Social capital II: determinants of economic connectedness

Many researchers and policy-makers have raised concerns that societies around the world are fragmented and polarized, with little interaction across racial, political and class lines. A lack of interaction between different groups is associated with worse economic and social outcomes; communities in which people with low socioeconomic status (SES) interact less with people with high SES exhibit less upward income mobility across generations. This Article analyses the determinants of social interactions across class lines, examining why people with higher SES tend to have more high-SES friends. Building on previous research, the authors distinguish two channels that can generate differences in an individual’s share of high-SES friends: (1) differences in high-SES exposure, the share of high-SES members in the groups to which people belong (for example, schools or religious organizations), and (2) differences in friending bias, the rate at which people befriend high-SES individuals conditional on exposure. Distinguishing between these channels is critical for designing interventions to increase economic connectedness—if differences in exposure are central, efforts to increase socioeconomic integration may be key; if friending bias is central, one must focus on increasing cross-SES interaction within existing groups. The study uses data on the social networks of 70.3 million Facebook users aged 25–44 in the United States to construct and publicly release privacy-protected measures of exposure and friending bias for each high school, college and ZIP code. Social network data are used as a proxy for real-world friendships rather than online interactions themselves. Socioeconomic status is measured by combining several indicators (such as neighborhood income and educational attainment) into a single index using a machine learning algorithm. Friendships are allocated to the groups in which they were formed using Facebook profile and group information. The goal is to determine the factors that generate the observed under-representation of high-SES friends among low-SES individuals and to inform targeted policies to increase cross-class interaction.

The study builds on a broad literature on social fragmentation and homophily across racial, political and class lines and their consequences for public goods, polarization and mobility (for example, Fischer & Mattson; Smith et al.; Doob; Putnam; Alesina et al.). It extends prior work using Facebook and other large-scale data to study social networks as a proxy for offline friendships and their economic effects (Bailey et al.; Gentzkow & Shapiro). The analysis advances beyond measures of segregation based on co-location or mobile device data by distinguishing exposure (group composition) from actual interaction (friendship) within groups. The conceptual framework draws on classic theories of intergroup contact and social structure (Allport; Blau & Schwartz) and economic models of friendship formation with homophily (Currarini, Jackson & Pin). It connects to research on neighborhoods and mobility (Chetty & Hendren; Opportunity Atlas) by examining how exposure and friending bias relate to upward income mobility.

Data: The authors analyze 70.3 million Facebook users aged 25–44 residing in the United States who were active in the prior 30 days, have at least 100 US-based friends, and a non-missing ZIP code. They further require that at least one friendship can be allocated to a setting/group using Facebook profiles and memberships. The final sample corresponds to 82% of the US population aged 25–44; benchmarking suggests reasonable representativeness. Measures: Socioeconomic status (SES) is computed by combining multiple indicators (e.g., neighborhood average income, educational attainment) into a single machine-learning-based index. Economic connectedness (EC) is defined for low-SES individuals as twice the fraction of their friends who have above-median SES; EC=1 implies equal numbers of high- and low-SES friends; EC=0 implies no cross-SES connections. Settings and group assignment: Friendships are allocated to six settings where friendships are commonly formed: high schools, colleges, religious groups, recreational groups, workplaces, and neighborhoods (ZIP codes). For each individual and setting, the authors estimate: (i) exposure: two times the share of group members with above-median SES; (ii) friending bias: one minus the ratio of the share of high-SES friends formed in that group to the share of high-SES peers in the group; (iii) friending shares: the share of an individual’s total friends formed in each setting. Exposure and friending bias are aggregated to setting×SES and to geographic levels (ZIP code, county) using individual-level weights (friendship shares per setting). Privacy protection: Differential privacy-style methods add noise to published statistics. Decomposition analyses: To quantify determinants of differences in EC, the authors conduct counterfactual exercises that sequentially equate (a) friending shares across settings, (b) exposure within settings, and (c) friending bias within settings between comparison groups. Two decompositions are performed: (1) between individuals with low vs high SES; (2) between ZIP codes in the bottom vs top quintiles of EC for low-SES residents. These exercises isolate the contribution of each component to EC gaps. Geographic analysis: Individual-level exposure and friending bias are aggregated to counties and ZIP codes to map the geography of exposure and bias and to assess their covariance and spatial patterns. Association with upward mobility: The study regresses log upward income mobility (Opportunity Atlas) on log EC and on its components, log exposure and log(1−friending bias), at ZIP-code and county levels, with and without county fixed effects. It also uses counties’ causal effects on mobility from movers-based estimates. Group-level (school/college) estimation: For high schools and colleges, exposure and friending bias are estimated using parental SES and (separately) own adult SES for birth cohorts 1990–2000. Reliability is assessed via split-sample methods. Quasi-experimental designs:

• Cross-cohort fluctuations: Within a school, random-like year-to-year variation in the share of high-SES students is used to estimate the causal effect of exposure on EC among low-SES students, focusing on within-cohort friendships and peers (parental SES bottom vs top quintiles).
• Regression discontinuity (RD): School entry birthdate cutoffs generate discontinuities in cohort composition (high-SES share). RD estimates the impact of exposure on EC around the cutoff, comparing effects across schools with low vs high friending bias. Total contribution to EC (TCEC): For schools, TCEC equals the product of the share of low-SES students and their average EC, capturing a school’s accounting contribution to the total number of cross-SES friendships, holding enrollment and friend counts fixed.

