Chapter 9.4 · Coda

Places that beat the model

Five census measures account for 44.0% of the variation in tract upward mobility, counties below that prediction cluster in the South and the largest counties above it lie around New York, and counties above it lost tax filers on net in filing years 2022 and 2023, a correlation of −0.44.

Author: Vishal Singh
Affiliation: NYU Stern
Published: 9 October 2026

Five census measures account for 44.0% (95% CI 41.8% to 46.1%) of the variation in upward mobility across 82,791 tracts, where upward mobility is the adult household income rank of children raised in low-income families.[1] The remaining 56.0% is not scattered at random. Averaged to counties, it lies below the prediction in the South and above it in the largest counties around New York, and the counties above the prediction lost tax filers on net in filing years 2022 and 2023, a correlation of −0.44 (95% CI −0.57 to −0.28) between the county residual and the net migration rate.

The chapter takes two steps. The first fits the model, averages its residuals to counties and describes the counties at both ends. The second follows tax filers between counties and asks whether net migration runs toward the places where children from low-income families rise furthest. The life-expectancy gradient that the book starts from is in Chapter 6.1, Place and the length of a life, and the Opportunity Atlas map of mobility is in Chapter 3.3, Race, income and mobility. A residual is the part of mobility that these five measures leave unexplained. It is not a measure of what a place does to a child.

A five-measure model of upward mobility

The outcome is the position in the national distribution of household income that children born in 1978 to 1983 to parents at the 25th income percentile had reached by 2014 to 2015, as the Opportunity Atlas estimates it. The predictors are the poverty rate, the share of adults with a bachelor's degree or more, the single-parent household share, the natural log of median household income and the Black share of residents (Black alone, of any ethnicity), all from the American Community Survey for 2019 to 2023. A weighted least-squares line, with each tract counted in proportion to its population, fits 82,791 tracts in 3,136 counties after leaving out tracts with 200 residents or fewer. It accounts for 44.0% (95% CI 41.8% to 46.1%) of the variance in tract mobility, and 56.0% (95% CI 53.9% to 58.2%) remains. The interval comes from a bootstrap that resamples whole counties.

The outcome and the predictors describe different decades. The children grew up in the 1980s and 1990s, and the predictors describe 2019 to 2023, so each prediction asks what a tract's census profile today implies for children raised there a generation ago. Figure 1 plots the actual rank against the prediction. Tenths of tracts ranked by prediction lie close to the diagonal, so the line is not tilted at either end, and the dots spread around it with a population-weighted standard deviation of 5.0 percentile points. The gap between a tract's actual and predicted rank is its residual.

Figure 1

Five census measures account for 44.0% of the variation in tract upward mobility, and the population-weighted standard deviation of the gap between actual and predicted rank is 5.0 percentile points

Adult household income rank (percentile) of children raised in low-income families against the rank predicted from poverty, college, single-parent share, income and Black share, one dot per tract (random sample)

Loading data

Dots spread on both sides of the diagonal, which marks a perfect prediction, and the markers for tenths of tracts ranked by prediction sit close to it. Dots are a random sample of 4,000 of the 82,791 tracts in the model, and 132 tracts fall outside the plotted range; markers are population-weighted mean actual ranks in tenths of all tracts, with 95% intervals from a bootstrap over counties. The vertical distance from the diagonal is the residual.Note: The dots are a random sample of the tracts inside the plotted range, drawn with a fixed seed (stated in the methods note); the R-squared, the residuals and the markers use every tract in the model.Note: Predictors are the poverty rate, the share of adults with a bachelor's degree or more, the single-parent household share, the natural log of median household income and the Black share of residents (Black alone, of any ethnicity), all from the ACS for 2019 to 2023. Tracts with 200 residents or fewer are left out, and each tract is weighted by its ACS population.Note: The outcome describes children raised in the 1980s and 1990s and the predictors describe 2019 to 2023, so the residual includes whatever changed in between. Ranks are percentiles of the national distribution of household income for children born in 1978 to 1983, measured in 2014 to 2015. The Opportunity Atlas adds noise to its estimates to protect privacy and the staged values are never clipped. The file on 2020 tracts is interpolated from 2010 tracts by land area.Source: Opportunity Atlas tract outcomes, 1978-1983 birth cohorts (2020 tracts), and ACS 2019-2023 predictors; population-weighted (ACS tract population), modeled estimates

A tract's residual mixes things the data cannot separate. The Atlas estimate for a tract rests on a modest number of children and carries added privacy noise. The five measures omit much that differs between places, such as schools and local labor markets. And the outcome and the predictors describe different decades. The next section asks whether residuals averaged to counties behave like noise.

