are difficult to interpret, or worse, misleading. First, it is not clear what is or should be in the denominator for heritability estimates: the variance of individuals’ liability or the variance of couples’ liability? (Twin studies implicitly assume the latter). Second, individual-level estimates can be confounded by effects of the spouse (Eaves et al., Reference Eaves, Silberg and Maes2005), especially if the individual-level characteristics underlying the couple-shared outcome are subject to assortative mating.
Do individuals who are likely to divorce marry each other? If men and women with similar individual-level liabilities pair up more frequently than expected by chance, then variation in couples’ combined divorce liability will be greater than it otherwise would have been (per the sum of variances law), and high-liability individuals will be at risk of divorce in part because they typically marry other high-liability individuals. Assortative mating may thus be essential for a full understanding of divorce. Yet, spousal similarity is not straightforward to study for couple-shared outcomes. Unlike individual-level outcomes, where spousal similarity can be directly gauged with a correlation, couple-shared outcomes have no variation within couples and therefore no covariation between spouses either. Nevertheless, spouses do correlate on a range of divorce-related traits, including education, personality, mental health, and parental divorce history (Golovina et al., Reference Golovina, Gutvilig, Niemi and Hakulinen2025; Horwitz et al., Reference Horwitz, Balbona, Paulich and Keller2023; Kailaheimo-Lönnqvist et al., Reference Kailaheimo-Lönnqvist, Fasang, Jalovaara and Struffolino2021; Torvik et al., Reference Torvik, Sunde, Cheesman, Eftedal, Keller, Ystrom and Eilertsen2024). Spouses are therefore likely similar in their divorce liabilities too.
If there is assortative mating for liability to divorce, an individual’s liability will predict divorce partly because it is correlated with the spouse’s liability. This could plausibly account for some puzzling findings, such as McGue and Lykken’s (Reference McGue and Lykken1992) seminal twin study that reported divorce liability to be about 52% heritable for both men and women (see also Jocklin et al., Reference Jocklin, Mcgue and Lykken1996). Naively, this finding is logically impossible, because female and male factors combined would account for more than 100% of variation in couples’ divorce liability. (It was admittedly not significantly higher than 50%, and the authors did note that 50% heritability was the theoretical upper bound if both spouses contributed equally. Yet, their results implied that divorce was near perfectly heritable when combining female and male factors, which we find implausible.) It is plausible that these individual-level estimates are capturing effects of the spouse. McGue and Lykken (Reference McGue and Lykken1992) discounted assortative mating by pointing to a lack of spousal similarity for parental divorce history. Their reported contingency table implies a tetrachoric correlation of .04 (SE = .05), which is small and statistically insignificant (Supplementary Note 1). However, it is not significantly smaller than one could reasonably expect considering how indirectly it measures a person’s own divorce liability. For comparison, the parent-offspring correlation was only .17. A larger and more recent Finnish study did find significant associations between spouses’ parental divorce histories (equivalent to a tetrachoric correlation of .07, SE = .01) (Kailaheimo-Lönnqvist et al., Reference Kailaheimo-Lönnqvist, Fasang, Jalovaara and Struffolino2021).
We have identified two limitations in the literature: First, divorce has been implicitly treated as an individual-level outcome despite being a couple-shared outcome. Second, liability to divorce may be subject to assortative mating, which must be accounted for when estimating individual-level effects on couple-shared outcomes. In this article, we frame divorce as a couple-level outcome comprising female, male, and couple-specific (residual) liabilities, and investigate whether individuals with similar liabilities tend to marry each other. To accomplish this, we show how chain-linking couples and relatives can provide the necessary information to estimate spousal similarity in divorce liability. We then fit two models to data on extended families of twins and siblings in the Norwegian population registers. First, we fit a classic twin model, which we use as a point of comparison. We then extend the model by (1) including information on the original twins’ and siblings’ respective siblings-in-law (i.e., affines), and (2) change the specification so that phenotypic variance is the sum of female and male familial factors.
