Available via license: CC BY 4.0
Content may be subject to copyright.
A comprehensive meta-analysis of human
assortative mating in 22 complex traits
Tanya Horwitz ( taho7982@colorado.edu )
University of Colorado Boulder
Matthew Keller
University of Colorado
Article
Keywords:
Posted Date: March 21st, 2022
DOI: https://doi.org/10.21203/rs.3.rs-1467426/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
1
1
A comprehensive meta-analysis of human assortative mating in 22 complex traits
2
3
Tanya B Horwitz1,2* and Matthew C Keller1,2*
4
5
1Institute for Behavioral Genetics, University of Colorado Boulder, Boulder, CO, United States 6
of America. 7
2Department of Psychology and Neuroscience, University of Colorado Boulder, Boulder, CO,
8
United States of America. 9
10
11
2
Abstract 12
Assortative mating (AM) occurs when the correlation for a trait between mates is larger than 13
would be expected by chance. AM can increase the genetic and environmental variation of traits, 14
can increase the prevalence of disorders in a population, and can bias estimates in genetically 15
informed designs. In this study, we conducted the largest set of meta-analyses on human AM 16
published to date. Across 22 traits, meta-analyzed correlations ranged from r = .08 to r = .58, 17
with social attitude, substance use, and cognitive traits showing the highest correlations and 18
personality, disorder, and biometrical traits generally yielding smaller but still positive and 19
nominally significant (p < .05) correlations. We observed high between-study heterogeneity for 20
most traits, which could have been the result of phenotypic measurement differences between 21
samples and/or differences in the degree of AM across time or cultures. 22
23
24
25
26
27
28
29
3
A comprehensive meta-analysis of human assortative mating in 22 complex traits
30
31
Assortative mating (AM) is the phenomenon whereby individuals with similar trait 32
values mate with one another at levels higher than expected by chance1. Contrary to the maxim 33
“opposites attract,” nonzero phenotypic correlations between human2–21 and nonhuman1 mates 34
are overwhelmingly in the positive direction, with only a handful of examples of disassortative 35
mating, or negative mate correlations, reported in the literature1,4,8,20,22–29. Several potential 36
mechanisms of AM in humans have been described, although they are not mutually exclusive 37
because multiple mechanisms can simultaneously be responsible for observed correlations. 38
Phenotypic homogamy (also known as primary phenotypic assortment) occurs when mates 39
match directly on the trait of interest30. While phenotypic homogamy is often conceptualized as 40
mates actively preferring similarity, this type of homogamy can also be a function of indirect 41
selection, such as when mates are chosen from among strata that are partially determined by 42
individuals’ phenotypic values (e.g., AM for educational attainment arising as an indirect 43
consequence of mate choice occurring within job occupations). Social homogamy, on the other 44
hand, occurs when individuals match within strata that are determined by non-heritable 45
background social factors18,31, such as within social class in cultures where class is not 46
genetically influenced. At the other end of the spectrum, genetic homogamy is the mechanism 47
whereby mates correlate more genetically than phenotypically for a trait; this can occur when 48
there is phenotypic homogamy on a trait that is more correlated genetically than environmentally 49
with the trait of interest30,32. Finally, convergence occurs when mates become more similar over 50
time3,8, either due to direct (reciprocal or one-way) phenotypic influences on one another or to 51
the mutual influence of shared environmental factors. 52
4
Social scientists and quantitative geneticists care about the mechanisms and the strength 53
of AM because both influence parameters of interest and impact how various estimates in the 54
literature should be interpreted. Phenotypic and genetic homogamy on heritable traits increase 55
correlations between and within causal loci, which in turn increases the genetic covariance 56
between relatives and the trait’s phenotypic and genetic variation. Such an increase in variation 57
could manifest as increased prevalence rates of dichotomous traits such as psychiatric 58
disorders18,33, although this effect should only be pronounced in rare, highly heritable disorders 59
under strong AM18. Social homogamy can also increase trait variation when parental phenotypic 60
values for sociocultural traits are inherited by offspring via vertical transmission34. Failing to 61
account for AM can lead to biases in estimates from genetically informed designs, including the 62
association statistics from genome-wide association studies35, heritability estimates from 63
twin/family designs and from single nucleotide polymorphisms36, and the strength of estimated 64
causal associations in Mendelian randomization studies37. 65
Given that the genetic consequences of AM and the impacts of not accounting for it in 66
certain genetically informed designs are non-negligible, it is important to understand the strength 67
of AM for traits commonly investigated in human genetics. The strength and breadth of AM is 68
also of interest to investigators of human mating in psychology, sociology, and economics. 69
While many studies have reported estimates of AM in humans, we are aware of no study that has 70
meta-analyzed AM on a large number of phenotypically diverse traits. In the current report, we 71
use stringent methodology to meta-analyze and compare partner correlations for 22 commonly 72
investigated complex traits. These results are the most comprehensive set of meta-analyses on 73
human AM to date, and should shed light on contemporary human mating trends, help with the 74
5
interpretation of heritability estimates, motivate studies into the various causes of AM across 75
traits, and aid in the choice of design in genetic studies. 76
77
Results 78
Meta-analysis 79
We meta-analyzed partner concordance rates for 22 traits. While AM has been analyzed 80
