Chapter 7.1 · Alone together

Civic life, loneliness and partnership

Voter turnout and census self-response fall across deciles of county deprivation, from 72.1% to 56.4% and from 68.6% to 47.6%, and Gallup respondents who are partnered rate their lives 0.50 ladder points higher than others, in counties with low and high social capital alike.

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

Voter turnout in the 2020 presidential election was 72.1% among adult citizens in the tenth of counties with the lowest Area Deprivation Index scores and 56.4% in the tenth with the highest, and the share of households that answered the 2020 census on their own fell from 68.6% to 47.6% across the same ten groups.[1] Both measures are lower in the most deprived tenth than in the least deprived, with point estimates that fall at every step, and neither keeps a link to upward mobility once county income, racial composition and poverty are held fixed.

Gallup respondents who were married or in a domestic partnership rated their lives 0.50 (95% CI 0.49 to 0.51) ladder points higher than other adults, and the difference was no smaller in counties with more social capital. Across ZIP Code Tabulation Areas, modeled loneliness is lower where more residents are married, a correlation of −0.65 (95% CI −0.69 to −0.62), but the model that produces those estimates already uses age, so part of that gradient is built in.

The results describe three different units. County results describe places, the Gallup results describe people interviewed in 2009 to 2017, and the loneliness estimates describe ZCTAs, the Census Bureau's approximation of ZIP codes. A pattern in one unit is not a finding about another, and the chapter keeps them apart.

Participation falls with county deprivation

Counties are ranked on the Area Deprivation Index, where each county takes the mean national percentile rank of its tracts weighted by tract population, and are cut into ten groups with equal numbers of the 3,128 counties that have a rank. County Health Rankings supplies two measures of showing up. Voter turnout is the share of adult citizens who voted in the 2020 presidential election, and census self-response is the share of all households that answered the 2020 census online, by mail or by phone without a visit from a census taker.[1] Both fall from the least to the most deprived tenth, and their point estimates are lower at every step (Figure 1). Between the least and the most deprived tenth, turnout is 15.6 (95% CI 13.9 to 17.7) points lower and self-response 21.0 (95% CI 19.0 to 22.9) points lower, so the self-response gradient is the steeper of the two.

Figure 1

Voter turnout and census self-response fall with county deprivation, and across counties they correlate less with upward mobility than family structure and friendship across incomes do

County Health Rankings 2024 participation measures by decile of county deprivation, and county measures of civic life against the adult income rank of children born 1978 to 1983 to parents at the 25th percentile

Voter turnout fell from 72.1% to 56.4% and census self-response from 68.6% to 47.6% across deciles of county deprivation

Population-weighted mean voter turnout (citizens aged 18 and over, 2020 election) and census self-response (households, 2020 census) by decile of county deprivation

Decile of county deprivation (1 = least deprived tenth of counties, 10 = most deprived)

Both lines fall from the least to the most deprived tenth, their point estimates fall at every step, and the self-response line falls further. Each point is a population-weighted mean over about a tenth of counties, with a 95% interval from a bootstrap over counties.

Show data

Data for Voter turnout and census self-response fall with county deprivation, and across counties they correlate less with upward mobility than family structure and friendship across incomes do, Participation by deprivation
conf_high	conf_low	decile	estimate	measure	n_units
73.80	70.58	1	72.06	Voter turnout, 2020	307
71	67.93	2	69.33	Voter turnout, 2020	304
68.58	64.62	3	66.59	Voter turnout, 2020	309
65.57	62.07	4	63.95	Voter turnout, 2020	309
64.15	61.43	5	62.84	Voter turnout, 2020	310
62.08	59.10	6	60.63	Voter turnout, 2020	310
60.94	58.84	7	59.85	Voter turnout, 2020	313
60.52	56.90	8	58.66	Voter turnout, 2020	315
59.01	57.18	9	58.06	Voter turnout, 2020	311
57.59	55.37	10	56.43	Voter turnout, 2020	310
69.94	67.39	1	68.60	Census self-response, 2020	312
68.15	65.91	2	66.97	Census self-response, 2020	312
68.04	65	3	66.45	Census self-response, 2020	312
66.25	63.61	4	64.94	Census self-response, 2020	312
64.80	62.56	5	63.80	Census self-response, 2020	314
62.88	60.24	6	61.63	Census self-response, 2020	311
61.85	59.72	7	60.76	Census self-response, 2020	313
58.76	55.75	8	57.15	Census self-response, 2020	315
55.21	53.18	9	54.20	Census self-response, 2020	312
49.12	46.11	10	47.63	Census self-response, 2020	312

The first tab shows participation by deprivation decile, and the second ranks county measures by their correlation with upward mobility. Counties are the unit in both tabs, so the patterns describe places and not the people who live in them.Note: Deciles hold equal numbers of counties. County deprivation is the mean of the tract Area Deprivation Index national ranks (1 = least deprived), weighted by tract population. Turnout counts only citizens and depends on state voting laws; self-response is low where census takers counted households in person.Note: Intervals resample whole counties. Resampling whole states gives wider intervals for every point, and the decile intervals still separate the first from the last decile.Note: Counties with a missing measure drop out of that point only, so the number of counties differs by measure. The JEC collective-efficacy subindex is measured at the county level by violent-crime rates, and the JEC institutional-health and community-health subindexes include state-level inputs.Source: County Health Rankings 2024 (voter turnout and census participation, 2020), Area Deprivation Index 2023 and ACS 2019-2023 county population, population-weighted; second tab adds Social Capital Atlas 2022, JEC Social Capital Index 2018 release and Opportunity Atlas county outcomes (1978-1983 birth cohorts), population-weighted by 2018 county population

These are means weighted by county population. Counting each county once puts turnout at 76.9% in the least deprived tenth and 58.8% in the most deprived, and self-response at 65.0% and 46.3%. In every tenth the unweighted turnout is higher than the weighted one and the unweighted self-response lower, so small counties vote more and respond less on their own than the counties where most people live. The gradient points the same way under either weighting.

