IQ and relationships: partners match, but only partly

0 The short answer

Partners are similar in measured IQ, with correlations that run from about 0.2 to about 0.45 depending on the study, but that similarity describes populations of couples and says almost nothing about any single relationship. A 2023 meta-analysis found 0.44 across 17 samples, while the UK Biobank, using a brief 13 item test, found 0.23. Cohort studies link higher childhood IQ with a lower chance of later divorce, and the evidence on marrying at all differs by cohort and by sex. Dating studies mostly measured stated preferences, partner ratings and choices at speed dating events, not the IQ scores of the people being chosen.

1 What Is Assortative Mating, and How Is a Partner Correlation Measured?

Assortative mating means that partners are more alike on a trait than two people paired at random would be, and a partner correlation is the number that summarizes how much more alike they are. The most complete recent synthesis, a 2023 paper in Nature Human Behaviour by Horwitz, Balbona, Paulich and Keller, describes the pattern this way: nonzero correlations between mates are overwhelmingly positive, and only a handful of significant negative correlations, called disassortative mating, appear in the literature. The phrase "opposites attract" does not describe the average pattern for the 22 traits they reviewed.

To compute a partner correlation, researchers measure the same trait in both members of many couples and correlate the two columns of scores. Each couple contributes one pair of numbers. A correlation of 0 would mean that knowing one partner's score tells you nothing about the other's. A correlation of 1 would mean that the second score could be read off the first. The figure is a Pearson correlation for continuous traits, and the unit is not points of IQ. It is the slope of the relationship once both scores are expressed in standard deviation units, which is why the page on the standard deviation of 15 is the right background for reading it.

Horwitz and colleagues also separate several mechanisms that can produce a similar number. Phenotypic homogamy occurs when partners match directly on the trait, for instance by preferring similarity, or indirectly, when partnerships form inside strata that carry trait values. Social homogamy occurs when partners match on non-heritable aspects of a trait, for example through the social networks they belong to. Genetic homogamy occurs when partners match on heritable aspects. Convergence is different: partners become more alike over time, and the authors state that it is not a form of assortative mating because it does not come from initial matching. A single observed correlation cannot say which of these produced it, and the mechanisms can operate together.

Two further features of a partner correlation matter for everything that follows. First, it is measured only among couples that exist. People who never partnered, or whose partnerships ended before the study, are not in the data, so the figure describes who ended up together rather than who was available. Second, it depends on the population. The same trait can show a different correlation in different samples and populations, and the same Horwitz paper reports high heterogeneity between samples for most of the traits it meta-analyzed, which the authors attribute to both systematic differences between samples and true differences in partner correlations across populations.

For intelligence in particular, "the trait" is whatever the test measures. The page on what IQ measures explains why a full scale score blends several abilities, and the page on how IQ scores are normed explains the reference group behind a score. The next section sets out the main estimates and the reasons they differ.

2 How Alike Are Partners in IQ? The Main Estimates Compared

The best available estimates put the partner correlation for IQ between about 0.2 and about 0.45, higher than for height or personality and lower than for education or political attitudes. The meta-analysis by Horwitz and colleagues searched ScienceDirect, PubMed and Google Scholar and incorporated 480 partner correlations from 199 peer reviewed studies of co-parents, engaged pairs, married pairs and cohabiting pairs published on or before August 16, 2022. For the trait labeled IQ score, the random effects estimate was 0.44, with a Bonferroni adjusted 95 percent interval of 0.23 to 0.61, from 17 samples drawn from 13 studies and 5,672 pairs. The heterogeneity statistic I squared was 0.94 and the prediction interval, the range in which a future meta-analysis of the trait is expected to fall, ran from minus 0.03 to 0.74.

The same paper analyzed up to 79,074 male-female couples in the UK Biobank. There, fluid IQ, evaluated with a brief thirteen item test, showed what the authors call a somewhat modest correlation of 0.23, while educational attainment between partners reached 0.48. The 0.23 sits inside the 0.44 estimate's prediction interval, which is our observation rather than a statement by the authors. It suggests the two figures are not necessarily in conflict, but they are not the same measurement either: one pools published samples with different tests, and the other uses one short fluid reasoning test in one large cohort.

