Chapter 6.1 · Two addresses

Place and the length of a life

Life expectancy at birth is lower in more deprived neighborhoods, 74.0 against 82.3 years in the most and least deprived tenths of census tracts for deaths in 2010 to 2015, and county averages conceal differences within counties.

Author: Vishal Singh
Affiliation: NYU Stern
Published: 9 October 2026
Updated: 11 October 2026

Life expectancy at birth was 82.3 years in the least deprived tenth of census tracts and 74.0 years in the most deprived, for deaths recorded in 2010 to 2015.[1] Both are population-weighted averages of tenths cut from the 77,936 non-institutional tracts in the states and DC that have a life expectancy estimate and a deprivation rank. The difference between them is 8.3 (95% CI 7.9 to 8.7) years, with an interval from resampling whole counties. This chapter asks how steadily life expectancy falls with deprivation, what a county average leaves out, how far apart nearby neighborhoods are, whether the ordering repeats inside metros and counties, and what an experiment adds about moving. The introduction places the comparison among the book's other outcomes and sources.

A life expectancy estimate summarizes the death rates of a place during a period. It is the average age at death of newborns who would meet, at every age, the death rates that the tract recorded in 2010 to 2015, and it does not predict the lifespan of a particular child. Deprivation is measured separately, by the Area Deprivation Index (ADI), a national percentile built from census measures of income, education, employment and housing.[2] The ADI describes 2019 to 2023, after the deaths behind the life expectancy estimates, so the opening comparison sets an earlier pattern of mortality beside a later description of neighborhood conditions.

The distance between the two averages calls for a closer look. Comparing only the ends of a scale leaves the middle unseen. Comparing distant places mixes differences in their residents with differences in their surroundings, and a detailed map can exaggerate a contrast when each neighborhood's estimate is uncertain. The sections below take up those problems in turn.

A national pattern needs a closer look

Arranging the tracts from least to most deprived and cutting the ranking into tenths gives a mean life expectancy that is lower in every tenth than in the one before (Figure 1). The means are weighted by tract population. The ordering holds along the whole scale, and the size of the step changes along the way. The step between the two least deprived tenths is 1.3 (95% CI 1.0 to 1.6) years and the step between the two most deprived is 1.9 (95% CI 1.6 to 2.1) years, and the steps are largest at the two ends. Between neighboring ADI percentiles the line sometimes rises by a fraction of a year. In the units of the index itself, life expectancy is 0.84 (95% CI 0.75 to 0.92) years lower for each ten-point step in the ADI percentile across all tracts.

Tracts with similar ranks still differ. A tract's estimate has a median standard error of 1.7 years, so part of the spread among tracts at one rank is estimation error, and part is differences that the deprivation rank does not summarize. The line describes an ordering of averages and leaves much about individual places unexplained.

Figure 1

Life expectancy is lower in each higher tenth of tracts by deprivation, and smoking rises with it while the annual checkup rises little

Tract life expectancy at birth by Area Deprivation Index percentile, and two modeled health measures by tenth of tracts

Life expectancy at birth falls from 82.3 years in the least deprived tenth of tracts to 74.0 years in the most deprived

Life expectancy at birth against the Area Deprivation Index national percentile, one dot per sampled tract, with the population-weighted mean and the interquartile range of tracts at each percentile, deaths in 2010 to 2015

Loading data

The cloud and the line slope down from the least deprived tracts on the left to the most deprived on the right; the line is lower in each tenth than in the one before, and between neighboring percentiles it sometimes rises by a fraction of a year. The line is the population-weighted mean of tract life expectancy at each ADI percentile, the band is the interquartile range of tracts at that percentile, which is spread and not an interval, and the markers are tenths of tracts with 95% intervals from a bootstrap over counties.

The first tab shows the gradient in life expectancy across 77,936 tracts, and the second compares smoking with the annual checkup, the measure that follows the gradient least. Chapter 5.1 follows the other measures.Source: USALEEP, 2010-2015; Area Deprivation Index, 2023; CDC PLACES 2024 release (BRFSS 2022); population-weighted, modeled estimates

The figure keeps these ideas apart. The line is the mean at each ADI percentile. The shaded band is the interquartile range of tract estimates, a spread of places that includes estimation error and spans several years at every percentile. The markers on the tenths carry intervals from a bootstrap over counties, which describe uncertainty about those averages. Averaging over thousands of tracts shrinks the error of a mean to a small interval, so a wide band of neighborhoods can surround a precisely estimated line, and the band is not the uncertainty of the line. A single dot in the first tab says little on its own.

County averages leave differences out

A county is a useful administrative unit, and it can contain neighborhoods with different incomes, housing and health. New York City makes the problem visible, because each of its five boroughs is also a county. Figure 2 shades the 2,074 non-institutional tracts of the city that have an estimate and an ADI rank. In every borough the 5th to 95th percentile of tract life expectancy spans at least 9.3 years, while the borough means lie within 4.2 years of one another.

The shades need care. The estimates describe deaths in 2010 to 2015 and were moved by area from the 2010 tract boundaries onto the 2020 outlines drawn here. The published standard errors, which understate the noise, put estimation error at 37% of the variance of these estimates, so the map tends to exaggerate the contrast between neighbors. It gives a view inside the borough averages, and the edge between two shades is not a precise boundary in residents' prospects. A gray tract has no estimate or was excluded, and it is not a zero.

Figure 2

Within every New York City borough, the 5th to 95th percentile of tract life expectancy spans at least 9.3 years, while the borough means lie within 4.2 years of one another

Life expectancy at birth of 2,074 census tracts in the five boroughs of New York City, deaths in 2010 to 2015, on 2020 tracts; 250 tracts are gray because they have no value

Loading tract outlines

Life expectancy at birth (years), deaths in 2010 to 2015

Lower life expectancyHigher life expectancy

No data (0 tracts)

Select a class to show only those places.

Find a tract

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

Tracts in several classes of the key appear inside every borough: the tracts of each span at least 9.3 years between the 5th and 95th percentile, 6.9 years net of estimation error, and the five borough means lie within 4.2 years of one another. Each tract is shaded by a modeled estimate with a median standard error of 2.0 years, so part of every difference between shades is estimation error; the tooltip gives the 95% interval, the standard error and the ADI national percentile. Gray tracts have no value, and the tooltip says why.Note: The unit is the 2020 census tract. Life expectancy is the USALEEP estimate for deaths in 2010 to 2015, which the National Center for Health Statistics built on 2010 tracts and which is area-weighted onto 2020 tracts here, so the estimate for a 2020 tract blends the 2010 tracts that overlap it and its standard error is mapping grade, which understates the error. Every shade is a modeled estimate with a standard error (the median is 2.0 years), and the interval in the tooltip is a normal-approximation 95% interval built from the standard error.Note: Taking the published standard errors at face value, estimation error makes up 37% of the variance of these estimates. If the errors of neighboring tracts are independent, the difference between two neighbors carries the sum of their error variances, so a map of noisy estimates exaggerates the contrast between neighbors. That is why the pairs figure chooses its pairs on the ADI, which shares no sampling error with life expectancy. This map marks no tract.Note: 2,074 tracts are shaded: the non-institutional tracts of the five counties of New York City (Bronx, Kings, New York, Queens and Richmond) that have a life expectancy estimate, an ACS 2019-2023 population above zero and an ADI rank, which is the sample of the gradient. 250 are gray because the tract has no residents in the survey, no published estimate, an institutional population or no ADI rank; the tooltip names which.Note: Ranges are population-weighted 5th and 95th percentiles of tract estimates, and borough means are population-weighted. A range of tract estimates and a spread of borough means are not the same kind of contrast, and the range includes estimation error. The range net of error shrinks each tract estimate toward its borough mean by the square root of one minus the borough's share of variance that is estimation error, as in the county spread figure, and is an upper bound because the 2020-tract errors are mapping grade.Source: USALEEP life expectancy at birth, 2010-2015, area-weighted onto 2020 tracts; Area Deprivation Index, 2023; ACS 2019-2023 population; population-weighted ranges; modeled estimates

A comparison of nearby neighborhoods needs a selection rule as well. Tract life expectancy is a modeled estimate with a standard error, and picking the pairs with the largest apparent gaps would also pick the largest estimation errors. The pairs in Figure 3 are chosen instead by their difference in deprivation, in the 20 most populous counties only, subject to requirements on distance, population, boundary stability and precision. The rule never compares the life expectancy of the tracts in a pair; it uses the estimate only to require one with a small enough standard error. Each pair uses the native estimates for deaths in 2010 to 2015 on the 2010 tract boundaries and the ADI for 2019 to 2023, so the deprivation measures date from about a decade after the deaths.