Counties above and below the prediction

The residual of a county is the mean of its tract residuals, weighted by tract population. A county needs at least 8 tracts in the model, which leaves 1,576 counties that hold 94% of the model's population (Figure 2). County residuals run from −4.8 percentile points at the fifth percentile of counties to +5.0 at the ninety-fifth.

Ranking counties on a noisy estimate can produce a list of extremes that is mostly noise, so two checks come first. The typical standard error of a county residual is 0.7 percentile points, small beside that spread, though it treats the tracts of a county as independent and leaves out errors that all of them share. The second check splits the tracts of each county at random into two subsets, ranks counties on one subset and reads the residual off the other. Across counties the residuals of the two subsets correlate +0.85 (95% CI 0.80 to 0.88), and the tenth of counties highest in the first subset has a mean residual of +5.0 (95% CI 3.7 to 6.1) in the second, against −4.4 (95% CI −5.0 to −3.4) for the lowest tenth. The ranking is therefore not an artifact of the noise in a subset. The floor matters for the extremes. A floor of 5 tracts maps 2,140 counties and puts the ninety-fifth percentile of county residuals at +6.6, and a floor of 15 maps 944 and puts it at +3.9. The correlation of the residual with net migration, below, keeps its sign at both floors.

Among the 281 counties with at least 250,000 residents, the largest residuals, in percentile points of adult income rank, are in Richmond County, New York, which is Staten Island (+6.2 (95% CI 5.3 to 7.2)), Nassau County, New York (+6.1 (95% CI 5.4 to 6.8)), Bergen County, New Jersey (+6.1 (95% CI 5.5 to 6.7)) and Queens (+5.6 (95% CI 5.2 to 6.1)). Of the 10 counties of that size with the largest residuals, 9 lie in New York, New Jersey, Pennsylvania, Massachusetts or Virginia. The smallest residuals are in St. Johns County, Florida (−5.8 (95% CI −7.3 to −4.4)), Denver (−5.4 (95% CI −6.1 to −4.6)) and Buncombe County, North Carolina, which holds Asheville (−5.4 (95% CI −6.1 to −4.6)), and 8 of the 10 smallest lie in the South, a Census region that includes the District of Columbia.

Across the mapped counties, the South holds the bottom tenth. Of its counties, 81% lie in the South, which holds 46% of all mapped counties, and the share is 83% with a floor of 15 tracts. The top tenth depends on the floor. With the floor of 8 tracts, 54% of the top tenth lie in the Midwest, which holds 28% of the counties, and 14% in the Northeast, which holds 13%. With a floor of 15 tracts the Midwest share is 37% against 23% of the counties, and the Northeast share is 28% against 17%. The Midwestern part of the top tenth is therefore made of counties with few tracts, and the largest counties above the prediction are the ones near New York.

Figure 2

Counties below the prediction cluster in the South, and the largest counties above it lie around New York

Population-weighted mean residual of tract upward mobility (actual minus predicted rank, percentile points) by county, counties with at least 8 tracts in the model

Loading county values

Mean residual of upward mobility (percentile points, actual minus predicted)

Below the predictionAbove the prediction

Values outside −5.0 to 5.0 take the end colors.

No data (3,143 counties)

Select a class to show only those places.

Find a county

Hover or tap a county for its value, or search by name. Zoom with the buttons on the map, then drag to move it.

Teal counties lie above the prediction and brown counties below it. The South sits toward the low end of the key, and the largest counties toward the high end are in the New York area. Among smaller counties the high end is in the Midwest under the floor used here and moves toward the Northeast under a higher floor. The map plots the mean tract residual of each of 1,576 counties, and 148 of them lie beyond the 5 points at either end of the key. The 1,567 gray counties have fewer tracts than the floor or none in the model; tooltips give a 95% interval from the spread of tract residuals within the county.Note: A county is mapped when at least 8 of its tracts are in the model, which covers 94% of the model's population. Each gray county gives its reason in the tooltip.Note: The interval in the tooltip treats a county's tracts as independent draws, so it leaves out errors that all tracts of a county share. The typical standard error is 0.7 percentile points.Note: Connecticut is drawn as its eight legacy counties; its tracts are assigned to them exactly. Alaska is drawn. Puerto Rico is left out of the model and is not drawn. Ranks are percentiles of the national distribution of household income for children born in 1978 to 1983, measured in 2014 to 2015. The Opportunity Atlas adds noise to its estimates to protect privacy and the staged values are never clipped. The file on 2020 tracts is interpolated from 2010 tracts by land area.Source: Opportunity Atlas tract outcomes, 1978-1983 birth cohorts, and ACS 2019-2023; population-weighted county means of model residuals, modeled estimates

Counties that gained many college-educated adults after the 1990s have high predictors for children who grew up there earlier, and they would sit below the line for that reason alone. The staged data hold no census measures from before 2019, so the chapter cannot test that explanation.