Materials and Methods
Sample and Measures
Our study is based on the Norwegian population register, which contains basic demographic information on all individuals living in Norway since 1960 (N = 8,589,458). The register includes individual-level data on births, deaths, marriage status, and parentage, among other things. Data on marriage status were available from 1975 to 2023, and contained information on whether someone was nonmarried, married, separated, divorced, or widowed, as well as the identifier for the spouse, if any. By comparing records against the previous year’s record, we identified all new marriages between 1976 and 2008. By linking individuals with their spouse, we were able to identify 592,637 complete, opposite-sex couples where both individuals were alive and living in Norway 15 years after their wedding, and where neither were first-generation immigrants (which we excluded because of incomplete information on marriage history and available relatives). The same individuals could at this stage form part of multiple marriages (see below). A total of 169,130 (28.5%) marriages ended in divorce or separation within 15 years.
We identified all pairs of twins and full siblings within our sample of marriages. Full siblings were identified by having the same mother and father, and twins were identified as having the same mother and birth month. Information about zygosity of twin pairs were obtained from the Norwegian Twin Register (Nilsen et al., Reference Nilsen, Brandt and Harris2019). Same-sex twins with unknown zygosity were not included. To create extended family units, we took the already identified pairs of twins and siblings and cross-joined all eligible same-sex full siblings of their respective spouses. (We only attached same-sex full siblings to limit the number of unique family structures, which drastically simplified the implementation of the models). This created 886,339 possible extended family units each comprising up to four marriages: The marriages of the two twins/siblings and the marriages of their respective siblings-in-law. We then pruned the data to make sure that no individual formed part of more than one extended family unit. To do this, we sorted the possible extended family units by priority and removed rows that contained individuals that had appeared before. We first prioritized monozygotic (MZ) twin units, then dizygotic (DZ) twin units, and then full sibling units. Within each priority level, we prioritized more complete units but otherwise randomized the order. Our final sample consisted of 124,544 extended family units comprising 353,210 marriages, out of which 86,401 (24.5%) ended in divorce or separation with 15 years (Table 1). There were 1196 units with MZ twins.
Table 1. Sample size and prevalence of divorce across extended family units
Note: Zygosity Group and Sex describe the relation between the middle couples (i.e., couples 2 and 3), which defines the type of extended family unit. The total number of units within each group can be read from the columns for couples 2 and 3, as they are always complete. Each cell presents the following numbers: Divorced/Total (Prevalence, Standard Error).
Note that each extended family unit included individuals who are socially very distant: For example, relative to the leftmost wife in Figure 1 (
${F_1}$
), her extended family unit will include information on her husband’s brother’s wife’s twin sister’s husband’s brother’s marriage. We use the term first-order affines to refer to the relations between couples 1–3 and 2–4 (i.e., your sibling’s spouse’s sibling), and second-order affines to refer to the relation between couples 1–4 (i.e., your sibling’s spouse’s sibling’s spouse’s sibling). The relations between couples 1–2, 2–3, and 3–4 are always full siblings or twins.
Figure 1. Path diagram for the extended twin model. The model includes observations for four couples, which are thought to result from female (F, red), male (M, blue), and idiosyncratic or couple-specific (E) factors. The familial factors are in turn thought to comprise additive genetic factors (A), dominant genetic factors (D) and — omitted to reduce clutter — shared environmental factors (C). The copath coefficient (μ) denotes assortment between female and male familial factors. The brackets on top explain the relations between the various couples. Couples 2–3 include a pair of twins or full siblings, and couples 1 and 4 are their respective siblings-in-law. The covariance between siblings depends on their genotypic correlations (monozygotic twins: r A = r D = 1; dizygotic twins and full siblings: r A = 0.50, r D = 0.25). See methods for more details. The dashed line surrounds the part of the model that corresponds to the classic twin model (excluding idiosyncratic environmental factors). This specific path diagram describes extended families where couples 2–3 are same sex female siblings/twins, and couples 1–2 and couples 3–4 are same-sex male siblings. For other family constellations, the path diagram will look slightly different (e.g., M and F can flip side).
