Short answer. People in couples are, across many measured traits, more likely to be similar to each other than would be expected if people paired up at random. The match is especially strong for education, political and religious views, and some substance-use habits. For most personality traits and physical characteristics, the relationship is weaker. That is the result of a study published in Nature Human Behaviour in 2023: the authors combined data from 199 published papers on 22 traits and separately analyzed 133 traits in pairs from UK Biobank.
This result does not mean that everyone looks for a copy of themselves, that differences harm relationships, or that similarity necessarily caused the couple to meet. The study measured correlations within already existing couples. Those correlations can be influenced by partner choice, shared environment, age, social circles, and changes over years of living together. Let’s break down the numbers, methods, limitations, and how to read this scientific finding without popular but misleading simplifications.
What exactly the researchers studied
The work was conducted by Tanya Horwitz, Jared Balbona, Katie Paulych, and Matthew Keller. The researchers addressed two related questions: how similar partners are on the same characteristic, and whether the strength of similarity differs across characteristics. The subjects were men and women in couples: married spouses, cohabiting partners, engaged people, and parents of a shared child. That definition matters: it does not include all dating, crushes, or potential partners, and the main findings cannot automatically be extended to couples with a different composition. In the source article the inclusion criteria are described in detail.
The study had two parts.
- Systematic review and meta-analysis. The authors identified 199 peer-reviewed publications released no later than August 16, 2022. From them, they extracted 480 estimates of similarity for 22 traits. For each trait, they calculated a pooled correlation that accounted for differences between samples.
- UK Biobank analysis. The authors identified up to 79,074 presumed couples among UK Biobank participants. Initially, they considered 140 traits; for 133 there was enough data for the main analysis. The number of couples varied by trait and did not equal 79,074 in every result row.
These are two lines of evidence, not one giant survey in which all 155 parameters were measured for everyone. The meta-analysis summarizes studies from different eras and populations; the biobank allows comparison of a broad set of measures within one large data set. Reading them together makes the overall picture more convincing, but it does not remove the limitations of either part.
What partner correlation means
For each trait, the researchers matched one person’s value with the value of the same trait for their partner. In simple terms: if, in a sample, people with higher educational attainment are more likely to be paired with people with higher educational attainment, the correlation is positive. If “morning people” are more likely to be matched with “night owls” under the chosen chronotype coding, it may be negative. If values within a couple are barely related, the estimate is close to zero.
The correlation coefficient r ranges from −1 to +1. A positive sign indicates statistical similarity on a measure, a negative sign indicates statistical difference, and the magnitude shows the strength of the relationship in a specific sample and under a specific measurement approach. For example, r = 0.08 and r = 0.58 are both positive values, but the first indicates a much weaker relationship. Neither one means “8%” or “58% identical people.” You also cannot convert r into the probability that two people will like each other.
There is an additional nuance: for continuous, ordinal, and binary measures, the authors used different kinds of correlation — Pearson, Spearman, and tetrachoric. So comparing the numbers requires attention to how the trait was defined and measured. The tables in the publication show sample size and estimate uncertainty; a single coefficient without that context does not. Methods and the results table are available in the full article.
Main takeaway: similarity is more common than difference
In the meta-analysis, 18 of the 22 average correlations were positive and statistically significant after correction for multiple testing. For the other four traits, estimates were based on only four samples each, and uncertainty was higher. In the UK Biobank analysis, 118 of 133 correlations were significant after the same correction: 115 positive and three negative. In other words, the phrase “partners are similar on 89% of traits” describes the share of tested measures with a statistically significant association in this sample, not the share of matching traits in any specific couple. The authors present the distribution of results in the UK Biobank section.
The average correlation across the 133 traits in UK Biobank was 0.19. That hides a wide range: some measures were clearly associated, others only barely, and some pointed in the opposite direction. So it is more accurate to say “similarity is widespread, but its strength depends on the trait” than to say “people choose their doubles.”
Where similarity is strongest
In the pooled results, the following traits stood out in particular:
| Trait in the meta-analysis | Average correlation between partners | How to read the result |
|---|---|---|
| Political views | 0.58 | One of the strongest relationships among the traits studied |
| Religiosity | 0.56 | Clear similarity, but with a wide uncertainty interval |
| Educational attainment | 0.55 | Substantial similarity across different samples |
| IQ score | 0.44 | The relationship was estimated from fewer couples than education |
| Smoking status | 0.49 | Noticeable match in habit |
| Height | 0.24 | Positive but moderate or weak relationship |
| Extraversion | 0.08 | Weak positive relationship |
These are rounded estimates from Table 1 of the original study. They come from different sets of studies and should not be read as a ranking of “what matters in love.” A stronger correlation does not prove that people consciously look for that exact trait when they meet.