Neither measure is a pure gauge of civic willingness. Turnout counts only citizens, depends on state voting laws, and leaves out people who have lost their voting rights.[1] Self-response depends on how a place was counted. Census takers counted households in person instead of inviting them to respond in remote Alaska, in remote parts of northern Maine and southeast Alaska, and optionally in tribal areas, so a low rate there reflects the operation as well as the residents.[2]

Figure 2 maps self-response by county. With counties weighted by population, the mean is 66.0%, the fifth percentile 51.8% and the ninety-fifth percentile 77.5%. State averages are highest in Minnesota, Wisconsin and Michigan and lowest in Alaska, West Virginia and Maine, and the four lowest county values are all in Alaska.

Figure 2

Census self-response in 2020 was highest in Minnesota, Wisconsin and Michigan and lowest in Alaska, West Virginia and Maine

Percent of households that answered the 2020 census online, by mail or by phone without a visit from a census taker, by county

Loading county values

Households answering the 2020 census on their own (%)

Values outside 50 to 80 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.

Counties toward the high end of the key are common in the upper Midwest, and the four lowest values are in Alaska. The map plots each county's self-response rate; the mean over counties weighted by population is 66.0%, and the 3 gray counties have no published rate. State comparisons in the title average county rates weighted by county population.Note: Each gray county has its reason in the tooltip. Puerto Rico is not in County Health Rankings and is not drawn. Connecticut appears as its eight legacy counties.Note: Census takers counted households in person instead of inviting them to respond in remote Alaska, in remote parts of northern Maine and southeast Alaska, and optionally in tribal areas, so a low self-response rate there reflects how the count was run as well as the residents.Note: The tooltip also gives voter turnout in the 2020 presidential election, which depends on state voting laws, counts only citizens and divides ballots by an ACS estimate of adult citizens.Source: County Health Rankings 2024 release, census participation (2020 census; full count of households) and voter turnout (2020; ballots over an ACS 5-year estimate of adult citizens)

Voting tracks mobility less than family structure and friendship do

The second tab of Figure 1 asks which county measures of civic life go with the upward mobility of children from low-income families. Mobility is the mean adult income rank of children born in 1978 to 1983 to parents at the 25th percentile of the national income distribution, which the Opportunity Atlas reports by county (Chapter 3.3, Race, income and mobility describes it). Each measure is correlated with mobility across counties weighted by their 2018 population, with intervals from a bootstrap over counties.

Voter turnout correlates +0.35 (95% CI 0.28 to 0.42) with mobility and census self-response +0.32 (95% CI 0.26 to 0.37). The organization measures sit lower. Civic groups per capita, from the Social Capital Atlas, correlate +0.06 (95% CI −0.01 to 0.14), membership associations −0.05 (95% CI −0.12 to 0.02) and religious congregations −0.10 (95% CI −0.15 to −0.04), while membership in volunteering groups reaches +0.19 (95% CI 0.12 to 0.25). Putnam's account of declining social capital gave associations a central place,[3] yet across counties the density of associations says little about where children from low-income families rise, and congregations run the other way. The measures come from different sources and years (Facebook groups, business records and a 2010 religion census), which limits what a gap between two of them can show.

The Joint Economic Committee's county Social Capital Index has four parts. Family unity uses births to unmarried mothers, children in single-parent families and the share of women in midlife who are married. Community health uses nonprofits, congregations and a state-level measure of civic engagement. Institutional health uses presidential voting, census mail response and state-level confidence in institutions. Collective efficacy is measured at the county level by violent-crime rates.[4] Family unity has the highest correlation with mobility of any measure, +0.71 (95% CI 0.67 to 0.74), and economic connectedness, the measure of friendships across income lines that Chapter 3.3, Race, income and mobility examines,[5] is +0.69 (95% CI 0.64 to 0.74), with an overlapping interval. The overall index is +0.61 (95% CI 0.53 to 0.68), collective efficacy (low violent crime) +0.51 (95% CI 0.43 to 0.58), institutional health +0.24 (95% CI 0.16 to 0.33) and community health +0.19 (95% CI 0.10 to 0.28). Of the four parts, the two that describe families and safety sit well above the two that describe participation and associations.

Collective efficacy in the county index is a violent-crime rate, so its correlation with mobility is a statement about crime and not about residents acting together.

Counties with high turnout also differ in income, racial composition and poverty, so Figure 3 enters those characteristics. With the four civic measures alone in the model, turnout has a standardized coefficient of 0.24 (95% CI 0.06 to 0.42) on mobility. With the log of median household income, the Black share of residents and the poverty rate added, it falls to 0.00 (95% CI −0.15 to 0.16) and the coefficient on census self-response is −0.08 (95% CI −0.21 to 0.05). Once economic connectedness is added, its own coefficient is 0.63 (95% CI 0.47 to 0.79), the largest in the model, while the coefficients on census self-response and volunteering become negative, −0.15 (95% CI −0.26 to −0.05) and −0.16 (95% CI −0.22 to −0.09). These are partial associations among measures that move together, and the change of sign says that each coefficient depends on which others share the model.