Figure 3

In 17 large counties, the pair of tracts within 5 kilometers with the widest ADI difference had higher life expectancy in the less deprived tract in 15, and a median gap over the pairs of 5.4 years

Life expectancy at birth in the two tracts of one pair per county, the pair with the widest ADI difference, in 17 of the 20 most populous counties, native 2010 tracts, deaths in 2010 to 2015, 95% intervals

More deprived tract
Less deprived tract

Show data

Data for In 17 large counties, the pair of tracts within 5 kilometers with the widest ADI difference had higher life expectancy in the less deprived tract in 15, and a median gap over the pairs of 5.4 years
County and tracts	Life expectancy, more deprived tract (years)	Life expectancy, less deprived tract (years)	Gap (years)	ADI, more deprived tract	ADI, less deprived tract	Distance (kilometers)
Cook County, IL (tracts 4910 and 7201)	65.8	78.2	12.4	94	24	5.0
Harris County, TX (tracts 4222 and 4207)	74.7	80.9	6.2	100	14	4.8
Maricopa County, AZ (tracts 4204.01 and 4204.02)	73.2	80.4	7.2	94	14	4.7
San Diego County, CA (tracts 202.09 and 207.06)	79.2	78.5	−0.7	69	5	4.4
Orange County, CA (tracts 748.02 and 751)	76.2	82.3	6.1	61	10	2.2
Miami-Dade County, FL (tracts 1.24 and 11.04)	75.9	81.3	5.4	88	9	4.5
Kings County, NY (tracts 596 and 422)	78.5	81.0	2.5	67	1	3.1
Riverside County, CA (tracts 422.09 and 306.03)	75.8	80.1	4.3	90	11	4.7
Dallas County, TX (tracts 136.21 and 136.11)	77.6	80.6	3.0	72	6	2.0
Queens County, NY (tracts 240 and 1333)	76.8	83.7	6.9	66	5	3.9
King County, WA (tracts 308.01 and 299.02)	72.8	79.4	6.6	89	14	4.1
San Bernardino County, CA (tracts 87.10 and 87.08)	74.2	79.1	4.9	98	24	2.1
Clark County, NV (tracts 29.65 and 29.47)	79.1	79.6	0.5	86	18	3.8
Tarrant County, TX (tracts 1025 and 1054.05)	69.1	79.1	10.0	82	6	2.4
Broward County, FL (tracts 104.02 and 105.02)	79.9	79.9	0.0	92	20	3.1
Bexar County, TX (tracts 1702 and 1902)	76.7	81.0	4.3	96	30	4.6
Wayne County, MI (tracts 5015 and 5506)	70.9	78.3	7.4	94	10	4.0

In most rows the filled dot lies right of the open dot, and in 12 rows the 95% interval for the gap excludes zero. Each row is the pair of tracts in a county, no more than 5 kilometers apart, whose ADI differs most, among tracts with at least 2,000 residents and a standard error of at most 2.0 years. The open dot is the more deprived tract, the filled dot the less deprived, and the last column is the difference.Note: The two tracts of a pair are chosen on the ADI, not on life expectancy, so the differences are not the largest that could be found among noisy estimates. A county with no pair that reaches an ADI difference of 50 percentile points is left out (3 of the 20 counties).Note: Life expectancy is the native estimate for the 2010 tract, used only where the boundary did not change by 2020, meaning that the 2010 and 2020 tracts with one GEOID share at least 99% of the land area of each, so the ADI and the ACS population of the 2020 tract describe the same ground. Distances are straight lines between the Census internal points of the tracts, and the data hold no transit routes. Intervals are normal-approximation 95% intervals from the standard error; the gap treats the two errors as independent. Institutional tracts are excluded.Source: USALEEP life expectancy at birth on 2010 tracts, 2010-2015, with standard errors; Area Deprivation Index, 2023; ACS 2019-2023; modeled estimates

The pairs are selected by deprivation differences rather than by their life expectancy gaps, so the selection rule does not directly choose the largest outcome estimates.

How the neighborhoods were paired

A tract qualifies if its boundary did not change between the 2010 and 2020 censuses, it is not an institutional tract, it has at least 2,000 residents and the standard error of its estimate is at most 2.0 years. Among qualifying tracts within 5 kilometers of each other, a county's pair is the one whose ADI differs most, and it counts if the difference reaches 50 percentile points.

Full procedure in Methods

In 15 of the 17 pairs the less deprived tract has the higher estimate, and in 12 the 95% interval for the gap lies above zero. The median gap over all 17 pairs, including those whose gap is not positive, is 5.4 (95% CI 3.0 to 6.9) years, with an interval from a bootstrap over pairs. The gap treats the two standard errors as independent. This is a result about the pairs that meet the rule in the large counties searched. It does not estimate the gap between a typical pair of neighbors in the country, and it depends on the settings of the rule, above all the radius. Pairs chosen within 3 kilometers have a median gap of 2.7 years (see alternative choices in Methods for the other settings).

Showing the complete set keeps the widest contrast from standing in for every comparison. The widest pair is in Cook County, Illinois, and its tracts lie 5.0 kilometers apart. The more deprived one (ADI 94) had a life expectancy of 65.8 (95% CI 62.5 to 69.1) years and the less deprived one (ADI 24) 78.2 (95% CI 74.7 to 81.7) years, a gap of 12.4 (95% CI 7.6 to 17.2) years. The widest pair of a set is also the one most likely to have been widened by estimation error.

Wider gaps may exist among tracts the rule does not pair, because the ADI ranks some tracts alike that differ widely in life expectancy, as it does in high-cost cities (Chapter 10.2, The anatomy of a deprivation index), and the largest gaps among noisy estimates are the least reliable, so the rule does not seek them.

Most variation lies between tracts in the same county

Splitting the variance of tract life expectancy into a part between states, a part between counties in a state and a part between tracts in a county gives 13.7%, 15.9% and 70.3% (Figure 4). The split comes from every one of 78,008 tracts in 3,112 counties and Connecticut planning regions and has no sampling interval.

The within-county share includes estimation error. The tract estimates are modeled, and their mean squared standard error equals 26.7% of the total variance, all of it inside the within-county share. Taking the published errors at face value, 60% of the variance that is not estimation error lies within counties. Because the 2020-tract errors are mapping grade and understate the noise, the true share is lower.

The ADI is the contrast. Only 34% of its variance lies within counties, so the index that ranks tracts for the gradient places most of its variation between counties, while life expectancy places most of its variation inside them. Chapter 10.2, The anatomy of a deprivation index examines what the ADI is made of.

