↖︎ Vishal Singh
The ZIP Code Destiny·Data Story № 15

The Places That
Beat the Model

Five census facts predict almost half the variation in neighborhood upward mobility. This story is about the other half — the counties whose poor children climb five or six percentiles higher than their demographics predict, and the sunbelt boomtowns that fall just as far short.

The ZIP Code Destiny · One national model, 81,885 tracts, county residuals · June 2026
R² = 0.44
five variables, national model
+6.3
Staten Island: percentiles above prediction
−5.4
Denver: percentiles below
density, overperformers vs under

Every model is a foil. Regress each tract's upward mobility — the adult income rank of children raised there in low-income families — on five ordinary census facts (poverty, college share, single-parent share, log income, Black share) and you explain 44 percent of the variance. That is the boring half of the story, mostly restating earlier chapters.

The interesting half is the disagreement. Where do children systematically out-climb what those five facts predict? A residual is not noise when it has a geography — it is a list of places that know something the model doesn't.

01

Prediction and escape

Each dot is a tract: the model's prediction against what actually happened to the children. The diagonal is perfect prediction; the vertical scatter is everything five census facts cannot see — schools, networks, churches, transit, safety, the local labor market, and luck.

Actual vs. predicted mobility, 4,000-tract sample. Population-weighted national WLS; predictors: poverty, college, single-parent, log income, Black share. Opportunity Atlas × ACS.
02

The geography of overperformance

Average the residuals by county and the disagreement organizes itself. The green belt of overperformance runs through the New York orbit — Staten Island (+6.3), Nassau, Bergen, Queens — and the immigrant-dense, union-dense, transit-dense Northeast generally, plus the rural Plains. The red belt of underperformance is the fast-growing Sunbelt and the scenic boomtowns: Denver (−5.4), Asheville's Buncombe County (−5.3), Knoxville, Durham, St. Johns FL. Places that look demographically alike launched their poor children very differently.

−5 under+5 over county residual, percentiles
County mean residual (population-weighted; counties with ≥8 measured tracts). Green: children out-performed the model's prediction. As above.

BEAT THE MODEL (≥250k people)

FELL SHORT (≥250k people)

03

What the winners share

Compare the top and bottom tenth of counties. The overperformers are five times denser, more immigrant, less smoking-prone — and, crucially, they are old places: established urban regions whose 1980s neighborhoods still resemble themselves. The underperformers are disproportionately places that boomed after these children grew up. That timing is a warning and a finding at once: part of the residual is history — the model sees today's census; the children lived in yesterday's county. But the New York orbit's edge survives every cut we tried, echoing the Atlas literature's own conclusion that dense immigrant metros punch far above their poverty rates.

Tract traits of the bottom vs. top residual decile (population-weighted). ACS; CDC PLACES.
04

Residuals as a research agenda

A residual map is the most honest to-do list a book can have. Each green county is a candidate mechanism — networks, transit, parishes, public sectors, immigrant scaffolding — and each red one a candidate failure worth a chapter visit. The model's job was never to be right; it was to subtract the obvious so that the non-obvious has somewhere to stand.

Notes & data