Do Opposites Attract? The Largest Study of Couples, and the Number It Actually Reports
The study everyone quotes measured 133 traits in up to 79,074 couples. Its mean partner correlation is 0.19, and not one page on the first screen reports it.
Before the article. If you are here because you and someone you care about are unalike in a way that has started to feel like a verdict on the relationship, nothing on this page is a diagnosis and no correlation in it is a statement about the two of you. If the difference that worries you is really about one person being told who they may see, where they may go or how they may spend money, that is a different page and a different kind of help: the section near the end says why, and Find A Helpline in the sources lists free and confidential services by country.
With that said. Almost every page answering this question is reporting the same study — Horwitz, Balbona, Paulich and Keller in Nature Human Behaviour, 2023 — and almost none of them reports its numbers. Nine pages from the first screen of results were downloaded and searched on 7 November 2026. None gives the sample size of the analysis they are describing. None quotes a single correlation from it. None links the paper. What they all quote instead is a pair of percentages from the university's press release.
The paper is free to read, and the number it actually leads with is a mean partner correlation of 0.19 across 133 traits. That is a real population-level tendency for partners to resemble each other, and it is small. Underneath it sit several things the coverage skips: that the large correlations are social facts like education and politics rather than anything about temperament, that personality assorts at about the level of a coin flip, that the authors list four mechanisms that could produce their finding and can separate none of them, that the couples were inferred rather than recorded, and that a follow-up on same-sex couples came out in 2025 and nobody is citing it.
This article goes through all of that, and then the separate question of whether resembling your partner predicts anything good — where the evidence genuinely contradicts itself, and says so.
Two different questions are wearing the same four words
If you are reading this because you and someone you care about are unalike in a way that has started to feel like a verdict, a conversation with a therapist or counsellor is a reasonable thing to arrange and not an overreaction. Nothing on this page is a diagnosis, and no correlation in it is a statement about you.
With that said, the first useful thing to do with do opposites attract is to notice that it is two questions pretending to be one, and that they have different answers from different data.
The first is a descriptive question about who ends up with whom. Across the population, do partners turn out to be alike or unalike? That one has been measured on an enormous scale, and the answer is clear in direction and much smaller in size than the headlines suggest.
The second is a causal question about what difference does. Does being unlike someone make them more interesting, more attractive, better to live with? That question has barely been tested at all, and the paper most often credited with the idea is from the 1950s.
Almost every page answering this query answers the first question and then quietly reports it as an answer to the second. They are not the same. Two people can resemble each other for reasons that have nothing to do with either of them choosing it — they grew up in the same town, went to the same kind of school, met at the same workplace, or spent thirty years becoming more alike. A correlation between partners is a fact about a population. It is not evidence about what pulled two particular people together.
That distinction is the whole article, and it is also the distinction this site keeps running into. What a question can and cannot predict about compatibility covers the related problem on the individual scale: matching people on their stated positions has been experimentally tested and did nothing measurable. This page is about the population scale, and the data are much bigger.
Where the phrase comes from, and why you cannot read the paper that started it
The idea that opposites attract is older than psychology, but the version that gets presented as scientific has a name attached. Robert F. Winch proposed the theory of complementary needs in mate selection in the American Sociological Review in 1954, tested one form of it in the same journal in 1955, and published a book-length version, Mate-Selection: A Study of Complementary Needs, in 1958. The claim was that people pair with partners whose psychological needs fit theirs the way a key fits a lock: the person who needs to dominate finds the person who needs to defer.
It is worth being exact about what can and cannot be said here, because this is the point at which most pages start paraphrasing something they have not read. All three of Winch's publications are behind paywalls. So is the 1963 review in Psychological Bulletin — Roland Tharp's Psychological patterning in marriage — that assessed the complementary-needs literature as it stood. So is the 1997 paper in the Journal of Personality and Social Psychology that carries the question as its title: D. Christopher Dryer and Leonard Horowitz, When do opposites attract? Interpersonal complementarity versus similarity. None of the four has a publicly readable abstract. This article therefore does not quote a single number from any of them, and no page that does so without saying where it got the number should be believed.
What can be checked is the shape of the literature that followed, and it is lopsided. Searched on 7 November 2026, PubMed returns 765 records with assortative mating in the title or abstract restricted to human studies, against 15 for disassortative mating — the term for partners being systematically unalike. Europe PMC returns 2,237 and 136 for the same two phrases without the species restriction. Whatever else is true, the question has been asked overwhelmingly in one direction, and the direction it has been asked in is similarity.
And the phrase itself has largely left relationship science. Of the 58 PubMed records with opposites attract in the title or abstract, most are about magnets, antibodies, nanoparticles, plant hormones, the Casimir force, and — twice — children who have swallowed magnets. The largest research programme in the world actually called Opposites Attract is a cohort study of HIV transmission in serodiscordant couples. The idea survives as a headline pun more comfortably than as a hypothesis.
The study everyone is quoting, and what it actually did
Nearly every page on the first screen of results for this question is reporting the same piece of work, usually without naming it in full.
It is Tanya Horwitz, Jared Balbona, Katie Paulich and Matthew Keller, published in Nature Human Behaviour in 2023. There are two separate studies inside it, and conflating them is the most common error in the coverage.
The meta-analysis. The authors searched ScienceDirect, PubMed and Google Scholar for partner correlations published on or before 16 August 2022, and pooled 480 correlations from 199 peer-reviewed studies across 22 traits. The samples are co-parents, engaged pairs, married pairs and cohabiting pairs. The number of partners behind a single trait ranges from 2,527 for generalised anxiety to 2,727,151 for diabetes — which is where the phrase millions of couples in the headlines comes from, and it applies to that one trait, not to the study as a whole. The oldest study in the pile is Karl Pearson and Alice Lee's 1903 paper in Biometrika, which is also roughly where the statistical machinery for measuring this was invented.
