Researchers measured every kind of social capital a county can have — volunteering, civic clubs, tight-knit networks, cross-class friendship. Asked which one predicts whether poor children rise, the answer is brutally specific: only the friendships that cross class lines.
"Social capital" has been the great vague hope of American social policy since Bowling Alone: rebuild the leagues, the churches, the PTAs, and community will heal the rest. In 2022, Raj Chetty's team turned the privacy-protected Facebook graph of 72 million American adults into the first atlas of social capital actually measured — who is friends with whom, in every ZIP code and county.
The atlas distinguishes things the word "community" blurs together. Economic connectedness: do people with below-median incomes have friends with above-median incomes? Cohesion: are networks tight and cliquish, do friends support one another? Civic engagement: do people volunteer and join? These turn out to be different properties of different places — and only one of them tracks the chance that a poor child climbs.
Each dot is a county (50,000+ residents). The horizontal axis is economic connectedness: the share of high-income friends among low-income adults, scaled so 1.0 means friendships are fully integrated across the income distribution. The vertical axis is the Opportunity Atlas measure this book keeps returning to — the average adult income rank of children raised in poor families. The relationship is among the strongest county-level correlations in social science: r = 0.69.
Now run every form of social capital against the same mobility measure. Cross-class friendship and exposure to high-income people dominate. The classic civic measures — volunteering rates, the density of civic organizations — barely correlate. Network cliquishness does nothing. The support ratio runs negative: tight-knit mutual support is, if anything, a feature of places poor children struggle to leave. Community in general predicts nothing; cross-class contact predicts almost everything.
The obvious objection: connected counties are rich counties, and rich counties produce mobile children, so friendship is a proxy for income. Put them in the same regression and the objection fails — economic connectedness keeps the largest standardized coefficient, ahead of income itself. (In Chetty's tract-level work, controlling for connectedness roughly halves the predictive role of neighborhood income; the county pattern here agrees.)
Low connectedness has two distinct anatomies. In some counties, poor residents simply never encounter rich ones — there are few to encounter (low exposure, left side). In others, rich and poor share schools and churches and still don't befriend one another (high friending bias, top). The distinction matters for policy: exposure problems are about segregation between places; bias problems are about what happens inside them.
Zoom to ZIP codes and the uncomfortable underside appears: economic connectedness is itself stratified by income (r = 0.82 across ~19,000 ZIPs). In the richest ZIP codes, a below-median-income adult's friend group is dominated by above-median friends; in the poorest, such friendships are scarce. The asset that best predicts a poor child's rise is rationed by the same map as everything else — though the scatter shows real exceptions in both directions, and those exceptions are where the book's later chapters will dig.
Connectedness has a geography of its own: high across the Mountain West, the upper Midwest and rural New England — and low across the Southeast and the industrial cities, tracing (not coincidentally) the same arc as this book's mobility and life-expectancy maps.
The friendships were measured in 2022; the children whose adult incomes anchor the mobility measure grew up in the 1980s and 90s. The correlation works because both the friendship structure of places and their economic character are remarkably persistent — but it is a correlation, buttressed by the movers designs and school-cohort evidence in Chetty's papers rather than by anything experimental in this chart. And Facebook's graph is a proxy: friendships of adults 25–44 who use the platform, with SES inferred by model.
What the data rule out is cheaper than what they prove, and it is still a lot: the comfortable idea that any kind of "community" helps poor children equally. Bowling leagues do not move the needle. Neither does neighborly support. What moves it is the specific, awkward, hard-to-engineer resource of knowing people unlike yourself — and that resource is distributed by address, like everything else in this book.