Both folk beliefs are half right, in a way that makes them useless. You did marry someone remarkably like you — on dimensions you never consciously picked. And you did marry someone with a different temperament, which is the difference you notice daily.
In 2023 researchers published the most comprehensive test of “opposites attract” ever run: a meta-analysis of 480 partner correlations from 199 studies covering 22 traits, plus a fresh analysis of 133 traits in up to 79,074 couples.
Of 118 traits showing a statistically significant partner correlation, 115 were positive. Three were negative. The authors wrote that partner correlations are “overwhelmingly in the positive direction.”
That is the headline, and it is the least interesting thing in the data.
The gap in the middle of that chart
Look at where the bars collapse. Partners are strongly matched on year of birth, political values, religiosity, education and IQ. They are barely matched at all on the Big Five personality traits — openness 0.21, conscientiousness 0.16, neuroticism 0.11, extraversion 0.08.
This is not new. A 2004 study of 291 newlywed couples found age 0.77, religiousness 0.75, political conservatism 0.63, IQ 0.42 — and then neuroticism 0.02, openness 0.04, conscientiousness 0.02, agreeableness −0.07, and extraversion at −0.17, significantly negative and replicated in spouse ratings.
So when you say “we’re opposites,” you are almost certainly talking about temperament. And on temperament, couples genuinely are close to unrelated. The belief is not baseless. It is just describing the one domain where matching does not happen, while ignoring the domains where it happens powerfully.
The single most negative correlation across all 133 traits was chronotype at −0.18 — morning people pairing with night people. It is one trait out of 133, and plausibly a matter of household scheduling rather than attraction.
You did not grow alike
A study of 1,296 married couples, average marriage 19.8 years, tested whether personality similarity increases over time. It found evidence of convergence “inconsistent across analyses,” arguing against it. Whatever similarity exists was there at the start.
The twist: none of it predicts happiness
Here is where the whole debate turns out to be asking the wrong question.
Assortative mating is real. Similarity making couples happy is not supported — and the evidence against it is unusually strong, because it converges across three completely different designs.
The 2004 newlywed study found similarity effects on satisfaction in about 6 of 40 analyses, with inconsistent direction. Satisfaction was mainly a function of a person’s own traits.
A 2023 study of 1,294 couples — average relationship 30 years — used response surface analysis, which avoids the statistical artefacts that plague older similarity research. Across 20 traits and facets, it found zero similarity effects meeting criteria for relationship satisfaction, and zero for life satisfaction. The apparent similarity effects turned out to be both partners being high on a desirable trait, not matching each other.
And a machine-learning analysis pooling 43 longitudinal datasets from 29 laboratories found that individual differences and partner reports “had no predictive effects beyond actor-reported relationship-specific variables alone.” How each person felt about the relationship predicted everything. The fit between them added nothing.
Compatibility may not be predictable at all
The most striking result comes from speed dating. Researchers collected more than 100 self-report measures from participants before they met anyone, then used machine learning to predict who would desire whom.
They could predict how much a person would like others in general. They could predict how much a person would be liked in general. But the dyadic component — the specific chemistry between two particular people — came out at negative explained variance. Worse than chance. The models “were consistently unable to predict any variance in relationship desire.”
That is a real problem for the entire premise of trait-based matching.
What does predict attraction: feeling alike
A meta-analysis of 313 studies separated actual similarity from perceived similarity. The pattern is stark:
- In paper-and-pencil studies with no interaction, actual similarity predicted attraction at r = 0.59.
- After a short interaction, 0.21.
- In existing relationships, 0.08 — not significant.
- Perceived similarity held at around 0.39 throughout, including inside established relationships.
In a speed-dating study of 2,184 four-minute dates, actual similarity across 26 measured constructs produced a coefficient of 0.005. Perceived general similarity: 0.75.
What predicts attraction and satisfaction is not being alike. It is feeling alike.
Is there any case for complementarity?
There is one, and it is genuinely contested. A 2007 study found couples reporting the highest relationship quality were similar in warmth but dissimilar in dominance — the classic complementarity finding.
But a 2024 set of four studies across Germany, Israel and the US, total N = 2,010, explicitly found no support for complementary dominance patterns — and found dominance was simply bad for relationship quality on both sides. A 2024 study of 173 couples in observed conflict found dominance behaviour unrelated to how well the conflict resolved.
The honest position is that the question is live. The best-powered evidence currently points against it.
Why the belief survives
Two findings explain it neatly.
First, similarity only drives attraction when it is made salient. In real life your shared politics, background and education are invisible wallpaper. A difference in temperament is the argument you had on Tuesday.
Second, an experiment on 150 people found that learning you are similar to someone increases liking — but learning you are dissimilar produces no repulsion at all. Differences do not push people apart. So opposite-seeming couples form, persist, and become the anecdote everyone remembers.
What this cannot tell you
Several of the meta-analytic figures have very wide confidence intervals — religiosity’s 0.56 spans 0.25 to 0.77. A correlation is not a choice: matching on politics may reflect who you meet as much as who you pick. The biobank sample is older, healthier, more educated and overwhelmingly British and White, and covers only mixed-sex couples. And extraversion genuinely conflicts between the large meta-analysis and the newlywed study, unexplained.
The useful version
You almost certainly partnered with someone who shares your background, your politics and your education, and you did not experience that as a choice. You may well have partnered with someone whose temperament differs from yours, and you notice it constantly.
Neither fact tells you whether you will be happy. What predicts that is how each of you feels — and whether you feel like the same kind of person, which is not the same as being one.