for hundreds of traits, we focused on those most studied in the AM literature as well as some less 81
commonly studied dichotomous traits that have important health implications. The total number 82
of partner pairs for each trait ranged from 2,270 (for drinking quantity) to 1,533,956 (for 83
substance use disorder); effective sample sizes for dichotomous traits (see Methods) ranged from 84
721 (for alcohol use disorder) to 241,817 (for substance use disorder). Supplementary Tables S1 85
and S2 show all studies that we included in our meta-analysis for continuous and dichotomous 86
traits, respectively, as well as the effect sizes for each sample. For comparability across traits, we 87
focus here on Pearson and tetrachoric correlations for continuous and dichotomous traits, 88
respectively. Supplementary Table S2 also includes an alternative metric of partner concordance 89
for dichotomous traits, the odds ratio (OR), which is the odds of a participant possessing a trait 90
given that their partner has it divided by the odds of a participant possessing the trait given that 91
their partner does not have it. Supplementary Table S3 lists studies excluded from our meta-92
analysis along with the reasons for their exclusion. 93
Fig. 1 displays the meta-analyzed random effects correlations for all traits along with 94
their 95% confidence intervals. The meta-analyzed correlations were greater than zero at the 95
nominal significance level (p < .05) for all traits. The point estimates for fourteen traits were also 96
6
significant at the Bonferroni-corrected (p < .05/22 = 0.00227) significance level. Cognitive and 97
social attitude traits showed the highest correlations (.39 ≤ rmeta ≤ .58); personality, 98
anthropometric traits, substance use disorders, and other disorders showed the lowest (.08 ≤ rmeta 99
≤ .29); and correlations for non-pathological substance use traits typically lay between these two 100
sets (.24 ≤ rmeta ≤.54) (see Table 1). Fig. S1 displays forest plots for all the traits we analyzed 101
with publications ordered by year and color-coded by region. The meta-analyzed fixed effects 102
results for each trait (Fig. S2) were qualitatively similar to the random effect results. Fig. S5 103
shows the number of studies included and excluded for each trait. 104
Table 2 summarizes each trait’s heterogeneity estimates and the prediction intervals of 105
future studies’ effect sizes. We quantified heterogeneity using the Higgins & Thompson’s I2 106
metric, which represents the percentage of variance resulting from between-study heterogeneity 107
in effect sizes rather than within-study sampling error38. Higgins and Thompson (2002)39 108
classified I2 values of 25%, 50%, and 75% as low, medium, and high heterogeneity, respectively. 109
Across traits in our 22 meta-analyses, the median Higgins & Thompson I2 statistic was 87.5%, 110
reflecting very high heterogeneity in AM estimates for most traits. However, a high I2 reflects not 111
only high between-study heterogeneity in estimated effect sizes but also low within-study 112
heterogeneity due to highly precise estimates of individual studies. Thus, these high I2 values 113
may in part be due to the high precision of estimates afforded by the large sample sizes of many 114
of the studies included in our analyses. An alternative metric of heterogeneity that is unaffected 115
by the precision of estimates of individual studies, τ2, represents the estimated variance of the 116
true effect size under a random effects model. The estimated standard deviations of true effects 117
(τ) were large relative to the meta-analyzed correlation values for many traits. The median 118
coefficient of variation
was .41, and the coefficient of variation was above .50 for 119
7
intelligence quotient (IQ), drinking quantity, agreeableness, conscientiousness, extraversion, 120
body mass index (BMI), and generalized anxiety disorder (GAD). However, for some traits, such 121
as EA (rmeta = .53 +/- τ = .10), political values (rmeta = .58 +/- .08), and depression (rmeta = .14 +/- 122
.02), the estimated standard deviation of true effects was not very large compared to the meta-123
analyzed estimate. Overall, our results suggest that AM is characterized by substantial 124
differences in the strength of true effect across populations differentiated by place or time. 125
For each trait, we also created Graphic Display of Heterogeneity (GOSH) plots (Fig. 126
S4)40, which are scatterplots of the meta-analyzed correlations for all possible 2k-1 combinations 127
of k studies of size 2 through k (up to 1 million combinations) on the x-axis and the I2 values of 128
these combinations on the y-axis. Two or more distinct clusters anywhere in the plot may 129
indicate subpopulations that differ in their average effect size40, although a smear of points along 130
the bottom of GOSH plots is caused by two or more study results that happen to be similar 131
(thereby producing I2 values near 0) and is typically not of interest. For most traits plotted in Fig. 132
S4, there are no obvious clusters. However, for IQ and conscientiousness, there do appear to be 133
two clusters, one made up of study combinations that have higher heterogeneity and higher 134
average correlations, and another with lower heterogeneity and lower average correlations. The 135
two clusters in the GOSH plot for IQ may have resulted from an outlier reported in a 1938 study 136
that found a partner correlation of .8141, which is substantially greater than the meta-analyzed 137
estimate we report for this trait. 138
Because AM studies ostensibly focus more on effect size than hypothesis testing, we 139
expected that publication bias was unlikely to be a major factor for the study results we meta-140
analyzed. Nevertheless, we created funnel plots (Fig. S3), which plot study effect size (Fisher Z 141
transformed correlations here) on the x-axis against standard error on the y-axis, to visually 142
8
inspect whether there was evidence for asymmetry, a potential indicator of publication bias. 143
Overall, there was no obvious asymmetry across the funnel plots. Only for IQ and drinking 144
