EPC 2026 · Poster companion Jesper Lindmarker · Linköping University

Swipe to match.

★ EPC 2026 poster 01 / 04
The EPC 2026 poster

Do we like similar partners, or just happen to work with them?

Lindmarker & Jarvis · educational matching in Sweden

The study behind the poster. Swipe right for the question, the data, the method, and what we found.

Working paper · 1991 to 2022 02 / 04

Three Decades of Ethnic Assortative Mating in Sweden

Jesper Lindmarker

If interethnic unions keep rising, are ethnic boundaries actually softening? Or has the partner market simply shifted?

Sweden's foreign-born share of the population rose from about 5 percent in 1981 to over 20 percent in 2022
Interactive · Companion tools 03 / 04

Three Simulations of Opportunity and Partner Sorting

Built by Jesper Lindmarker

Pull the sliders. Watch couples form. See how segregation, attraction, and distance sensitivity produce the sorting we see in real Swedish data.

A simulated city: two groups cluster into neighbourhoods while couples form mostly nearby
Preprint · SocArXiv 04 / 04

However Far Away? The Spatial Contingencies of Assortative Mating

Jesper Lindmarker & Benjamin F. Jarvis

Residential proximity mediates 20 to 40 percent of ethnic endogamy. Segregation narrows the partner market before anything else gets a say.

Endogamy odds by ancestry group, with and without a residential proximity term
The work behind the poster

Do we like similar partners, or do we just happen to work with them?

The full EPC 2026 poster: The Opportunity Structure of Educational Homogamy, by Jesper Lindmarker and Benjamin F. Jarvis. Tap to enlarge ⤢
The setup · before any model

Universities and workplaces sort by education

No gymnasium Gymnasium Post-sec short Post-sec long
0% 25% 50% University Workplace Rest of country % with ego's level of education
Educational sorting of partner pools, by setting. For each ego we use the registers to sample individuals from the population who, in the same year, (1) co-attended the same university, (2) worked at the same workplace, or (3) lived in the rest of the country. For each pool we then report the share of sampled individuals who share the ego's level of education, plotted separately by the ego's own education. Sorting by education would show up as a higher same-education share in the university and workplace pools than in the rest-of-country pool. Swedish registers, 1990 to 2019.
Q1The dose-response

How much more likely are you to partner someone you shared a setting with?

Shared university · relative probability of a union

0 years
1.0×
1 year
2.3×
2 years
3.7×
3+ years
7.6×

Shared workplace · relative probability of a union

0 years
1.0×
1 year
11×
2 years
24×
3+ years
46×
Relative probability of forming a union, by years of co-membership, from the conditional logit. Each setting enters as a categorical dose (none, 1, 2, 3 or more years) rather than a linear term, net of the sampled partner market and controls for age, cohort, and region. Each extra year multiplies the relative probability: a clear dose-response, not a faint nudge. Reference category for each setting: zero shared years (1.0×). Bars are scaled within each setting; the workplace effects are several times larger than the university effects. Estimated on the full Model 4.

See the full coefficient plot, all settings and both sexes →

Q2Interactive

How much of educational homogamy is left after we account for sorting?

Each dot already holds the partner market fixed: it compares a real partner against a pool of singles the ego could plausibly have met, so the supply of similarly-educated partners, plus age, cohort, and region, is already accounted for. The hollow dot is that baseline. Drag the slider to add the settings two people actually shared, and watch the filled dot slide left toward 1×, the rate you would see if those settings explained all of it.

baseline after sorting 1× = different education

Swedish registers, 842,486 first unions with women as ego (men estimated separately), 1991 to 2019. Each dot pair shows the odds of partnering within rather than across education, net of the sampled choice set and controls for age, cohort, and region; 1× means education carried no weight. The slider adds only shared institutions and residential distance, in the order people meet them. What remains is residual assortativity, an upper bound on what choice could explain, not proof of preference. KHB decomposition on the Model 4 scale.

