Socio-economic Segregation in a Population-Scale Social Network

• Social segregation can extend beyond neighborhoods into social circles, education, lifestyles, and careers, forming socio-economic bubbles of interaction with similar others.Source: §1 · p. 2
• The study asks whether segregation is more pronounced in social networks than in spatial neighborhoods and seeks to uncover this potentially hidden segregation across social contexts.Source: §1 · p. 2
• The analysis represents 17.2 million Netherlands residents through around 1.3 billion register-based ties, including approximately 914 million household ties.Source: §1 · p. 2
• Compared with surveys, online networks, and mobile communication records, register data better support population coverage, tie-type distinctions, and analysis of multi-layered social segregation.Source: §1 · pp. 2–3
• Population-scale register data distinguish formal ties such as family, school, neighborhood, and work connections while providing individual attributes without imposing a priori spatial assumptions.Source: §1 · p. 3
• The register-based approach cannot measure tie strength, communication intensity, or whether relationships are positive or negative, and its inputs are difficult to tailor to specific research questions.Source: §1 · p. 3

• Socio-economic segregation may persist because homophilous social networks cluster people with similar attributes and regulate access to information, financial opportunities, education, and mobility.Source: §2 · p. 4
• The study defines its target as output homophily: how strongly the population-scale social network’s social opportunities are segregated by socio-economic status.Source: §2 · pp. 4–5
• The paper compares network segregation with administrative-neighborhood indicators rather than identifying the causal mechanisms generating observed segregation.Source: §2 · p. 5
• Neighborhoods remain relevant for social and economic opportunities, but mobility, declining neighborhood tie production, and online networks disperse social connections beyond residential areas.Source: §2 · p. 5
• Households are used as the unit of socio-economic status because they integrate members into shared economic units and capture choices such as housing and education.Source: §2 · p. 6
• The Netherlands provides a case combining relatively high wealth concentration, moderate income inequality, and comprehensive state-administered data for constructing a large-scale social network.Source: §2 · p. 6

Data

• 17.2 million registered residents are linked through 1.3 billion ties in Dutch population-register data, with pseudonymized attributes analyzed in secure Statistics Netherlands environments.The ties cover family, household, colleague, next-door-neighbor, and classmate relationships.Source: §3 · p. 6

Network construction

• The person-level network is aggregated into an unweighted multilayer household network, preserving separate family, neighborhood, school, and work links between households.Institutional households such as care homes, prisons, and orphanages are excluded.Source: §3 · pp. 6–7

Node attributes

• Household income, education, ethnic group, and migration background are assigned from household members’ founding adults, with combined or mixed categories where appropriate.Income is equivalised for household size and composition, while income groups are divided into ten equally populated deciles.Source: §3 · pp. 7–9

Network layers

• Four social layers capture family, school, work, and next-door-neighbor ties, which can span large distances without relying on predefined spatial units.Work ties are sampled among geographically closest colleagues at large workplaces, while neighbor ties include at most ten geographically closest households.Source: §3 · pp. 8–9

Mixing matrices

• Mixing matrices normalize each row to show how contacts from each subgroup are distributed across all subgroups, with diagonal cells representing within-group contacts.These matrices provide the basis for visualizing and quantifying aggregated social mixing patterns.Source: §3 · p. 9

Segregation measurement

• Assortativity coefficient ρ measures attribute similarity among connected households, with positive values indicating stronger mixing among similar groups and zero indicating random mixing.Income uses scalar correlation, whereas education, ethnic, and migration characteristics use discrete assortativity based on identical categories.Source: §3 · pp. 9–11

Spatial comparison

• The analysis compares link assortativity among direct multilayer-network neighbors with assortativity among residents sharing the same administrative “buurt” neighborhood.The neighborhood perspective uses the smallest administrative neighborhood borders in the Netherlands, typically covering a few thousand households.Source: §3 · pp. 9–10

Measurement limitations

• Nominal assortativity is constrained by group-size distributions, degree distributions, network topology, and possible asymmetric mixing, so empirical values may occupy a narrower range than −1 to 1.The authors also caution that scalar and discrete assortativity are not directly equivalent because income differences are quantitatively ordered whereas nominal categories are not.Source: §3 · pp. 10–11

Dimensions of segregation in the Netherlands

• Income assortativity ranges from 0.08 to 0.28 across municipalities and remains high for most of them.The analysis therefore focuses on income as the main dimension across which segregation manifests.Source: §4 · p. 12, Figure 2 · p. 12