quantity did it appear that there may be a systematic bias of larger studies having smaller effect 145
sizes, but both were based on 10 or fewer studies, which can lead to apparent asymmetry by 146
chance38,42. The more obvious pattern observed in most funnel plots was the large number of 147
points that were outside the expected triangular region, again reflecting the high heterogeneity in 148
correlations observed across studies. 149
150
Discussion 151
In this study, we collated and synthesized the results from a large number of studies on 152
human AM to provide a better understanding of which traits mates assort on and how strong the 153
assortment is. To our knowledge, this is the largest and most comprehensive set of meta-analyses 154
on human AM to date. We found the highest levels of AM for political and religious values, 155
educational attainment, IQ, and some substance use traits; partner correlations for other traits 156
were smaller. Nevertheless, we found nominally significant (p < .05) evidence for AM for every 157
trait investigated. More than half of the meta-analyzed correlations were also significant at the 158
Bonferroni-corrected level. Whether these correlations are due to convergence or to initial 159
nonrandom mating based on phenotypic, social, or genetic homogamy remains to be determined, 160
though some research has attempted to investigate which of these mechanisms is responsible for 161
observed AM for particular traits. 162
The two social attitude traits that we examined—political attitudes and religiosity—163
showed the highest levels of AM of all the traits we assessed. For these traits, we examined 164
continuous measures of attitudes toward political issues and self-report of multiple religious 165
9
ideas/practices. Interestingly, despite clear geographical stratification of religious and voting 166
trends apparent in countries such as the United States, most studies to date investigating the 167
cause of mate similarity on political and religious attitudes have suggested that the data is most 168
consistent with phenotypic rather than social homogamy, and there is no compelling evidence of 169
substantial convergence for either trait4,43–46. This may be relevant to current events because, to 170
the degree that social attitudes are genetically or socially heritable, AM on them may contribute 171
to heightened political and cultural polarization. 172
We also found a high partner correlations for educational attainment (EA) (rmeta = .53), 173
and only one sample47 out of 27 reported a correlation under .30. Thus, there is consistent 174
evidence for strong AM on EA across recent decades and across cultures in which the trait has 175
been studied. Robinson et al. (2017)32 found that the implied phenotypic correlation for EA 176
between partners in the UK Biobank, extrapolated from the observed correlation between 177
partners' trait-associated loci, was .65. This value was substantially larger than the phenotypic 178
correlation they observed for EA in the same sample and exceeds the upper limit of our 179
confidence interval for the meta-analyzed EA partner correlation. This suggests that AM for EA 180
is consistent with genetic homogamy, and that mates may be assorting on some trait that is more 181
genetically than environmentally correlated with EA. Contrary to Robinson et al.’s (2017)32 182
finding, Torvik et al. (2022)48 did not find evidence for genetic homogamy in educational 183
attainment in a sample of partners, siblings, and in-laws in Norway. Instead, they found evidence 184
that AM on EA was due to a mix of both social homogamy and phenotypic homogamy. Whether 185
this discrepancy is due to differences in EA AM between Norway and the UK or to differences 186
in sample characteristics (e.g., ascertainment) is an open question. 187
10
The meta-analyzed partner correlation coefficients for substance use/abuse traits ranged 188
from rmeta = .24 to rmeta = .54. Interestingly, some (but not all49,50) studies that have examined 189
mechanisms of assortment in drinking and smoking have reported evidence of convergence for 190
these behaviors6,8,12,51, making these traits amongst the only ones to show support for 191
convergence in the literature. 192
We observed substantial between-study heterogeneity in partner correlations for most 193
traits. A large degree of between-study heterogeneity would certainly be problematic in fixed 194
effects meta-analyses that assume a single underlying effect. However, even for random effects 195
meta-analyses, which are viewed as more appropriate when heterogeneity is present, high levels 196
of heterogeneity suggest caution should be used in interpretation of results. Random effects 197
meta-analyses assume an underlying (normal) distribution of true effects across the studies’ 198
sampled populations, and the meta-analytic result is the estimated mean of those true effects. 199
Thus, the estimates we present here cannot be interpreted as estimates of a single true level of 200
AM for a given trait, but rather estimates of the typical level of AM across many possible levels 201
that might be observed at different times or locations. 202
There are several possible causes of the high levels of heterogeneity in AM we observed 203
across studies within the same trait. Most obviously, it is possible that the true degree of AM 204
varied across populations due to cultural differences in mating systems or preferences. This 205
seems plausible; AM involves mate preferences, social stratification, and/or couple dynamics, 206
and so it is unlikely to be consistent across different cultural contexts. Differences in population 207
size, mobility, and/or education across populations may impact the pool of a person’s potential 208
mates and thereby the degree to which preferences can be acted on. However, there was 209
insufficient cultural diversity within traits to test whether there were significant differences in 210
11
partner concordance across cultures. Similarly, we determined that publication year was too 211
coarse a metric of the year in which mates were married, and too many studies failed to report 212