Figure 3

Once income, race and poverty are in the model, voter turnout has a coefficient of 0.00 on county mobility, and economic connectedness has 0.63

Standardized coefficients on the adult income rank of children from low-income families (born 1978 to 1983) in three county regressions, weighted by 2018 county population

Civic measures only
Plus income, percent Black and poverty
Plus economic connectedness

Show data

Data for Once income, race and poverty are in the model, voter turnout has a coefficient of 0.00 on county mobility, and economic connectedness has 0.63
conf_high	conf_low	estimate	model	n_units	term
0.420	0.058	0.239	Civic measures only	2,915	Voter turnout
0.327	−0.007	0.160	Civic measures only	2,915	Census self-response
0.127	−0.150	−0.012	Civic measures only	2,915	Civic groups per capita
0.253	−0.016	0.118	Civic measures only	2,915	Volunteering group membership
0.161	−0.152	0.004	Plus income, percent Black and poverty	2,915	Voter turnout
0.047	−0.212	−0.083	Plus income, percent Black and poverty	2,915	Census self-response
0.160	−0.103	0.028	Plus income, percent Black and poverty	2,915	Civic groups per capita
0.091	−0.061	0.015	Plus income, percent Black and poverty	2,915	Volunteering group membership
0.593	0.238	0.415	Plus income, percent Black and poverty	2,915	Log median household income
−0.425	−0.608	−0.517	Plus income, percent Black and poverty	2,915	Percent Black
0.076	−0.158	−0.041	Plus income, percent Black and poverty	2,915	Poverty rate
0.101	−0.121	−0.010	Plus economic connectedness	2,915	Voter turnout
−0.053	−0.256	−0.155	Plus economic connectedness	2,915	Census self-response
0.049	−0.146	−0.049	Plus economic connectedness	2,915	Civic groups per capita
−0.094	−0.220	−0.157	Plus economic connectedness	2,915	Volunteering group membership
0.790	0.467	0.628	Plus economic connectedness	2,915	Economic connectedness
0.323	−0.053	0.135	Plus economic connectedness	2,915	Log median household income
−0.339	−0.480	−0.409	Plus economic connectedness	2,915	Percent Black
0.145	−0.074	0.036	Plus economic connectedness	2,915	Poverty rate

Turnout's coefficient falls to about zero when the county controls enter, the census coefficient has an interval that includes zero in the first two models, and economic connectedness stands apart once it is added. Each dot is a standardized coefficient with a 95% interval, from a weighted regression with errors clustered on state; the three models use the same counties.Note: Every variable is a z-score using the Atlas's 2018 county population as the weight. The controls are the log of median household income, the percent Black and the percent below poverty from the ACS 2019-2023 tract file, averaged to counties; Connecticut drops out because the ACS tract file uses planning-region codes.Note: Intervals use t distributions on the state clusters. These are partial associations among correlated county measures, and a coefficient can change sign when its neighbors change.Source: County Health Rankings 2024, Social Capital Atlas 2022, Opportunity Atlas county outcomes (1978-1983 birth cohorts) and ACS 2019-2023 county aggregates; population-weighted by 2018 county population

The measures also have different dates. Mobility is for children born in 1978 to 1983, the participation measures are for 2020, the Facebook measures for 2022 and the Joint Economic Committee components for 2010 to 2016. Using recent civic life to describe where children rose decades ago assumes that county differences persist, which these data do not test.

The partner premium in the ladder score

The Gallup U.S. Daily survey has 2,189,201 interviews from 2009 to 2017, in 3,140 counties, with both a marital status and a score on the Cantril ladder, which asks people to place their life on a ladder from the worst possible life to the best possible one. Adults who were married or in a domestic partnership placed their lives at 7.28 (95% CI 7.27 to 7.29) on average, and adults who were never married, separated, divorced or widowed placed theirs at 6.78 (95% CI 6.77 to 6.79). The never married (6.77) and the separated, divorced or widowed (6.78) score the same, so the chapter compares the partnered with everyone else. The difference, 0.50 (95% CI 0.49 to 0.51) ladder points, is the partner premium. The 2008 interviews are left out. In 2008 Gallup asked political questions immediately before the ladder, which lowers the ladder score, and it removed them from part of the sample in January 2009 and added a transition question in April 2009.[6] Including 2008 would put the group means at 7.22 and 6.73 and leave the premium at 0.50 (95% CI 0.49 to 0.51), and leaving out 2009 as well gives 0.50 (95% CI 0.49 to 0.51). The item was asked of nearly all respondents in 2010 to 2012, of 85% in 2009 and of 49% from 2013 to 2017, so 61% of these interviews are from 2009 to 2012. Every Gallup figure here is unweighted, because the extract has no survey weights, and intervals are clustered by the respondent's county, so they describe sampling variation in these interviews and not national prevalence.

The premium is not an artifact of where partnered people live. Comparing each respondent only with others in the same county (county fixed effects) gives 0.50 (95% CI 0.49 to 0.50), the same as the national difference. It is larger once age and sex are taken into account: 0.64 (95% CI 0.63 to 0.65) within county and year, with a fixed effect for each year of age and for sex, so the unadjusted difference is lowered by the age mix of the two groups. A fixed effect for each year of age is used because the ladder is U-shaped in age and the premium varies with it, and a single linear age term gives a lower premium. Among the interviews that report household income and education, adding both lowers the premium to 0.31 (95% CI 0.30 to 0.32). Household income includes a partner's earnings, so partnered adults are more likely to live in households with higher incomes, and this model compares adults within the same income and education group. By age band, with the same adjustment inside each band, the premium is smallest in the youngest band (0.15 (95% CI 0.12 to 0.18)) and next smallest in the oldest (0.35 (95% CI 0.33 to 0.37)), and it is largest in the middle of adulthood, 0.80 (95% CI 0.78 to 0.82) at its highest.

