Similar partners

Why do partners so reliably come from similar backgrounds even when nobody is looking for that?

Ask almost anyone what drew them to their partner and you will hear about temperament, humour, a particular evening. You will not hear "we had the same number of years of education", and yet if you tabulate a few thousand couples that is one of the first regularities to appear. The folk theory says opposites attract. The data say the opposite, consistently. So how does a population produce sorted couples out of individuals who were not sorting?

Split the question in two, because the answer has two parts that are easy to conflate. There is who you choose, and there is who you had the chance to choose from. Preference and opportunity produce the same pattern in the final data, and only one of them is visible from the inside.

Start with opportunity, since it is the part nobody experiences as a constraint. Where do people actually meet? School, university, work, a neighbourhood, a congregation, a sport, a friend of a friend. Now ask what each of those places already is. A university has selected on prior attainment before you arrive. A workplace has selected on qualification. A neighbourhood has selected on what its housing costs. A congregation has selected on belief. Every one of these is a filter that ran before you walked in, and the pool you meet from is its output.

So consider a thought experiment: let everyone choose completely at random from the people they actually encounter. What comes out? Sorted couples — because the encountering was sorted. No preference for similarity is needed to generate the correlation. It is already sitting in the structure of where people are.

Does that mean preference does nothing? No, and it would be too tidy to claim so. People do prefer similarity, particularly on values, politics and religion, and they act on it: they filter, they decline, they notice mismatch. Preference and opportunity compound rather than compete.

Now look at what the similarity is strongest on, because the pattern of strengths is informative. Spousal correlations are large for education, political attitudes and religiosity, moderate for age, and small for personality traits — typically only a little above zero. Ask which of those the meeting places actually sort on. Schooling, income, faith, locality: exactly the traits at the top of that list. Personality, which no institution sorts by, sits at the bottom. That fit is the strongest single argument that opportunity structure is doing much of the work.

One correction before the picture is complete. Some of what we measure in a couple was not there at the start. Partners converge over years together, especially in attitudes and habits, so a correlation measured among married couples overstates how alike they were when they met. It is a real effect and it inflates the raw numbers, but it does not explain them away — the sorting is visible at the point of pairing too.

And the folk theory? A recent large-scale analysis pooling many traits across many samples found the correlations overwhelmingly positive, with essentially nothing supporting systematic complementarity. Opposites, as a general rule about who pairs with whom, does not survive contact with the data.

Assortative mating is a description of an outcome — similarity between partners — and not, by itself, a claim about a cause. Two very different processes deliver it, and they are hard to separate because they predict the same tables.

That distinction matters for what follows from it. Sorting on education compounds across a generation: two graduate incomes in one household, two non-graduate incomes in another, and the gap between households widens even if nothing about individual earnings has changed. Whether that is chiefly a story about choosier people or about a labour market that concentrates graduates in the same firms and cities determines what, if anything, could change it.

The honest caveats are worth keeping visible. The similarity is a tendency, not a rule — these are correlations well below one, and plenty of couples sit far off the diagonal. Effect sizes vary sharply by trait and by country. Separating the preference channel from the opportunity channel is difficult in observational data, and estimates of the split are not settled. And the structure is not fixed: online meeting has become a leading way couples form, which removes the geographic filter while adding explicit ones people set themselves — age bands, education, faith — so whether it loosens or tightens the sorting is still being measured, and the answer looks different trait by trait.

• How educational assortative mating feeds into household income inequality across a generation.
• Whether meeting online genuinely widens the pool, given that its filters are explicit rather than incidental.

Every term the collection defines is gathered in the glossary.