Figure 4

70.3 percent of the variance of tract life expectancy lies between tracts in the same county, 26.7 points of it estimation error, against 34% percent for the ADI

Share of the population-weighted variance of tract life expectancy and of the ADI national percentile that lies between states, between counties in a state and between tracts in a county, percent

Between states
Between counties in a state
Between tracts in a county
Estimation error within counties

Show data

Data for 70.3 percent of the variance of tract life expectancy lies between tracts in the same county, 26.7 points of it estimation error, against 34% percent for the ADI
level	measure	share
Between states	Tract life expectancy	13.75
Between counties in a state	Tract life expectancy	15.92
Between tracts in a county	Tract life expectancy	43.62
Estimation error within counties	Tract life expectancy	26.71
Between states	ADI national percentile	36.64
Between counties in a state	ADI national percentile	29.38
Between tracts in a county	ADI national percentile	33.98

The life expectancy bar is mostly the two right-hand segments, which together are the share within counties, and the ADI bar is mostly the other two. Each bar sums to the whole variance of its measure; the shares come from every tract in the sample and have no sampling interval.Note: Tract life expectancy is modeled. The mean squared standard error of the tract estimates equals 26.7 percent of the total variance and sits inside the within-county share, as the two right-hand segments show; the 2020-tract standard errors are mapping grade and understate the noise, so the share of the true differences that lies between tracts in a county is lower than the net share.Note: The life expectancy bar uses 78,008 tracts in 3,112 counties and Connecticut planning regions; the ADI bar uses the tracts of chapter 10.1's index. Connecticut's planning regions stand in for its counties.Source: USALEEP life expectancy, 2010-2015; ACS 2019-2023; Area Deprivation Index, 2023; population-weighted decomposition of the variance, full counts of tracts

Figure 5 shows the range inside the 30 most populous counties. The 5th to 95th percentile range of tract life expectancy has a median of 10.5 years across them, and the widest, 13.9 years in Cook County, is wider than the 8.3-year difference between the means of the nation's least and most deprived tenths, though a range of tract estimates and a difference between means are not the same kind of contrast.

Part of every range is estimation error. Shrinking each tract estimate toward its county mean approximates the range net of that error and gives a median range of 7.7 years. The approximation takes the published errors at face value and does not recover the true differences between tracts exactly, and because those errors understate the noise, the shrunken range is an upper bound. The median share of the variance within a county that is estimation error is 43% across these counties, against 38% for all counties together, and each estimate is shrunk by the square root of one minus its county's share.

Figure 5

In the 30 most populous counties, tract life expectancy spans a median of 10.5 years between the 5th and 95th percentile, 7.7 net of estimation error

Population-weighted mean life expectancy at birth of each county (dot) and the 5th to 95th percentile of its tracts (whisker), sorted by the width of the range, deaths in 2010 to 2015

Life expectancy at birth (years; dot: county mean; whisker: 5th to 95th percentile of tracts)

Show data

Data for In the 30 most populous counties, tract life expectancy spans a median of 10.5 years between the 5th and 95th percentile, 7.7 net of estimation error
County	Mean life expectancy (years)	5th percentile of tracts	95th percentile of tracts	Range (years)	Range net of estimation error (years)	Estimation error, share of variance (%)	Tracts	Residents
Cook County, IL	78.5	70.7	84.6	13.9	12.1	24	1,221	4,864,100
Wayne County, MI	75.4	68.9	82.2	13.3	11.7	23	555	1,707,514
New York County, NY	82.4	75.6	88.6	13.0	10.8	31	250	1,405,675
Palm Beach County, FL	80.6	74.0	86.9	12.9	10.8	30	330	1,399,959
Philadelphia County, PA	75.4	69.0	81.5	12.5	10.9	25	364	1,502,783
Hillsborough County, FL	78.2	71.4	83.4	12.0	9.7	35	287	1,317,655
Miami-Dade County, FL	80.5	74.3	86.1	11.8	9.9	30	655	2,544,531
Kings County, NY	79.9	73.8	85.3	11.6	8.8	42	734	2,521,198
Alameda County, CA	81.0	75.3	86.5	11.2	8.6	41	362	1,598,446
Broward County, FL	79.5	73.6	84.7	11.1	8.7	39	392	1,847,256
Dallas County, TX	77.8	72.0	82.8	10.9	8.4	40	561	2,324,374
King County, WA	81.3	76.1	86.9	10.8	8.1	44	475	2,184,985
Harris County, TX	78.3	72.7	83.5	10.8	8.1	44	1,000	4,292,763
Queens County, NY	81.7	76.6	87.1	10.5	7.3	52	656	2,250,510
Orange County, FL	79.3	73.6	84.1	10.5	8.0	42	248	1,339,363
Los Angeles County, CA	80.6	75.4	85.9	10.5	7.8	45	2,350	9,461,723
Clark County, NV	77.7	72.2	82.5	10.3	7.2	52	475	2,040,360
Bronx County, NY	78.3	73.5	83.9	10.3	7.4	49	326	1,369,766
Riverside County, CA	79.4	74.1	84.3	10.2	7.6	45	492	2,352,240
San Diego County, CA	80.9	75.9	85.8	9.9	7.2	47	681	3,038,849
Middlesex County, MA	81.8	77.4	87.1	9.7	7.6	40	338	1,544,211
San Bernardino County, CA	78.4	73.5	83.2	9.7	7.1	47	447	2,093,159
Sacramento County, CA	78.8	74.2	83.9	9.7	7.6	39	343	1,515,289
Bexar County, TX	78.2	73.1	82.8	9.7	7.3	43	334	1,812,879
Maricopa County, AZ	78.8	73.9	83.5	9.6	6.8	50	865	3,980,946
Tarrant County, TX	78.0	73.0	82.4	9.4	6.5	52	402	1,931,602
Orange County, CA	81.7	77.6	86.1	8.5	4.9	67	575	2,964,928
Nassau County, NY	82.4	78.6	87.0	8.4	6.3	43	272	1,358,476
Santa Clara County, CA	82.8	78.8	87.1	8.3	4.2	75	393	1,835,064
Suffolk County, NY	80.7	77.0	84.5	7.5	5.0	55	359	1,400,449

Every whisker spans several years, in counties with high means and low means alike. The dot is the county mean and the whisker is the spread of its tracts, which describes the tracts and is not an interval for the mean. The second view shrinks the tract estimates toward the county mean by the share of their variance that is estimation error.Note: Tract estimates are modeled, so the range of the estimates is wider than the range of the true values. The second view takes the published standard errors at face value: the median share of the variance within a county that is estimation error is 43%, against 38% for all counties together. The shrinkage multiplies the distance of the 5th and 95th percentiles from the county mean by the square root of one minus that share. The 2020-tract errors are mapping grade, so the range net of error is an upper bound. The range of tract estimates is not like-for-like with a difference between means of tenths.Note: Counties are chosen by population and not by range. With 15 counties the median range is 10.5 years and with 60 it is 10.5. Institutional tracts are dropped. Connecticut's planning regions stand in for its counties.Source: USALEEP life expectancy, 2010-2015; ACS 2019-2023 population; population-weighted percentiles of tract values, modeled estimates

A map that paints each county one color shows the county mean and hides the range around it. Chapter 10.3, A deprivation index at two zooms builds the deprivation index at two zooms to see how much a county-level measure keeps.

Repeating the comparison changes the question

The national line compares neighborhoods across the country, and it could arise from where deprived tracts are, with the least deprived concentrated in some regions and the most deprived in others. A different comparison asks whether the ordering also appears among neighborhoods of the same area.

Fitting the line separately in each metro, meaning a core-based statistical area of the Census Bureau, tests that. The 267 metros that have at least 40 tracts with an estimate and an ADI rank hold 82% of the residents in the sample. All of them show lower life expectancy in more deprived tracts (Figure 6), and the 95% interval lies above zero in 260 of them. The median metro has life expectancy 1.20 (95% CI 1.14 to 1.23) years lower per ten-point step. At the 5th percentile of metros the difference is 0.70 years and at the 95th 1.67.