Figure 4

The partner premium is 0.50 ladder points overall and 0.50 within counties, and runs from 0.15 in the youngest age band to 0.80 in its largest band

Difference in mean Cantril ladder score between adults who are married or in a domestic partnership and all others, under successive adjustments and by age band, Gallup U.S. Daily 2009 to 2017

All adults
By age band, adjusted for single-year age and sex

Show data

Data for The partner premium is 0.50 ladder points overall and 0.50 within counties, and runs from 0.15 in the youngest age band to 0.80 in its largest band
conf_high	conf_low	estimate	group	label	n
0.506	0.489	0.498	All adults	Difference of means, no controls	2,189,201
0.504	0.488	0.496	All adults	Within county	2,189,199
0.650	0.630	0.640	All adults	Within county and year, adjusted for single-year age and sex	2,163,354
0.318	0.301	0.310	All adults	Same, adding household income and education	1,786,303
0.183	0.121	0.152	By age band, adjusted for single-year age and sex	Ages 18 to 24	147,891
0.632	0.586	0.609	By age band, adjusted for single-year age and sex	Ages 25 to 34	226,935
0.809	0.762	0.785	By age band, adjusted for single-year age and sex	Ages 35 to 44	268,006
0.817	0.779	0.798	By age band, adjusted for single-year age and sex	Ages 45 to 54	377,207
0.787	0.758	0.772	By age band, adjusted for single-year age and sex	Ages 55 to 64	461,305
0.609	0.579	0.594	By age band, adjusted for single-year age and sex	Ages 65 to 74	394,603
0.368	0.333	0.350	By age band, adjusted for single-year age and sex	Ages 75 and over	286,964

The premium is almost the same within counties as across them, rises to 0.64 with an adjustment for single-year age and sex, falls to 0.31 when household income and education enter, and is largest in the middle age bands. Each dot is the coefficient on partnership with a 95% interval clustered by county; the number of interviews is in the tooltip.Note: The last all-adults row uses the interviews that report household income and education, a smaller sample than the other rows. Household income includes a partner's earnings. The rows are statistical adjustments, and none identifies an effect of partnership; the age-band rows are separate regressions with county, year, single-year age and sex fixed effects. A linear age term fits poorly because the ladder is U-shaped in age, and it gives a lower premium.Note: Partnered means married or in a domestic partnership, which is how Gallup's three-way marital status counts domestic partners. Adults who live with a partner and call themselves single are in the other group.Note: Gallup's extract has no survey weights, so every estimate is unweighted. County comes from the respondent's ZIP code and is approximate near county lines.Source: Gallup U.S. Daily, 2009-2017; unweighted, intervals clustered by county

Study evidenceLongitudinal panel following the same people over time

Whether partnership moves with the ladder score

The interviews compare different people, and happier people may be more likely to marry and to stay married. Two studies follow the same people over time in the German Socio-Economic Panel. Stutzer and Frey compared the life satisfaction of single people who later married with that of people who stayed single, and found that happier singles were more likely to marry, so part of a cross-sectional difference reflects who marries.[7] Lucas and colleagues followed a large panel through changes in marital status and found that, on average, life satisfaction reacted to marriage and then moved back toward its earlier level, with large differences between individuals.[8]

Neither study randomizes marriage. Lucas and colleagues compare people with themselves over time, which removes differences that do not change. Stutzer and Frey also compare people who later married with people who stayed single, which is a comparison between persons before marriage. Neither speaks to the Gallup interviews, which have no before and after.

If community could stand in for a partner, the premium would be smaller in counties with more of it. Figure 5 sorts interviews into ten groups by the Joint Economic Committee's Social Capital Index of the respondent's county, a standardized score from the 2018 release whose components date from 2010 to 2016, and compares the partnered with everyone else in each group. The premium was 0.45 (95% CI 0.43 to 0.48) ladder points in the lowest group and 0.49 (95% CI 0.47 to 0.51) in the highest, with a largest value of 0.53 (95% CI 0.51 to 0.55) in one of the upper groups, so it did not shrink. The point estimates are smaller in the three lowest groups than in the other seven, although some intervals overlap. The mean ladder score of the partnered is 7.25 (95% CI 7.21 to 7.30) in the lowest group and 7.35 (95% CI 7.33 to 7.37) in the highest, and that of the others 6.80 (95% CI 6.75 to 6.84) and 6.86 (95% CI 6.83 to 6.89).

Figure 5

The partner premium does not shrink where county social capital is higher: 0.45 ladder points in the lowest decile and 0.49 in the highest

Cantril ladder scores of partnered and other adults by decile of the JEC Social Capital Index of the respondent's county, and the change in the premium per standard deviation of seven county measures

Ladder scores differ little between the lowest and highest decile of county social capital, 7.25 and 7.35 among partnered adults and 6.80 and 6.86 among others

Unweighted mean Cantril ladder score by decile of the JEC Social Capital Index of the respondent's county, Gallup U.S. Daily interviews 2009 to 2017

Decile of the county Social Capital Index (1 = lowest tenth of interviews, 10 = highest)

The two lines run roughly parallel and flat, with the partnered line higher at every decile. Each point is an unweighted mean with a 95% interval clustered by county.

Show data

Data for The partner premium does not shrink where county social capital is higher: 0.45 ladder points in the lowest decile and 0.49 in the highest, Ladder scores
conf_high	conf_low	decile	estimate	group	n
7.30	7.21	1	7.25	Married or domestic partner	112,277
6.84	6.75	1	6.80	Single, separated, divorced or widowed	106,246
7.30	7.23	2	7.26	Married or domestic partner	119,477
6.85	6.77	2	6.81	Single, separated, divorced or widowed	97,396
7.30	7.24	3	7.27	Married or domestic partner	117,601
6.85	6.78	3	6.81	Single, separated, divorced or widowed	101,443
7.28	7.22	4	7.25	Married or domestic partner	125,048
6.77	6.71	4	6.74	Single, separated, divorced or widowed	92,165
7.29	7.23	5	7.26	Married or domestic partner	126,801
6.78	6.71	5	6.74	Single, separated, divorced or widowed	91,539
7.28	7.22	6	7.25	Married or domestic partner	131,967
6.79	6.70	6	6.75	Single, separated, divorced or widowed	90,649
7.29	7.23	7	7.26	Married or domestic partner	129,786
6.77	6.70	7	6.74	Single, separated, divorced or widowed	83,949
7.30	7.25	8	7.28	Married or domestic partner	134,032
6.79	6.73	8	6.76	Single, separated, divorced or widowed	81,701
7.35	7.29	9	7.32	Married or domestic partner	135,115
6.83	6.76	9	6.79	Single, separated, divorced or widowed	82,489
7.37	7.33	10	7.35	Married or domestic partner	141,007
6.89	6.83	10	6.86	Single, separated, divorced or widowed	79,202