Older housing and fewer smokers mark the counties above the prediction

Figure 3 compares the tenth of counties with the largest residuals, home to 18.1 million people, with the tenth with the smallest, home to 17.2 million people. The table under the figure gives each top minus bottom difference with an interval from draws shared by the two groups. Five differences have intervals that exclude zero. Housing built before 1960 makes up 38% (95% CI 28% to 47%) of housing units in the top tenth and 18% (95% CI 15% to 22%) in the bottom tenth, a difference of +20.1 (95% CI 9.8 to 29.2) points. The smoking share of adults is 13.5% (95% CI 12.6% to 14.8%) in the top tenth and 15.9% (95% CI 15.2% to 16.6%) in the bottom tenth, a difference of −2.4 (95% CI −3.5 to −0.9) points. The Hispanic share differs by +5.2 (95% CI 0.7 to 9.3) points and the poverty rate by −1.9 (95% CI −3.4 to −0.2) points. The average resident of the top tenth lives at 9,090 (95% CI 2,838 to 18,233) people per square mile, against 2,286 (95% CI 1,149 to 3,952) in the bottom tenth, a difference of +6,803 (95% CI 97 to 15,794), whose lower limit is close to zero. The differences in the Black share, in ownership and in college include zero.

Figure 3

Counties in the top tenth by residual have more pre-1960 housing, a larger Hispanic share, less poverty and fewer smokers than the bottom tenth, and the differences in the Black share, ownership and college include zero

Population-weighted tract means in the bottom and top tenths of counties by mean residual (equal numbers of counties), with 95% intervals from a bootstrap over counties

Bottom tenth of counties (below the prediction)
Top tenth of counties (above the prediction)

Show data

Data for Counties in the top tenth by residual have more pre-1960 housing, a larger Hispanic share, less poverty and fewer smokers than the bottom tenth, and the differences in the Black share, ownership and college include zero
Trait	Bottom tenth	Top tenth	Top minus bottom (points)	Lower limit of the 95% interval	Upper limit of the 95% interval
Housing built before 1960 (share of housing units)	18.1	38.3	+20.1	+9.8	+29.2
Hispanic residents (share of residents)	10.5	15.7	+5.2	+0.7	+9.3
Black-alone residents (share of residents)	11.8	8.6	−3.2	−7.2	+0.3
Adults who currently smoke (crude prevalence, share of adults)	15.9	13.5	−2.4	−3.5	−0.9
Owner-occupied homes (share of households)	70.2	68.2	−1.9	−8.4	+4.0
Adults with a bachelor's degree or more (share of adults)	33.9	37.6	+3.6	−2.5	+9.2
Residents below the poverty line (share of residents)	12.6	10.7	−1.9	−3.4	−0.2

The open dot is the bottom tenth of counties, below the prediction, and the filled dot is the top tenth. The last column is the top minus bottom difference, and the table under the figure gives it with a 95% interval from draws shared by the two groups: the interval excludes zero for pre-1960 housing, the Hispanic share, poverty and smoking and includes zero for the Black share, ownership and college. Residents of the bottom tenth number 17.2 million and of the top tenth 18.1 million.Note: Tenths hold equal numbers of counties ranked on the mean residual, and the tracts of those counties form the groups. The groups are fixed before the draws, so the intervals leave out the uncertainty of the ranking.Note: Race shares are race alone and include Hispanic residents of any race. Smoking is the PLACES crude prevalence among adults, a model-based estimate built partly from ACS inputs. The pre-1960 share is from EJScreen, which draws it from the ACS for 2018 to 2022. Tracts without a value drop out of that row.Note: The poverty, college and Black-share rows are predictors in the model, so their gaps are differences between whole counties; inside a tract the residual is unrelated to them by construction.Source: ACS 2019-2023, CDC PLACES 2024 release (smoking, modeled estimates) and EPA EJScreen 2024 (housing age); population-weighted (ACS tract population) tract means by group of counties

The same five measures account for 44.2% (95% CI 42.0% to 46.5%) of the variance of tract life expectancy at birth for deaths in 2010 to 2015. County residuals for the two outcomes move together. Across counties, the residual of mobility correlates +0.39 (95% CI 0.27 to 0.51) with the residual of life expectancy, so counties above the prediction for children raised in the 1980s tend to be above it for the life expectancy of residents in 2010 to 2015. What the five measures leave out is partly shared between the two outcomes, and these data do not say what it is.