Those percentiles overstate the real spread of slopes, and a ranking of local slopes would put the noisiest estimates at the extremes. Estimation error makes up 48% of the variance of the estimated slopes, and the order of the dots in the figure puts the least precise slopes at both ends, so no metro is named steepest or flattest. Chapter 10.1, Gradients, slopes and variance shrinks the estimates and reports the same metro slopes.

Figure 6

Life expectancy is lower in more deprived tracts in all 267 metros with enough tracts, by a median of 1.20 years per ten-point step

One regression of tract life expectancy on the ADI national percentile in each metro with enough tracts, shown by metro and as a distribution, deaths in 2010 to 2015

Life expectancy is lower in more deprived tracts in every metro, from 0.70 years per ten-point step at the 5th percentile of metros to 1.67 at the 95th

Slope of tract life expectancy at birth on the ADI national percentile within each metro, years lower per ten-point step, one dot per metro labeled by its most populous county, deaths in 2010 to 2015

HC3 errors
Clustered on counties

Every dot lies right of zero. Sorting puts the metros with the noisiest slopes at both ends, so the order of neighbors carries no information. Each dot is a metro slope with a 95% interval, and the key names the kind of error: clustered on counties, or HC3 where the metro has fewer than five counties.

Show data

Data for Life expectancy is lower in more deprived tracts in all 267 metros with enough tracts, by a median of 1.20 years per ten-point step, Slope by metro
conf_high	conf_low	errors	estimate	metro	n
1.46	1.09	HC3 errors	1.27	Summit, OH	167
1.73	0.711	HC3 errors	1.22	Dougherty, GA	43
1.55	0.922	Clustered on counties	1.24	Albany, NY	232
1.17	0.770	HC3 errors	0.968	Bernalillo, NM	223
1.90	0.894	HC3 errors	1.39	Rapides, LA	42
1.13	0.802	HC3 errors	0.965	Lehigh, PA	188
1.30	0.637	Clustered on counties	0.970	Randall, TX	64
1.84	1.07	HC3 errors	1.45	Anchorage, AK	79
1.71	1.15	HC3 errors	1.43	Washtenaw, MI	86
1.63	0.507	HC3 errors	1.07	Outagamie, WI	52
0.934	0.363	HC3 errors	0.649	Buncombe, NC	112
1.59	0.881	HC3 errors	1.23	Clarke, GA	53
1.19	0.949	Clustered on counties	1.07	Fulton, GA	1,354
1.18	0.542	HC3 errors	0.859	Atlantic, NJ	69
1.06	−0.236	HC3 errors	0.411	Lee, AL	43
1.26	0.917	Clustered on counties	1.09	Richmond, GA	134
0.846	0.511	Clustered on counties	0.679	Travis, TX	442
1.16	0.751	HC3 errors	0.957	Kern, CA	212
1.67	1.14	Clustered on counties	1.40	Baltimore, MD	670
2.99	0.846	HC3 errors	1.92	Barnstable, MA	52
1.87	0.796	Clustered on counties	1.33	East Baton Rouge, LA	192
1.49	0.976	HC3 errors	1.23	Jefferson, TX	105
1.60	0.018	HC3 errors	0.807	Whatcom, WA	50
2.18	0.665	HC3 errors	1.42	Deschutes, OR	42
2.00	0.799	HC3 errors	1.40	Yellowstone, MT	42
1.88	0.700	HC3 errors	1.29	Broome, NY	64
1.45	1.17	Clustered on counties	1.31	Jefferson, AL	289
1.63	0.624	HC3 errors	1.13	McLean, IL	40
1.17	0.843	Clustered on counties	1.00	Ada, ID	190
1.80	0.798	Clustered on counties	1.30	Middlesex, MA	1,045
1.98	0.652	HC3 errors	1.31	Boulder, CO	64
2.98	1.76	HC3 errors	2.37	Kitsap, WA	53
1.28	0.983	HC3 errors	1.13	Western Connecticut, CT	212
1.73	1.00	HC3 errors	1.36	Cameron, TX	118
1.49	1.17	HC3 errors	1.33	Erie, NY	307
1.76	0.633	HC3 errors	1.20	Chittenden, VT	50
1.84	1.31	HC3 errors	1.58	Stark, OH	95
1.02	0.583	HC3 errors	0.799	Lee, FL	190
1.63	0.653	HC3 errors	1.14	Linn, IA	65
2.34	1.40	HC3 errors	1.87	Champaign, IL	43
1.29	0.654	Clustered on counties	0.972	Kanawha, WV	76
1.06	0.708	HC3 errors	0.884	Charleston, SC	159
1.28	0.986	Clustered on counties	1.13	Mecklenburg, NC	599
1.76	0.435	Clustered on counties	1.10	Albemarle, VA	51
1.60	1.01	Clustered on counties	1.31	Hamilton, TN	130
1.23	1.06	Clustered on counties	1.15	Cook, IL	2,183
2.11	1.14	HC3 errors	1.63	Butte, CA	51
1.69	1.45	Clustered on counties	1.57	Hamilton, OH	522
1.22	0.711	HC3 errors	0.964	Montgomery, TN	52
1.46	1.19	Clustered on counties	1.33	Cuyahoga, OH	589
1.16	0.549	HC3 errors	0.854	Brazos, TX	60
1.81	1.09	HC3 errors	1.45	El Paso, CO	155
1.46	0.677	HC3 errors	1.07	Boone, MO	40
1.33	0.942	Clustered on counties	1.14	Richland, SC	205
1.53	1.34	Clustered on counties	1.43	Muscogee, GA	87
1.74	0.972	Clustered on counties	1.36	Franklin, OH	474
1.18	0.518	HC3 errors	0.848	Nueces, TX	103
1.33	0.752	HC3 errors	1.04	Okaloosa, FL	63
0.986	0.834	Clustered on counties	0.910	Dallas, TX	1,501
1.36	0.726	HC3 errors	1.04	Baldwin, AL	41
1.51	0.487	HC3 errors	0.996	Scott, IA	107
1.87	1.45	HC3 errors	1.66	Montgomery, OH	209
1.57	0.953	HC3 errors	1.26	Volusia, FL	143
1.43	0.939	Clustered on counties	1.19	Denver, CO	635
1.38	0.681	Clustered on counties	1.03	Polk, IA	154
1.40	1.20	Clustered on counties	1.30	Wayne, MI	1,238
1.51	0.670	HC3 errors	1.09	St. Louis, MN	93
1.75	0.308	Clustered on counties	1.03	Durham, NC	131
1.79	−0.045	HC3 errors	0.871	Monroe, PA	46
2.17	1.36	HC3 errors	1.77	Elkhart, IN	44
0.851	0.416	HC3 errors	0.633	El Paso, TX	175
2.08	0.935	HC3 errors	1.50	Erie, PA	65
1.91	0.820	HC3 errors	1.36	Lane, OR	86
2.13	1.46	HC3 errors	1.80	Vanderburgh, IN	88
2.06	1.28	HC3 errors	1.67	Cass, ND	58
1.65	0.765	HC3 errors	1.21	Cumberland, NC	107
0.985	0.401	HC3 errors	0.693	Benton, AR	109
1.57	1.07	HC3 errors	1.32	Genesee, MI	124
1.62	0.646	HC3 errors	1.13	Florence, SC	52
2.02	0.669	HC3 errors	1.34	Lauderdale, AL	40
2.56	−0.146	HC3 errors	1.21	Larimer, CO	76
1.79	0.738	HC3 errors	1.27	Sebastian, AR	63
1.96	1.20	HC3 errors	1.58	Allen, IN	100
1.26	0.827	HC3 errors	1.04	Fresno, CA	218
1.77	1.21	HC3 errors	1.49	Alachua, FL	60
1.19	0.487	HC3 errors	0.840	Hall, GA	48
1.26	0.142	HC3 errors	0.701	Warren, NY	42
1.24	0.877	HC3 errors	1.06	Kent, MI	228
1.53	0.618	HC3 errors	1.07	Weld, CO	69
1.57	0.572	HC3 errors	1.07	Brown, WI	69
1.64	1.20	HC3 errors	1.42	Guilford, NC	169
1.36	0.988	HC3 errors	1.18	Greenville, SC	221
1.26	0.390	HC3 errors	0.827	Harrison, MS	137
1.98	1.13	HC3 errors	1.56	Washington, MD	65
1.55	0.981	HC3 errors	1.27	Dauphin, PA	126
1.29	0.997	HC3 errors	1.15	Capitol, CT	272
1.66	0.607	HC3 errors	1.13	Forrest, MS	45
1.36	0.395	HC3 errors	0.879	Catawba, NC	78
1.52	0.392	HC3 errors	0.955	Hawaii, HI	43
1.36	0.503	HC3 errors	0.930	Beaufort, SC	42
1.51	0.360	HC3 errors	0.936	Citrus, FL	41
1.13	0.623	HC3 errors	0.874	Terrebonne, LA	59
0.892	0.758	Clustered on counties	0.825	Harris, TX	1,457
2.12	0.348	Clustered on counties	1.24	Cabell, WV	102
1.53	1.08	HC3 errors	1.30	Madison, AL	108
1.48	1.18	Clustered on counties	1.33	Marion, IN	453
1.59	0.656	HC3 errors	1.12	Jackson, MI	45
1.50	1.00	Clustered on counties	1.25	Hinds, MS	149
1.83	0.819	HC3 errors	1.32	Madison, TN	50
1.53	1.23	Clustered on counties	1.38	Duval, FL	319
1.47	0.562	HC3 errors	1.02	Washington, TN	53
2.26	0.816	HC3 errors	1.54	Jasper, MO	42
2.53	0.560	HC3 errors	1.54	Maui, HI	42
1.55	0.897	HC3 errors	1.22	Kalamazoo, MI	63
1.50	1.05	Clustered on counties	1.27	Jackson, MO	556