The first tab shows the two groups' ladder scores, the second the premium between them, and the third how the premium changes with each of seven county measures. Respondents are people, but the social capital measures describe their counties, so the comparison says how the premium differs across kinds of place and not what a county does for a person.Note: Gallup's extract has no survey weights, so every estimate is unweighted and levels can differ from the weighted figures Gallup publishes. County comes from the respondent's ZIP code and is approximate near county lines. The ladder is the evaluative measure, and daily feelings are not used here.Note: Deciles hold about a tenth of interviews each, and a county sits in one decile. The JEC index combines family unity, community health, institutional health and collective efficacy, and family unity includes the share of women who are married, so the Atlas measures serve as a check that does not depend on marriage.Note: The premium in the second tab is largest in one of the upper deciles; the highest of the ten estimates is selected for being the highest.Note: The third tab clusters errors on the respondent's state, because the JEC community-health and institutional-health subindexes include state-level inputs; the first two tabs cluster on county.Source: Gallup U.S. Daily, 2009-2017, unweighted, intervals clustered by county (by state in the third tab); JEC Social Capital Index 2018 release (components 2010-2016) and Social Capital Atlas 2022 joined on the county of the interview

A regression with an interaction between partnership and the county index puts the premium 0.023 (95% CI 0.012 to 0.034) ladder points higher for each standard deviation of the index. The same coefficient has a positive point estimate for the index's four parts and for the Social Capital Atlas measures of civic groups and volunteering, from 0.007 (95% CI −0.005 to 0.019) to 0.035 (95% CI 0.019 to 0.051). These intervals cluster on the respondent's state, because the community-health and institutional-health subindexes include state-level inputs, and 3 of the seven include zero, those for community health, institutional health and civic groups. The data support "the premium does not shrink" and do not establish that it grows. The index includes family structure, which moves with the share of partnered adults in a county, and the Atlas measures do not, so they serve as a check on that overlap.

The comparison says that the premium is not smaller in places with more social capital by these measures. It does not say that community adds to a partner's value, because counties differ in age, income and other ways that this comparison does not hold fixed, and the county measures describe places while the ladder scores describe people.

Modeled loneliness follows who lives in a ZCTA

CDC PLACES publishes modeled estimates of the share of adults who say they feel lonely always, usually or sometimes for each ZCTA. The measure comes from an optional module of the 2023 Behavioral Risk Factor Surveillance System that only some states asked, so the estimates cover 39 states, those in which at least 90% of ZCTAs have an estimate, and ZCTAs that hold 75% of the residents of ZCTAs in the PLACES file.[9] The values are modeled estimates: a multilevel regression predicts a respondent's answer from age, sex, race and ethnicity and education, with the county poverty rate and state and county effects, and the prediction is applied to the 2020 census population of each ZCTA.[10][11] The county input is the share of adults below 150% of the federal poverty level. Marital status and population density are not inputs. A ZCTA estimate therefore reports who lives there, as the survey answers by age, sex, race and education and county poverty imply, and a gradient against a characteristic that moves with those inputs, including age and the poverty rate, is partly built in.

Across the 19,684 ZCTAs with an estimate and at least 500 residents, weighted by adult population, modeled loneliness averages 34.4%; 68 of those ZCTAs are assigned to other states. It is lower where more adults are married, a correlation of −0.65 (95% CI −0.69 to −0.62), and lower where the population is older, −0.57 (95% CI −0.60 to −0.54) with median age and −0.48 (95% CI −0.51 to −0.45) with the share of older residents (Figure 6). It is higher where the poverty rate is higher, +0.59 (95% CI 0.55 to 0.63), which is partly built in through the county poverty input, and where population density is higher, +0.41 (95% CI 0.31 to 0.49) in logs, and lower where median household income is higher, −0.48 (95% CI −0.54 to −0.42) in logs. Counting each ZCTA once gives −0.64 (95% CI −0.66 to −0.62) for the married share and +0.52 (95% CI 0.49 to 0.55) for poverty. The residents floor matters for the unweighted figures: with no floor, the weighted correlation with the married share is −0.65 (95% CI −0.69 to −0.62) and the unweighted one −0.56 (95% CI −0.58 to −0.54). The Area Deprivation Index is the exception. Its correlation is +0.02 (95% CI −0.13 to 0.20) when weighted and +0.16 (95% CI 0.03 to 0.27) when each ZCTA counts once, so the book's deprivation measure says little about modeled loneliness, although the poverty rate says a good deal. The ZCTA values of the index are an area-weighted approximation, and the index understates deprivation in high-cost cities.