A review by Plomin and Deary in Molecular Psychiatry (2015) reports a similar order of magnitude. They state that assortative mating for intelligence is about 0.40, compared with about 0.20 for height and weight and about 0.10 for personality, and that verbal intelligence shows more assortative mating (about 0.50) than nonverbal intelligence (about 0.30). Their suggested reason is that it is easier to gauge someone's vocabulary than their spatial ability. That is a hypothesis, flagged with "perhaps" in the review, and it is consistent with the distinction between crystallized and fluid abilities explained on the page on fluid versus crystallized intelligence.

Several readings follow from the table. Intelligence ranks near the top of the cognitive and physical traits, below education and political and religious attitudes. The meta-analytic IQ estimate is also far more uncertain than the education estimate, because it rests on 5,672 pairs rather than nearly two million. Chronotype, the tendency to be a morning or evening person, has the lowest of the 133 UK Biobank correlations, which makes it a useful reminder that similarity is not universal; the page on night owls and IQ takes up what that trait does and does not relate to.

Why do the estimates differ? Horwitz and colleagues attribute heterogeneity in general to differences between samples and true differences between populations. For IQ specifically, three further reasons are plausible, and they are our reasoning rather than findings of the paper. A brief test has more measurement error than a full battery, and error in either partner's score pulls the observed correlation toward zero. Tests that weight verbal knowledge would be expected to show higher similarity, in line with the Plomin and Deary figures. And the studies pooled in a meta-analysis come from different decades and countries, so they reflect different marriage markets. None of these explanations has been tested by the sources cited here.

3 What a Correlation of 0.4 Means for One Couple: Our Arithmetic

A partner correlation of 0.4 is high compared with most other traits and still a weak predictor for any one pair, because most of the variation in a partner's IQ is left unexplained. The arithmetic below is ours, not a finding of any study. It assumes that both partners' scores follow a normal distribution with a mean of 100 and a standard deviation of 15, that the two scores are bivariate normal with the stated correlation, and that scores are measured without error, so real couples would show a looser pattern.

Squaring the correlation gives the share of variance in one partner's score that a straight line from the other partner's score accounts for: 0.16 for a correlation of 0.40, 0.19 for 0.44 and 0.05 for 0.23. Put plainly, about 84 percent of the variation in a partner's IQ remains unexplained at a correlation of 0.40. For a person whose own IQ is 130, the expected score of a partner is 100 plus 0.40 times 30, which is 112, and the spread of partner scores around that expectation is about 13.7 points, only a little smaller than the 15 points in the general population.

The last column compares all couples regardless of own score, using the standard deviation of the difference between two scores. Even at a correlation of 0.44, about one couple in five would differ by 20 points or more, which is more than a standard deviation. At the same time, a person at 130 is several times more likely to have a partner at or above 130 than random pairing would imply: 10.6 percent against 2.3 percent. Both statements are true together. A correlation moves the odds without fixing the outcome.

This is also why a single couple cannot be read from the correlation. The figure describes a pattern across many pairs. It does not say that your partner is near your score, that a gap predicts trouble or that a match predicts happiness. For the rarity of any single score, the calculator for how rare an IQ is shows the percentile and the proportion above it, and the page on IQ score versus percentile explains why the two cannot be averaged or compared as if they were the same unit.

4 Education, Selection and Convergence: How Partners Come to Match

Partners appear to match on IQ largely through selection at the start of a relationship, with education acting as one channel, rather than by growing more alike after they meet. Plomin and Deary state that assortative mating for intelligence is caused by initial selection of a mate rather than by couples becoming more similar after living together, and they note that, in part, spouses select each other for intelligence on the basis of education: spouses correlate about 0.60 for years of education, and education correlates about 0.45 with intelligence. The Horwitz meta-analysis gives 0.55 for educational attainment across 86 samples and 1,911,720 pairs, third highest of its 22 meta-analyzed traits after political values and religiosity, and 0.48 in the UK Biobank.