sufficient information for us to formally assess changes in AM over time. 213
It is also possible that some of the heterogeneity in AM effect sizes was due to 214
differences in how constructs were measured across studies—for example, differences in the 215
measurement batteries used, differences in participants’ interpretations of battery items, or 216
differences in the clinical thresholds employed. Potentially consistent with this possibility, we 217
observed that the prevalence rates of dichotomous traits varied greatly in supposedly non-218
ascertained samples, which may have contributed to the heterogeneity we observed in our 219
correlation coefficients. Nevertheless, we observed high levels of heterogeneity even for traits—220
such as height and BMI—measured in standardized ways, suggesting that differences in how the 221
constructs were measured is unlikely to be a complete explanation. Finally, it is possible that 222
publication bias led to heterogeneity, particularly if studies that found AM results that were 223
substantially different from those already published in the literature were more likely to be 224
submitted and published—a kind of "novelty bias." However, it is also possible that a 225
"conformity bias" exists in the opposite direction and has led to downwardly biased estimates of 226
heterogeneity. While we could not test and therefore cannot rule out either possibility, we find 227
them unlikely given that the incentives for both seem dubious. 228
Although we initially gathered data on AM for rare psychiatric disorders, we did not 229
formally meta-analyze the tetrachoric correlations for these traits because too few studies met 230
our inclusion criteria as a result of unspecified sample sizes, the use of longitudinal rather than 231
cross-sectional measurements of concordance, and small expected cell frequencies (see 232
Supplementary Table S2 and S3). Nevertheless, studies that have provided robust estimates of 233
12
partner concordance for psychiatric disorders have suggested low to moderate AM, both within 234
and across disorders18,21,52,53. For example, based on data from Swedish population registers that 235
included more than 700,000 unique cases—originally analyzed by Nordsletten et al. (2016)54--236
Peyrot et al. (2016)18 estimated ascertainment-corrected tetrachoric correlation coefficients of .26 237
for schizophrenia, .10 for bipolar disorder, .28 for autism spectrum disorder, and .31 for 238
attention-deficit/hyperactivity disorder. 239
There are several implications for the consistent evidence of AM across traits we 240
documented in this meta-analysis. First, as noted above, AM can increase the genetic variance and 241
the prevalence of a disorder. Although the increase in prevalence for common disorders may not 242
be large (e.g., ~10%), the levels of AM observed for rare traits of high heritability, such as autism, 243
could lead to a ~1.5-fold prevalence increase after one generation, and an even higher increase 244
(~2.4-fold) over many generations18. Second, AM can create biases in estimates of interest in 245
genetically informative designs, such as estimates based on twin studies10,54, genome-wide 246
association studies (GWAS)35, Mendelian randomization37, and SNP-heritability36. Finally, to the 247
degree that the heterogeneity in AM we observed was due to true differences in the strength of 248
AM rather than differences in measurement, our estimates of the strength of AM may not 249
generalize to other populations. While estimates for some traits, such as height, were based on a 250
geographically and ethnically diverse set of samples, most of the samples included in our meta-251
analyses were drawn from Europe, North America, and Australia, and Asia. For example, all 252
estimates of AM for religiosity came from samples in the United States. 253
In summary, we conducted the largest and most comprehensive set of meta-analyses of 254
human AM to date. Our estimates were based on nearly a century of research and millions of 255
partner pairs. We found high partner correlations for traits related to substance use, IQ, EA, and 256
13
social attitudes, and smaller but nominally significant (p < .05) correlations for personality, 257
anthropometric, and disorder traits. However, we also observed high levels of heterogeneity in 258
AM estimates across studies for most traits investigated, suggesting that AM may differ across 259
time or place and that a single estimate of AM cannot typically be assumed for a given trait 260
across populations. 261
262
Methods 263
Inclusion and exclusion criteria 264
We conducted a systematic review of English-language studies that examined AM based 265
on partners’ continuous and dichotomous self-reports on the same complex traits. All included 266
studies were published in peer-reviewed journals on or before December 22, 2021. To conduct 267
this review, we searched for words pertaining to the traits of interest in conjunction with the 268
terms assortative mating, assortative marriage, partner concordance, partner correlation, 269
nonrandom mating, homogamy, marital resemblance, and marital homophily in Google Scholar, 270
and we checked relevant papers cited in these studies for adherence to our criteria. We restricted 271
our analysis to studies of opposite-sex co-parents, engaged pairs, married pairs, and/or 272
cohabitating pairs (referred to as “partners” hereafter), with a few studies containing a small 273
number of divorced couples; we excluded same-sex partners because same-sex and opposite-sex 274
pairs show different patterns of assortment for some traits55,56, because there is less data on the 275
former, and because same-sex assortment does not have the same implications for genetic 276
studies. With the exception of studies that intentionally ascertained partners for the trait of 277
interest, we excluded studies in which pairs had a characteristic that deviated from the norm in 278
the general population in a way that might have affected the magnitude of concordance (e.g., a 279
14