Statistical Analysis
Our main analyses involved fitting three sets of structural equation models to our data, described below. We included year of marriage and year of marriage squared as covariates in all models. Because divorce is a dichotomous outcome, we used liability threshold models where we assumed unit variance for liability to divorce (
$Var\left( Y \right) = 1$
) and equal thresholds across groups. We calculated p values by performing log-likelihood ratio tests comparing the fit of the full model to a model where the parameters in question were fixed to their null hypothesis value. We calculated 95% confidence intervals using the Wald method (
$\pm {z_{\alpha /2}} \times SE$
). All models were fit using full information maximum likelihood in OpenMx 2.21.8 (Neale et al., Reference Neale, Hunter, Pritikin, Zahery, Brick, Kirkpatrick, Estabrook, Bates, Maes and Boker2016) in R 4.2.3 (R Core Team, 2022).
Model 0: Correlations. To estimate tetrachoric correlations between the various types of relatives, we applied a constrained saturated model to the data. This is a multigroup structural equation model where all equivalent correlations (e.g., male full siblings) are constrained to be equal across groups. (This did not result in significantly worse fit than a fully saturated model where all correlations were freely estimated, ΔLL = 24.3, Δdf = 23, p = .39). We also checked whether correlations that involved DZ twins were significantly different from correlations that involved full siblings. Constraining those sets of correlations to be equal did not result in significantly worse fit (ΔLL = 2.9, Δdf = 10, p = .98). We therefore combined DZ twin units with full sibling units when applying the twin models.
Model 1: The classic twin model. We estimated a classic twin model where we assumed that variation in liability to divorce (
$Y$
) was the result of additive genetic factors (
$A$
), dominant genetic factors (
$D$
), environmental factors shared by twins and siblings (
$C$
), and idiosyncratic factors (
$E$
).
$$Var\left( {{Y_*}} \right) = {V_{{A_\ast}}} + {V_{{D_\ast}}} + {V_{{C_\ast}}} + {V_{{E_\ast}}}$$
Here, the asterisk (*) is a placeholder for female-specific (
$f$
) and male-specific (
$m$
) components. To estimate this model, we assumed that correlations between same-sex pairs of relatives is a function of shared genetic and environmental factors:
$${r_{{Y_*}}} = {r_A}{V_{{A_\ast}}} + {r_D}{V_{{D_\ast}}} + {V_{{C_\ast}}}$$
Here,
${r_A}$
denotes the additive genotypic correlation for pairs of relatives (MZ twins:
${r_A} = 1$
; DZ twins and full siblings:
${r_A} = .50$
), and
${r_D}$
denotes the dominant genotypic correlation (MZ twins:
${r_D} = 1$
; DZ twins and full siblings:
${r_D} = .25$
). Dominant genetic effects (
$D$
) and shared environmental effects (
$C$
) cannot be estimated at the same time (Neale & Maes, Reference Neale and Maes2004). We therefore estimated both ACE and ADE models, where either
${V_D}$
or
${V_C}$
, respectively, were fixed to zero. We also estimated AE models where both
${V_D}$
and
${V_C}$
were fixed to zero. Because the ACE model would, under some specifications, result in negative variance components, we focus primarily on results from the ADE model. However, results from all models are reported.
