Intersectional inequalities in social networks

The question

Most studies of social inequality look at one dimension at a time — gender, ethnicity, class, or age in isolation. But real people inhabit multiple identities simultaneously, and the disadvantages they face are not always well described by adding the effects of each axis independently. We ask: under what conditions does belonging to several minority groups produce more disadvantage than the sum of its parts, and what network-formation mechanisms can explain this?

Data

Both papers combine analytical results on synthetic networks with large-scale empirical validation. Synthetic networks span a wide range of group sizes, homophily levels, and population structures, allowing controlled exploration of when intersectional disadvantage emerges.

Empirical validation draws on two sources. The Science Advances paper calibrates the model against US high-school friendship networks from the AddHealth study (National Longitudinal Study of Adolescent to Adult Health), finding that closed-form predictions of degree inequality align remarkably well with observed patterns. The Communications Physics paper (introducing the MAPS framework) uses two systems: 70 AddHealth school friendship networks covering 41,880 students with grade, ethnicity, and gender attributes; and marriage data from the IPUMS USA database for the 50 most populous US cities, drawn from the 2013–2017 and 2017–2021 American Community Survey 5-year pooled samples.

Approach

We model network formation in populations stratified along multiple identity dimensions. Each individual is described by a vector of group memberships, and tie formation depends on a homophily preference that aggregates across these dimensions. We then study how a simple preference aggregation rule — applied at the level of individual choice — produces macroscopic patterns of segregation and compounded inequality at the network level.

The framework lets us reason analytically about which population structures and homophily levels lead to intersectional disadvantage, and which interventions could mitigate it.

What we found

Science Advances (2025) derives a closed-form expression for degree inequality in networks with multi-attribute homophily. The key result: the degree penalty for sitting at the intersection of two minority groups is multiplicative, not additive. A double-minority individual ends up with far fewer cross-group ties than (minority-A effect) + (minority-B effect) would predict — and in many parameter regimes, fewer connections overall. As the CSH write-up put it: double disadvantage hurts more than twice as much.

Communications Physics (2026) introduces the MAPS (Multi-Attribute Preference Similarity) framework — a general model that extends the analytical results to arbitrary numbers of identity attributes and population structures. Validated on 70 US school friendship networks (41,880 students, with grade, ethnicity, and gender attributes) and marriage data from 50 US cities (IPUMS, 2013–2021 ACS), the framework predicts observed patterns of structural segregation from individual-level preference data with high accuracy.

What this means

Single-axis interventions — promoting gender diversity without attending to ethnicity, or vice versa — leave intersectional disadvantage largely intact. Because the compounding effect arises from the interaction of homophily preferences across dimensions, reducing one dimension of bias while leaving others unchanged produces only partial relief for those at the intersection.

More broadly, the compounding effect is not a quirk of any specific dataset. It is a structural consequence of multi-attribute homophily: in any population where people prefer similar others across multiple dimensions simultaneously, those who differ on all dimensions will be systematically more isolated than a dimension-by-dimension accounting would suggest.

Related publications

Press

• Infobae — May 2026
• Radio Eins / RBB — Nov 2025
• ORF Science — Nov 2025