sample of only adoptive parents was excluded), and we only included studies where the sample 280
size was reported or could be inferred. For example, if only percentages were reported for each 281
cell of a contingency table, the sample size of each cell could be inferred as the percentage 282
multiplied by N. 283
We restricted our analysis to studies with sample sizes greater than 100. For dichotomous 284
traits, we restricted our analysis to studies with expected contingency table cell frequencies of 285
five or greater and observed cell frequencies greater than zero. When the samples in multiple 286
studies that were appropriate for our meta-analysis overlapped or were likely to have overlapped 287
based on information provided in the publication, we only used the study with the largest sample 288
size. We calculated effect sizes from the data reported in primary studies rather than relying on 289
effect size estimates from other published meta-analyses. If a study reported partner concordance 290
rates for multiple independent samples, each was included as a separate entry. When studies 291
reported partner correlation at different waves, we reported the results from the first wave. 292
When studies reported both the raw correlation and the partial correlation(s) controlling 293
for covariates (such as age), we included the raw correlation for consistency across studies. For 294
studies that only reported partial correlations, we used the estimate with the fewest number of 295
covariates. For ordinal and continuous traits, studies typically reported Spearman’s rho or 296
Pearson’s r but at times reported polychoric correlations. We excluded polychoric correlations 297
reported for such traits in order to avoid pooling two classes of correlation for the same meta-298
analyzed effect size. Because polychoric correlations occurred rarely, we do not anticipate a 299
large loss of power as a result. Because AM for height has already been meta-analyzed 300
extensively by Stulp et al. (2017)9, we re-analyzed studies from the paper’s supplement in the 301
15
same way we analyzed other continuous traits, after eliminating studies from this meta-analysis 302
in accordance with our exclusion criteria. Finally, we restricted our meta-analysis to traits for 303
which there were at least three samples that met our criteria. 304
Dichotomous traits
305
For dichotomous traits, we primarily considered studies that examined pairs in non-306
ascertained community samples or national registers as well as those from samples that 307
ascertained probands. Most ascertained studies were ultimately excluded because probands were 308
typically in clinical settings (e.g., hospitalized), whereas partners of probands with the disorder 309
typically were not. Although such ascertainment can be dealt with if all the applicable 310
populations’ (i.e. inpatient, outpatient, and those who have never received treatment) prevalence 311
rates are known, it was typically impossible to know all of these rates. We eliminated any 312
ascertained studies in which there was a >~two-fold difference in male and female prevalence if 313
there was not enough information to divide discordant couples based on sex. Simulation results 314
suggested that mixing individuals of different sexes when prevalence rates were more discrepant 315
than this would lead to unacceptable levels of bias. Because of possible differences in the 316
strength of AM implied from concordance of male probands versus that implied from female 317
probands, we excluded studies that only included single-sex probands. When both male and 318
female proband data was available (only a single study52), estimates based on each proband 319
(female and male) were included as separate results. 320
We only used cross-sectional measures of partner concordance and therefore excluded 321
studies that used longitudinal metrics such as morbidity risks57, hazard ratios, and incidence 322
ratios. We required that either odds ratios (ORs), risk ratios (RR), phi coefficients (Φ), 323
contingency tables, or—if the study was not ascertained (see below)–tetrachoric correlations, 324
16
were reported for dichotomous traits. Concordance rates captured by any of the first four of these 325
measures were then converted to tetrachoric correlations for consistency. When the contingency 326
table was unknown but the OR was reported, we first inferred the contingency table using an R 327
function described in the supplementary methods of Peyrot et al. (2016)18 (provided to us by the 328
authors) and then estimated the tetrachoric correlation. When the contingency table was 329
provided, we calculated the OR and tetrachoric correlation (using the polychoric() function from 330
the “polycor” package58) in R ourselves, and thus the effect size we used in our analysis was 331
sometimes different than that reported in the original study. When the contingency table was 332
unknown but Φ was reported, Φ was converted to a tetrachoric correlation using the phi2tetra() 333
function from the “psych” package59 in R. The prevalence rates for each sex used for these 334
conversions (from Φ and the OR) are reported in Supplementary Table S2. No studies that we 335
included in our final analysis reported an RR. 336
For studies where probands were ascertained, we used the OR, which is not influenced by 337
ascertainment, along with estimates of sex-specific prevalence rates from the country or region 338
the sample came from, to calculate tetrachoric correlations. To do this, we used the 339
aforementioned R function provided to us by Peyrot and colleagues, which produces the 340
population (non-ascertained) contingency table that is implied given the observed OR in the 341
ascertained sample and the assumed population prevalence in each sex. We then used this 342
implied contingency table to estimate the underlying (non-ascertained) tetrachoric correlation in 343
the population. This correction is necessary because the liability in the ascertained sample, where 344
the case to control ratio is usually higher than that in the population, is different than the liability 345