Figure 6

Modeled loneliness is lower in ZCTAs with more married adults, a correlation of −0.65, and higher where poverty is higher, +0.59

Correlation of CDC PLACES modeled loneliness (percent of adults, BRFSS 2023) with seven ZCTA characteristics, ZCTAs with an estimate and at least 500 residents

Weighted by adult population
Each ZCTA counted once

Show data

Data for Modeled loneliness is lower in ZCTAs with more married adults, a correlation of −0.65, and higher where poverty is higher, +0.59
conf_high	conf_low	estimate	group	label	n_units
−0.621	−0.694	−0.653	Weighted by adult population	Married share of adults	19,684
−0.619	−0.658	−0.639	Each ZCTA counted once	Married share of adults	19,684
−0.543	−0.605	−0.571	Weighted by adult population	Median age	19,683
−0.520	−0.563	−0.541	Each ZCTA counted once	Median age	19,683
−0.450	−0.512	−0.480	Weighted by adult population	Share aged 65 and over	19,684
−0.434	−0.480	−0.458	Each ZCTA counted once	Share aged 65 and over	19,684
0.487	0.312	0.408	Weighted by adult population	Population density (log)	19,665
0.422	0.289	0.356	Each ZCTA counted once	Population density (log)	19,665
−0.419	−0.542	−0.478	Weighted by adult population	Median household income (log)	19,355
−0.361	−0.462	−0.410	Each ZCTA counted once	Median household income (log)	19,355
0.633	0.550	0.587	Weighted by adult population	Poverty rate	19,538
0.545	0.490	0.517	Each ZCTA counted once	Poverty rate	19,538
0.201	−0.128	0.025	Weighted by adult population	Area Deprivation Index	19,622
0.266	0.034	0.156	Each ZCTA counted once	Area Deprivation Index	19,622

The married share, median age and share of older residents sit left of zero, income sits left of it too, poverty and density sit right of it, and the Area Deprivation Index sits near zero. Each dot is a correlation across ZCTAs with a 95% interval from a bootstrap over counties, weighted by adult population and with each ZCTA counted once.Note: PLACES estimates come from a model fed by age, sex, race and ethnicity, education and the county share of adults below a multiple of the poverty line, so a gradient against characteristics that move with those inputs, including age, poverty and income, is partly built in. Marital status and population density are not inputs.Note: The ZCTA characteristics come from an ACS-based table whose year is not documented in the project files, and the married share is labeled a share of adults because its age universe is not documented either. ZCTAs are not ZIP codes.Note: The Area Deprivation Index at ZCTA grain is an area-weighted approximation and understates deprivation in high-cost cities.Source: CDC PLACES 2025 release (BRFSS 2023), ZCTA modeled estimates, with ACS-based ZCTA characteristics from the project's ZCTA atlas and the Area Deprivation Index 2023; modeled estimates, population-weighted by adults

Sorted into ten groups by married share of adults, ZCTAs run from a mean of 28.8% to 67.4%, and modeled loneliness falls from 38.7% (95% CI 38.1% to 39.3%) to 30.2% (95% CI 29.7% to 30.7%). Sorted by median age, from 29.6 years to 55.8 years, it falls from 38.1% (95% CI 37.3% to 39.0%) to 29.4% (95% CI 29.0% to 29.9%) (Figure 7). The two gradients overlap, because ZCTAs with more married adults are older, and age is an input to the model.

Figure 7

Modeled loneliness is higher in ZCTAs with fewer married adults and a younger population

CDC PLACES modeled loneliness (percent of adults, BRFSS 2023) by decile of ZCTA married share and of ZCTA median age, population-weighted

Modeled loneliness falls from 38.7% in the tenth of ZCTAs with the lowest married share to 30.2% in the tenth with the highest

Population-weighted mean of CDC PLACES modeled loneliness by decile of ZCTA married share, ZCTAs with an estimate and at least 500 residents

The point estimates fall at every step, steeply over the first deciles. Each point is a weighted mean over about a tenth of ZCTAs, plotted at that tenth's mean married share, with a 95% interval from a bootstrap over counties.

Show data

Data for Modeled loneliness is higher in ZCTAs with fewer married adults and a younger population, By married share
conf_high	conf_low	decile	estimate	n_units	x
39.35	38.07	1	38.70	1,968	28.82
37.63	35.65	2	36.56	1,968	40.27
35.91	34.49	3	35.19	1,969	45.59
34.31	33.40	4	33.83	1,968	49.30
33.67	32.96	5	33.32	1,969	52.05
33.01	32.14	6	32.58	1,968	54.53
32.47	31.71	7	32.09	1,968	56.94
31.89	31.20	8	31.53	1,969	59.48
31.08	30.44	9	30.76	1,968	62.45
30.74	29.75	10	30.22	1,969	67.43

The first tab sorts ZCTAs by married share and the second by median age, and loneliness falls across both. The two gradients overlap, because ZCTAs with more married adults are older.Note: PLACES uses age as an input to the model, so the age gradient is partly built in, and the married-share gradient moves with age. The estimates describe ZCTAs, not the people in them.Note: Deciles hold equal numbers of ZCTAs. The ZCTA characteristics come from an ACS-based table whose year is not documented in the project files. ZCTAs are not ZIP codes.Source: CDC PLACES 2025 release (BRFSS 2023), ZCTA modeled estimates, and ACS-based ZCTA characteristics from the project's ZCTA atlas; modeled estimates, population-weighted by adults

Density shows that the model does not produce one gradient for every outcome. Modeled loneliness is 32.7% (95% CI 32.6% to 32.9%) in ZCTAs with fewer than 100 people per square mile and 37.9% (95% CI 36.6% to 39.0%) in those with 5,000 or more, while modeled depression falls from 23.5% (95% CI 23.4% to 23.7%) to 19.6% (95% CI 18.8% to 20.6%) (Figure 8). Depression here is the share of adults ever told by a health professional that they have a depressive disorder, which depends in part on access to care, and both measures come from the same model inputs.