1.21	0.656	HC3 errors	0.934	Benton, WA	63
1.00	0.383	HC3 errors	0.694	Bell, TX	91
2.04	0.991	Clustered on counties	1.51	Sullivan, TN	77
1.88	0.315	HC3 errors	1.10	Ulster, NY	41
1.59	1.29	Clustered on counties	1.44	Knox, TN	198
1.44	0.826	Clustered on counties	1.13	Lafayette, LA	125
1.98	0.909	HC3 errors	1.44	Tippecanoe, IN	49
1.48	0.741	HC3 errors	1.11	Calcasieu, LA	52
1.29	0.490	HC3 errors	0.892	Mohave, AZ	55
1.51	0.881	HC3 errors	1.20	Polk, FL	162
1.57	0.750	HC3 errors	1.16	Lancaster, PA	112
1.33	0.912	HC3 errors	1.12	Ingham, MI	135
1.34	0.641	HC3 errors	0.991	Webb, TX	61
0.970	−0.013	HC3 errors	0.478	Doña Ana, NM	57
1.03	0.735	HC3 errors	0.885	Clark, NV	475
1.53	0.543	HC3 errors	1.03	Grafton, NH	66
1.84	0.989	Clustered on counties	1.41	Fayette, KY	124
1.65	0.994	HC3 errors	1.32	Lancaster, NE	74
1.82	0.949	Clustered on counties	1.38	Pulaski, AR	177
1.98	0.429	HC3 errors	1.20	Laurel, KY	43
1.02	0.470	HC3 errors	0.744	Gregg, TX	64
1.45	1.15	HC3 errors	1.30	Los Angeles, CA	2,923
1.73	1.29	Clustered on counties	1.51	Jefferson, KY	320
1.25	0.097	HC3 errors	0.675	Lubbock, TX	97
1.90	0.911	Clustered on counties	1.40	Bedford, VA	64
1.68	1.12	Clustered on counties	1.40	Bibb, GA	64
1.15	0.419	HC3 errors	0.782	Dane, WI	123
1.55	0.681	HC3 errors	1.12	Hillsborough, NH	99
0.888	0.374	HC3 errors	0.631	Hidalgo, TX	208
1.83	0.570	HC3 errors	1.20	Jackson, OR	51
1.45	1.12	Clustered on counties	1.29	Shelby, TN	316
1.32	0.358	HC3 errors	0.839	Merced, CA	60
0.922	0.725	HC3 errors	0.823	Miami-Dade, FL	1,377
1.32	1.08	HC3 errors	1.20	Milwaukee, WI	374
1.52	0.938	Clustered on counties	1.23	Hennepin, MN	847
1.61	1.02	HC3 errors	1.31	Mobile, AL	128
1.35	0.694	HC3 errors	1.02	Stanislaus, CA	111
1.93	1.04	HC3 errors	1.49	Ouachita, LA	57
2.14	1.11	HC3 errors	1.62	Monroe, MI	40
1.49	0.964	HC3 errors	1.23	Montgomery, AL	94
1.98	0.493	HC3 errors	1.24	Skagit, WA	41
1.92	1.19	HC3 errors	1.56	Muskegon, MI	44
1.15	0.593	HC3 errors	0.872	Horry, SC	110
1.38	0.423	HC3 errors	0.902	Collier, FL	94
1.55	0.985	Clustered on counties	1.27	Davidson, TN	384
1.31	0.818	HC3 errors	1.06	South Central Connecticut, CT	128
1.38	0.948	Clustered on counties	1.16	Jefferson, LA	379
1.12	0.421	Clustered on counties	0.771	Kings, NY	4,563
1.92	1.12	HC3 errors	1.52	Berrien, MI	51
1.19	0.714	HC3 errors	0.953	Sarasota, FL	204
1.66	0.818	HC3 errors	1.24	Southeastern Connecticut, CT	62
1.27	0.541	HC3 errors	0.906	Marion, FL	77
1.96	1.32	HC3 errors	1.64	Davis, UT	134
1.49	1.11	Clustered on counties	1.30	Oklahoma, OK	380
2.18	1.21	HC3 errors	1.69	Thurston, WA	60
1.31	0.980	Clustered on counties	1.14	Douglas, NE	247
1.03	0.689	HC3 errors	0.859	Orange, FL	460
1.49	0.593	HC3 errors	1.04	Ventura, CA	182
1.47	1.01	HC3 errors	1.24	Brevard, FL	140
1.23	0.379	HC3 errors	0.803	Bay, FL	46
1.73	1.17	HC3 errors	1.45	Escambia, FL	106
1.41	1.07	Clustered on counties	1.24	Peoria, IL	104
1.27	1.13	Clustered on counties	1.20	Philadelphia, PA	1,455
0.989	0.800	HC3 errors	0.894	Maricopa, AZ	948
1.21	0.893	Clustered on counties	1.05	Allegheny, PA	674
1.45	0.677	HC3 errors	1.06	Cumberland, ME	127
1.51	1.21	Clustered on counties	1.36	Multnomah, OR	537
1.67	1.00	HC3 errors	1.34	St. Lucie, FL	102
2.22	1.04	HC3 errors	1.63	Schuylkill, PA	40
1.02	0.313	HC3 errors	0.666	Orange, NY	162
1.26	0.223	HC3 errors	0.740	Yavapai, AZ	70
1.46	1.07	Clustered on counties	1.26	Providence, RI	357
1.53	0.768	HC3 errors	1.15	Utah, UT	140
1.79	1.14	HC3 errors	1.47	Pueblo, CO	55
1.58	0.643	HC3 errors	1.11	Charlotte, FL	40
1.42	0.528	HC3 errors	0.976	Racine, WI	41
1.04	0.745	HC3 errors	0.890	Wake, NC	268
1.39	0.967	HC3 errors	1.18	Berks, PA	103
1.73	0.732	HC3 errors	1.23	Shasta, CA	49
1.67	0.812	HC3 errors	1.24	Washoe, NV	117
1.36	1.12	Clustered on counties	1.24	Chesterfield, VA	298
1.15	0.890	HC3 errors	1.02	Riverside, CA	939
1.82	0.799	Clustered on counties	1.31	Roanoke, VA	75
1.41	0.373	HC3 errors	0.892	Olmsted, MN	52
1.39	1.05	Clustered on counties	1.22	Monroe, NY	288
1.91	1.01	HC3 errors	1.46	Winnebago, IL	88
1.61	1.27	HC3 errors	1.44	Sacramento, CA	519
2.09	1.35	HC3 errors	1.72	Saginaw, MI	59
1.86	0.503	HC3 errors	1.18	Stearns, MN	43
1.43	1.07	Clustered on counties	1.25	St. Louis, MO	669
2.15	0.667	HC3 errors	1.41	Marion, OR	78
1.58	0.320	HC3 errors	0.950	Monterey, CA	96
1.21	0.759	HC3 errors	0.986	Sussex, DE	105
1.51	0.894	HC3 errors	1.20	Salt Lake, UT	247
1.05	0.647	Clustered on counties	0.849	Bexar, TX	453
1.34	0.826	HC3 errors	1.08	San Diego, CA	681
3.04	1.67	Clustered on counties	2.36	Alameda, CA	1,044
1.09	0.161	HC3 errors	0.624	Santa Clara, CA	405
1.48	−0.040	HC3 errors	0.721	San Luis Obispo, CA	61
1.73	−0.548	HC3 errors	0.592	Santa Cruz, CA	63
1.03	0.374	HC3 errors	0.703	Santa Fe, NM	52
1.64	0.787	HC3 errors	1.22	Santa Barbara, CA	98
2.12	0.547	HC3 errors	1.34	Sonoma, CA	116
1.28	0.806	HC3 errors	1.04	Chatham, GA	101
1.64	1.08	HC3 errors	1.36	Luzerne, PA	162
1.79	1.34	HC3 errors	1.56	King, WA	829
1.27	0.555	HC3 errors	0.914	Indian River, FL	41
1.51	1.01	HC3 errors	1.26	Caddo, LA	114
1.87	0.709	HC3 errors	1.29	Minnehaha, SD	60
1.51	1.05	HC3 errors	1.28	St. Joseph, IN	92
1.63	0.905	HC3 errors	1.27	Spartanburg, SC	85
1.84	1.26	HC3 errors	1.55	Spokane, WA	140
2.00	1.14	HC3 errors	1.57	Sangamon, IL	56
1.67	1.27	HC3 errors	1.47	Hampden, MA	151
2.08	0.848	Clustered on counties	1.46	Greene, MO	111
2.70	1.57	HC3 errors	2.13	Clark, OH	41
1.76	1.15	HC3 errors	1.46	San Joaquin, CA	169
1.43	0.879	HC3 errors	1.15	Onondaga, NY	183
1.22	0.803	HC3 errors	1.01	Leon, FL	82
1.10	0.888	HC3 errors	0.995	Hillsborough, FL	692
2.40	1.28	Clustered on counties	1.84	Vigo, IN	53
1.64	0.254	HC3 errors	0.949	Bowie, TX	42
1.42	1.05	HC3 errors	1.23	Lucas, OH	213
1.32	1.12	Clustered on counties	1.22	Shawnee, KS	59
1.42	0.785	HC3 errors	1.10	Grand Traverse, MI	43
1.58	1.11	HC3 errors	1.34	Mercer, NJ	79
1.30	0.873	HC3 errors	1.08	Pima, AZ	246
1.48	0.994	Clustered on counties	1.24	Tulsa, OK	304
1.84	0.561	HC3 errors	1.20	Lee, MS	40
1.28	0.561	HC3 errors	0.920	Tuscaloosa, AL	69
1.45	0.295	HC3 errors	0.875	Smith, TX	53
1.57	0.764	HC3 errors	1.16	Honolulu, HI	286
2.16	1.08	HC3 errors	1.62	Oneida, NY	83
2.86	1.47	HC3 errors	2.16	Solano, CA	96
1.68	1.04	Clustered on counties	1.36	Virginia Beach, VA	436
1.44	0.731	HC3 errors	1.08	Tulare, CA	101
1.71	0.963	HC3 errors	1.34	McLennan, TX	55
1.37	0.235	HC3 errors	0.804	Houston, GA	44
1.34	0.861	Clustered on counties	1.10	Fairfax, VA	1,350
1.34	0.879	HC3 errors	1.11	Naugatuck Valley, CT	95
1.67	0.512	HC3 errors	1.09	Black Hawk, IA	46
2.51	0.881	HC3 errors	1.69	Marathon, WI	41
2.19	1.04	HC3 errors	1.61	Belmont, OH	48
1.51	1.07	HC3 errors	1.29	Sedgwick, KS	164
2.15	0.833	HC3 errors	1.49	Wichita, TX	44
1.27	0.688	HC3 errors	0.978	New Hanover, NC	61
1.59	0.891	Clustered on counties	1.24	Forsyth, NC	161
1.24	0.783	HC3 errors	1.01	Worcester, MA	184
1.41	0.812	HC3 errors	1.11	Yakima, WA	56
1.75	1.10	HC3 errors	1.43	York, PA	105
2.14	1.27	HC3 errors	1.70	Mahoning, OH	155
0.485	−0.212	HC3 errors	0.137	Yuma, AZ	60