distribution in the population, which would lead to upwardly biased estimates if the tetrachoric 346
correlation was estimated based on just the sample contingency table. 347
17
We used the metacor() function from the “meta” package in R60 to conduct both random 348
and fixed effects meta-analyses using inverse-variance weighting of the Fisher z transformed 349
correlations. For continuous traits, we used the Knapp-Hartung adjustment61,62 to calculate the 350
variance of point estimates and restricted maximum-likelihood (REML) to estimate τ2, the 351
variance of the true overall effect size under random effects63,64. For binary traits, we used the 352
Paule-Mandel estimator65 to estimate τ2 and applied the Knapp-Hartung adjustment61,62 to our 353
calculation of the variance of the point estimate. We conducted a Monte Carlo analysis to 354
determine how best to pool information for different studies in a meta-analysis. While the “true” 355
base spousal correlation varied across simulated meta-analyses, the population-level spousal 356
correlation across “studies” within the same meta-analysis was consistent (in order to establish a 357
true rate of spousal concordance against which to compare our point estimates). However, 358
prevalence rates were allowed to vary across populations in the same simulated meta-analysis 359
(see Supplementary Table S4 for the results of each method used in conjunction with various 360
parameter estimates). We found that calculating tetrachoric correlations for each sample and then 361
meta-analyzing them provided more accurate point estimates than pooling contingency tables 362
and then calculating tetrachoric correlations. Thus, we followed this procedure for binary traits 363
throughout. The metacor() function internally calculates the expected variance of correlations 364
based on sample sizes and assumes they are Pearson correlations, which would be incorrect for 365
tetrachoric correlations. Thus, we needed to input effective (rather than actual) sample sizes for 366
tetrachoric correlations. For non-ascertained studies, we estimated the effective sample sizes by 367
using the standard error calculated in the polychor() package and solving for n in the equation 368
For ascertained studies examining dichotomous traits, we created bootstrapped 369
contingency tables, each of size n (the number of partners) and sampled from the study’s (raw, 370
18
ascertained) contingency table with replacement. We followed the procedure described above to 371
convert the ascertained contingency table to a tetrachoric correlation corrected for ascertainment. 372
We repeated this process 1,000 times, calculated the standard error by estimating the standard 373
deviation of the 1,000 bootstrapped tetrachoric correlations, and used this standard error to 374
calculate the effective sample size as described above. 375
Four of the traits in our supplementary tables—bipolar disorder, schizophrenia, panic 376
disorder, and phobia—posed a problem because they were rare (bipolar disorder and 377
schizophrenia) or have not been studied in sufficiently large samples (panic disorder and phobia). 378
This resulted in contingency tables with zero frequency cells or with expected cell frequencies 379
that were less than five. As a result, there was not a sufficient number of studies meeting our 380
inclusion criteria to justify formally meta-analyzing these four traits, though we included the 381
results from studies that otherwise met our criteria for these traits in Supplementary Table S2. 382
383
Data availability 384
Studies included in the meta-analysis are listed in Supplementary Tables S1 and S2, and studies 385
excluded from the meta-analysis are listed in Supplementary Table S3. 386
387
Code availability 388
The code for the analyses and simulations is available from the authors upon request. 389
19
References 390
1. Jiang, Y., Bolnick, D. I. & Kirkpatrick, M. Assortative mating in animals. Am. Nat. 181, 391
E125–E138 (2013). 392
2. Luo, S. Assortative mating and couple similarity: patterns, mechanisms, and consequences. 393
Soc. Personal. Psychol. Compass 11, e12337 (2017). 394
3. Burgess, E. W. & Wallin, P. Homogamy in social characteristics. Am. J. Sociol. 49, 109–124 395
(1943). 396
4. Alford, J. R., Hatemi, P. K., Hibbing, J. R., Martin, N. G. & Eaves, L. J. The politics of mate 397
choice. J. Polit. 73, 362–379 (2011). 398
5. Jurj, A. L. et al. Spousal correlations for lifestyle factors and selected diseases in Chinese 399
couples. Ann. Epidemiol. 16, 285–291 (2006). 400
6. Ask, H., Rognmo, K., Torvik, F. A., Røysamb, E. & Tambs, K. Non-random mating and 401
convergence over time for alcohol consumption, smoking, and exercise: the Nord-Trøndelag 402
Health Study. Behav. Genet. 42, 354–365 (2012). 403
7. Rawlik, K., Canela-Xandri, O. & Tenesa, A. Indirect assortative mating for human disease 404
and longevity. Heredity 123, 106–116 (2019). 405
8. Price, R. A. & Vandenberg, S. G. Spouse similarity in American and Swedish couples. 406
Behav. Genet. 10, 59–71 (1980). 407
9. Stulp, G., Simons, M. J. P., Grasman, S. & Pollet, T. V. Assortative mating for human 408
height: a meta-analysis. Am. J. Hum. Biol. 29, e22917 (2017). 409
10. Allison, D. B. et al. Assortative mating for relative weight: genetic implications. Behav. 410
Genet. 26, 103–111 (1996). 411
20
11. Mascie-Taylor, C. G. N. Spouse similarity for IQ and personality and convergence. Behav. 412
Genet. 19, 223–227 (1989). 413
12. Meyler, D., Stimpson, J. P. & Peek, M. K. Health concordance within couples: a systematic 414
review. Soc. Sci. Med. 64, 2297–2310 (2007). 415
13. Stimpson, J. P. & Peek, M. K. Concordance of chronic conditions in older Mexican 416
American couples. Prev. Chronic. Dis. 2, (2005). 417
14. Jeong, S. & Cho, S.-I. Concordance in the health behaviors of couples by age: a cross-418
sectional study. J. Prev. Med. Pub. Health 51, 6 (2018). 419
15. Di Castelnuovo, A., Quacquaruccio, G., Donati, M. B., de Gaetano, G. & Iacoviello, L. 420
Spousal concordance for major coronary risk factors: a systematic review and meta-analysis. 421
Am. J. Epidemiol. 169, 1–8 (2009). 422