We included opposite-sex DZ twins and full siblings. These can provide information on qualitative differences between female and male familial factors. Because there are no opposite-sex MZ twins, there is only one extra degree of freedom, which means one cannot investigate differences in both genetic and environmental factors separately (Neale & Maes, Reference Neale and Maes2004). Traditional twin models typically only investigate differences in one factor (usually additive genetic factors) and assume the other factors are perfectly correlated across the sexes. However, because genetic correlations tend to track phenotypic correlations (Sodini et al., Reference Sodini, Kemper, Wray and Trzaskowski2018), we employed an alternative assumption where we investigated differences in the summed female and male familial components. This is equivalent to assuming the same cross-sex correlation for all genetic and shared environmental components, denoted by the parameter
${r_s}$
:
$${r_{{Y_{OS}}}} = {r_s}\left( {{r_A}\sqrt {{V_{Af}}{V_{Am}}} + {r_D}\sqrt {{V_{Df}}{V_{Dm}}} + \sqrt {{V_{Cf}}{V_{Cm}}} } \right)$$
Model 2: The extended twin model. The path diagram for the extended twin model is shown in Figure 1. Compared to the classic twin model, we (1) included information on the respective siblings-in-law of the original (middle) twins and siblings, and (2) changed the expected variance for liability to divorce to be a combination of both female and male familial factors. We also included a copath denoted
$\mu $
between the female and male familial factors. A copath represents similarity attributable to matching instead of a shared cause, and comes with special path tracing rules (Cloninger, Reference Cloninger1980; Sunde et al., Reference Sunde, Eftedal, Cheesman, Corfield, Kleppesto, Seierstad, Ystrom, Eilertsen and Torvik2024). Because all valid chains between the respective relatives of husbands and wives must go through this copath, the correlation between affines can be used to estimate the copath coefficient, which in turn can be used to calculate the implied correlation between female and male familial factors. A correlation between female and male familial factors will, in turn, increase the variance of divorce liability per the sum of variances law:
$$Var\left( Y \right) = {V_F} + {V_M} + 2\mu {V_F}{V_M} + {V_E}$$
Here,
${V_F} = {V_{Af}} + {V_{Df}} + {V_{Cf}}$
is the female familial factors and
${V_M} = {V_{Am}} + {V_{Dm}} + {V_{Cm}}$
is the male familial factors. We cannot distinguish female and male idiosyncratic factors (
$E$
), so these have been combined. The expected covariances between all observations are listed in Supplementary Table S1.
Assortative mating changes the expected covariance between siblings and other relatives in two ways: First, all expected covariances increases by factor of
${\left( {1 + \mu {V_*}} \right)^2}$
owing to correlations between individuals’ familial factors and the familial factors of their spouses. Second, if there has been genetic homogamy in earlier generations (meaning that heritable components of the phenotype have been subject to assortative mating), siblings will be more genetically similar than normal (Fisher, Reference Fisher1918; Sunde et al., Reference Sunde, Eftedal, Cheesman, Corfield, Kleppesto, Seierstad, Ystrom, Eilertsen and Torvik2024). The magnitude of this increase depends on the number of generations of assortment and the genotypic correlation between spouses, which in turn depends on the genetic architecture of the traits spouses are assorting on, and how they are associated with divorce (Sunde et al., Reference Sunde, Eilertsen and Torvik2025). Here, we estimate and report results using two opposing assumptions: In the first set of analyses, we assume a long history of genetic homogamy where genetic similarity between siblings have reached a stable equilibrium (
${r_A} = \left( {1 + {\rho _g}} \right) \times 0.50$
, where
${\rho _g}$
is the implied genotypic correlation between spouses under direct assortment on the familial factors associated with divorce). In the second set of analyses, we assume no genetic homogamy in earlier generations, meaning siblings are no more genetically similar than under random mating (
${r_A} = 0.50$
).
Ethics
The study was approved, and participant consent was waived by the Norwegian Regional Committee for Medical and Health Research Ethics (project #2018/434).
Results
We analyzed 124,544 extended family units comprising 353,210 marriages, out of which 86,401 (24.5%) ended in divorce or separation with 15 years. The proportion of marriages ending in divorce was reasonably stable across groups (Table 1). Marriage rates decreased across the observation years, while the proportion of marriages ending in divorce varied between 18% and 29%, with a small inverse-U-shaped trend across time (Supplementary Figure S1).
Correlations Between Relatives and Affines
Figure 2 and Supplementary Table S2 present tetrachoric correlations for divorce between various extended family members, adjusted for both year of marriage and year of marriage squared. Correlations involving dizygotic twins were not significantly different from correlations involving full siblings (p = .98, Supplementary Table S3). Female full siblings were more highly correlated than male full siblings (r = .18, SE < .01; vs. r = .14, SE < .01), with opposite-sex siblings about as correlated as male full siblings (r = .14, SE = .01). Monozygotic twins were more highly correlated than full siblings, again with higher correlations between female twins than male twins (r = .38, SE = .07; vs. r = .28, SE = .07; respectively).
Correlations between affines in twin families were not estimated with sufficient precision to afford any remarks. In full sibling families, first-order affines (e.g., you and your brother’s wife’s sister) were significantly correlated (e.g., r = .06, SE = .01). This correlation was stable across the sex-composition of the mediating pairs. Second-order affines were also significantly correlated (e.g., r = . 03, SE = .01).