Net income moved toward the South in filing years 2022 and 2023

The Internal Revenue Service matches the individual returns filed in 2022 with those filed in 2023 on the filer's taxpayer identification number, assigns each return a county from the full ZIP code on it, and counts a return as a migrant when the county differs between the two filings.[2] The unit is the filer, usually a household, and not the person; people who do not file are missing, and the address on a return is a mailing address. Arrivals include moves from abroad and departures include moves abroad. Arriving returns equal 5.6% of the returns filed for tax year 2022. Adjusted gross income is the amount on the 2023 return, and net income is arriving minus leaving.

Of the 3,086 counties with both directions published, 1,719 gained net income and 1,367 lost it, and the gaining counties together received 71.6 billion dollars net (Figure 4). Among counties of 50,000 residents or more, San Francisco lost the most per resident, 2,488 dollars, and Collier County, Florida, gained the most, 5,793 dollars. Pooled by Census region, net income per resident was 270 (95% CI 22 to 528) dollars in the South, −332 (95% CI −475 to −54) in the Northeast and −172 (95% CI −296 to −60) in the Midwest; the interval for the West includes zero.

Figure 4

Net adjusted gross income was positive in 1,719 counties and negative in 1,367, with the largest losses per resident in San Francisco and Boston's Suffolk County and the largest gains in Florida

Net adjusted gross income (arriving minus leaving returns) in dollars per resident by county, returns filed in 2022 matched with returns filed in 2023

Loading county values

Net adjusted gross income per resident (dollars)

Net outflow of incomeNet inflow of income

Values outside −2,000 to 2,000 take the end colors.

No data (3,143 counties)

Select a class to show only those places.

Find a county

Hover or tap a county for its value, or search by name. Zoom with the buttons on the map, then drag to move it.

Teal counties received more income than they lost and brown counties lost more than they received. Among counties of 50,000 residents or more, the largest per-resident losses are in San Francisco, Suffolk and Arlington counties and the largest gains in Collier, Walton and Nassau counties in Florida. The key ends at 2,000 dollars either way and 55 counties lie beyond it; 57 gray counties have no net figure.Note: Arrivals and departures are tax returns matched across the 2022 and 2023 filings on the filer's taxpayer identification number, with a county assigned from the ZIP+4 on each return, so the unit is the filer and not the person. People who do not file are missing, and the IRS withholds small flows; a withheld direction leaves the county gray, never at zero. Income is adjusted gross income on the 2023 return, converted from thousands of dollars.Note: Arrivals include moves from abroad, departures include moves abroad, and both include moves to and from the territories. Net income is divided by the ACS 2019-2023 population, which counts people who do not file.Note: Connecticut's IRS flows are published by planning region and are allocated to its legacy counties in proportion to 2020 population, which is approximate. Alaska is drawn. Puerto Rico is not drawn.Source: IRS Statistics of Income county-to-county migration data, filing years 2022 and 2023, and ACS 2019-2023 population; full counts of returns, not weighted

Net migration is lower where childhood mobility is higher

Figure 5 sets each county's net migration rate, arriving minus leaving returns per thousand returns filed, against the Atlas mobility rank of children raised there. Across 3,086 counties the correlation is −0.19 (95% CI −0.31 to −0.02) when each county counts in proportion to its population and −0.23 (95% CI −0.33 to −0.10) when each counts once. The slope of the fitted line is −0.39 (95% CI −0.68 to −0.05) returns per thousand for each percentile point of the mobility rank. The sign is the same against life expectancy for deaths in 2010 to 2015, −0.25 (95% CI −0.37 to −0.09) weighted and −0.13 (95% CI −0.23 to −0.01) unweighted, so counties where children from low-income families rose further and adults lived longer gained fewer filers.

The tenths show where the association sits. The two tenths highest on mobility had net rates of −5.2 (95% CI −7.9 to −1.6) and −3.6 (95% CI −4.7 to −1.9) returns per thousand, with intervals below zero. The lowest tenths gained in the point estimates, +0.9 (95% CI −2.7 to 3.6) and +3.6 (95% CI −0.7 to 7.3), but their intervals include zero. Deprivation gives a less consistent picture, since the correlation is +0.19 (95% CI −0.05 to 0.34) weighted and −0.19 (95% CI −0.28 to −0.09) unweighted. The most deprived tenth of counties had a net rate of −4.5 (95% CI −6.7 to −2.3) returns per thousand, the least deprived tenth had −2.4 (95% CI −4.6 to 1.5), and the second least deprived had +2.6 (95% CI −0.1 to 5.1), so net out-migration of filers shows at the most deprived end with an interval below zero and at the least deprived end only as a point estimate.