Views and education
Political and religious attitudes matter for everyday decisions: who people socialize with, where they live, how they spend time, and which rules they consider important. People often meet in settings already selected by such traits: in school, in professional circles, in local communities, or in religious organizations. So the observed similarity can arise both from preferences and from the structure of social connections. The original meta-analysis does not separate the contribution of these causes for each specific measure.
Education is a special case. Couples with similar educational levels may come together because people meet while studying, work in similar fields, live in the same neighborhoods, or build a shared life around similar plans. Direct preferences are also plausible. A correlation of 0.55 does not let us choose one explanation and declare it the main one. It shows a stable statistical similarity across the studies collected.
Habits: smoking and alcohol use
The correlation by smoking status in the meta-analysis was 0.49. The correlation for actually starting smoking was 0.38. For several measures of alcohol and other substance use, the estimates were also positive, but some of the corresponding meta-analyses were small or had wide confidence intervals. The authors specifically note: not all estimates in this group survived correction for multiple comparisons, despite relatively high point estimates. See the results and discussion.
A couple’s shared lifestyle can gradually strengthen similarities in habits, or the resemblance may result from people selecting partners from the same environment. Observing an association does not establish who influenced whom. More importantly, it does not prove that a shared habit is inevitable: group statistics describe a trend, not the behavior of every individual.
Personality: there is similarity, but it is weaker
The five well-known personality dimensions in the meta-analysis did not show similarity as pronounced as education or attitudes. For extraversion, the average correlation was 0.08; for neuroticism and agreeableness, 0.11 each; for conscientiousness, 0.16; and for openness to experience, 0.21. This is an important clarification to the intuitive phrase “we’re very similar in character.” People may feel close in values or overall lifestyle while still differing in temperament. All five estimates are given in the original source.
A weak correlation does not mean that personality is unimportant for relationships. The study answers a different question: how similar existing partners are on the measured trait. It did not test which traits contribute to relationship satisfaction, help people get through conflict, or predict breakup. Treating the extraversion figure as a measure of relationship quality would be a logical error.
Appearance and health
For height, the average correlation in the meta-analysis was 0.24; for body mass index, 0.16; and for waist-to-hip ratio, 0.17. For depression and diabetes, the corresponding estimates were 0.15 each. These associations are smaller than those for attitudes and education, but in large datasets they are often different from zero. In the UK Biobank, the results for height, BMI, and waist-to-hip ratio were also positive: 0.27, 0.25, and 0.19, respectively. Comparing the two parts of the study shows the overall trend, but not a perfect match for every estimate.
Health requires particular caution. If a certain condition is more common in both partners, that may reflect shared age, living conditions, social environment, mutual influence, characteristics of study selection, or a combination of factors. You cannot conclude that the illness was “passed” from one partner to the other, and you cannot use such correlations to diagnose a couple. A meta-analysis describes population-level associations, not medical causation.
Do “opposites attract” really hold true?
The popular saying is too broad to test honestly with a single number. If taken literally — as if stable couples, on average, form between people with opposite values on most traits — the study does not support that picture. Positive associations were much more common than negative ones. But if the idea is simply that two specific people may like their differences, the data do not rule that out.
A couple may differ in extraversion, daily routine, hobbies, or communication style and still share important views. The reverse can also be true: people may answer questionnaires similarly but behave differently in relationships. Group-level correlation does not capture the full complexity of individual compatibility. Moreover, statistical similarity is not the same as attractiveness: the study examined established couples, not an experiment in which people were randomly shown candidates and their desire to meet was measured.
This topic has another subtle point. If partners are similar on some trait, that does not necessarily result from a preference for “someone with the same trait.” For example, people may have met in the same professional circle, and similarity in income, schedule, and views may have emerged as a result of that shared circle. Therefore, the scientific term “assortative mating” should not be automatically applied to every correlation that is found.
Exception: chronotype and other negative associations
The most notable negative correlation in the UK Biobank analysis concerned chronotype — the tendency to be a “morning” or “evening” person. In the authors’ chosen coding, it was −0.18. The researchers identified two more statistically significant negative associations among 133 measures, while 115 significant associations were positive. These figures are published in the biobank analysis results.
A negative sign does not justify advising people to choose a partner with the opposite sleep schedule. It also does not prove that contrasting routines improve relationships. A correlation can have several explanations, and the measurement method and sample composition can change the estimate. In everyday life, different routines may be convenient for some couples and inconvenient for others. The study did not measure that convenience.
The very high positive correlation of 0.87 referred to the year of birth. It looks impressive, but its meaning is fairly simple: partners are often close in age. This measure should not be placed on the same level as “similarity in character” as evidence of near-total psychological overlap.
Why partners may become similar
The same outcome — similar indicators for two people — can arise in several ways. The authors of the original source describe mechanisms that can act simultaneously.