The UK Biobank analysis. Separately, the authors calculated partner correlations for 133 traits using up to 79,074 male-female couples in the UK Biobank. These are 59 continuous traits, 24 ordinal and 50 dichotomous, and they include things nobody had looked at before — hours of television, how many sexual partners someone has had, whether they were breastfed.
The headline result is not in dispute and this article is not disputing it. Across both halves, partner correlations are overwhelmingly positive. In the paper's own words, the literature contains only a handful of examples of significant negative partner correlations. People end up with people like them, on average, on most measurable things.
The interesting part is everything underneath that sentence, and it starts with a number that the coverage leaves out entirely.
The number the paper leads with, and nobody repeats: 0.19
Nine pages from the first screen of results for this question could be downloaded as plain text and searched on 7 November 2026. Among them: the Jerusalem Post, Earth.com, Live Science, UPI, BetterHelp, Yahoo News, a hospital-system health blog and a radio station's entertainment page.
Across all nine: zero mention the figure 79,074. Zero report a single correlation coefficient from the paper. Zero link the DOI. Zero contain the string et al. The one number that would let a reader judge the size of the finding appears nowhere.
Here it is. Of the 133 adequately powered traits in the UK Biobank analysis, the mean partner correlation was 0.19.
A correlation of 0.19 means that knowing one partner's score on a trait leaves you knowing about four per cent of what you would need to know to predict the other's. It is a real, reliably non-zero, population-level tendency for partners to resemble each other. It is not a rule, it is not close to a rule, and it does not describe any individual couple. In the same dataset the figures run from −0.18 for chronotype — being a morning or an evening person — to 0.87 for year of birth.
The authors themselves say so. The press release from the University of Colorado Boulder, which is where most of the coverage was sourced, quotes them cautioning that the correlations they found were fairly modest and should not be overstated, and the first author adds a reason that is worth repeating: assortative-mating research was historically co-opted by the eugenics movement, and overstating resemblance between partners is not a harmless error.
This site keeps arriving at the same place from different directions. The pattern in the three-day rule and the 3-6-9 rule was a number with no source. The pattern here is the opposite and just as misleading: an excellent source, with its central number left out.
Where the 82% and the 89% actually come from
If you have read anything about this study you have met the figure 82% to 89% of traits. It is in the Jerusalem Post, in Earth.com, in UPI, in Yahoo, in Live Science. It is often the only number in the article.
It is not in the paper. It is a sentence from the CU Boulder press release: for between 82% and 89% of traits analyzed … partners were more likely than not to be similar, with a companion sentence that for only 3% of traits, in one part of the analysis, people tended to partner with someone different.
What the paper reports instead, in its own results section, is this. Of the 133 UK Biobank traits, 118 (88.72%) were statistically significant after correction for multiple testing and 15 (11.28%) were not. Of those 118, 115 (97.46%) were positive and three (2.54%) were negative.
Those are compatible numbers, not contradictory ones. But notice what gets lost in the translation. Significant and similar are different claims: a correlation of 0.04 for overall neuroticism was significant in a sample that large, and it is not similarity in any sense a reader would recognise. A percentage of traits tells you how often the needle moved. It tells you nothing about how far.
This is the shape of the whole genre's problem with this study. The press release was accurate and careful; the coverage kept the headline percentage and dropped the caution, the effect sizes, the sample, and the authors' own warning about overstatement. Nine pages later, the public version of the finding is science proves opposites don't attract, which is both stronger and vaguer than anything the researchers wrote.
The same transformation is traceable in detail elsewhere on this site: what the research on the five love languages actually measured follows one popular matching idea from the primary studies to the quiz, and the real numbers on modern dating does it for the statistics that circulate without a denominator.
The traits where partners really do resemble each other are social, not personal
The single most useful thing in the 2023 paper is not that correlations are positive. It is which ones are large, because they are not the ones the folklore is about.
From the meta-analysis, with the pooled correlation for each trait:
- Political values: .58
- Religiosity: .56
- Educational attainment: .55 — across 86 correlations and 1,911,720 partners
- Smoking status: .49
- IQ score: .44
- Drinking: .42
- Smoking initiation: .38
And in the UK Biobank, educational attainment came out at .48, with year of birth at .87 and number of children at .70.
Every one of those is a social fact before it is a personal one. Where you went to school, what you believe about politics and religion, whether you drink, how old you are: these are the variables that decide which rooms you are in. Two people with the same education, in the same city, of the same age, with the same attitude to a pub, have a far higher chance of meeting at all than two people without that. The correlation is partly a measurement of who gets introduced to whom.
The authors do not claim otherwise. They set out four mechanisms that can produce a partner correlation, and only one of them is people prefer similarity. A later section goes through all four, because it is the part the coverage skips entirely.
Two things follow for a reader. First, when someone says couples who have a lot in common do better, check what the lot in common consists of — the large correlations are demographic, and nobody chooses their birth year. Second, the one demographic difference that gets treated as a verdict more often than any other does not behave the way people assume: in the largest American study to ask, the size of an age gap predicted nothing about whether a marriage ended, which what an age gap does and does not predict sets out with its limits.
Personality — the thing people actually mean — barely assorts at all
When somebody asks whether opposites attract, they are almost never asking about educational attainment. They are asking about temperament: whether the quiet one and the loud one work, whether two anxious people are a disaster, whether a planner should be with someone who improvises.
That is the part of the dataset where almost nothing happens.
From the meta-analysis, the Big Five pooled correlations between partners:
- Extraversion: .08, across 40 correlations and 32,729 partners
- Agreeableness: .11
- Neuroticism: .11
- Conscientiousness: .16
- Openness: .21
In the UK Biobank, the overall neuroticism score came out at .04. Among the twelve yes/no items that make up the neuroticism scale, correlations ranged from −.09 to .16 — some slightly negative, some slightly positive, all of them tiny.
The physical traits are not much larger: height .24, body mass index .16, waist-to-hip ratio .17 in the meta-analysis.