Figure 8

Modeled loneliness rises from 32.7% to 37.9% across density classes of ZCTAs while modeled depression falls from 23.5% to 19.6%

Population-weighted mean of two CDC PLACES modeled measures (percent of adults, BRFSS 2023) by population density of the ZCTA, ZCTAs with both estimates and at least 500 residents

Feeling lonely
Depression diagnosis

Show data

Data for Modeled loneliness rises from 32.7% to 37.9% across density classes of ZCTAs while modeled depression falls from 23.5% to 19.6%
conf_high	conf_low	estimate	label	measure	n_units
32.91	32.60	32.75	Under 100 people per square mile	Feeling lonely	9,891
23.68	23.36	23.52	Under 100 people per square mile	Depression diagnosis	9,891
33.36	32.81	33.08	100 to 999 people per square mile	Feeling lonely	5,136
23.51	23.02	23.26	100 to 999 people per square mile	Depression diagnosis	5,136
34.60	33.59	34.09	1,000 to 4,999 people per square mile	Feeling lonely	3,219
22	20.86	21.42	1,000 to 4,999 people per square mile	Depression diagnosis	3,219
39.04	36.63	37.87	5,000 and over people per square mile	Feeling lonely	1,438
20.63	18.78	19.58	5,000 and over people per square mile	Depression diagnosis	1,438

The point estimates of loneliness are higher in each denser class and those of depression lower, and the intervals of the sparsest and densest classes are apart, so the two modeled measures do not share one gradient. Each dot is a weighted mean with a 95% interval from a bootstrap over counties.Note: Feeling lonely is the share of adults who say they feel lonely always, usually or sometimes. Depression is the share ever told by a health professional that they have a depressive disorder, which depends on access to diagnosis.Note: Both are modeled from the same inputs. Density classes use ACS-based ZCTA population per square mile from a table whose year is not documented in the project files. Lower cuts are inclusive.Source: CDC PLACES 2025 release (BRFSS 2023), ZCTA modeled estimates, and ACS-based ZCTA population density from the project's ZCTA atlas; modeled estimates, population-weighted by adults

A regression of ZCTA loneliness on median age, the share of older residents, race and ethnicity shares and the share with a college degree, weighted by adults, shows how much of the pattern lines up with composition. It has an R-squared of 66% (95% CI 63% to 70%), and adding the married share raises it by 3.5 (95% CI 2.4 to 4.8) points of R-squared. Most of the variation in the estimates lines up with the characteristics the model works from, and the married share adds little beyond them.

The ladder scores in the Gallup survey and the modeled loneliness in PLACES are different measures at different grains. The first is how a person rates their life, estimated from interviews, and the second is the modeled share of a ZCTA's adults who feel lonely. Partnered adults rate their lives higher and ZCTAs with fewer married adults have higher modeled loneliness, but the first statement is about people and the second about places built from a model, and one does not stand in for the other. Chapter 7.2, Life evaluation and daily feelings covers the ladder and daily feelings in full.

What these data cannot show

The county and ZCTA patterns are ecological. A county with low turnout need not be one where the same people stayed home and had fewer chances to rise, and a ZCTA with high modeled loneliness need not hold lonelier people than its neighbors. Deaths are outside this chapter (Chapter 6.3, Overdose, alcohol and suicide maps overdose, alcohol and suicide deaths).

The county measures are dated 2010 to 2022 and mobility refers to children born in 1978 to 1983, so the comparisons assume that county differences persist. Census self-response reflects the operation as well as the residents in places counted in person, and turnout is limited to citizens and shaped by state law. The Joint Economic Committee index is a compiled snapshot of inputs from different years, and the Atlas measures describe active Facebook users in 2022, not a census of ties. The regressions are partial associations among correlated measures, with errors clustered on state.

The Gallup partner premium is a comparison between people at one time, in a repeated cross-section and not a panel. Selection into partnership is the main limit, and the panel studies above find evidence of it. Partnered means married or in a domestic partnership, so adults who live with a partner and call themselves single are in the comparison group. County comes from the respondent's ZIP code and is approximate near county lines. The extract is unweighted, so the levels can differ from the weighted figures Gallup publishes, and the evaluative ladder is the only Gallup measure used here.

The PLACES loneliness estimates are modeled, exist only for the states that asked the question, and do not include marital status as an input. The ZCTA characteristics come from an ACS-based table whose year the project files do not document, and the age universe of its married share is not documented either, so the measure is labeled a share of adults. ZCTAs are not ZIP codes. The intervals in the figures describe the bootstrap over counties and do not carry the model's own uncertainty for each ZCTA. Puerto Rico is not in County Health Rankings and is not drawn, and Connecticut appears under its legacy counties in the county files and drops out of the regressions that need the ACS tract file, which uses planning-region codes.

Data sources

County Health RankingsCounty Health Rankings 2024: voter turnout (2020) and census participation (2020 census self-response), membership associations (County Business Patterns, 2021)Years: 2024 release · Weights: 2019-2023 county population for decile means; 2018 county population for correlations · License: CC BY 4.0 (University of Wisconsin Population Health Institute)
Area Deprivation IndexArea Deprivation Index 2023, national rank, tract ranks 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
American Community SurveyACS 2019-2023 tract file: tract population as the weight for county deprivation, and county controls (income, percent Black, poverty) for the regressionsYears: 2019-2023 (5-year) · Weights: 2019-2023 tract population · License: Public domain (U.S. Census Bureau)
Opportunity AtlasCounty outcomes of children born 1978 to 1983 to parents at the 25th percentile (mean adult household income rank, as a percentile)Years: 1978-1983 birth cohorts (2018 release) · Weights: 2018 county population (Social Capital Atlas) · License: Public use (Opportunity Insights and the U.S. Census Bureau)
Social Capital AtlasEconomic connectedness, civic groups per capita and volunteering group membership by county, 2022Years: 2022 (friendship snapshot) · Weights: 2018 county population (Social Capital Atlas) · License: Open for research use (Opportunity Insights)
JEC Social Capital IndexJEC Social Capital Index and subindices by county, 2018 release with components from 2010 to 2016, and religious congregations per 1,000 (2010)Years: 2018 release (components from about 2010-2016) · Weights: 2018 county population (Social Capital Atlas) · License: Public domain (U.S. Government work, Joint Economic Committee)
Gallup U.S. Daily PollGallup U.S. Daily tracking poll, 2009-2017 (the 2008 interviews are left out because of a question-order artifact in the ladder): Cantril ladder score and marital status; aggregates onlyYears: 2008-2017 · Weights: none (unweighted) · License: Proprietary Gallup microdata; respondent-level rows are not published
CDC PLACESCDC PLACES 2025 release, ZCTA estimates of loneliness and depression (BRFSS 2023); ZCTA characteristics are ACS-based values from the project's ZCTA atlas, whose year is not documentedYears: 2024 release (tracts); 2025 release (ZCTAs, 2023 data year) · Weights: 2020 census adult population of the ZCTA · License: Public domain (CDC, U.S. Department of Health and Human Services)