Direct selection by trait
A person may prefer a partner with similar views, habits, or life goals. Such initial matching is called phenotypic assortative mating: similarity appears already when the couple is formed. However, from data on an existing couple, it is usually impossible to know how much role was played by conscious preference and how much by the circumstances of meeting.
Shared social environment
People do not meet through a random sample of the entire population. Their social circles are shaped by age, place of residence, education, work, language, interests, and income. This structure itself creates statistical similarity, even if a person never stated requirements for a partner’s education or occupation. This mechanism is often discussed as social homogamy. It helps explain why a correlation does not have to reflect a romantic “selection criterion.”
Similarity through a related trait
Sometimes partners are similar on a measure indirectly: both entered the same environment for a different reason. For example, agreement on one habit may accompany similarity in age, education, or living conditions. The observed association between partners on the first characteristic is then partly driven by the second. To understand the causal structure, you need special studies, not just a list of correlations.
Convergence after the relationship begins
Living together, people can change habits, routines, and how they answer questions. Shared circumstances and mutual influence shape them. This is called convergence. It differs from initial selection: at the start of a relationship, partners may have been more different. A cross-sectional snapshot does not reliably separate convergence from similarity that existed before people met.
A separate 2023 study by Syaarda and Kutalik using UK Biobank data examined possible reasons for similarity and showed the role of selection, shared factors, and, for some measures, within-couple dynamics. But the authors also stress the limits of data on time spent together: without measurements before the couple formed, it is hard to distinguish the mechanisms precisely. This is consistent with the cautious interpretation of the large meta-analysis.
What the study does not answer
The scale of the work is impressive, but it does not turn correlations into universal laws. To use the findings responsibly, it helps to recognize several boundaries.
It cannot predict the compatibility of specific people. An average association with education or beliefs does not provide an individual “compatibility score.” Two people can build a strong relationship despite differences, and conflict despite many similarities. The study discussed here did not build or test a couple-matching algorithm.
It cannot establish causation. We see similarity among existing partners, but this study does not show the full path from meeting to the time of measurement. Some characteristics may have existed before the union, others may have changed within it, and still others may be tied to shared circumstances.
It cannot claim that similarity improves relationships. That would require separate data on satisfaction, relationship stability, communication quality, and other pre-defined outcomes. The authors discuss the literature on these questions, but their main analysis measures trait similarity, not the benefits of similarity.
It cannot assume that all results are equally precise. In meta-analyses across different traits, the number of couples and studies varied widely. Some confidence intervals are broad. An estimate based on only a few samples requires more caution than a result built on many independent data sets.
It cannot automatically be generalized to all countries and generations. Most of the included studies involved participants from Europe and the United States. Representation from other regions was much smaller. UK Biobank reflects a specific group of volunteers and certain age cohorts. Even if the overall tendency toward similarity is replicated, the strength of a specific association may differ in another society or historical period. The authors list the limitations in detail.
Why a large sample still requires caution
It is easy for readers to think that if a study includes tens of thousands of couples, the result must be definitive. A large sample reduces random error within the selected group, but by itself it does not correct for systematic differences between the group and the general population. UK Biobank recruited volunteers of a certain age; participants as a whole differed from the UK population on a number of social, behavioral, and medical characteristics. This is shown in the UK Biobank representativeness study.
In the study analyzed here, couples in the biobank were identified using indirect data. These were presumed partners, not all couples in the country. Moreover, both people had to be among the biobank participants. Such a selection scheme can affect the estimates. The researchers discuss this risk explicitly.
The meta-analysis, in turn, combined different periods, societies, and measurement methods. The median value of the heterogeneity statistic I² across 22 traits was 0.915, indicating substantial differences between samples. In simple terms, the average correlation for a trait is not a universal constant. It summarizes results that can vary notably from place to place. The authors also note that part of the heterogeneity is linked to the high precision of large studies and to measurement features; a single I² value is not enough to explain the reasons for the differences.
Why “statistically significant” does not mean “strong”
In large samples, a researcher can detect even a small association with high confidence. That is why a significant correlation of 0.08 for extraversion and a significant correlation of 0.58 for political views answer two very different practical questions. The first says the estimate is likely not zero under the analysis assumptions. The second also shows much stronger similarity. Mixing up these meanings creates the false impression that all 115 positive significant associations in UK Biobank are equally important.
The authors tested many traits at the same time. The more tests you run, the greater the risk of getting a “significant” result by chance. The article used the Bonferroni correction for confidence intervals and p-values. That makes the criterion stricter. But the correction does not answer the question of why the association exists, and it does not turn a small coefficient into an effect that matters in everyday life. The statistical method is described in the publication.
Why confidence intervals matter
A point estimate is the best calculation based on the available data, but it is not exact knowledge about all possible couples. A confidence interval shows the range of values compatible with the data under the statistical model. For example, the meta-analytic estimate for extraversion was 0.08 with an adjusted 95% interval from 0.03 to 0.13. For political views, the estimate was 0.58 with an interval from 0.50 to 0.64. These intervals help show not only the direction, but also the degree of uncertainty. The numbers are taken from Table 1.