Figure 5

Net migration of tax filers is lower in counties with higher childhood mobility, a population-weighted correlation of −0.19, and the two highest tenths by mobility lost filers on net

Net migrating returns per 1,000 returns filed for tax year 2022, by county, against the Opportunity Atlas upward-mobility rank, returns filed in 2022 matched with returns filed in 2023

Counties with a higher upward-mobility rank had lower net migration, a slope of −0.39 returns per thousand for each percentile point

Loading data

Dots drift down from left to right, and the fitted line falls across the range of mobility ranks. The line is the population-weighted least-squares fit, with a 95% band from a bootstrap over states.

The first tab shows counties sloping down to the right, and the second and third tabs group counties into tenths. Each dot in the first tab is a county sized by population, and 21 counties fall outside the plotted range; the fitted line is population-weighted, with a 95% band from a bootstrap over states. The bars in the other tabs are pooled rates for tenths of counties, with 95% intervals from the same bootstrap.Note: The net migration rate is arriving minus leaving returns per thousand returns filed in the county for tax year 2022. Returns are filers, not people; arrivals include moves from abroad and departures include moves abroad.Note: The scatter shows 3,086 counties. The correlation counts each county in proportion to its population. Counting each county once gives −0.23.Note: Tenths hold equal numbers of the counties in the analysis. A bar is the sum of net returns over the counties of a tenth divided by the sum of their returns filed, so large counties count more. The intervals resample states, whose counties resemble each other.Source: IRS Statistics of Income county migration data, filing years 2022 and 2023, IRS SOI county file, tax year 2022, Opportunity Atlas county outcomes (1978-1983 birth cohorts) and ADI 2023; population-weighted fit, full counts of returns

The county residual from the model moves with net migration more strongly than mobility itself does. Across the counties with a residual, the correlation of the residual with the net migration rate is −0.44 (95% CI −0.57 to −0.28) weighted and −0.39 (95% CI −0.49 to −0.27) counting each county once. Part of that link is built in by timing. The predictors describe 2019 to 2023, after many movers had arrived, so a county that gained college-educated residents has high predictors and, if its children's outcomes predate those residents, a negative residual. The data cannot separate that mechanical link from any other.

Home values and net income per resident

Figure 6 plots net income per resident against county home value. Across 3,076 counties the correlation is −0.37 (95% CI −0.50 to −0.13) weighted by population and +0.11 (95% CI −0.01 to 0.24) counting each county once, and the second interval includes zero. The two differ because the weighted version counts each county in proportion to its population. The negative association is concentrated at the top of the home-value scale. In the top tenth of counties by home value, net income was −266 (95% CI −385 to −31) dollars per resident, and in the lowest tenth −91 (95% CI −119 to −60), while the ninth tenth received 300 (95% CI 74 to 482). Net income per resident did not decline steadily with home value. It was negative in the cheapest tenth and the most expensive tenth and positive in the ninth. The relation of life expectancy to home value, and the years of life that go with each doubling of it, are in Chapter 4.1, Home value, income and redlining.

Figure 6

Net income per resident is lower in the counties with the highest home values, a population-weighted correlation of −0.37 that is +0.11 when each county counts once

Net adjusted gross income per resident in dollars against county home value in dollars (log scale), one dot per county sized by population, with pooled means for tenths of counties

Loading data

Only the markers for the lowest and the highest tenth of counties by home value have intervals entirely below zero, so the loss in the highest tenth carries the negative correlation. Dots are counties sized by population; the markers pool the counties of each tenth, with 95% intervals from a bootstrap over states. 6 counties fall outside the plotted range.Note: The county home value is the population-weighted mean of the median home values of the ZIP code areas in the county, taken from the ZCTA atlas, whose ACS vintage is not stated upstream. Connecticut's legacy counties are included, and the Census Bureau top-codes home values.Note: The correlation counts each county in proportion to its population. Counting each county once, it is +0.11, and its interval includes zero. Both use 3,076 counties.Note: Tenths hold equal numbers of counties, and each marker pools the counties of a tenth, so large counties count more. Intervals resample states.Source: IRS Statistics of Income county migration data, filing years 2022 and 2023, ACS 2019-2023 population and ZIP-area median home values from the ZCTA atlas (ACS-based, vintage not stated upstream); population-weighted, full counts of returns

Pooled by deprivation, arriving and leaving returns report similar incomes

One account of these flows is income sorting, in which richer filers replace poorer ones. Figure 7 compares the average adjusted gross income of arriving and leaving returns, pooled over tenths of counties by deprivation. Both averages fall by nearly the same amount from the least to the most deprived tenth, and the arriving minus leaving gap runs from −2,110 to 2,486 dollars across the tenths, with an interval that includes zero in 9 of the ten, among them the most deprived tenth, where it is −337 (95% CI −1126 to 559) dollars. Only the ninth tenth has arriving returns that report more than leaving returns by an amount the interval separates from zero.