The education channel has a genetic signature that helps to interpret it. The authors cite earlier evidence of genetic homogamy for educational attainment in the UK Biobank, which they take to suggest that partners are assorting heavily on a different trait, intelligence quotient score, that is more genetically than environmentally correlated with education level. That is an inference from the pattern, not a direct measurement of IQ in those couples. The Norwegian Mother, Father and Child Cohort Study analysis by Torvik and colleagues points the same way. Using polygenic scores and phenotypic data from 26,681 pairs of partners, the authors found genetic similarity between partners for educational attainment (a genetic correlation of 0.37), height (0.13) and depression (0.08), and concluded that indirect assortment on secondary traits accounted for partner similarity in education and depression, though not in height. Their comparisons also indicated that genetic variances were in intergenerational equilibrium, and that assortative mating has taken place for several generations.

One reading of these findings, which is our inference and not a result of the papers, is that a partner correlation for IQ is built from several routes at once: education, income and social circles correlate with test scores, and people meet and pair inside those strata.

What about convergence? The evidence cited above favors selection for intelligence, but the picture is not identical for every trait. In a study of 291 newlyweds by Luo and Klohnen, couples showed substantial similarity on attitude-related domains and little on personality-related domains, and the authors reported that similarity was not due to social homogamy or convergence. That abstract does not report a result for cognitive ability, so it supports the general point that similarity can exist at the outset of a marriage, not a specific claim about IQ. Plomin and Deary also note that for a few other traits, such as social attitudes, smoking and drinking, convergence might play a part.

Putting these studies together yields a cautious description. Partners match on schooling and on correlated characteristics, and a large part of the IQ resemblance probably arrives with the couple. It is not evidence that people consciously screen for test scores. Nobody in these datasets was shown to choose a partner by IQ; the results are consistent with sorting through education, occupation, neighborhood and social networks as well as through direct perception of ability. The page on average IQ by education level and the page on IQ and academic achievement cover the link between schooling and test scores that makes education such a strong sorting variable, and the page on IQ and income covers the economic side.

5 What Assortative Mating Does to Genetic Estimates

Assortative mating matters for behavioral genetics because it changes the genetic structure of a population and can bias the estimates that genetic designs produce, but the direction of the bias depends on the design. Plomin and Deary explain that assortative mating on a polygenic trait increases additive genetic variance in the next generation, because children of two parents who are both high or both low on the trait differ more from the average than they would under random mating. Their worked example, taken from the quantitative genetics text of Falconer and Mackay, is that if the heritability of intelligence under random mating were 0.40, additive genetic variance would increase by one-quarter at equilibrium given assortative mating of 0.40. That is a textbook calculation for an illustrative value, not a measured result for the population.

The Horwitz paper lists the estimates that can be affected: genetic correlations, latent variances from twin and family studies, Mendelian randomization and SNP heritability. The authors say that, to the degree that observed partner similarity is due to these mating mechanisms, some degree of bias from assortative mating may exist in those estimates. The word "may" matters. Whether a given estimate is pushed up or down depends on the design and on the assumptions it makes about random mating, which is why the page on what heritability actually means treats the number as a property of a population and a design rather than a fixed fact about a person.

A concrete case is the correlation between height and IQ. In a study of monozygotic and dizygotic twins, their siblings and their parents (total N = 7,905), Keller and colleagues modeled the covariation of the two traits in a design that accounts for assortative mating. After that adjustment, they found the correlation between height and IQ to be almost entirely genetic in nature, and both pleiotropy (the same genes affecting both traits) and assortative mating contributed significantly and about equally to the genetic correlation. The page on height and IQ covers that finding in its own right. For this page, the point is narrower: partner similarity is not only a social curiosity. It is a parameter that researchers must model if they want to separate shared genes from correlated mating choices.

None of this speaks to whether a given child will resemble a given parent, and none of it says that matched partners produce brighter or duller children. The sources describe population level variance, not family level prediction.