The first author's own summary of the extraversion result, in the press release, is the plainest statement of it anywhere: people have theories about extroverts liking introverts or extroverts liking other extroverts, but the fact of the matter is that it's about like flipping a coin.
This is the result that should have been the headline, and it is not opposites attract either. It is that on the dimension the question is really about, neither folk theory is true. Personality does not sort people into couples in any direction worth noticing. The dichotomy — birds of a feather versus opposites attract — is a false one, and the honest answer to the question as most people mean it is that the variable does not do much.
Which is consistent with what shows up when the question is asked the other way round. The largest study ever run on what predicts relationship quality found that individual traits added nothing once a person's own experience of their relationship was in the model; what builds attraction over time works through that finding, its 11,196 couples, and what it means for the idea that becoming a certain sort of person is a relationship strategy.
The three negative correlations, and the one the paper names
In the whole UK Biobank analysis — 133 traits, up to 79,074 couples — exactly three correlations were both significant and negative. That is the entire measured evidence for opposites attracting in the largest dataset ever assembled on the question.
The paper's results section names one of them. Chronotype, dichotomised into morning person versus evening person, at r = −0.18, with a Bonferroni-adjusted confidence interval of [−0.20, −0.16]. It is described as the largest significant negative correlation across all UK Biobank traits.
The other two are named only in the CU Boulder press release, which lists the negatives as chronotype, tendency to worry and hearing difficulty. Those are not in the paper's main text; they would be in the supplementary tables, which are not openly readable. So that is how they are reported here: two trait names attributed to the university's own announcement rather than to a table anybody can check.
Now the part worth sitting with. Several articles describe the chronotype result as night owls pairing up with early risers. At −0.18, that is not what it says. It says that across tens of thousands of couples there is a small tendency in that direction, about the same size as the average positive correlation in the dataset pointed the other way, and that it explains roughly three per cent of the variation. In any given couple it predicts essentially nothing.
And even taken at face value, look at what the finding is about. The best-supported instance of opposites attracting in human partnerships is what time two people prefer to go to bed. It is not temperament, not values, not life goals, not any of the things the phrase is used to justify. The authors say more work is needed to unpack it, and they are right: a negative correlation on sleep timing has at least one obvious non-romantic explanation, which is that two people sharing a bed and a schedule may answer a question about being a morning person differently depending on who gets up with the alarm.
Resembling each other is not the same as choosing each other
This is the section the coverage omits, and it is in the paper's introduction, before any result.
The authors list the distinct processes that can produce a correlation between partners on a trait:
- Phenotypic homogamy. Partners match directly on the trait itself — the thing people mean by preferring similarity. The authors note this can also happen indirectly, through selection on something correlated with it.
- Social homogamy. Partners match because of where and among whom they were available to meet, not because of anything either of them weighed up.
- Genetic homogamy. The genetic correlation between partners is higher than the trait correlation would predict, because they are actually sorting on something else that is more strongly genetic. The authors give educational attainment as the example: partners in the UK Biobank appear to be assorting heavily on IQ, which is more genetically than environmentally correlated with level of education.
- Convergence. Partners become more alike over time, through influencing each other or through sharing an environment.
Then the sentence that should be in every article on this subject: convergence is not assortative mating at all, because it is not a consequence of initial matching. And the authors are explicit that these mechanisms are not mutually exclusive and that their data cannot separate them. Their discussion says that future studies will be necessary to understand the specific processes underlying partner similarity.
So the honest reading of the headline figure is narrower than it looks. Partners resemble each other is established. People are attracted to people like themselves is one of four candidate explanations for it, and this study does not test it.
It matters practically, not just pedantically. A couple who have grown alike over fifteen years and a couple who selected each other for similarity on day one produce the same number in a spreadsheet, and the advice you would draw from the two cases is completely different. The same confusion between what a measurement contains and what a reader assumes it contains runs through whether dating apps' matching actually does anything, where the algorithms are sold on a prediction nobody has demonstrated.
The couples in the headline were not couples on record
Here is the methodological fact that is both the most surprising thing in this literature and entirely absent from every page covering it. The UK Biobank does not record who is partnered with whom.
It is a cohort of roughly half a million people in the United Kingdom, recruited between the ages of 40 and 69, and the relationships between participants are not reported. Every study of couples in that dataset — including the 2023 one — is working with putative couples: pairs of participants who were inferred to be a couple, by living at the same address, saying they live with a spouse or partner, not being genetically related, and giving matching answers about their household.
The chain in the 2023 paper runs from 502,414 participants down through each of those filters. Along the way it drops people who report an unrelated housemate, keeps only pairs at the same address, removes related pairs, requires agreement on household size, on owning versus renting, and on the deprivation index of the area, removes same-sex pairs, and removes pairs whose reported household income differs by more than one bracket.
The same lab has now published a paper about exactly this problem. Kristen Kelly and colleagues, in Social Science & Medicine in 2025, point out that because the colocation data is not available to new researchers, different teams have used different criteria, producing highly discrepant sample sizes across papers — estimates of the number of couples in the UK Biobank have ranged from under 19,000 to over 105,000. They developed and released a standardised method, which identifies 92,510 putative couples geographically against 89,278 from the colocation data.
And they measured what the old criteria do. The pairs excluded by an income-matching rule — 15,295 of them — are less similar on age, educational attainment, physical activity, television hours and diet than the pairs it keeps. In other words a rule applied to make the sample cleaner also removes the couples who resemble each other least, on several of the very traits that carry the headline. They report the same for pairs with missing income data and for pairs whose time at their current address differs by five years or more.
None of this makes the 2023 finding wrong. Kelly and colleagues say directly that the correlations from their samples are consistent with it. What it does is put a ceiling on how confidently anyone should recite a specific number, and it is the kind of detail that only becomes visible if you read the paper rather than the press release.
The 2025 follow-up nobody is citing: same-sex couples
The 2023 study excluded same-sex couples. The press release said the authors were studying those patterns separately. That paper came out in August 2025, it is open access, and not one page in the search results for this question mentions it.