16. Hippisley-Cox, J. Married couples’ risk of same disease: cross sectional study. BMJ 325, 423
636–636 (2002). 424
17. Galbaud Du Fort, G., Bland, R. C., Newman, S. C. & Boothroyd, L. J. Spouse similarity for 425
lifetime psychiatric history in the general population. Psychol. Med. 28, 789–802 (1998). 426
18. Peyrot, W. J., Robinson, M. R., Penninx, B. W. J. H. & Wray, N. R. Exploring boundaries 427
for the genetic consequences of assortative mating for psychiatric traits. JAMA Psychiatry 428
73, 1189–1195 (2016). 429
19. McLeod, J. D. Social and psychological bases of homogamy for common psychiatric 430
disorders. J. Marriage Fam. 201–214 (1995). 431
20. Maes, H. H. et al. Assortative mating for major psychiatric diagnoses in two population-432
based samples. Psychol. Med. 28, 1389–1401 (1998). 433
21
21. Mathews, C. A. & Reus, V. I. Assortative mating in the affective disorders: a systematic 434
review and meta-analysis. Compr. Psychiatry 42, 257–262 (2001). 435
22. Watson, D. et al. Match makers and deal breakers: analyses of assortative mating in 436
newlywed couples. J. Pers. 72, 1029–1068 (2004). 437
23. Botwin, M. D., Buss, D. M. & Shackelford, T. K. Personality and mate preferences: five 438
factors in mate selection and marital satisfaction. J. Pers. 65, 107–136 (1997). 439
24. Eysenck, H. J. & Wakefield, J. A. Psychological factors as predictors of marital satisfaction. 440
Adv. Behav. Res. Ther. 3, 151–192 (1981). 441
25. McCrae, R. R. et al. Personality trait similarity between spouses in four cultures. J. Pers. 76, 442
1137–1164 (2008). 443
26. Markey, P. M. & Markey, C. N. Romantic ideals, romantic obtainment, and relationship 444
experiences: the complementarity of interpersonal traits among romantic partners. J. Soc. 445
Pers. Relatsh. 24, 517–533 (2007). 446
27. Stimpson, J. P., Masel, M. C., Rudkin, L. & Peek, M. K. Shared health behaviors among 447
older Mexican American spouses. Am. J. Health Behav. 30, 495–502 (2006). 448
28. Al-Sharbatti, S. S., Abed, Y. I., Al-Heety, L. M. & Basha, S. A. Spousal concordance of 449
diabetes mellitus among women in Ajman, United Arab Emirates. Sultan Qaboos Univ. Med. 450
J. 16, e197 (2016). 451
29. Ober, C. et al. HLA and mate choice in humans. Am. J. Hum. Genet. 61, 497–504 (1997). 452
30. Keller, M. C., Medland, S. E. & Duncan, L. E. Are extended twin family designs worth the 453
trouble? A comparison of the bias, precision, and accuracy of parameters estimated in four 454
twin family models. Behav. Genet. 40, 377–393 (2010). 455
22
31. Cloninger, C. R. Interpretation of intrinsic and extrinsic structural relations by path analysis: 456
theory and applications to assortative mating. Genet. Res. 36, 133–145 (1980). 457
32. Robinson, M. R. et al. Genetic evidence of assortative mating in humans. Nat. Hum. Behav. 458
1, 1–13 (2017). 459
33. Yengo, L. et al. Imprint of assortative mating on the human genome. Nat. Hum. Behav. 2, 460
948–954 (2018). 461
34. Cloninger, C. R., Rice, J. & Reich, T. Multifactorial inheritance with cultural transmission 462
and assortative mating. II. a general model of combined polygenic and cultural inheritance. 463
Am. J. Hum. Genet. 31, 176–198 (1979). 464
35. Lee, J. J. et al. Gene discovery and polygenic prediction from a genome-wide association 465
study of educational attainment in 1.1 million individuals. Nat. Genet. 50, 1112–1121 466
(2018). 467
36. Border, R. et al. Assortative mating biases marker-based heritability estimators. Nat. 468
Commun. 13, 660 (2022). 469
37. Hartwig, F. P., Davies, N. M. & Davey Smith, G. Bias in Mendelian randomization due to 470
assortative mating. Genet. Epidemiol. 42, 608–620 (2018). 471
38. Harrer, M., Cuijpers, P., Furukawa, T. A. & Ebert, D. D. Doing Meta-Analysis in R. (2019). 472
39. Higgins, J. P. & Thompson, S. G. Quantifying heterogeneity in a meta-analysis. Stat. Med. 473
21, 1539–1558 (2002). 474
40. Olkin, I., Dahabreh, I. J. & Trikalinos, T. A. GOSH – a graphical display of study 475
heterogeneity. Res. Synth. Methods 3, 214–223 (2012). 476
41. Cattell, R. B. & Willson, J. L. Contributions concerning mental inheritance: I. of intelligence. 477
Br. J. Educ. Psychol. 8, 129–149 (1938). 478
23
42. Sterne, J. A. C. et al. Recommendations for examining and interpreting funnel plot 479
asymmetry in meta-analyses of randomised controlled trials. BMJ 343, d4002 (2011). 480
43. Watson, D. et al. Match Makers and Deal Breakers: Analyses of Assortative Mating in 481
Newlywed Couples. J. Pers. 72, 1029–1068 (2004). 482
44. Feng, D. & Baker, L. Spouse similarity in attitudes, personality, and psychological well-483
being. Behav. Genet. 24, 357–364 (1994). 484
45. Zietsch, B. P., Verweij, K. J., Heath, A. C. & Martin, N. G. Variation in human mate choice: 485
simultaneously investigating heritability, parental influence, sexual imprinting, and 486
assortative mating. Am. Nat. 177, 605–616 (2011). 487
46. Martin, N. G. et al. Transmission of social attitudes. Proc. Natl. Acad. Sci. 83, 4364–4368 488
(1986). 489
47. Procidano, M. E. & Rogler, L. H. Homogamous assortative mating among Puerto Rican 490
families: intergenerational processes and the migration experience. Behav. Genet. 19, 343–491
354 (1989). 492
48. Torvik, F. A. et al. Modeling assortative mating and genetic similarities between partners, 493
siblings, and in-laws. Nat. Commun. 13, 1108 (2022). 494
49. Treur, J. L., Vink, J. M., Boomsma, D. I. & Middeldorp, C. M. Spousal resemblance for 495
smoking: underlying mechanisms and effects of cohort and age. Drug Alcohol Depend. 153, 496
221–228 (2015). 497
50. Agrawal, A. et al. Assortative mating for cigarette smoking and for alcohol consumption in 498
female Australian twins and their spouses. Behav. Genet. 36, 553–566 (2006). 499
51. Reczek, C., Pudrovska, T., Carr, D., Thomeer, M. B. & Umberson, D. Marital histories and 500
heavy alcohol use among older adults. J. Health Soc. Behav. 57, 77–96 (2016). 501
24
52. Nordsletten, A. E. et al. Patterns of nonrandom mating within and across 11 major 502
psychiatric disorders. JAMA Psychiatry 73, 354–361 (2016). 503
53. Chou, I.-J. et al. Familial aggregation and heritability of schizophrenia and co-aggregation of 504