Model 1: The Classic Twin Design
Figure 3A and Supplementary Tables S5 to S7 present results from applying the classic twin model to divorce. Additive heritability estimates were higher for women (
${V_{Af}}$
= .38, SE = .08) than for men (
${V_{Am}}$
= .27, SE = .08). The between-sex correlation for familial factors was estimated to r = 0.89 (SE = .12) and was not significantly different from unity (ΔLL = 1.8, Δdf = 1, p = .18). Equating female and male factors resulted in significantly worse fit (ΔLL = 15.6, Δdf = 3, p = .001). In other words, we find evidence of quantitative sex differences (female factors explain more variance than male factors), but not of qualitative sex differences (female factors and male factors appear to be the same).
Figure 2. Correlations between members of extended family units. Color and shape indicate zygosity group, meaning the relationship between couple 2 and couple 3 in the extended family unit (FS = Full siblings, DZ = Dizygotic twins, MZ = Monozygotic twins). The y-axis describes the relation between who is being correlated. For genetic relatives, it describes the zygosity group and sex composition of the pair (m = male, f = female, mf = opposite-sex). For affines, it describes the composition of the intermediate pairs. For example, ‘MZm – FSf’ describes the correlation between a male index person and their same-sex monozygotic twin’s (MZm) wife’s sister (FSf). Correlations are adjusted for year of marriage and year of marriage squared. Unadjusted correlations are available in Supplementary Table S2. Error bars are 95% confidence intervals.
Figure 3. Model estimates (with 95% confidence intervals) of variance components (ADE-specification) for liability to divorce. Panel A presents results from the classic twin model; Panel B presents results from the extended twin model when assuming genetic homogamy in earlier generations, r A = (1+ρg)/2; Panel C presents results from the extended twin model when assuming no genetic homogamy in earlier generations r A = 1/2; The suffixes denote whether variance components are female (f, red) or male (m, blue) factors. V A = additive genetic effects, V D = dominant genetic effects, V E = idiosyncratic environmental effects. The variance component attributable to covariance between female and male factors is denoted 2μV F V M. Estimates are also available in Supplementary Tables S7, S9, and S11, and as .csv-files in the Supplementary Files.
The dominance effects were estimated to practically zero for both men and women, and were not significant (ΔLL = .003, Δdf = 2, p = .99). The same was true when we estimated shared environmental effects instead of dominance effects (Supplementary Figure S2).
Model 2: The Extended Twin Model
Figure 3B and Supplementary Tables S8 and S9 presents results from applying the extended twin model to divorce while assuming genetic homogamy in earlier generations. The estimated correlation between spouses’ familial factors was large (r = .60, SE = .10). This, in turn, meant that 16% of the variance in divorce liability was attributed to the correlation between female and male factors (
$2\mu {V_F}{V_M}$
= .16, SE = .04). Female familial factors explained 18% of the variance (
${V_F}$
= .18, SE = .05) whereas male familial factors explained 10% of the variance (
${V_M}$
= .10, SE = .03).
Results about spousal similarity and sex differences were similar for both ACE and ADE specification, as was the overall contribution of familial factors. However, the dissection into subcomponents varied. This was because the additive heritability estimate had to compensate for either negative shared environmental effects or relatively large dominance effects. In the ADE model, the additive heritability estimates were lower (
${V_{Af}}$
= .12, SE = .02;
${V_{Am}}$
= .07, SE = .01) and accompanied by dominance estimates of
${V_{Df}}$
= .05 (SE = .05) and
${V_{Dm}}$
= .03 (SE = .04). In the ACE model, the additive heritability estimates were higher (
${V_{Af}}$
= .22, SE = .09;
${V_{Am}}$
= 0.11, SE = .05) and accompanied by negative shared environmental effects estimates of
${V_{Cf}}$
= –0.04 (SE = .04) and
${V_{Cm}}$
= –0.01 (SE = .02). Estimating an AE model with neither C nor D did not result in significantly worse fit (ΔLL = 2.01, Δdf = 2, p = .37). No heritability estimates were as high as those from Model 1.