• Average connectedness: For low-SES individuals in the U.S., EC is 0.776, indicating a 22.4% under-representation of high-SES friends relative to population share.
• Exposure vs friending bias: Roughly half of the social disconnection between low- and high-SES individuals is due to differences in exposure and half to friending bias. Decomposition across SES shows that equating friending shares closes 12% of the EC gap, equating exposure (plus friending shares) closes 54%, and equating friending bias (plus friending shares) closes the remaining 46%.
• Variation across settings: Friending bias for low-SES individuals is highest in neighborhoods (mean ~0.17) and lowest in religious groups (mean ~−0.03). Holding exposure fixed, low-SES people are about 20% more likely to befriend a given high-SES person in religious groups than in neighborhoods. Despite low friending bias, religious groups’ high segregation limits their EC.
• Geographic decomposition: Between bottom- and top-quintile ZIP codes (by EC for low-SES residents), 73% of the EC difference is explained by exposure differences, 16% by friending bias, and 11% by friending shares across settings.
• Spatial patterns: Exposure is generally higher in higher-income coasts; friending bias is lowest in the Midwest/Great Plains. Exposure and friending bias are mildly negatively correlated across counties (≈−0.2).
• Upward mobility associations: Elasticity of upward mobility w.r.t. EC ≈ 0.24 at ZIP codes. When decomposed, elasticities are ~0.25 for exposure and ~0.19 for 1−friending bias; within-county estimates remain near 0.25. At counties, results are qualitatively similar, with less precision for friending bias. Using counties’ causal mobility effects, both exposure and friending bias remain strongly predictive (elasticities ~0.116 for exposure and ~0.339 for 1−friending bias).
• School/college heterogeneity: Substantial variation in both exposure and friending bias across high schools and colleges, even among nearby institutions with similar SES composition (e.g., ETHS vs Payton). Friending bias is higher with more academic tracking, larger group size, greater SES and racial diversity; lower in smaller and less diverse schools.
• Causal effects of integration: Cross-cohort analysis shows slope ≈0.89 between cohort exposure and EC changes among low-SES students (marginal friending bias ≈0.11), aligning with cross-sectional bias. The effect of exposure on EC is attenuated in schools with higher friending bias; a 1 pp increase in mean friending bias in other cohorts reduces the exposure effect by ~0.61 pp (adjusted ~0.91 pp when accounting for measurement error). RD analyses around school-entry cutoffs confirm that increased exposure raises high-SES friendships more in low-bias schools; the jump in high-SES friendships is ~0.06 units larger in low-bias than high-bias schools.
• Policy implications: Socioeconomic integration increases EC especially where friending bias is low. Where friending bias is high, reducing within-group segregation and structural barriers is crucial. Reducing friending bias always increases both individual EC and total contribution to connectedness (TCEC); exposure has non-monotonic effects on TCEC and redistributes opportunities across schools.
• Public data: The study releases privacy-protected measures of EC, exposure, and friending bias for U.S. ZIP codes, high schools, and colleges.

The findings demonstrate that interaction across class lines depends on both exposure (group composition) and friending bias (propensity to form cross-SES ties conditional on exposure). While traditional policies—school desegregation, zoning, college admissions reforms—target exposure, the results show that even perfect socioeconomic integration would leave about half of the EC gap intact due to friending bias. Friending bias varies across settings and appears shaped by institutional and structural factors—such as group size, tracking, and racial diversity—rather than fixed individual preferences, implying it can be influenced through policy. The strong associations of both exposure and reduced friending bias with upward mobility—and their predictive power for counties’ causal mobility effects—suggest that it is interaction with higher-SES peers, not simply their presence, that most strongly relates to economic opportunity. Quasi-experimental evidence indicates that increases in exposure causally raise cross-SES friendships, especially where friending bias is low, validating the use of observed bias to forecast policy impacts. Therefore, effective strategies to boost economic connectedness and mobility should combine efforts to increase integration with targeted interventions that reduce friending bias within groups.

This paper decomposes social disconnection across class lines into exposure and friending bias using large-scale Facebook network data and shows that both components contribute roughly equally to the EC gap between low- and high-SES individuals. Exposure differences primarily drive geographic variation in EC, while friending bias is more setting-specific and shaped by institutional structures. The study documents substantial heterogeneity in exposure and friending bias across schools and colleges and provides quasi-experimental evidence that socioeconomic integration increases cross-SES friendships, especially in low-bias environments. The authors release privacy-protected measures of EC, exposure, and friending bias for U.S. ZIP codes, high schools, and colleges to guide policy. Future research should evaluate causal impacts of specific interventions that reduce friending bias (e.g., changes in tracking, group sizes, spatial design, and new interaction programs) and study how institutional reforms can durably shift friending patterns while balancing system-wide effects of reassigning exposure across schools.

• Data proxy: Facebook friendship networks are used as a proxy for real-world friendships; the analysis does not identify effects of online social networks themselves.
• Representativeness: Although coverage is broad (70.3M users, ~82% of U.S. ages 25–44) and benchmarking suggests reasonable representativeness, Facebook users with required activity and friend counts may differ from the full population.
• Measurement choices: The distinction between exposure and friending bias depends on the level of aggregation (e.g., school-level vs within-school tracking), making it policy-dependent; unobserved within-group segregation can be absorbed into friending bias.
• Privacy noise and measurement error: Published statistics include added noise for privacy; friending bias estimates (e.g., parental SES-based) have imperfect reliability (e.g., ~0.58), introducing attenuation.
• Scope limitations: Group-specific estimates are released only for high schools, colleges, and ZIP codes; sample sizes are too small for reliable group-level estimates for religious organizations, recreational groups, and employers.
• Observational components: While quasi-experimental designs support causal interpretations for exposure effects, many analyses (e.g., associations with upward mobility) remain observational and subject to omitted-variable concerns.
• Age and geography scope: Focus on U.S.-based users aged 25–44; results may not generalize to other ages or countries.
• Friend count differences: High-SES individuals make more friends on average; assumptions and adjustments are needed when interpreting bias under non-random friendship formation.