6 Do People With Higher IQ Scores Marry More? Two Scottish Cohorts

Cohort studies that measured IQ in childhood and marital status decades later give results that differ by sex and by birth cohort, so there is no single answer to whether higher IQ goes with marrying. The first study linked the Scottish Mental Survey 1932 to the Midspan studies. Taylor and colleagues examined the influence of IQ at age 11 on marital status by mid-adulthood in 883 subjects. They found an interaction between sex and marital status (p = 0.0001). Women who had ever married had lower mean childhood IQ scores than women who had never married (p < 0.001), whereas there was a trend for ever-married men to have higher childhood IQ scores than never-married men (p = 0.07). Per standard deviation of childhood IQ, the odds ratio of ever marrying was 1.35 in men (95 percent interval 0.98 to 1.86) and 0.42 in women (0.27 to 0.64).

Mid-life social class showed an association with marriage too, with women in more professional jobs and men in more manual jobs less likely to have ever married. Adjusting for social class and height attenuated the IQ associations somewhat, and the authors concluded that IQ, height and social class acted partly independently. In other words, the women's result cannot be reduced to one explanation in this dataset, and the paper does not test why it occurred. A reader should note that the participants were tested as children in 1932, so the marriage patterns belong to a particular historical period.

The second study is the Aberdeen Children of the 1950s cohort, analyzed by von Stumm, Batty and Deary in a sample of 9,614. Intelligence at age 11 and occupational social status at midlife (ages 46 to 51) were examined against marital status and reproduction. In this cohort, no meaningful difference in childhood intelligence was found between ever-married and never-married individuals, which differs from the Scottish 1932 result for women. By our arithmetic the Aberdeen children, born between 1950 and 1956, were tested roughly three decades after the 1932 survey, though the abstract does not offer that gap as an explanation.

The comparison teaches a methodological lesson. Both studies are observational, both rely on a single childhood test, and both treat marriage as an outcome that depends on social norms, cohabitation, employment and who was available to marry, so a result from one decade cannot be carried to another. Who has children is a separate question, which the page on IQ and fertility examines.

7 Divorce and Relationship Quality: What Cohorts and Couple Studies Report

Childhood IQ is modestly associated with a lower chance of being divorced at midlife in one large cohort, but a study of couples found no link between intellectual compatibility and satisfaction. In the Aberdeen cohort, male and female divorcees had lower childhood intelligence test scores than their married counterparts. Higher intelligence scores were associated with being married rather than divorced at midlife, with odds ratios of 0.86 (95 percent interval 0.76 to 0.99) in men and 0.87 (0.77 to 0.98) in women. The abstract does not state how many IQ points a unit represents, so the figures should be read as modest associations rather than as a point-for-point conversion. The intervals come close to 1.0, and the sample is a single Scottish cohort.

A paper that combines theory with an observational analysis, Holley, Yabiku and Benin (2006) in the Journal of Family Issues, proposed three hypotheses for a link between intelligence and divorce, based on ascribed statuses, achieved statuses and direct mechanisms. Their abstract reports that the results are consistent with a direct influence of intelligence on divorce, net of ascribed and achieved statuses, and explains the observed effect through three aspects of intelligence: direction, adaptation and criticism. Two cautions apply. "Consistent with" is a weaker phrase than "shows", and any analysis that adjusts for background statuses cannot rule out unmeasured differences between people who divorce and people who do not. We cite the abstract only, which carries no figures.

When the unit changes from individuals to couples, the finding becomes less favorable to the idea of intellectual compatibility. Gignac and Zajenkowski (2019) studied 218 couples, measured IQ with Advanced Progressive Matrices and asked each person to estimate their own and their partner's IQ. People overestimated their own IQ by about 30 points and their partner's by 38 points for women and 36 for men. Both sexes predicted the partner's IQ with some accuracy (women r = 0.30, men r = 0.19), and the difference was not statistically significant. The degree of intellectual compatibility, whether objectively or subjectively assessed, failed to correlate significantly with relationship satisfaction for both sexes. A later study of 148 couples by Górniak and colleagues found that women perceived angrier men as less intelligent, even after controlling for the men's objective intelligence, and that perceived partner intelligence mediated the link between a man's anger and satisfaction for both sexes. In these data, how smart a partner seems appears to matter more than how smart the partner tests, though a mediation analysis on one sample cannot establish a causal chain.