Counties taken one by one give a different answer from pooling. The correlation of the county gap with deprivation is +0.02 (95% CI −0.13 to 0.13) weighted by population, a value near zero, and −0.21 (95% CI −0.29 to −0.14) counting each county once, so across counties arriving returns report less than leaving returns where deprivation is higher. In the median county, arriving returns report 54,902 (95% CI 52,722 to 57,194) dollars of income and leaving returns 52,444 (95% CI 50,430 to 54,356), a median gap of 2,514 (95% CI 1,466 to 3,433) dollars. Similar incomes therefore hold for the pooled tenths and the population-weighted correlation, and not for the median county or the unweighted correlation. The averages describe returns, and they hide who within a county moves.

Figure 7

In pooled tenths of counties by deprivation, returns arriving report about the same income as returns leaving, though the median county's arriving returns report 2,514 dollars more

Average adjusted gross income per return of arriving and of leaving returns in dollars, by tenth of counties ranked on the Area Deprivation Index (1 = least deprived), returns filed in 2022 matched with 2023

Average adjusted gross income per return (dollars)

Show data

Data for In pooled tenths of counties by deprivation, returns arriving report about the same income as returns leaving, though the median county's arriving returns report 2,514 dollars more
conf_high	conf_low	decile	estimate	n_units	series
101858	89963	1	96045	308	Returns arriving
100399	88774	1	95914	308	Returns leaving
81801	71936	2	77560	308	Returns arriving
81840	69588	2	75074	308	Returns leaving
67604	62926	3	65531	308	Returns arriving
70059	62338	3	66572	308	Returns leaving
61978	58963	4	60616	308	Returns arriving
65788	59809	4	62568	308	Returns leaving
57473	53277	5	55646	308	Returns arriving
60851	54038	5	57756	308	Returns leaving
54966	51571	6	53303	308	Returns arriving
55552	51644	6	53594	308	Returns leaving
52838	49110	7	50810	308	Returns arriving
51851	48448	7	50127	308	Returns leaving
51626	47803	8	49699	308	Returns arriving
49903	46304	8	48257	308	Returns leaving
47507	44521	9	45946	308	Returns arriving
45728	43273	9	44383	308	Returns leaving
42970	40300	10	41459	308	Returns arriving
43325	40678	10	41796	308	Returns leaving

The two lines follow each other down the deprivation scale and cross. Each point pools the counties of a tenth (summed income over summed returns), with a 95% interval from a bootstrap over states; the arriving minus leaving gap has an interval that includes zero in 9 tenths, among them the most deprived, where it is −337. Across counties, the gap shows no trend with deprivation when counties count in proportion to their population (+0.02) and falls with deprivation when each counts once (−0.21).Note: Adjusted gross income is the amount on the 2023 return of the matched pair of returns. A return is a filer, usually a household, and the averages weight each return equally.Note: Counties need both directions published, so the 57 map units without a net figure are left out. The median county is the middle county when each counts once.Note: Pooled averages hide who moves within a county, and pooling (which weights counties by their returns) gives a different answer from counting each county once.Source: IRS Statistics of Income county migration data, filing years 2022 and 2023, and Area Deprivation Index 2023; pooled over counties with both directions published, full counts of returns

Causal evidenceMovers design; Bartik-type comparisons of state income convergence, split by predetermined land-use regulation and by land availability

Whether places change children's incomes and whether housing costs steer migration

Two studies ask the causal questions that this chapter's comparisons point to.

Chetty and Hendren followed children whose families moved between counties and commuting zones, using variation in the age of the child at the move, including siblings in the same family. The outcomes of children who moved to a county with higher incomes among its permanent residents moved toward those of permanent residents in proportion to the years of childhood they spent there.[3] The county-level companion paper estimates the causal contribution of each county and identifies affordable counties with strong outcomes.[4] The design identifies how much a move changes a child's outcome. It does not identify why a county sits above or below a census-based prediction.

Ganong and Shoag count the state supreme and appellate court cases per capita that mention land use, as a measure of land-use regulation, and relate the convergence of state incomes to it. To check that regulation is not a result of slower convergence, they split states by the same measure for 1956 to 1965, before the slowdown, and by land availability, a Bartik-type use of historical differences across places with a national change. Convergence slowed where regulation was tighter, and rising housing prices in high-income states deterred the migration of low-skill workers toward them.[5] The study covers states from 1940 to 2010. It does not test whether housing cost steers the county flows of 2022 and 2023.

What these data cannot show

None of the residuals measures what a place does for a child. A county residual combines noise in the Atlas estimates, measures the model omits and the gap between the decades of the outcome and of the predictors. The standard error of a county residual leaves out errors that all of a county's tracts share, and the floor of 8 tracts excludes 1,567 map units, so the map says nothing about them. The traits in Figure 3 describe counties selected on the residual, and their intervals treat the two groups as fixed.