Results based on a small number of samples should be read with extra caution. A large overall study size does not mean the same level of precision for each of the 22 meta-analyses. For some traits, the data are more limited and the intervals are wider. That is one reason not to build a flashy ranking using point estimates alone.
Why correlations cannot be turned into a “love formula”
Suppose a reader wants to choose a partner based on the three highest rows in the table: political views, religiosity, and education. That approach is flawed for several reasons. First, these characteristics may be related to one another and to the social environment, so their effects cannot simply be added together. Second, the study did not test whether matches improve relationships. Third, personal preferences, life circumstances, and how people negotiate cannot be reduced to measured scales. The table shows which pairs were more common in the data, not which pairs should be created.
You also should not draw the opposite conclusion: if similarity is weak for most psychological traits, that does not mean personality plays no role. Perhaps what matters is not having the same level of extraversion, but a person’s attitude toward differences, communication rules, and support in difficult situations. But those are separate hypotheses that need to be tested in other studies. The meta-analysis discussed here does not directly confirm or refute them.
What changed in research after publication
The main 2023 article focused on heterosexual couples. In in a separate paper by the same researchers, published in 2025, trends in similarity were compared in presumed same-sex and different-sex pairs in UK Biobank. This broadens the scope of the research, but it does not make the earlier sample “inclusive of everyone”: each result must be read within its own design and pair-definition criteria.
For the reader, this is a reason to be more cautious with broad statements like “everyone chooses similar partners.” The scientific question is not just whether there is an average correlation. It is important to understand for which people, by what measurement method, in what era, and under what social conditions . it is observed.
How to use the findings in everyday life
The study does not provide a ready-made list of must-have matches for a happy relationship. It can help you ask better questions about yourself and your relationship.
- Separate values from personality traits. Similar views and education are found more often between partners than identical levels of extraversion. So the phrase “we have different personalities” by itself says nothing about whether life goals align.
- Take the context of the relationship into account. People who met within the same social circle may already have shared circumstances. Similarity after meeting does not always mean each person was specifically looking for someone similar.
- Do not turn statistics into a prescription. If a couple has different sleep schedules or one person is more outgoing, that is not a scientific diagnosis of incompatibility. It is more important to discuss specific inconveniences and agreements.
- Do not confuse similarity with relationship quality. Having shared habits can make some decisions easier and others harder. Data on a couple’s actual outcomes are needed to judge the benefit of a particular match.
- Look at the size of the differences. A moderate correlation across thousands of couples still leaves enormous room for variation in individual relationships.
This is not a psychological test or a clinical recommendation. The study works better as a correction to popular beliefs: couples are not random combinations of people, but their similarity is more complex than the sayings about “opposites” or “two halves.”
Frequently asked questions
Is it true that partners match 89% of traits?
No. In UK Biobank, 118 of the 133 examined correlations were statistically significant after correction for multiple comparisons — that is 88.72% of the indicators tested. Most significant associations were positive. This share does not show how similar an individual couple is or how many specific answers matched. The original source gives both the share and the coefficients themselves.
Which traits most often match between spouses?
In the data reviewed, the strongest statistical similarity was observed for political and religious views, education, and some substance-use characteristics. Height and other physical measures also more often showed positive associations, but usually weaker ones. For Big Five personality traits, the average correlations were relatively small.
Does this mean opposites cannot be together?
No. The study describes average patterns in groups of established couples. It does not impose a ban on relationships between people with different traits and does not measure every form of attraction. Strong statistical similarity on one trait does not rule out differences on another.
Why are partners similar: do they choose each other or change together?
Both mechanisms are possible. In addition, the shared social circle, place of living, education, and other circumstances play a role. The main meta-analysis cannot precisely break down the observed correlation into these causes. That requires data on people before they meet and following them over time.
Can correlation be used to estimate the chance of a successful marriage?
No. Correlation between partners on a trait and forecasting relationship quality are different tasks. This study did not develop an individual prediction of satisfaction or relationship duration. A table of scientific coefficients should not be turned into a “perfect partner” questionnaire.
Conclusion
A large review of studies and the UK Biobank analysis provide a consistent overall picture: partners are more often similar than opposite on many traits; this is especially noticeable for education, views, and some habits. However, the strength of similarity varies widely, depends on the trait and the sample, and the correlation found does not by itself reveal the mechanism behind it. The most accurate reading of the study is that similarity between couples is common, but it is not a rule for every couple and it is not a measure of relationship quality.
Main source: Horwitz T. B. et al., 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, DOI: 10.1038/s41562-023-01672-z. The data and wording in the article were checked as of September 24, 2026.