Figure 3C and Supplementary Tables S10 and S11 present results from applying the extended twin model to divorce while assuming no genetic homogamy in earlier generations. Results were overall very similar (r = .60, SE = .08;
$2\mu {V_F}{V_M}$
= .16, SE = .03;
${V_F}$
= .18, SE = .05;
${V_M}$
= .10, SE = .03), although the specific dissection into subcomponents differed. In the ADE model, the additive heritability was higher than under the alternative assumption (
${V_{Af}}$
= .20, SE = .03;
${V_{Am}}$
= .11, SE = .02), accompanied by smaller and reversed dominance estimates of
${V_{Df}}$
= –.02 (SE = .06) and
${V_{Dm}}$
= –.02 (SE = .04). In the ACE model, the additive heritability estimates were smaller than under the alternative assumption (
${V_{Af}}$
= .17, SE = .07;
${V_{Am}}$
= .09, SE = .05) and accompanied by shared environmental effects estimated to near-zero (
${V_{Cf}}$
= .01, SE = .05;
${V_{Cm}}$
= .01, SE = .02).
Discussion
In this study, we have linked twins and siblings with their affines in the Norwegian population register to estimate assortative mating on liability to divorce. We found that affines display measurable similarity in divorce liability, consistent with assortative mating for traits associated with risk of divorce. When we incorporated this information into an extended twin model, we estimated that spouses’ familial factors were highly correlated (r = .60), which accounted for a relatively large part of the variance (16%) in couples’ liability to divorce. We also found that after accounting for spousal similarity in an extended twin model, familial factors explained less variance compared to a classic twin model, which is as expected if estimates from the classic twin model are confounded by the effects of the spouse. These results demonstrate that spousal similarity is an important and previously ignored source of variation in divorce liability.
Assortative mating for divorce is not too surprising given that assortative mating on traits linked to divorce risk is well documented (Golovina et al., Reference Golovina, Gutvilig, Niemi and Hakulinen2025; Horwitz et al., Reference Horwitz, Balbona, Paulich and Keller2023; Kailaheimo-Lönnqvist et al., Reference Kailaheimo-Lönnqvist, Fasang, Jalovaara and Struffolino2021; Torvik et al., Reference Torvik, Sunde, Cheesman, Eftedal, Keller, Ystrom and Eilertsen2024). Spousal similarity in liability to divorce likely results as a necessary consequence of assortment on factors associated with variation in divorce risk — so-called indirect assortative mating (Sunde et al., Reference Sunde, Eilertsen and Torvik2025). Assortative mating for traits related to divorce increases the variance in the couple’s liability beyond the sum of the two individuals’ liabilities (per the sum of variance law). All else equal, this increases the number of couples with high divorce liabilities, and hence the number of couples who end up divorcing. If we remove the 16% of the variance attributable to spousal similarity while keeping the threshold for divorcing constant, the divorce rate would drop by almost 8% (i.e., 2 pp., Supplementary Note 2). Obviously, all else is not equal under random mating. In the counterfactual world with random mating, the variance in couples’ liabilities would be smaller, but more mismatched couples may increase the divorce rate. Investigating the effect of spousal similarity itself on marital quality and divorce risk remains a fascinating and understudied research question (e.g., Luo & Klohnen, Reference Luo and Klohnen2005).
If the mechanism of assortment involves genetic homogamy, meaning that the heritable components of the trait are subject to assortative mating, then there will also be genetic consequences. Genetic variants inherited from the mother will be correlated with those inherited from the father, which, over time, increases genetic variance in the population and can increase differences in divorce liability further (Lynch & Walsh, Reference Lynch and Walsh1998). Mapping the genetic consequences of assortment for divorce remains an outstanding research question.
Earlier twin models have generally not found notable sex differences in the heritability of divorce (e.g., D’Onofrio et al., Reference D’Onofrio, Turkheimer, Emery, Slutske, Heath, Madden and Martin2005; D’Onofrio et al., Reference D’Onofrio, Turkheimer, Emery, Harden, Slutske, Heath, Madden and Martin2007; Jocklin et al., Reference Jocklin, Mcgue and Lykken1996; McGue & Lykken, Reference McGue and Lykken1992). However, earlier samples have been relatively small and standard errors too wide to detect them. In this study, we included a large number of full siblings to improve precision and found that sisters were more correlated than brothers, and that female familial factors had a greater influence on the couple’s divorce liability than male familial factors. This aligns with research in family demography and sociology on gendered patterns of divorce initiation, showing that women more often take the initiative to divorce and that many risk factors such as financial problems are more strongly associated with the women’s decision to divorce compared to the men’s (Kalmijn & Poortman, Reference Kalmijn and Poortman2006).