The migration figures cover one pair of filing years. Internal migration in the United States has declined since the 1980s,[6] and arriving returns equal only 5.6% of the returns filed here, so most filers in every county did not change county. The two filings follow the pandemic years, in which demand for housing shifted from the centers of large cities toward their outskirts,[7] and one pair of years may not describe other periods. The unit is the county, so the correlations compare places and say nothing about why any household moved, and several of them change sign with the weighting. Connecticut's flows are published by planning region and are spread over its legacy counties by population, which is approximate. County home values come from the ZCTA atlas, whose ACS vintage is not stated upstream. Puerto Rico is left out of the model and the maps.

The two halves of the chapter do not measure the same people. The residual scores the adult outcomes of children raised in the 1980s and 1990s, and the migration data score the movements of filers in 2022 and 2023, so reading one against the other compares where two different populations sit on the same map. The adjusted gross income of a return is the amount on the 2023 return, for calendar 2022, and a move between the two filings can come after part or all of that income was earned.

Data sources

Opportunity AtlasAdult household income rank of children born 1978 to 1983 to parents at the 25th income percentile, 2020-tract file (the model's outcome) and county file (the migration comparison)Years: 1978-1983 birth cohorts (2018 release) · Weights: ACS 2019-2023 tract population; county population · License: Public use (Opportunity Insights and the U.S. Census Bureau)
American Community SurveyACS 2019 to 2023 tract measures (poverty, college, single-parent share, median household income, race and ethnicity, homeownership, population) and, from the ZCTA atlas, ZIP-area median home values (ACS-based; the vintage is not stated upstream)Years: 2019-2023 (5-year) · Weights: 2019-2023 tract population · License: Public domain (U.S. Census Bureau)
Census TIGER/Line tracts2020 tract land area, for residents per square mileYears: 2020 · Weights: none (land area) · License: Public domain (U.S. Census Bureau)
CDC PLACESCurrent smoking among adults, 2024 release, tractsYears: 2024 release (tracts); 2025 release (ZCTAs, 2023 data year) · Weights: 2019-2023 tract population · License: Public domain (CDC, U.S. Department of Health and Human Services)
EPA EJScreenShare of housing units built before 1960, 2024 release, tractsYears: 2024 (indicators); ACS 2018-2022 (demographics) · Weights: 2019-2023 tract population · License: Public domain (U.S. EPA); archived copy on Zenodo
USALEEP life expectancyLife expectancy at birth by tract, deaths in 2010 to 2015, area-weighted onto 2020 tracts, for the life-expectancy version of the model and county meansYears: 2010-2015 · Weights: 2019-2023 tract population · License: Public domain (NCHS, CDC)
Area Deprivation IndexArea Deprivation Index 2023, national rank, tracts averaged to countiesYears: 2023 release (version 4.0.1; built on ACS 2019-2023) · Weights: 2019-2023 tract population · License: Free for non-commercial use; cite the Neighborhood Atlas and Kind and Buckingham
IRS migration dataCounty-to-county migration, filing years 2022 and 2023: arriving and leaving returns and adjusted gross income, all origins and destinationsYears: 2022-2023 (filing years) · Weights: none (full counts of returns) · License: Public domain (IRS Statistics of Income)
IRS Statistics of IncomeIRS Statistics of Income county file, tax year 2022: returns filed, summed over the income brackets, the denominator of the net migration rateYears: tax year 2022 · Weights: none (full counts of returns) · License: Public domain (IRS Statistics of Income)

All data sources

Methods

A population-weighted line across 2020 census tracts predicts the Opportunity Atlas adult income rank of children born 1978 to 1983 from five ACS 2019-2023 measures, and a county's residual is the actual rank minus the prediction, averaged over its tracts and reported only for counties with at least 8 tracts in the model. The migration comparison uses IRS returns matched between filing years 2022 and 2023, so it describes returns that changed county, with moves from abroad and moves abroad in the county totals. County correlations are population-weighted and, where weighting changes the sign, also reported with each county counted once; intervals resample states, and the home-value measure comes from a ZIP-area source whose vintage is not stated upstream.

The tract model and its sample

A population-weighted least-squares line across 2020 census tracts regresses the Opportunity Atlas mean adult household income rank of children born 1978 to 1983 to parents at the 25th income percentile on five ACS 2019-2023 measures: the poverty rate, the share of adults with a bachelor's degree or more, the single-parent household share, the natural log of median household income and the Black share of residents. Tracts with 200 residents or fewer, tracts missing a variable and Puerto Rico are left out. The weight is the ACS tract population.