Geographical decomposition

• Larger municipalities exhibit higher income segregation in a log-linear relationship, including cities in the Randstad and other larger municipalities.This relationship is not explained by differences in degree distributions, which remain consistent across municipality sizes.Source: §4 · p. 12, Figure 3 · p. 13

Comparing the social network and the spatial approach

• 0.217 versus 0.103 income assortativity shows more than twice as much segregation in the social network as in administrative neighborhoods.Social-network ties concentrate more strongly along the income diagonal, whereas neighborhood contacts are comparatively uniform across income ranges.Source: §4 · pp. 12–13, Figure 4 · p. 13
• Social networks reveal moderate segregation across the whole income range, whereas spatial segregation is concentrated mainly among the poorest and richest deciles.The network perspective therefore captures segregation patterns that spatial aggregation can minimize.Source: §4 · pp. 13–14

The multi-layered nature of social segregation

• Next-door neighbors are the most segregated layer (ρ=0.34), followed by work (ρ=0.20), school (ρ=0.16), and family (ρ=0.14); together they yield ρ=0.22.Removing any one layer leaves the overall score between 0.19 and 0.24, while work contributes most when layers are added sequentially.Source: §4 · p. 14, Figure 5 · p. 15
• Family ties are least segregated overall, but same-income exposure exceeds the 10% baseline across the income range and reaches half of social exposure among the three most similar groups for the top decile.School ties show around 50% and 45% exposure within the income ranges of the top and bottom deciles, respectively; workplace similarity strengthens with income, while neighbor segregation is concentrated among the bottom 30% and top 10%.Source: §4 · pp. 14–15, Figure 5 · p. 15

• Social-network segregation is more than twice as high as segregation in administrative neighborhoods, indicating that neighborhood indicators can miss persistent inequalities.The study positions spatial and social measures as complementary rather than interchangeable.Source: §5 · p. 16
• Next-door neighbors and work colleagues provide the least diverse socio-economic exposure, whereas family and school layers are comparatively more diverse but remain income-segregated.Next-door-neighbor segregation is highest overall, while workplace segregation combines substantial segregation with many connections.Source: §5 · pp. 16–17
• Income assortativity of 0.22 is consistent with reported assortativity ranges in empirical social networks, supporting further comparisons across countries and time.Such comparisons must account for network structure and underlying income inequality.Source: §5 · p. 17
• Larger municipalities exhibit higher socio-economic segregation because greater choice among affiliations may enable stronger choice homophily despite more diverse contact opportunities.This finding challenges the view of cities as hubs for diverse socio-economic mixing.Source: §5 · p. 18
• Policy evaluation should consider both assortativity and the number of connections affected, so reducing workplace segregation may matter more than reducing more-assortative neighbor ties.Workplace interventions can affect more connections even when neighbor assortativity is higher.Source: §5 · p. 18
• Neighborhood diversity does not guarantee social contact across income groups, and longitudinal data are needed to connect changes in neighborhood composition with changes in social-network composition.Administrative neighborhoods can appear mixed while nearby social ties remain segregated.Source: §5 · pp. 18–19
• The register-based network supports broader social-capital and cross-country comparisons, but missing migrant family ties, bounded firm structures, and indistinguishable same-year school classes constrain interpretation.These data-quality issues are presented as inputs for improving future population-scale network datasets.Source: §5 · p. 19
• The network captures formal ties within the Netherlands, omits informal and cross-border ties, and uses an assortativity measure with known sensitivity to group sizes, degree distributions, and asymmetric mixing.Evenly sized income groups and the observed degree distribution reduce some degree-skew bias, while subgroup asymmetry remains an open issue.Source: §5 · pp. 19–20

• Pseudonymized register data were analyzed in secure Statistics Netherlands server environments, without personal details such as addresses or birth dates.Source: §6 · p. 20

• Supplementary figures compare overall household degree distributions with distributions for family, neighbor, school, and work layers.Source: Figure S2 · p. 27
• Household degree patterns are also separated between border regions and the rest of the country, and between the largest and smallest municipalities.Source: Figure S2 · p. 27
• The supplement reports significance tests for relationships between income decile and migrant generation, ethnic group, education level, and categorical attribute pairs.ANOVA tests numeric–categorical relationships, while chi-square tests examine relationships among categorical variables at a 99% confidence interval.Source: Table S1 · p. 27
• Additional figures relate income assortativity per network layer to municipal population size and assess each layer’s contribution through exclusion and consecutive inclusion.Source: Figure S3 · p. 28, Figure S4 · p. 28
• Mixing matrices are provided separately for elementary, secondary, vocational, and higher education.Source: Figure S5 · p. 29