Tanya Horwitz, Kristen Kelly and Matthew Keller, in Scientific Reports, compared partner correlations across 66 traits in the UK Biobank among 89,628 female-male, 236 female-female and 216 male-male putative couples.
The mean correlation was 0.19 for opposite-sex pairs, 0.15 for female-female pairs and 0.15 for male-male pairs. The medians were 0.14, 0.14 and 0.15 — nearly identical.
The difference in statistical significance is dramatic and is mostly not about effect sizes. 93.94% of the 66 correlations were significant in opposite-sex pairs, against 25.76% in male-male and 10.61% in female-female pairs — which the authors attribute mostly to the difference in sample size, since the same-sex groups are a few hundred pairs against ninety thousand. After correction for multiple testing, there was no significant difference in assortment between any of the groups for 64 of the 66 traits. The two exceptions, where opposite-sex pairs assorted more strongly, were year of birth and number of children.
No negative correlation reached significance in either same-sex group.
And the detail that makes the point better than any of the statistics. The traits with the highest concordance in male-male pairs were time spent watching television per day, alcohol intake frequency and year of birth, at correlations between .56 and .57, with the same three leading in female-female pairs at .54 to .59. In opposite-sex pairs the leaders were year of birth at .86, number of children at .70 and alcohol intake at .57.
The strongest measured resemblance between partners, once you set age aside, is how much television they watch and how often they drink. That is not a romance-shaped finding, and it is the best illustration in the dataset of why a partner correlation is a fact about shared circumstances at least as much as a fact about choosing.
What people say they want, and the one study that tested the decision rule
All of the above is about who people end up with. There is a separate question about what they go looking for, and it has been tested directly in a way that is free to read.
Peter Buston and Stephen Emlen, in the Proceedings of the National Academy of Sciences in 2003, gave a two-part questionnaire to 978 heterosexual residents of Ithaca, New York, aged 18 to 24. Participants first rated how important ten attributes were in a long-term partner, then rated themselves on the same ten.
The design is the useful part. Three different decision rules predict different patterns. An opposites-attract rule predicts that people who rate themselves low on an attribute will demand it more in a partner. A reproductive-potentials-attract rule predicts that everyone demands as much as possible, moderated by their own market value in general. A likes-attract rule predicts that self-rating on a specific attribute drives selectivity on that same attribute.
The third one won. Both women and men who rated themselves highly were significantly more selective. When the ten attributes were grouped into four categories — wealth and status, family commitment, physical appearance, sexual fidelity — the variation in selectivity within each category was best explained by self-perception in that same category. The authors conclude that people in this population use neither an opposites-attract nor a reproductive-potentials-attract rule, but a likes-attract rule.
The limits are substantial and worth stating: 978 young heterosexual adults in one American college town, reporting stated preferences on a form, more than twenty years ago. And stated preferences are a famously weak instrument — the meta-analysis of ideal partner preferences across 97 samples found that what people say they want predicts how they evaluate a profile and stops predicting once they have met someone, which is the finding at the centre of why capable people keep making the same choice badly.
So the most that can be said is that when people describe the partner they are looking for, the description points at themselves rather than at their complement. What that description is worth once an actual person is sitting opposite them is a different and much less flattering story.
Does resembling your partner predict a good relationship? Mostly, no
Suppose the descriptive finding is entirely true and you and your partner are unusually alike. Does that forecast anything?
The largest test is Portia Dyrenforth, Deborah Kashy, M. Brent Donnellan and Richard Lucas, in the Journal of Personality and Social Psychology in 2010. They used three nationally representative samples of married couples — Australia (N = 5,278), the United Kingdom (N = 6,554) and Germany (N = 11,418) — and separated three things personality can do: your own traits (actor effects), your partner's traits (partner effects), and how similar the two of you are (similarity effects).
The results, replicated across all three countries:
- Actor effects accounted for about 6% of the variance in relationship satisfaction.
- Partner effects — largest for agreeableness, conscientiousness and emotional stability — accounted for 1% to 3%.
- Couple similarity consistently explained less than 0.5%, after controlling for the other two.
Twenty-three thousand people in three countries, and how alike two partners are does essentially no work. Your own disposition matters six to twelve times more than the match between you.
David Watson and colleagues reached the same place from a different direction in a newlywed sample, reporting strong similarity on age, religiousness and political orientation, moderate on education and verbal intelligence, and little similarity in personality, emotional expression and attachment — and that marital satisfaction was primarily a function of the rater's own traits and showed little relation to spousal similarity.
And there is a third instance of the same pattern on a trait people assume is decisive. In a study of 167 couples with physical attractiveness rated from video by trained coders, how well the two partners matched on attractiveness had no association with relationship satisfaction for either of them. That study, and what it found about how matching decays with longer acquaintance, is covered in what builds attraction over time.
The consistent message across all three is unglamorous: whether you resemble your partner is a poor predictor of whether you are happy with them. What does predict is collected in what 137 studies found about couples who stay together, and none of it is a matching variable.
Where the evidence contradicts itself, which is worth saying out loud
Not every study agrees with the section above, and this site's editorial policy commits to saying so rather than picking the tidier result.
Shanhong Luo and Eva Klohnen published in the same journal as Dyrenforth, five years earlier, using 291 newlywed couples and a couple-centred method. They also found substantial similarity on attitude-related domains and little on personality-related ones — that part matches everyone else. But on whether similarity predicted marital quality, they report positive associations for personality-related domains but not for attitude-related ones, that similarity on attachment characteristics was the most strongly predictive of satisfaction, and robust curvilinear effects for husbands but not for wives.
That is close to the opposite conclusion on the question that matters, from a competent study in a top journal. And the awkward detail is that Luo and Klohnen's paper and Watson's came out of the same department at the University of Iowa, within a year of each other, partly overlapping in authorship, and landed on different answers.