For relationship quality in general, the stronger evidence concerns characteristics other than IQ. In the newlywed study by Luo and Klohnen, similarity was positively associated with marital quality for personality-related domains but not for attitude-related domains, similarity on attachment characteristics was most strongly predictive of satisfaction, and there were curvilinear effects for husbands but not for wives. The abstract does not list cognitive ability among the findings, so the study should not be read as evidence for or against IQ similarity. For the broader picture of intelligence and wellbeing, see the page on IQ and happiness. None of these studies randomized IQ or compatibility, so none can say that a match or a mismatch causes satisfaction or divorce.

8 What Dating Preference Studies Actually Measured

Studies of mate preferences measure what people say they want, how they rate partners and whom they choose at a short event, and none of those is the same as the IQ of the person chosen. This matters because headlines about dating often move from "people say they want intelligence" to "smart people are more attractive" to "smart people match with smart people", and each step uses a different measurement. The studies below are the ones we could verify, with the measurement named each time.

In Li, Bailey, Kenrick and Linsenmeier (2002), three studies tested the idea that people first secure necessities in a mate and only then consider luxuries. In two studies, participants designed ideal long-term mates by purchasing characteristics with three different budgets; the third used a mate-screening paradigm in which people asked first about hypothesized necessities. The authors report that physical attractiveness was a necessity to men, status and resources were necessities to women, and kindness and intelligence were necessities to both. The measure here is a design task, a hypothetical purchase, not an actual choice of a partner.

Fisman, Iyengar, Kamenica and Simonson (2006) used a speed dating experiment in which the researchers generated random matching of subjects and random variation in the number of potential partners, which let them observe individual decisions rather than only final matches. Their abstract reports that women put greater weight on the intelligence and the race of the partner, while men respond more to physical attractiveness, and that men do not value women's intelligence or ambition when it exceeds their own. Male selectivity was invariant to group size, while female selectivity increased strongly with it. We read only the abstract, which treats intelligence as an attribute of the partner that participants weighed, not as a test score; the abstract does not describe an IQ test. Differences in what men and women say they weigh in a partner are separate from differences in test scores, which the page on average IQ by gender covers.

Two studies show how far stated preferences can sit from choices. Todd, Penke, Fasolo and Lenton (2007) followed 46 adults at a speed dating event. Stated preferences did not predict the choices made during the speed dates; men chose women based on physical attractiveness, whereas women, who were generally much more discriminating, chose men whose overall desirability matched the women's self-perceived physical attractiveness. Eastwick and Finkel (2008) combined speed dating with longitudinal follow-up. Participants showed the traditional sex differences when they stated the importance of attractiveness and earning prospects, but the associations between romantic interest in real partners and those partners' attractiveness and earning prospects showed no sex differences, and ideal preferences assessed before the event failed to predict what inspired actual desire. Their traits were attractiveness and earning prospects, not intelligence, so the study bears on intelligence only by analogy: it warns that a survey answer about the ideal partner is a weak guide to choosing a real one.

Perceived and objective intelligence are different quantities, as the Gignac and Zajenkowski overestimates of about 30 points for self and 36 to 38 points for the partner show. If perceived intelligence drives choice, the partner correlation for tested IQ could arise even though nobody ever sees a test score. People may read cues such as vocabulary, education and occupation, which correlate with test scores; that is our reasoning, not a result of the study. The page on IQ and humor takes up one such signal, and the page on how accurate people are at estimating their own IQ covers the wider problem of self-estimates.