The first tab lists the metro slopes with their intervals, and the second shows how they are distributed. The slope points the same way in 267 of 267 metros, and the median metro has a slope of 1.20 years lower per ten-point step, with a bootstrap interval over metros. The estimates are the ones chapter 10.1 reports.Source: USALEEP life expectancy, 2010-2015; Area Deprivation Index, 2023; ACS 2019-2023; population-weighted regression slopes, modeled estimates

Comparing only tracts in the same county, with a separate baseline for each county (fixed effects), gives a slope of 1.12 (95% CI 1.07 to 1.16) years per ten-point step, against 0.84 (95% CI 0.75 to 0.92) across all tracts. The within-county slope is the larger of the two, and its interval lies entirely beyond the interval of the national slope, so the same difference in ADI goes with a larger difference in life expectancy inside a county than across the country. These comparisons describe the geography more closely and leave the causal question open, and Chapter 10.1, Gradients, slopes and variance sets out how the slopes and the split of variation are built (the procedures used here are in Methods).

Some outcomes follow a different pattern

Other chapters ask whether the ordering seen in longevity holds for other measures of life, and where it breaks down. The second tab of Figure 1 gives a first example. In the PLACES 2024 release, which models survey responses from 2022, the share of adults who smoke was 9.1% (95% CI 8.5% to 9.6%) in the least deprived tenth of tracts and 22.8% (95% CI 22.5% to 23.1%) in the most deprived, 13.7 (95% CI 13.1 to 14.4) points higher. The share who had a routine checkup in the past year was 74.2% and 77.8%, only 3.6 (95% CI 2.4 to 4.6) points higher in the most deprived tenth. The checkup records contact with a physician and not the state of anyone's health, and its small difference across the scale is the exception that Chapter 5.1, Health behaviors and the supply of care examines beside the dental visit, which falls with deprivation. The smoking and checkup lines are both modeled estimates from CDC PLACES, which draws on survey responses and on census and American Community Survey inputs,[3] and the ADI is built from the American Community Survey, so part of each of those two gradients is built in.

The choice of comparison matters as well. Traffic pollution is lower in more deprived tracts across the country and higher in more deprived tracts within a county (Chapter 5.3, Air pollution). The choice of deprivation measure matters too, because a national ranking combines different kinds of places, and its measures of housing cost can rank expensive neighborhoods in ways that obscure hardship (Chapter 10.2, The anatomy of a deprivation index).