It is not a mystery requiring a conspiracy. The two differ in sample size — 291 couples against more than 23,000 individuals in Dyrenforth — in how similarity was computed, which Luo and Klohnen show matters a great deal, and in what counts as marital quality. Profile-based similarity indices behaved differently from absolute difference scores in their own analysis, and that is a measurement decision taken before any result appears.
Which leaves a reader with a fair summary rather than a clean one. The best-powered tests find similarity doing almost nothing for relationship satisfaction. A smaller, more carefully specified test finds it doing something, in personality rather than in attitudes, with a shape that differs between husbands and wives. Nobody has resolved it. Anyone telling you that science has settled whether having things in common makes a relationship better is reporting one of those studies and not the other.
The same honesty is owed in the other direction, and is applied elsewhere on this site: the shared-novelty finding is one of the few pieces of relationship advice with an experiment behind it, and its one preregistered replication moved boredom without moving satisfaction.
Similarity predicts attraction right up until people meet
There is one more layer, and it is the one that explains why the folklore feels true in both directions at once.
R. Matthew Montoya, Robert Horton and Jeffrey Kirchner pooled 460 effect sizes from 313 investigations and separated two things that usually get merged: actual similarity, meaning the measured overlap between two people, and perceived similarity, meaning how alike one person believes the other to be. Across the whole pile, actual similarity correlated with attraction at r = .47 and perceived similarity at r = .39.
The split is the finding. Actual similarity mattered in studies where participants never interacted, and in studies with brief interactions, and its effect dropped significantly once you moved past no-interaction designs. In existing relationships, the effect of actual similarity was not significant. Perceived similarity predicted attraction at every stage.
Read that next to the population data and the whole picture resolves. Measured resemblance between partners is real and modest. Measured resemblance predicts how much you like a description of someone. Once there is an actual relationship, what keeps predicting is the sense of being understood — which is not the same quantity and is not produced by counting how many answers two people share.
So both halves of the folklore have something to hang on to, which is why neither dies. People do end up with people like them, and it is mostly about the rooms they were in. People do fall for someone who feels different, and feeling different and being different are not the same measurement. And the thing people describe as chemistry corresponds to nothing anyone has managed to predict in advance: in the best-resourced attempt, more than a hundred measures collected a week before a speed-dating event predicted essentially none of how much two specific people wanted each other, a result set out in full in what questions can and cannot reveal about compatibility.
The wedge between the person you have information about and the person you have filled in opens wider the less contact there is, which is the mechanism behind what distance does to attraction.
The one place difference has anything going for it
In fairness to the complementarity idea, there is a corner of the literature where something like it has been observed, and it is much narrower than the phrase implies.
Larissa Tiedens and Alison Fragale, in the Journal of Personality and Social Psychology in 2003, ran two studies on dominant and submissive nonverbal behaviour. Participants interacting with a confederate who took up more space tended to contract their own posture; participants facing a constricted confederate expanded. In the first study, participants whose response complemented rather than mirrored their partner liked them more and were more comfortable. In the second, complementarity and mimicry were manipulated directly, and complementarity again produced more liking and comfort.
Note what that is and is not. It is strangers, in a laboratory, for minutes, with posture as the variable — about hierarchy settling itself in a short interaction. It is not a finding about who should marry whom, and it does not license any inference about a relationship. The broader claim that mirroring is a technique that produces warmth is a separate idea with its own weak record, taken apart in what body language actually predicts.
The most recent paper to argue for complementarity in actual couples is worth naming precisely so its size is visible. Li and colleagues, in NeuroImage in 2025, studied 48 married couples, measuring grey-matter volume and gender-role questionnaires, and report a mixed answer: a complementarity pattern for masculine-typed characteristics and a similarity pattern for feminine-typed ones, with each correlated to marital satisfaction in opposite directions. Forty-eight couples, correlational, one country, brain morphometry as the measure. It is a hypothesis-generating study and should be read as one.
Set against 199 pooled studies, 79,074 couples and 23,250 married people in three countries, that is the state of the evidence for opposites attracting: one laboratory effect about posture in strangers, and a small imaging study that has not been replicated.
When “we are too different” is not the problem
This section exists because the compatibility frame is actively dangerous when it gets applied to the wrong thing, and we are just too different is one of the ways that happens.
Some of what gets described as a difference in style is not a difference in style. Being told who you may see or message. Having your phone checked, or being required to share your location when it is not mutual and not freely chosen. Being made to account for your time. Money controlled or withheld. Anger, punishment or silence arriving as a sanction when you say no to something. Being separated from friends and family a little at a time.
UK statutory guidance on controlling or coercive behaviour names isolating a person from sources of support, monitoring their time and movements, controlling access to money, and depriving them of independence as recognised markers of that pattern. None of it is a mismatch, and none of it is improved by the two of you having more in common.
The reason the framing matters is practical. Calling control an incompatibility invites you to negotiate with it — to compromise, to meet in the middle, to find the shared ground. There is no version of that list that a better match fixes, and no correlation coefficient on this page applies to it.
If any of that is recognisable, the useful next step is a person rather than an article. Find A Helpline, listed in the sources at the bottom of this page, lists free and confidential services by country. In the United States the National Domestic Violence Hotline is 1-800-799-7233, or text START to 88788, 24 hours a day.
Where the feeling underneath is jealousy rather than control, the line between the two is worked through in the psychology of jealousy, and the phone-specific version of the same distinction is in what phone behaviour on dates actually shows. And where the difficulty is a partner who withdraws rather than one who restricts, what the research supports about supporting an avoidant partner and the same for an anxious one are better starting points than anything about matching.
What follows from this if you are deciding about a person
Four things, and none of them is a compatibility score.
Stop treating difference as a verdict. The variable most people mean when they ask this question — personality — has a partner correlation of around 0.1 across tens of thousands of couples, which means the population contains enormous numbers of unalike pairs who stayed together long enough to be counted. And the best-powered tests of whether similarity predicts satisfaction put it below half a per cent of the variance. Being unlike someone is not evidence about your relationship.