psychiatric illnesses in affected families. Schizophr. Bull. 43, 1070–1078 (2017). 505
54. Coventry, W. L. & Keller, M. C. Estimating the extent of parameter bias in the classical twin 506
design: a comparison of parameter estimates from extended twin-family and classical twin 507
designs. Twin Res. Hum. Genet. 8, 214–223 (2005). 508
55. Schwartz, C. R. & Graf, N. L. Assortative matching among same-sex and different-sex 509
couples in the United States, 1990–2000. Demogr. Res. 21, 843 (2009). 510
56. Verbakel, E. & Kalmijn, M. Assortative mating among Dutch married and cohabiting same-511
sex and different-sex couples. J. Marriage Fam. 76, 1–12 (2014). 512
57. Strömgren, E. Zum Ersatz des Weinbergschen “abgekürzten Verfahrens “. Z. Für Gesamte 513
Neurol. Psychiatr. 153, 784–797 (1935). 514
58. polycor: Polychoric and Polyserial Correlations version 0.8-1 from R-Forge. 515
https://rdrr.io/rforge/polycor/. 516
59. Revelle, W. psych: Procedures for Psychological, Psychometric, and Personality Research. 517
(2021). 518
60. Balduzzi, S., Rücker, G. & Schwarzer, G. How to perform a meta-analysis with R: a 519
practical tutorial. Evid. Based Ment. Health 22, 153–160 (2019). 520
61. Knapp, G. & Hartung, J. Improved tests for a random effects meta-regression with a single 521
covariate. Stat. Med. 22, 2693–2710 (2003). 522
62. Hartung, J. & Knapp, G. On tests of the overall treatment effect in meta-analysis with 523
normally distributed responses. Stat. Med. 20, 1771–1782 (2001). 524
25
63. Harrer, M., Cuijpers, P., Furukawa, T. A. & Ebert, D. D. Doing Meta-Analysis in R. (2019). 525
64. Veroniki, A. A. et al. Methods to estimate the between-study variance and its uncertainty in 526
meta-analysis. Res. Synth. Methods 7, 55–79 (2016). 527
65. Paule, R. C. & Mandel, J. Consensus values and weighting factors. J. Res. Natl. Bur. Stand. 528
87, 377–385 (1982). 529
Author contributions statement 530
TBH contributed to study design, statistical analyses, manuscript writing, collection of studies to 531
be meta-analyzed, simulation, and creation of all figures and tables; MCK contributed to study 532
design, statistical analyses, manuscript writing, and simulation. 533
534
Additional information 535
The authors declare no competing interests. 536
537
26
Trait
r [CI]
K
N
Effective N
p-value
EA
.53 [.49; .56]
27
230,915
NA
< .0001
IQ
.39 [.21; .54]
10
2,561
NA
.0012
Political values
.58 [.53; .63]
9
10,694
NA
< .0001
Religiosity
.57 [.37; .72]
5
5,750
NA
.0024
AUD
.24 [.09; .38]
3
5,162
721
.0221
Drinking quantity
.41 [.11; .64]
6
2,270
NA
.0178
Smoking cessation
.54 [.31; .72]
4
3,613
1,426
.0066
Smoking initiation
.37 [.30; .43]
12
87,253
13,469
< .0001
Smoking quantity
.24 [.14; .34]
6
4,701
NA
.0020
Smoking status
.46 [.35; .56]
15
168,404
20, 584
< .0001
SUD
.29 [.29, .30]
3
1,533,956
241,817
< .0001
Agreeableness
.11 [ .05; .18]
11
10,347
NA
.0035
Conscientiousness
.16 [.10; .23]
11
10,347
NA
.0003
Extraversion
.08 [.05; .11]
29
22,483
NA
< .0001
Neuroticism
.10 [.07; .13]
30
23,154
NA
< .0001
Openness
.21 [.14; .28]
11
10,483
NA
< .0001
Body mass index
.16 [.12; .19]
31
131,079
NA
< .0001
Height
.23 [.21; .26]
74
299,763
NA
< .0001
Waist-to-hip ratio
.16 [.08; .24]
5
83,630
NA
.0050
Depression
.14 [.11; .17]
7
1,483,486
211,154
< .0001
Diabetes
.15 [.07; .23]
7
178,522
17,530
.0038
27
538
539
540
541
542
543
544
545
546
547
548
GAD
.14 [.04; .24]
6
116,911
5,284
.0180
Trait
I2 [CI]
τ
τ2 [CI]
Prediction
Interval
EA
93% [91%; 94%]
.100
0.0100 [0.0058; 0.0238]
[0.3568; 0.6607]
IQ
91% [86%; 95%]
.260
0.0675 [0.0288; 0.2524]
[-0.2220; 0.7772]
Political values
80% [62%; 89%]
.082
0.0067 [0.0018; 0.0343]
[0.4256; 0.7014]
Religiosity
95% [91%; 97%]
.204
0.0417 [0.0128; 0.3736]
[-0.0662; 0.8782]
AUD
0% [0%; 90%]
.000
0 [0.0000; 0.3788]
[-0.2221; 0.6153]
Drinking quantity
92% [86%; 96%]
.294
0.0862 [0.0301; 0.5821]
[-0.4228; 0.8671]
Smoking cessation
90% [77%; 96%]
.169
0.0285 [0.0069; 0.4410]
[-0.2102; 0.8928]
Smoking initiation
95% [93%; 97%]
.104
0.0108 [0.0046; 0.0355]
[0.1408; 0.5587]
Smoking quantity
68% [24%; 87%]
.084
0.0070 [0.0006; 0.0642]
[-0.0103; 0.4700]
Smoking status
98% [98%; 99%]
.227
0.0517 [0.0247; 0.1400]
[-0.0095; 0.7651]
SUD
0% [0%; 90%]
.000
0 [0.0000; 0.0404]
[0.2722; 0.3119]
Table 1. r = meta-analyzed random effects spousal correlation (Pearson’s r for continuous
traits; tetrachoric r for dichotomous traits), CI = confidence interval, K = number of samples
meta-analyzed, N = number of total spouse pairs meta-analyzed; EA = educational
attainment, IQ = intelligence quotient, AUD = alcohol use disorder, SUD = substance use
disorder, GAD = generalized anxiety disorder; Effective N =
+ 2 (rearranged from the
formula for the standard error estimate).
28
549
550
551
552
553
554
555
556
557
558
559
560
Agreeableness
88% [80%; 93%]
.086
0.0074 [0.0022; 0.0278]
[-0.0908; 0.3108]
Conscientiousness
90% [84%; 94%]
.093
0.0087 [0.0028; 0.0266]
[-0.0564; 0.3698]
Extraversion
68% [54%; 79%]
.068
0.0046 [0.0017; 0.0117]
[-0.0625; 0.2198]
Neuroticism
58% [37%; 72%]
.040
0.0016 [0.0004; 0.0073]
[0.0142; 0.1845]
Openness
87% [78%; 92%]
.090
0.0081 [0.0027; 0.0345]
[-0.0070; 0.4027]
Body mass index
96% [95%; 97%]
.086
0.0074 [0.0038; 0.0129]
[-0.0205; 0.3267]
Height
91% [89%; 92%]
.098
0.0096 [0.0069; 0.0167]
[0.0408; 0.4091]
Waist-to-hip ratio
68% [18%; 88%]
.052
0.0027 [0.0001; 0.0380]
[-0.0265; 0.3380]
Depression
55% [0%; 81%]
.022
0.0005 [0.0000; 0.0085]
[0.0728; 0.2052]
Diabetes
78% [55%; 90%]
.072
0.0052 [0.0005; 0.0445]
[-0.0531; 0.3391]
GAD
51% [0%; 80%]
.076
0.0058 [0.0000; 0.0734]
[-0.0987; 0.3607]
Table 2. Heterogeneity statistics for each trait’s meta-analysis. CI = confidence interval, I2 =
Higgins & Thompson’s I2 statistic, a measure of between-study heterogeneity, τ = the estimated
standard deviation of the true effect size, τ2 = the estimated variance of the true effect size; EA =
educational attainment, IQ = intelligence quotient, AUD = alcohol use disorder, SUD = substance
use disorder, GAD = generalized anxiety disorder.
29
561
562
563
564
The meta-analyzed random effects spousal correlations and 95% confidence
intervals for each trait.
Figure 1