9 Intelligence and Looks: A Contested Association

The claim that attractive people are more intelligent comes from correlational studies whose results differ widely by sample and measure, and a twin and sibling study of 1,753 people found no relationship. Kanazawa (2011) used two nationally representative samples. In the United Kingdom, the National Child Development Study followed all babies born in Great Britain during one week of March 1958; teachers described each child's appearance at ages 7 and 11, and a child was coded as attractive if two different teachers chose that adjective, a binary measure. Attractive children had a mean IQ of 104.2 against 91.8 for the rest, a difference of 12.4 points and a correlation of 0.381. In the United States, the Add Health sample measured intelligence with the Peabody Picture Vocabulary Test and used an interviewer's rating of attractiveness, and the correlation was 0.126. The association was stronger among men than among women in both nations and remained significant net of controls for social class, body size and health. Kanazawa describes the two samples as having complementary strengths: the British sample has one of the best measures of general intelligence but a comparatively weak measure of attractiveness, and the American sample has a stronger measure of attractiveness but a comparatively weak measure of intelligence.

Mitchem, Zietsch, Wright, Martin, Hewitt and Keller (2015) approached the question with highly reliable measures of facial attractiveness and IQ in two twin and sibling samples, the Longitudinal Twin Study (n = 399) and the Brisbane Adolescent Twin Study (n = 1,354). They found no evidence of a phenotypic correlation: the standardized regression coefficient was 0.018 (p = 0.50), similar in both sexes, and neither the genetic nor the environmental latent factor correlations were statistically significant. They also meta-analyzed past studies and reported evidence of publication bias. Their survey of earlier work describes the two prior meta-analyses as showing a small-to-moderate correlation in children (about 0.19, weighted by sample size) that diminished with age (about 0.02), and they report that estimated effect sizes tended to fall as sample size rose (r = minus 0.41 with log transformed sample size), a pattern consistent with a bias toward publishing small studies that overestimate the relationship.

Reading the two studies together requires attention to what differs. Kanazawa's UK measure is a teacher's one word judgment of children at ages 7 and 11, a social impression rather than a rating of how a child would appear to prospective partners. Mitchem and colleagues used highly reliable ratings of facial photographs in samples of twins and siblings. The evidence does not support a firm statement that smarter people are better looking, nor a firm denial in every setting; it supports the narrower statements that published estimates range from 0.018 to 0.381 depending on sample and measure, and that the study with the most reliable measures found nothing. If a cross-trait match between intelligence and appearance exists, the data verified here suggest it is smaller than the largest published estimate.

10 The Evidence by Design

Almost every study on this page is observational, and the verbs that fit them are "is associated with", "correlates with" and "was found among", not "causes", "produces" or "leads to". The table sets out each study by design, sample, measure and result so that the strength of the evidence can be read at a glance.

Three design points follow. First, a partner correlation, however large, is a descriptive fact. It says that couples are similar, not why. The studies that probe why, such as the Norwegian cohort and the genetic comparisons discussed in the Horwitz paper, point to education and other correlated characteristics and to selection at the start rather than convergence, but they are inferences from patterns in observational data. Second, the cohort studies of marriage and divorce measure IQ before the outcome, which is a strength for ordering in time, but they cannot remove everything that differs between people who marry, divorce or stay single, such as social class, which the Scottish study found to be partly independent of IQ. Third, the speed dating studies have an element of experimental control, because Fisman and colleagues randomized who met whom, but the randomized thing was the matching of strangers, not IQ, so they cannot show that IQ itself changes attraction.

The page on reliability and validity explains why measurement noise in a short test biases a partner correlation downward. The practical rule is to ask what was measured, in whom, and at what point in a relationship, before accepting a sentence of the form "smart people do X".

11 What the Evidence Supports, Stated Narrowly

The verified evidence supports a short list of statements, each with its design attached, and it does not support a general claim that matching IQ makes relationships work. The list below separates what the studies show from what they leave open.