Separate the differences that are facts from the differences that are style. Children or none, where you are willing to live, money, care of ageing parents, faith where it is practised: those are the shape of a life, and two incompatible answers stay incompatible however good the conversation is. A difference in energy, tidiness, sociability or sleeping hours is not in that category and is routinely treated as if it were. What questions can actually reveal goes through how to ask about the first kind early and plainly.
Notice when a resemblance is just a shared circumstance. A large part of the population-level similarity in this literature is education, age, region and habits — the rooms people were in. Discovering that you and someone went to the same kind of university and vote the same way is weaker information than it feels, because so did almost everyone you were ever likely to meet.
Ask the questions that point at the relationship, not at the person. What predicted relationship quality in the largest study of it were things like felt appreciation, perceived commitment and how conflict actually goes — none of them a trait either person carries in from outside. If you want a read on where a relationship currently sits rather than a forecast, the relationship satisfaction test reports commitment and satisfaction as separate scores and says plainly what it cannot tell you, and the conflict style test separates pressing for resolution from pulling away. The attachment style test covers the one individual difference that did carry weight.
What this evidence cannot tell you
The limits go here rather than in a footnote, because on this subject they are most of the honest answer.
The big dataset is one country and one age band. UK Biobank participants were between 40 and 69 when recruited, are disproportionately healthy, are less likely to live in deprived areas and less likely to smoke and drink daily than the general population. The authors raise this themselves, and add a point that cuts the other way: restricting the range like that should usually push correlations towards zero, so their estimates may understate resemblance in a broader population. Both directions of error are live.
The meta-analysis is Europe, the United States and some of East Asia. Samples from Sub-Saharan Africa, South Asia, Latin America and the Caribbean, Southwest Asia and North Africa, and Oceania each appear in only a handful of studies. The authors say they could not meaningfully test whether partner correlations differ across even broad geographical divisions, and could not assess change over time at all.
The heterogeneity is large. Because of it the authors state explicitly that their figures should not be read as a single true partner correlation for a trait, but as the typical level across the samples they had. Quoting .55 for education as though it were a constant is a misreading of their own caution.
The same-sex comparison rests on a few hundred pairs. 236 female-female and 216 male-male couples is enough to say the means are positive and similar in size to opposite-sex pairs; it is not enough to detect modest differences, and the authors say so.
And nothing here is about your relationship. Every number on this page is an average across a population, computed on couples who were inferred from an address, in a cohort that was never designed to study relationships. A correlation of 0.19 is compatible with any two people you can imagine. What predicts whether a particular relationship works is not a matching variable, and the honest version of that answer is in how many dates it takes before you know anything — which is also, usefully, more than people think.
Every empirical claim on this page carries a linked source, which is the promise made in our editorial policy; claims this site has had to retract are listed on the corrections page. The rest of this section is under Dating.
Pulling it together
The short answer is that opposites do not attract, and that the sentence is less useful than it sounds. Across the largest dataset ever assembled on the question, partners resemble each other on almost everything measurable, by an average correlation of 0.19 — a real tendency, and a small one. The large resemblances are education, politics, religion, age and habits, which is to say they are substantially a record of who was ever in a position to meet whom. On personality, the thing the question is actually about, the correlation is close to nothing in either direction, and the first author's own description of it is a coin flip.
Two further things are worth carrying away. Resemblance is not choosing: the same correlation is produced by preference, by social sorting, by assorting on a correlated trait, and by two people growing alike over decades, and this study cannot tell them apart. And resemblance is a poor forecast: across 23,250 married people in three countries, how similar two partners were explained less than half a per cent of their relationship satisfaction.
So if the question behind the question is whether being different from your partner is a problem, the data have very little to say about it, and what they do say is reassuring rather than otherwise.
Frequently asked questions
Do opposites attract?
Not as a general pattern. The largest study of the question — Horwitz and colleagues in Nature Human Behaviour, 2023 — pooled 480 partner correlations from 199 studies across 22 traits and separately analysed 133 traits in up to 79,074 UK Biobank couples. Correlations between partners were overwhelmingly positive: of the 133 traits, 118 were significant after correction for multiple testing, 115 of those positive and three negative. But the size is the part that gets dropped. The mean partner correlation was 0.19, which means knowing one partner's score tells you about four per cent of what you would need to predict the other's. The authors caution in their own university's announcement that the correlations were fairly modest and should not be overstated. So: partners resemble each other on average, by a little, and the finding does not describe any particular couple.
Is there any trait where opposites really do attract?
Three, out of 133 measured. In the UK Biobank analysis only three correlations were both significant and negative, and the paper's own results section names one: chronotype, being a morning person versus an evening person, at r = −0.18. The University of Colorado Boulder's announcement of the study names the other two as tendency to worry and hearing difficulty; those are not in the paper's main text, so they are reported here as the university's description rather than as something a reader can check in a table. And −0.18 is small — about three per cent of the variation, the same magnitude as the average positive correlation pointing the other way. Articles that turn it into night owls pair with early risers are overstating it considerably, and a negative correlation on sleep timing has obvious non-romantic explanations in two people sharing one alarm clock.
Do opposite personalities attract?
Neither opposite nor alike, as far as anyone can measure. In the 2023 meta-analysis the partner correlations for the Big Five were extraversion .08, agreeableness .11, neuroticism .11, conscientiousness .16 and openness .21. In the UK Biobank the overall neuroticism score came out at .04, and across the twelve items making up that scale the correlations ranged from −.09 to .16 — some slightly negative, some slightly positive, all tiny. The first author's summary of the extraversion result is that people have theories about extroverts liking introverts or liking other extroverts, but it is about like flipping a coin. This is the honest answer to the question as most people mean it: on temperament, both folk theories are wrong, because the variable barely sorts people into couples at all.
Do couples with more in common last longer or do better?