Chapter 5.2, Health behaviors, person by person changes the unit. A survey can ask the same respondent about household income and drinking instead of inferring that relationship from neighborhood averages, and among adults measured one at a time, binge drinking rises with household income except among non-Hispanic Black adults. The survey gains information about individuals and loses geographic detail, because its public files do not support the tract comparisons of the maps. PLACES draws on the same survey, so agreement between the two is not an independent replication.

Moving is a different question

The comparisons so far describe places. A family deciding whether to move faces another question, what would change for that family after the move, and the residents already living in a destination differ from potential movers in ways that a map cannot remove. This is the one section of the chapter that asks what moving does, and it reports a randomized experiment.

Causal evidenceRandomized housing-voucher experiment

Whether moving to a lower-poverty neighborhood changes health and later income

Every comparison above is between places, and none says whether a family that moved to a less deprived tract would live longer. The Moving to Opportunity experiment tested part of that question. From 1994 to 1998 the U.S. Department of Housing and Urban Development randomly assigned families in public housing in high-poverty census tracts to a voucher usable only in a low-poverty tract, to an unrestricted voucher, or to no voucher.[4]

In a survey in 2008 to 2010 of the women who had enrolled, the prevalence of a glycated hemoglobin level of 6.5% or more, a marker of diabetes, was 4.3 (95% CI 0.8 to 7.8) points lower in the group offered the low-poverty voucher than in the control group, and the prevalence of a body-mass index of 40 or more was 3.4 (95% CI 0.4 to 6.4) points lower. The unrestricted voucher showed no significant difference.[4]

Chetty, Hendren and Katz followed the children through tax records. Children assigned the experimental voucher before age 13 had an annual income 14% higher in their mid-twenties than the control group, an intention-to-treat comparison that includes families who did not move. Only 48% of the families took up the voucher, and scaling the comparison by take-up gives 31% for the children of families that did.[5]

Neither study measured life expectancy. The health results are for women who volunteered for the program from public housing in the poorest tracts, and the health results and the first income comparison are between groups assigned an offer of a voucher, which is not the same as every family moving. They do not give the size of the change along the full ADI scale.

The book sets this evidence beside its descriptive comparisons and does not treat the experiment as an explanation of the whole gradient.

Where to read next

The rest of this part follows deaths before old age by cause. Chapter 6.2, Gun homicide and the causes of early death follows firearm homicide into the few neighborhoods where it concentrates, Chapter 6.3, Overdose, alcohol and suicide maps overdose, alcohol-induced deaths and suicide by county, Chapter 6.4, Mortality trends, 2018 to 2024 follows those deaths from 2018 to 2024, and Chapter 6.5, Violence and the life-expectancy gap asks how much of this chapter's gap between the least and the most deprived tenths homicide accounts for.

Reading paths connect to this chapter's comparisons. The argument in eight chapters follows the main line through the volumes, differences inside one county follows the within-county comparison, and outcomes that do not follow the gradient collects the exceptions. The contents page lists every chapter with its finding.

What these data cannot show

Outside the box, every number here describes places. A tract with a lower life expectancy and a higher ADI is a tract whose residents, taken together, died younger in 2010 to 2015, and the data do not say which of them did, or whether a person who moved would have died later (ecological inference). The pairs show that nearby tracts differ, and the data cannot separate what the places do from who lives in them. The experiment in the box covers women and children who enrolled in one program, and it does not say how much of the gradient in the figures would change if a place changed.

The periods differ. The life expectancy estimates are modeled for deaths in 2010 to 2015 and were never updated. The ADI is a national percentile built on survey data for 2019 to 2023, and it may rank a tract differently from the way the same tract looked in 2010.

The geography is matched imperfectly. The 2020-tract file used for the gradient is area-weighted from 2010 tracts, and its errors are mapping grade, approximations that understate the uncertainty. Because the pairs are chosen on the ADI among large counties, their median describes that rule, as the sensitivity to the radius shows, and says nothing about counties that are small or rural. The counties left out are Los Angeles, Santa Clara and Alameda, all in California, and the ADI understates deprivation in high-cost cities, so a qualifying pair there may exist and not be found. The distances are straight lines and not rides. Puerto Rico has no life expectancy estimates and is not included, and Connecticut's planning regions stand in for its counties. Institutional tracts are dropped, and a tract without an estimate is absent and not zero.

The estimates are modeled. A tract's life expectancy has a median standard error of 1.7 years, which the intervals in Figure 3 show. The intervals in the figures come from resampling counties of published estimates, and they leave out error in the life expectancy model beyond its standard errors, error from moving estimates onto 2020 tracts and uncertainty in the ADI. The PLACES measures are modeled estimates built partly from the same census and survey inputs as the ADI, so the lines for smoking and the checkup show how the model maps those inputs into prevalence as well as how adults differ.

Data sources

USALEEP life expectancyUSALEEP life expectancy at birth with standard errors, 2010-2015: native 2010 tracts for the pairs, 2020 tracts (area-weighted from 2010 tracts, institutional tracts dropped) for everything elseYears: 2010-2015 · Weights: ACS 2019-2023 tract population · License: Public domain (NCHS, CDC)
Area Deprivation IndexNeighborhood Atlas Area Deprivation Index 2023, national percentile, 2020 tractsYears: 2023 release (version 4.0.1; built on ACS 2019-2023) · Weights: ACS 2019-2023 tract population, also for the ADI's split by county (over chapter 10.1's frame of tracts) · License: Free for non-commercial use; cite the Neighborhood Atlas and Kind and Buckingham
American Community SurveyACS 2019-2023 tracts: population (weights and the population floor of the pairs)Years: 2019-2023 (5-year) · Weights: tract population · License: Public domain (U.S. Census Bureau)
Census TIGER/Line tractsCensus 2020 tracts: internal points (distances in the pairs) and tract names; Census 2020 county file: the core-based statistical area of each countyYears: 2020 · Weights: none (locations) · License: Public domain (U.S. Census Bureau)
CDC PLACESCDC PLACES 2024 release, tracts: current smoking and routine checkup among adultsYears: 2024 release (tracts); 2025 release (ZCTAs, 2023 data year) · Weights: adult population of each tract (PLACES) · License: Public domain (CDC, U.S. Department of Health and Human Services)
Published estimates from cited studiesEstimates from Ludwig and coauthors (2011) and Chetty, Hendren and Katz (2016), quoted in the box on the Moving to Opportunity experimentYears: as reported in each study · Weights: as reported in each study · License: Quoted from published research for commentary

All data sources

Methods

This chapter compares neighborhood estimates of life expectancy with a deprivation index and repeats the comparison at smaller geographic scales. The deaths behind the estimates (2010 to 2015) precede the survey data behind the index (2019 to 2023) by about a decade, the tract estimates are modeled and the displayed intervals do not carry all of their error, and the nearby pairs come from a rule applied to the most populous counties only. No result estimates what moving would do to a person's lifespan; the box on the Moving to Opportunity experiment quotes cited studies.

Sample and period

Tract life expectancy at birth is the USALEEP estimate for deaths in 2010 to 2015. The gradient, the metro slopes, the variance split, the county spread and the New York City map use the file on 2020 tracts, with institutional tracts (at least 25 percent of residents in group quarters, or a known case) dropped and an ACS 2019-2023 population above zero, in the 50 states and DC. The gradient covers the 77,936 such tracts that have an estimate and an ADI rank.

The Area Deprivation Index is the 2023 release of the national percentile (1 least, 100 most deprived), built on the ACS for 2019 to 2023, joined by tract. The deaths therefore precede the ACS data the ADI uses by about a decade.

The nearby pairs use a separate sample, the native 2010 tracts of the most populous counties whose boundaries did not change (How the neighborhoods were paired, below). Smoking and the checkup are CDC PLACES 2024 estimates, which model survey responses from 2022. The box on the Moving to Opportunity experiment quotes estimates from the cited studies; they are not results of this build and keep the populations, outcomes and periods the studies report.