Unfold the study →

01 · The question
  • Similarly-educated people partner far more often than chance would predict. Why?
  • Two candidate mechanisms. Preference: people seek out partners like themselves. Opportunity: universities and workplaces route similar people into the same rooms, year after year.
  • Both produce the same thing, a homogamous union, so the outcome alone cannot tell them apart.
  • And does one answer hold across the whole education distribution, or only part of it?
02 · The theory
  • Three forces shape who we pair with: who we prefer, who we get the chance to meet, and who family and community steer us toward (Kalmijn 1998). This study isolates the second.
  • Blau: social structure sets the baseline odds of contact between groups.
  • Feld: everyday life is organised around shared foci, a campus, a workplace, a neighbourhood, that route people into repeated contact.
  • Identification problem: since opportunity and choice both yield homogamous unions, the strategy is to model the opportunity channels we can measure, then read whatever sorting remains as an upper bound on what choice could explain.
03 · The data
  • Swedish population registers (Statistics Sweden), full resident population, 1990 to 2022.
  • Unit: first opposite-sex unions, dated to the year two people move from separate addresses into a shared one. Cohabitation onsets 1991 to 2019; ego aged 17 to 45. Same-sex unions and migrate-to-cohabit cases excluded.
  • Cohorts born 1972 or later only. The employer register starts in 1990, so this floor guarantees a full workplace and organisation history from age 18 and avoids left-censoring the institutional channels.
  • 842,486 first unions with women as ego, a comparable count with men. Women and men estimated separately.
  • Each person's history is reconstructed: residence at 100-metre resolution, every employer and organisation, every university enrolment.
  • Covariates: education in 4 levels (no gymnasium / gymnasium / post-secondary short under 3 years / post-secondary long 3 or more years); couple homogamy in 5 categories (reference: different education); log residential distance; institutional overlap entered categorically as 0, 1, 2, or 3 or more shared years. Controls: age-difference bands, shared ancestry (country and region of origin, from parental birth country), and immigrant generation.
04 · The method
  • Conditional logit (McFadden's discrete-choice model). Each ego who formed a union "chooses" from their real partner plus about 100 single alternatives.
  • The alternatives are drawn by stratified sampling from the local market that year, with a McFadden-Manski offset correcting for unequal sampling rates. The baseline therefore already holds the partner supply fixed.
  • exp(coefficient) reads as the relative odds of choosing a partner with a given feature: same education, short distance, shared employer, shared campus.
  • Opportunity channels are added one at a time in life-course order: university, then workplace, then organisation (excluding shared workplace), then residential distance.
  • The KHB decomposition places every nested model on one (Model 4) scale and reports how much of the education coefficient each channel absorbs. Standard errors clustered on the ego; estimation via the fastclogit R package (github.com/jeppelina/fastclogit).
05 · Limitations
  • Unobserved confounders. Traits registers cannot see, sociability, lifestyle, ambition, could route people into both shared institutions and similar partners, which would overstate the opportunity channels.
  • Coarse measurement. Sharing an employer or a campus is not the same as sharing a team or a programme, so true co-presence is measured with error. This works the opposite way and understates the channels. The net bias is unknown, but the qualitative patterns (the gradient, the gender asymmetry, the dose-response) are robust.
  • Selection into strata. Education shapes which institutions a person enters, so the sampling strata themselves sit partly on the causal pathway from education to partner, rather than being a clean exogenous market. I probe how far this distorts a clogit decomposition with simulations in the dissertation kappa; they stress-test selection and confounding but do not fully resolve the exact stratified design used here.
  • Upper bound, not preference. Whatever sorting survives all the channels is residual assortativity, an upper bound on what choice could explain, not proof of preference.

Simulation code: github.com/jeppelina/clogit-validation-simulations.

This is the methodological cousin of the ethnic-boundaries work below: the same insistence that opportunity be modelled before behaviour gets named, applied to a different line of division. It extends an earlier study with Ben Jarvis, "However Far Away?", which showed residential proximity alone mediates 20 to 40 percent of ethnic endogamy in Sweden.

Where this fits

A research agenda on boundaries and choice

The dissertation behind this poster asks one question in four ways: how much of observed partner sorting is opportunity, and how much is residual assortativity, across ethnic and educational lines, in registers covering the full Swedish population, using counterfactual partner comparisons.

The bigger story

Rising intermarriage. Stable boundaries.

Sweden's immigrant population has grown dramatically since 1991. Ethnic diversity has reshaped the partner market. Over the same period, interethnic unions have become more common in absolute terms. That reads like integration: boundaries eroding, communities mixing, demographic progress.

Top: foreign-born share of Sweden's population rises from 9.3 percent in 1991 to 19.5 percent in 2019. Bottom: ethnic endogamy among Swedish-ancestry women falls from 85.6 percent to 81.6 percent over the same period.
Sweden, 1991 to 2019. The pool changed. So did the outcome. The question is which drove which.

When you decompose the trend, the picture inverts.

Conditional logit models that hold partner availability constant show that the residual assortativity for same-group partners has not weakened. In several configurations, particularly among Swedish-origin women, it has strengthened. The rise in interethnic unions is not a sign of softening boundaries. It is a mechanical consequence of a partner market reshaped by migration and segregation. More diversity makes same-group pairing structurally harder, even when the appetite for it is unchanged or stronger.

Waterfall decomposition of the 2.2 percentage point fall in endogamy among Swedish-ancestry women in Stockholm. Baseline 1991 to 1999 at 85.3 percent. A more diverse pool would have pulled endogamy down by 3.7 percentage points. Residual assortativity for same-group partners strengthened by 1.5 percentage points, partially offsetting the structural pull. Endpoint 2010 to 2019 at 83.2 percent.
The two forces pull in opposite directions. The pool change dominates the net trend. The behavioural change runs the other way.
"Without accounting for who is actually available, we mistake a structural shift for an attitudinal one. The Swedish case shows you cannot read boundary change off mixing rates alone."

This matters for how demographic trends get interpreted. Rising intermarriage is routinely read as evidence that integration is working. Without an opportunity-side counterfactual, that reading is unsafe. The same observed rise can come from openness or from arithmetic, and the policy implications differ.

The poster paper upstairs is the methodological cousin of this one: same logic, different dimension, same insistence that opportunity be modelled before behaviour gets named.

Play with it

Three small simulations

The opportunity-versus-assortativity distinction is easier to feel than to explain. Three browser simulations make the lever visible. Adjust segregation, in-group attraction, and distance sensitivity, then watch couples form and see how much sorting emerges from each.

All simulations → Inline simulation on the homepage →
Find me · cite me · talk to me

If any of this resonates

I'm a PhD candidate at the Institute for Analytical Sociology, Linköping University, defending in late 2026 and starting a Swedish Research Council postdoc in 2027. If you work on assortative mating, ethnic boundaries, segregation, or the methods used here, I'd like to talk.

Email me Google Scholar LinkedIn All papers on homepage