The best-powered tests say similarity does almost nothing, and one good study disagrees. Dyrenforth, Kashy, Donnellan and Lucas used three nationally representative samples of married couples — Australia 5,278, the UK 6,554, Germany 11,418 — and separated a person's own traits, their partner's traits and the similarity between them. Their own traits explained about 6% of relationship satisfaction, the partner's 1–3%, and couple similarity consistently less than 0.5%. Watson and colleagues found the same in newlyweds: satisfaction was mainly a function of the rater's own traits with little relation to spousal similarity. The disagreement is Luo and Klohnen, 291 newlywed couples, who found similarity on personality-related domains did predict marital quality, with attachment similarity the strongest. Nobody has resolved it, so anyone saying science has settled this is reporting one of those studies and not the others.
Where did the idea that opposites attract come from?
In its scientific form, from Robert F. Winch's theory of complementary needs, published in the American Sociological Review in 1954 and 1955 and as a book, Mate-Selection: A Study of Complementary Needs, in 1958. The idea was that people pair with partners whose psychological needs fit theirs like a key in a lock. Worth noting honestly: all three of Winch's publications are paywalled, as is Tharp's 1963 review of the literature in Psychological Bulletin and the 1997 Journal of Personality and Social Psychology paper literally titled When do opposites attract?. None has a publicly readable abstract, so this article quotes no figures from any of them. What can be checked is the shape of what followed, and it is lopsided: searched in November 2026, PubMed returns 765 human-study records for assortative mating against 15 for disassortative mating.
Does any of this apply to same-sex couples?
The 2023 study excluded them deliberately, saying the patterns may differ and that less data exists. The follow-up came out in August 2025, is open access, and is not cited by any page currently ranking for this question. Horwitz, Kelly and Keller compared 66 traits across 89,628 female-male, 236 female-female and 216 male-male putative couples in the UK Biobank. The mean correlation was 0.19 for opposite-sex pairs and 0.15 for both same-sex groups, with medians of 0.14, 0.14 and 0.15. After correction for multiple testing there was no significant difference in assortment between any of the groups for 64 of the 66 traits; the two exceptions, where opposite-sex pairs assorted more, were year of birth and number of children. Far fewer correlations reached significance in the same-sex groups, which the authors attribute mostly to having a few hundred pairs rather than ninety thousand.
If partners resemble each other, does that mean people choose partners like themselves?
Not necessarily, and the paper everyone is quoting says so in its introduction. It lists four processes that can produce a correlation between partners: phenotypic homogamy, where people match directly on the trait; social homogamy, where they match because of where and among whom they were available to meet; genetic homogamy, where they are actually sorting on something correlated with the trait; and convergence, where partners grow more alike over time. The authors state that convergence is not assortative mating at all, because it is not a consequence of initial matching, that the mechanisms are not mutually exclusive, and that future studies will be needed to tell them apart. It matters practically: a couple who grew alike over fifteen years and a couple who selected for similarity on day one produce the same number, and nothing in this dataset distinguishes them.
My partner and I are very different. Is that a problem?
The evidence gives surprisingly little reason to think so, with one category of exception. On personality the partner correlation across tens of thousands of couples is around 0.1, so the population is full of unalike pairs who stayed together long enough to be counted, and the best-powered tests put similarity below half a per cent of what explains relationship satisfaction. What does carry weight is a different kind of difference: children or none, where you are willing to live, money, care of ageing parents, faith where it is practised. Those are facts about the shape of a life and two incompatible answers stay incompatible however good the conversation is — what questions can actually reveal about compatibility covers how to ask early and plainly. And one category is not a difference at all: being monitored, restricted or isolated is not a mismatch, and no amount of having things in common addresses it.
Sources
What this article draws on. For the broader bibliography see our sources page.
- Horwitz, T. B., Balbona, J. V., Paulich, K. N. & Keller, M. C. (2023) — 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, 7, 1568-1583 — 480 partner correlations from 199 studies, plus up to 79,074 UK Biobank couples; mean UKB partner correlation 0.19, range −0.18 (chronotype) to 0.87 (birth year). The author manuscript is free to read in PubMed Central.
- Horwitz, T. B., Balbona, J. V., Paulich, K. N. & Keller, M. C. (2023) — free full text — The same paper as an author manuscript in PubMed Central, PMC10967253. Every figure quoted from the 2023 study on this page was read here rather than from a secondary account.
- Horwitz, T. B., Kelly, K. M. & Keller, M. C. (2025) — A comparison of partner resemblance trends across putative same-sex and opposite-sex couples in the UK Biobank. Scientific Reports, 15, 31598 — 66 traits in 89,628 female-male, 236 female-female and 216 male-male putative couples; mean correlations 0.19, 0.15 and 0.15. Open access.
- Kelly, K. M. and colleagues (2025) — Criteria for detection of married or cohabiting couples in the UK Biobank. Social Science & Medicine, 383, 118475 — the UK Biobank does not record who is partnered with whom; published estimates of the number of couples in it range from under 19,000 to over 105,000, and the paper measures how specific match-checking rules bias the resulting sample. Free full text in PubMed Central, PMC12810986.
- University of Colorado Boulder (31 August 2023) — News flash: Opposites don't actually attract — the press release that is the actual source of the 82%-89% figure quoted across the first page of results, of the extraversion "flipping a coin" quote, of the naming of tendency to worry and hearing difficulty as the two other negative correlations, and of the authors' caution that the correlations are fairly modest and should not be overstated.
- Buston, P. M. & Emlen, S. T. (2003) — Cognitive processes underlying human mate choice: The relationship between self-perception and mate preference in Western society. PNAS, 100(15), 8805-8810 — 978 heterosexual residents of Ithaca, New York, aged 18-24; concludes people use a "likes-attract" rather than an "opposites-attract" decision rule.