• Partners are positively correlated in measured IQ. A 2023 meta-analysis found 0.44 across 17 samples and 5,672 pairs, with a wide interval (0.23 to 0.61) and high heterogeneity, and a UK Biobank analysis of up to 79,074 couples found 0.23 with a brief fluid test. A 2015 review gives about 0.40. These are descriptions of populations of couples.
• The correlation is larger for education than for IQ and larger for verbal than for nonverbal ability. Educational attainment was 0.55 in the meta-analysis, and the review figures for verbal and nonverbal intelligence were about 0.50 and about 0.30.
• Partner similarity for IQ appears to come mostly from selection at the start, not from convergence. The review attributes it to initial selection, and the newlywed study found similarity at the outset on attitude domains that was not due to convergence. Education and other correlated characteristics appear to be one route.
• Assortative mating changes how genetic estimates should be read. The authors of the meta-analysis state that some bias may exist in genetic estimates, and a twin and sibling design that accounted for it found height and IQ correlated for genetic reasons, with pleiotropy and assortative mating about equally responsible.
• Childhood IQ is modestly associated with a lower chance of divorce in a large Scottish cohort, with odds ratios of 0.86 and 0.87 in men and women. That is one cohort, and its odds ratios are close to 1.
• The evidence on marrying at all is mixed. The 1932 Scottish cohort showed a sex interaction, with ever-married women scoring lower in childhood, while the Aberdeen cohort showed no meaningful difference between ever-married and never-married people.
• Intellectual compatibility was not related to relationship satisfaction in 218 couples, whether assessed with a test or with estimates.
• Dating preference studies measured preferences, ratings and speed dating decisions. They did not establish that people choose partners by tested IQ, and stated preferences predicted real choices poorly in the studies that checked.
• The link between intelligence and attractiveness is unsettled. One national study reported 0.381 in the UK and 0.126 in the US, and a twin and sibling study of 1,753 people found a coefficient of 0.018, with evidence of publication bias in the earlier literature.

What the evidence does not support is any claim that a score gap predicts a breakup, that a match predicts happiness, or that a partner's IQ can be estimated from one's own with useful precision. Those claims need designs that nobody has run in the work verified here.

12 What This Does Not Say About You

Group correlations between partners, and cohort associations between childhood scores and marital status, do not describe your relationship, your partner or your own prospects. A population correlation of 0.44 leaves about 80 percent of the variance in a partner's score unexplained, as the arithmetic above shows, and an odds ratio of 0.86 describes a shift in odds across thousands of people, with confidence intervals that nearly reach 1.0. A particular couple might be a close match with a stormy marriage, or a wide gap with a stable one. Neither would contradict a single result on this page.

Nor do these studies justify using a test score to choose or judge a partner, or to decide whether to stay. What did relate to satisfaction in the couple studies was similarity on attachment and personality-related characteristics, perceived partner intelligence and, in one sample, how angry a man was. A tested score is one narrow input, measured at one time with one instrument.

The measurement standards point the same way. The Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014), published collaboratively by the three organizations, frame validity as the degree to which evidence supports an interpretation of scores for a proposed use, which means the limits of an interpretation should be stated. The APA Ethical Principles of Psychologists and Code of Conduct, Standard 9.02, says that assessments should be used in a manner and for purposes appropriate in light of the evidence of their usefulness and proper application. A compatibility judgment about a couple, built from two IQ scores, would be a use for which the research above offers no support. An IQ score is not designed or validated to rate relationships, and the page on what IQ scores mean explains how a score, its confidence interval and its norm group should be read.

An ACIS report gives a Full Scale IQ and six index scores, with percentiles and a 95 percent confidence interval, for the person who takes the test. It is online and unsupervised, it is not a clinical or diagnostic instrument, it is not for hiring, school accommodations or admission to high IQ societies, and it does not produce a compatibility rating for two people. The technical manual documents how scores are built and what their limits are. If you are curious about your own profile, the page on choosing a test and reading the result explains the options, and the Full Scale IQ test page describes the form that covers all six domains.