The life-expectancy version fits the same five measures to USALEEP life expectancy at birth (deaths in 2010 to 2015, institutional tracts left out) and correlates county mean residuals across counties, weighted by county population, with an interval that resamples states.

Figure explained here: Figure 1

County residuals, floors and groups

The tract residual is the actual rank minus the prediction. A county's residual is the population-weighted mean of its tracts' residuals, reported for counties with at least 8 tracts in the model; a floor of 5 tracts would map 2,140 counties and a floor of 15 would map 944. Its standard error treats the tracts as independent draws within the county.

Connecticut tracts are assigned to its eight legacy counties through the staged tract crosswalk.

Groups are tenths of counties with equal numbers of counties; their traits are population-weighted tract means, with intervals that resample counties and treat the groups as fixed.

Figures explained here: Figure 2 and Figure 3

Migration measures and county variables

The IRS matches individual returns filed in 2022 (tax year 2021) with returns filed in 2023 (tax year 2022) on the filer's taxpayer identification number, assigns each return a county from its ZIP+4, and counts a return as a migrant when the county differs; the county totals used here are the IRS rows for all origins and destinations, so moves from abroad are in the arrivals and moves abroad in the departures, and adjusted gross income is the amount on the 2023 return, converted from thousands of dollars. Net is arriving minus leaving, computed from both sides, and a county with one direction withheld has no net figure.

The net migration rate divides net returns by the returns filed for tax year 2022 (IRS SOI county file, summed over the income brackets) and multiplies by 1,000; net income per resident divides by the ACS 2019-2023 population. Connecticut's planning regions are allocated to the legacy counties in proportion to 2020 population.

County mobility is the Atlas county estimate, deprivation the population-weighted mean of tract ranks, life expectancy the population-weighted mean of non-institutional tracts, and home value the population-weighted mean of the ZIP-area median home values in the county, from the ZCTA atlas (ACS-based; the vintage is not stated upstream). Tenths of counties are cut on the counties in the analysis.

Figures explained here: Figure 4, Figure 7, Figure 5 and Figure 6

Weights and uncertainty

The R-squared interval and the interval for every statistic of tracts come from a bootstrap that resamples whole counties.

Correlations and fits across counties are weighted by ACS population and, where weighting changes the sign, also reported with each county counted once; intervals resample states. Bars and markers for tenths pool the counties of a group (summed numerator over summed denominator). Differences between groups use the same resampled counties for both groups.

Split-half check, replicates and seeds

The split-half check assigns each tract of a county to one of two halves at random (seed 12345), ranks counties on one half and reads the residual off the other, for counties with at least 3 tracts in each half; intervals resample states.

The dots of the prediction scatter are a random sample of tracts drawn with seed 12345. Bootstrap intervals use 2,000 replicates and seed 12345.

References

Chetty, Raj; Friedman, John N.; Hendren, Nathaniel; Jones, Maggie R.; Porter, Sonya R. (2018). The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility. NBER Working Paper 25147. doi:10.3386/w25147
Internal Revenue Service, Statistics of Income Division (2025). 2022-2023 Migration Data Users Guide. IRS Statistics of Income, County-to-County Migration Data. www.irs.gov/pub/irs-soi/2223inpublicmigdoc.pdf
Chetty, Raj; Hendren, Nathaniel (2018). The Impacts of Neighborhoods on Intergenerational Mobility I: Childhood Exposure Effects. The Quarterly Journal of Economics 133(3): 1107-1162. doi:10.1093/qje/qjy007
Chetty, Raj; Hendren, Nathaniel (2018). The Impacts of Neighborhoods on Intergenerational Mobility II: County-Level Estimates. The Quarterly Journal of Economics 133(3): 1163-1228. doi:10.1093/qje/qjy006
Ganong, Peter; Shoag, Daniel (2017). Why Has Regional Income Convergence in the U.S. Declined?. Journal of Urban Economics 102: 76-90. doi:10.1016/j.jue.2017.07.002
Molloy, Raven; Smith, Christopher L.; Wozniak, Abigail (2011). Internal Migration in the United States. Journal of Economic Perspectives 25(3): 173-196. doi:10.1257/jep.25.3.173
Ramani, Arjun; Bloom, Nicholas (2021). The Donut Effect of COVID-19 on Cities. NBER Working Paper 28876. doi:10.3386/w28876

How to cite

@misc{singh2026placesthatbeatthemodel,
  author = {Singh, Vishal},
  title = {Places that beat the model},
  year = {2026},
  url = {https://vishalsingh.org/zipcode/places-that-beat-the-model},
  note = {In ZIP Code Destiny. Published 9 October 2026.}
}

Reading paths

The argument in eight chapters