- Dyrenforth, P. S., Kashy, D. A., Donnellan, M. B. & Lucas, R. E. (2010) — Predicting relationship and life satisfaction from personality in nationally representative samples from three countries: the relative importance of actor, partner, and similarity effects. Journal of Personality and Social Psychology, 99(4), 690-702 — Australia N = 5,278, UK N = 6,554, Germany N = 11,418; couple similarity explained less than .5% of the variance. Paywalled; the figures quoted here are from the published abstract.
- Watson, D., Klohnen, E. C., Casillas, A., Simms, E. N., Haig, J. & Berry, D. S. (2004) — Match makers and deal breakers: analyses of assortative mating in newlywed couples. Journal of Personality, 72(5), 1029-1068 — strong similarity in age, religiousness and political orientation, little in personality, emotional expression and attachment; marital satisfaction primarily a function of the rater's own traits. Paywalled; quoted from the published abstract.
- Luo, S. & Klohnen, E. C. (2005) — Assortative mating and marital quality in newlyweds: a couple-centered approach. Journal of Personality and Social Psychology, 88(2), 304-326 — 291 newlywed couples; the study that disagrees, finding similarity on personality-related domains associated with marital quality. Paywalled; quoted from the published abstract.
- Montoya, R. M., Horton, R. S. & Kirchner, J. (2008) — Is actual similarity necessary for attraction? A meta-analysis of actual and perceived similarity. Journal of Social and Personal Relationships, 25(6), 889-922 — 460 effect sizes from 313 investigations; actual similarity r = .47 and perceived r = .39 overall, with actual similarity not significant in existing relationships.
- Hunt, L. L., Eastwick, P. W. & Finkel, E. J. (2015) — Leveling the playing field: Longer acquaintance predicts reduced assortative mating on attractiveness. Psychological Science, 26(7), 1046-1053 — 167 couples; matching on attractiveness had no association with either partner's relationship satisfaction.
- Tiedens, L. Z. & Fragale, A. R. (2003) — Power moves: complementarity in dominant and submissive nonverbal behavior. Journal of Personality and Social Psychology, 84(3), 558-568 — the strongest laboratory evidence for anything like complementarity, in posture between strangers over minutes. Paywalled; quoted from the published abstract.
- Li, L., Huang, X., Zheng, Q., Xiao, J., Shan, X., Chen, H. & Duan, X. (2025) — Similarity or complementarity? Understanding marital relationships in terms of sexual dimorphism in brain morphometry and gender roles. NeuroImage, 305, 120974 — 48 married couples; the most recent paper to argue for complementarity in actual couples, and reported here with its sample size visible. Paywalled; quoted from the published abstract.
- Winch, R. F. (1955) — The theory of complementary needs in mate-selection: a test of one kind of complementariness. American Sociological Review, 20(1), 52 — the origin of the scientific version of "opposites attract", together with the 1954 companion paper and the 1958 book Mate-Selection: A Study of Complementary Needs. Paywalled, with no public abstract; no figure from it is quoted on this page.
- Tharp, R. G. (1963) — Psychological patterning in marriage. Psychological Bulletin, 60(2), 97-117 — the review that assessed the complementary-needs literature of the period. Paywalled, with no abstract indexed in PubMed; named here so it can be located, and not quoted.
- Dryer, D. C. & Horowitz, L. M. (1997) — When do opposites attract? Interpersonal complementarity versus similarity. Journal of Personality and Social Psychology, 72(3), 592-603 — the paper that carries the question as its title. Paywalled, with no abstract available from Crossref or PubMed; named and not quoted.
- Pearson, K. & Lee, A. (1903) — On the laws of inheritance in man: I. Inheritance of physical characters. Biometrika, 2(4), 357-462 — the oldest study included in the 2023 meta-analysis, and roughly where the statistical machinery for measuring partner resemblance was invented.
- Literature and page-one counts, run on 7 November 2026 — Written out so they can be re-run rather than taken on trust. PubMed, title-and-abstract: "assortative mating" restricted to human studies, 765 records; "disassortative mating" restricted to human studies, 15; "opposites attract", 58, of which the large majority are about magnets, antibodies, nanoparticles, plant hormones and physics. Europe PMC, title-and-abstract: "assortative mating" 2,237; "disassortative mating" 136. Nine pages from the first screen of results for "do opposites attract" were downloaded as plain text and searched: zero contain "et al", zero give the sample size of the study they describe, zero quote a correlation from it, and zero link its DOI.
- Home Office (UK) — Controlling or coercive behaviour in an intimate or family relationship: statutory guidance framework — the source for the markers listed in the section on when a difference is not the problem.
- Find A Helpline — Free, confidential support services listed by country — the referral in the section on control, which covers the situation that nothing else on this page addresses.
Keep reading
- DatingAge Gaps in Relationships: What the Research Predicts, and Where the Divorce Numbers Came FromThe 18%, 39% and 95% divorce figures were withdrawn in 2014 by the man who worked them out. The best US study found the gap itself predicted nothing.
- DatingBreadcrumbing in Dating: Where the 30% Figure Comes From, and What Was Actually MeasuredIt comes from 626 Spaniards surveyed in 2019, 86% of whom had never heard the word. In the same table, more said they had done it than had it done to them.
- DatingMicro-Cheating: What Has Actually Been Measured, and the Table of 27 Behaviours Nobody QuotesPubMed returns zero results for the word. The behaviours it names were measured in 456 people, and the ambiguous ones landed at a coin flip, 30 points apart.
- MindsetHow to Stop Being So Nice, and What the Research Says It CostsThe first validated measure of people-pleasing arrived in 2025, and what it tracks is not agreeableness. The fix everyone recommends has never been tested.
- MindsetTrauma Bonding: What the Term Actually Rests On, and Where the Seven Stages Came FromThe seven stages appear in no study. Four measures of trauma bonding do exist, and no page on the first results page mentions any of them.
- MindsetStaying Friends With an Ex: What the Research Measured, and Which Reasons Went BadlyThe pages that rank cite the right studies and not one reports a number. In the study that separated them, staying connected and watching ran opposite ways.