13 Sources Behind This Page

These references separate meta-analytic estimates, cohort studies, genetic designs and preference experiments, so that a result from one design is not mistaken for another. Correlations, odds ratios, sample sizes and intervals keep the context of the papers that report them. Where the page relied on an abstract, the text says so. Percentages that describe a hypothetical couple are our arithmetic and are not attributed to any author.

• Horwitz T, Balbona J, Paulich K and Keller M. Evidence of correlations between human partners based on systematic reviews and meta-analyses of 22 traits and UK Biobank analysis of 133 traits. Nature Human Behaviour, 2023, volume 7, issue 9, pages 1568 to 1583.
• Plomin R and Deary I. Genetics and intelligence differences: five special findings. Molecular Psychiatry, 2015, volume 20, issue 1, pages 98 to 108.
• Torvik F, Eilertsen E, Hannigan L, Cheesman R, Howe L, Magnus P and others. Modeling assortative mating and genetic similarities between partners, siblings, and in-laws. Nature Communications, 2022, volume 13, issue 1.
• Keller M, Garver-Apgar C, Wright M, Martin N, Corley R, Stallings M, Hewitt J and Zietsch B. The genetic correlation between height and IQ: shared genes or assortative mating? PLoS Genetics, 2013, volume 9, issue 4, e1003451.
• Taylor M, Hart C, Smith G, Whalley L, Hole D, Wilson V and Deary I. Childhood IQ and marriage by mid-life: the Scottish Mental Survey 1932 and the Midspan studies. Personality and Individual Differences, 2005, volume 38, issue 7, pages 1621 to 1630.
• von Stumm S, Batty G and Deary I. Marital status and reproduction: associations with childhood intelligence and adult social class in the Aberdeen children of the 1950s study. Intelligence, 2011, volume 39, issues 2 to 3, pages 161 to 167.
• Holley P, Yabiku S and Benin M. The relationship between intelligence and divorce. Journal of Family Issues, 2006, volume 27, issue 12, pages 1723 to 1748.
• Luo S and Klohnen E. Assortative mating and marital quality in newlyweds: a couple-centered approach. Journal of Personality and Social Psychology, 2005, volume 88, issue 2, pages 304 to 326.
• Gignac G and Zajenkowski M. People tend to overestimate their romantic partner's intelligence even more than their own. Intelligence, 2019, volume 73, pages 41 to 51.
• Górniak J, Zajenkowski M, Szymaniak K and Jonason P. Kindness or intelligence? Angry men are perceived as less intelligent by their female romantic partners. Evolutionary Psychology, 2024, volume 22, issue 3.
• Li N, Bailey J, Kenrick D and Linsenmeier J. The necessities and luxuries of mate preferences: testing the tradeoffs. Journal of Personality and Social Psychology, 2002, volume 82, issue 6, pages 947 to 955.
• Fisman R, Iyengar S, Kamenica E and Simonson I. Gender differences in mate selection: evidence from a speed dating experiment. The Quarterly Journal of Economics, 2006, volume 121, issue 2, pages 673 to 697.
• Todd P, Penke L, Fasolo B and Lenton A. Different cognitive processes underlie human mate choices and mate preferences. Proceedings of the National Academy of Sciences, 2007, volume 104, issue 38, pages 15011 to 15016.
• Eastwick P and Finkel E. Sex differences in mate preferences revisited: do people know what they initially desire in a romantic partner? Journal of Personality and Social Psychology, 2008, volume 94, issue 2, pages 245 to 264.
• Kanazawa S. Intelligence and physical attractiveness. Intelligence, 2011, volume 39, issue 1, pages 7 to 14.
• Mitchem D, Zietsch B, Wright M, Martin N, Hewitt J and Keller M. No relationship between intelligence and facial attractiveness in a large, genetically informative sample. Evolution and Human Behavior, 2015, volume 36, issue 3, pages 240 to 247.

The Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014) and the APA Ethical Principles of Psychologists and Code of Conduct, Standard 9.02, provide the interpretive framework: evidence must support the use of a score, and limits must be stated. Citing them does not imply that their authors endorse ACIS or any other product. Page read on October 6, 2026.

14 Frequently Asked Questions