The New York City map draws the tracts of the gradient sample that lie in the five counties of New York City (Bronx, Kings, New York, Queens and Richmond) on the 2020 tract outlines of the city, and every other tract of the outlines is gray with a reason: no residents in the ACS, no life expectancy estimate, an institutional tract or no ADI rank. No tract is chosen or marked.

Figures explained here: Figure 1, Figure 2, Figure 6, Figure 4 and Figure 5

Matching places across years

USALEEP built its estimates on 2010 tracts. The file on 2020 tracts used for the gradient and the comparisons of scale is area-weighted from the 2010 tracts, and its standard errors are mapping grade, so they understate the uncertainty of an estimate on a 2020 tract.

A pair uses the native estimate and standard error for the 2010 tract, not the interpolated file, and only for tracts whose boundary did not change. The 2010 and 2020 tracts with one GEOID must share at least 99% of the land area of each (Census 2010-to-2020 relationship file), so the ADI and the ACS population of the 2020 tract describe the same ground.

Weights and comparisons

Weights are the ACS 2019-2023 tract population for the means of tenths and percentiles, the metro slopes, the county spread and the borough figures, so those results describe residents, and so are the variance splits of life expectancy and of the ADI. Medians across metros, counties and boroughs count each place once, and each pair counts once in the median gap over pairs.

Tenths of tracts are rank bins of the ADI with tracts of tied rank in the same tenth, so they hold from 7,613 to 8,031 tracts. The dots in the first tab of the gradient figure are a random sample of 4,000 tracts drawn with a fixed seed; the line, the band and the markers use every tract.

Smoking and the checkup are chapter 5.1's estimates, from its function and its tenths of tracts, weighted by adults in each tract, with county-bootstrap intervals. PLACES values are modeled estimates built partly from the ACS, as the ADI is, so part of each gradient is built in.

Uncertainty

Each tract estimate has a published standard error, with a median of 1.7 years. The interval of a tract is 1.96 standard errors, and the interval of a gap treats the two errors as independent.

Intervals for the means of tenths and percentiles come from repeatedly drawing whole counties with replacement and recomputing the mean, a bootstrap over counties; they are percentile intervals from 2,000 draws with seed 12345. The median gap over pairs has a bootstrap interval over pairs, with the same draws and seed. Metro slopes have errors clustered on counties where a metro has at least five and HC3 errors otherwise.

The band in the gradient figure is the population-weighted interquartile range of tracts at each ADI percentile. It describes the spread of places, includes estimation error and is not an interval for the line.

The bootstrap resamples counties of published estimates. It does not add the error of the USALEEP model beyond the published standard errors, the error of moving estimates onto 2020 tracts or uncertainty in the ADI.

Figures explained here: Figure 1 and Figure 3

How the neighborhoods were paired

Each of the 20 most populous counties, by ACS population of the tracts in the sample, is searched. A tract enters if its boundary did not change between the 2010 and 2020 censuses (Matching places across years, above), if it is not an institutional tract such as a prison or a dormitory, if it has at least 2,000 residents and if the standard error of its life expectancy is at most 2.0 years. 4,917 of the 13,625 non-institutional tracts of these counties that have a life expectancy estimate qualify.

Among qualifying tracts whose Census internal points are at most 5 kilometers apart (a straight line, projected to EPSG:5070), the county's pair is the one whose ADI differs most, with ties going to the shorter distance, and it counts if the difference reaches 50 percentile points. No pair reaches that difference in 3 counties, and they are left out. The rule never looks at life expectancy.

The data hold no transit routes, so the pairs are a short distance apart and not a measured ride. The pairs are the result of a rule applied to large counties and are not a sample of neighborhoods.

Figure explained here: Figure 3

How the local slopes and variance split were calculated

The metro slopes and the fixed-effects ladder are the functions of chapter 10.1. A core-based statistical area is coded by the Census county file, a metro needs at least 40 tracts, and each slope is a population-weighted least-squares fit. The within-county slope adds a separate baseline for each county.

The variance split is b2.canon.variance_split(), the function chapter 10.1 calls, and the split of the ADI is chapter 10.1's. Estimation error is the population-weighted mean squared standard error of the tract estimates as a share of the total variance; the 2020-tract standard errors are mapping grade and understate it.

The county spread gives, for the 30 most populous counties, the population-weighted 5th and 95th percentiles of tract life expectancy and the population-weighted mean. The view net of estimation error shrinks each tract estimate toward the county mean by the square root of one minus the county's share of variance that is estimation error. The shrinkage is an approximation that takes the published errors at face value; it does not recover the true differences between tracts.

For the New York City map, the 5th to 95th percentile of tract life expectancy in a borough and the mean of a borough are population-weighted, and the shrinkage toward the borough mean and the share of the variance that is estimation error are the rules of the county spread, applied to the five boroughs and taking the published standard errors at face value.

Figures explained here: Figure 6, Figure 4, Figure 5 and Figure 2

How the results change under alternative choices

The median gap over pairs depends on the settings of the rule, above all the radius. Pairs chosen within 3 kilometers have a median gap of 2.7 years, and within 8 kilometers it is 5.1, against 5.4 within 5 kilometers. Raising the population floor to 4,000 residents gives 5.5.

A cap of 1.5 years on the standard error gives a median of 4.3 over 15 pairs, a cap of 3.0 years gives 6.2, and no cap gives 4.6 over 18 pairs.

The number of counties in the county spread matters little. The median range is 10.5 years among the 15 most populous counties and 10.5 among the 60 most populous, against 10.5 among 30.

Figures explained here: Figure 3 and Figure 5

References

Arias, Elizabeth; Escobedo, Luis A.; Kennedy, Jiaquan; Fu, Chenxi; Cisewski, Jodi (2018). U.S. Small-area Life Expectancy Estimates Project: Methodology and Results Summary. Vital and Health Statistics, Series 2, Number 181, National Center for Health Statistics. www.cdc.gov/nchs/data/series/sr_02/sr02_181.pdf
Kind, Amy J. H.; Buckingham, William R. (2018). Making Neighborhood-Disadvantage Metrics Accessible: The Neighborhood Atlas. New England Journal of Medicine 378(26): 2456-2458. doi:10.1056/NEJMp1802313
Greenlund, Kurt J.; Lu, Hua; Wang, Yan; Matthews, Kevin A.; LeClercq, Jennifer M.; Lee, Benjamin; Carlson, Susan A. (2022). PLACES: Local Data for Better Health. Preventing Chronic Disease 19: 210459. doi:10.5888/pcd19.210459
Ludwig, Jens; Sanbonmatsu, Lisa; Gennetian, Lisa; Adam, Emma; Duncan, Greg J.; Katz, Lawrence F.; Kessler, Ronald C.; Kling, Jeffrey R.; Lindau, Stacy Tessler; Whitaker, Robert C.; McDade, Thomas W. (2011). Neighborhoods, Obesity, and Diabetes: A Randomized Social Experiment. New England Journal of Medicine 365(16): 1509-1519. doi:10.1056/NEJMsa1103216
Chetty, Raj; Hendren, Nathaniel; Katz, Lawrence F. (2016). The Effects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment. American Economic Review 106(4): 855-902. doi:10.1257/aer.20150572

How to cite

@misc{singh2026placeandthelengthofalife,
  author = {Singh, Vishal},
  title = {Place and the length of a life},
  year = {2026},
  url = {https://vishalsingh.org/zipcode/place-and-the-length-of-a-life},
  note = {In ZIP Code Destiny. Published 9 October 2026; updated 11 October 2026.}
}

Same question at another grain

These chapters take up the same axis with a different unit of analysis.

Well-being and life expectancy
- Chapter 7.2, Life evaluation and daily feelings Individuals