This study was made possible by linking the Norwegian Twin Register (Nilsen et al., Reference Nilsen, Brandt and Harris2019) to the population register, which enabled us to chain-link affines and estimate the implied spousal similarity in couple-shared outcomes. This approach is rarely feasible outside large, register-based settings. For researchers without access to such data, we still offer a novel suggestion, namely combining female and male familial factors in the expected variance of couple-shared outcomes. Affines are only needed to account for spousal similarity.
In this article, we made the simplifying assumption that spouses were matching directly on the familial factors relevant to divorce risk and estimated models with two alternative assumptions about genetic homogamy in earlier generations. While it is technically possible to adapt the model to estimate relative degree of genetic and social homogamy (Sunde et al., Reference Sunde, Eilertsen and Torvik2025), lack of statistical power makes it unfeasible here, and assumptions would still have to be made about its history. Nevertheless, we found that the results about spousal similarity and the overall importance of female and male familial factors were robust to different assumptions about the history and mechanism of assortative mating. The specific estimates of additive heritability versus dominance effects (or shared environmental effects), on the other hand, were sensitive to these assumptions. Furthermore, it is a known limitation that twin models cannot estimate shared environmental and nonadditive genetic effects at the same time. If they are both nonzero, they may cancel each other out, resulting in a pattern of correlations where additive genetic factors appear sufficient to explain similarity between relatives. This is further complicated by assortative mating (Eftedal et al., Reference Eftedal, Eilertsen, Sunde, Kleppestø, Ystrom and Czajkowski2025). Because of these limitations, we caution against overinterpreting the decomposition of the familial factors.
To conclude, our results highlight the importance of analyzing divorce as a couple-level outcome and show that ignoring spousal similarity can produce misleading estimates of individual-level effects. Assortative mating on liability to divorce appeared substantial, and once modelled explicitly, the contribution of familial factors was markedly smaller. The same framework can be applied to any couple-shared outcome, such as fertility outcomes or joint health behaviors, to separate sex-specific sources of variation and account for their correlation. Finally, future work can use measured traits to identify the mechanisms that generate spousal similarity and test how these processes evolve across cohorts and contexts.
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/thg.2026.10050.
Data availability
The raw data are protected and are not available due to data privacy laws. The data for this study encompasses demographic information for entire cohorts of the Norwegian population. Researchers can access the data by applying to the Norwegian Regional Committees for Medical and Health Research Ethics and the data owners (Statistics Norway and the Norwegian Institute of Public Health). The authors cannot share these data with other researchers.
Acknowledgements
This work was performed on the TSD (Tjenester for Sensitive Data) facilities, owned by the University of Oslo, operated and developed by the TSD service group at the University of Oslo, IT-Department (USIT). Data on twin zygosity were obtained from the Norwegian Twin Registry, Norwegian Institute of Public Health.
Author contributions
H.F.S. and P.D. conceived of the idea. H.F.S did the analyses and wrote the first draft of the manuscript. Both authors provided critical feedback, discussed the results, helped shape the manuscript, and approved the final manuscript.
Financial support
This work was partly supported by the Research Council of Norway through its Centres of Excellence funding scheme (grant number 262700), and through the PARMENT project (334093). This work was co-funded by the European Union (ERC, BIOSFER, 101071773). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. P.D. was supported by the Strategic Research Council (SRC), FLUX consortium, decision numbers 364374 and 364375; by grants to the Max Planck—University of Helsinki Center from the Max Planck Society (5714240218), Jane and Aatos Erkko Foundation (210046), Faculty of Social Sciences at the University of Helsinki (77204227), and Cities of Helsinki, Vantaa and Espoo.
Competing interests
The authors declare no conflicts of interest.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