All data sources

Methods

This chapter compares voter turnout and census self-response across ten groups of counties ordered by deprivation, the Gallup life-evaluation score of partnered and unpartnered adults, and modeled loneliness across ZCTAs. All Gallup estimates are unweighted, the 2008 interviews are left out, and the ladder was asked of different shares of respondents in different years, so the pooled figures weight 2009 to 2012 most. Loneliness and depression are modeled estimates, and the ZCTA characteristics are ACS-based values whose year is not documented upstream. Gallup respondent rows are read inside Box and aggregated; nothing at respondent level is published.

County deprivation, civic measures and intervals

County deprivation is the mean of the Area Deprivation Index 2023 national ranks of a county's tracts, weighted by 2019-2023 tract population (Connecticut tracts move to the eight legacy counties; Puerto Rico is excluded), and counties are cut into ten groups of equal numbers of counties.

Decile means of voter turnout and census self-response (County Health Rankings 2024) are weighted by county population, with 95% percentile intervals from 2,000 draws of whole counties with replacement (seed 12345); the same routine gives the correlations across counties, which are weighted by the Social Capital Atlas 2018 population as in the chapter on race, income and mobility, and the differences between deciles with shared draws.

The three county regressions use z-scores with the same weight, and errors clustered on state with t intervals.

Figures explained here: Figure 2, Figure 1 and Figure 3

The Gallup sample is unweighted and its years differ by question coverage

Gallup respondent rows are read inside Box and aggregated; nothing at respondent level is published. All Gallup estimates are unweighted.

The 2008 interviews are left out because Gallup asked political questions immediately before the ladder in 2008, which lowers the ladder score (Deaton 2012); including them leaves the premium at 0.50. The ladder was asked of nearly all respondents in 2010 to 2012, of 85% in 2009 and of 49% from 2013 to 2017, so 61% of the analysis interviews are from 2009 to 2012 and the pooled figures weight those years most.

Partnered means married or in a domestic partnership.

Figure explained here: Figure 4

Adjusted comparisons and their standard errors

Group means and differences have errors clustered on the respondent's county (derived from the ZIP code), and the fixed-effect models are linear regressions with county, year, single-year age and sex fixed effects (and household income and education where stated) and errors clustered on county, with t intervals; a fixed effect for each year of age is used because the ladder is U-shaped in age and a linear age term fits poorly.

Figure explained here: Figure 4

The county Social Capital Index of the respondent's county is joined in Box on the county code and respondents are cut into ten groups of about equal numbers of interviews.

The interaction models cluster on the respondent's state, because the JEC community-health and institutional-health subindexes include state-level inputs.

Figure explained here: Figure 5

Modeled loneliness and depression across ZCTAs

Loneliness and depression are PLACES 2025 release ZCTA estimates (BRFSS 2023), modeled by multilevel regression and poststratification; ZCTAs enter with at least 500 residents, and correlations, decile means and the R-squared are weighted by the ZCTA's adult population and use bootstrap intervals over counties with the same draws.

The R-squared and its gain come from the sums of the weighted regression within each county, resampled with the same draws.

ZCTA characteristics are ACS-based values whose year is not documented upstream, and the table codes a missing value as minus one, which is set to missing.

Figures explained here: Figure 8, Figure 7 and Figure 6

References

University of Wisconsin Population Health Institute (2024). County Health Rankings & Roadmaps Technical Documentation, Version 2. University of Wisconsin Population Health Institute (updated March 18, 2024). www.countyhealthrankings.org/sites/default/files/media/document/2024%20CHRR%20Technical%20Document_1.pdf
U.S. Census Bureau (2024). 2020 Census Operational Timeline and Adjustments Due to COVID-19. U.S. Census Bureau web page, last revised September 16, 2024. www.census.gov/programs-surveys/decennial-census/decade/2020/planning-management/operational-adjustments.html
Putnam, Robert D. (2000). Bowling Alone: The Collapse and Revival of American Community. Simon & Schuster, New York.
U.S. Congress Joint Economic Committee, Social Capital Project (2018). The Geography of Social Capital in America. Social Capital Project Report No. 1-18, Joint Economic Committee, U.S. Congress. www.jec.senate.gov/public/index.cfm/republicans/2018/4/the-geography-of-social-capital-in-america
Chetty, Raj; Jackson, Matthew O.; Kuchler, Theresa; Stroebel, Johannes; et al. (2022). Social capital I: measurement and associations with economic mobility. Nature 608: 108-121. doi:10.1038/s41586-022-04996-4
Deaton, Angus (2012). The financial crisis and the well-being of Americans: 2011 OEP Hicks Lecture. Oxford Economic Papers 64(1): 1-26. doi:10.1093/oep/gpr051
Stutzer, Alois; Frey, Bruno S. (2006). Does marriage make people happy, or do happy people get married?. The Journal of Socio-Economics 35(2): 326-347. doi:10.1016/j.socec.2005.11.043
Lucas, Richard E.; Clark, Andrew E.; Georgellis, Yannis; Diener, Ed (2003). Reexamining adaptation and the set point model of happiness: Reactions to changes in marital status. Journal of Personality and Social Psychology 84(3): 527-539. doi:10.1037/0022-3514.84.3.527
Centers for Disease Control and Prevention (2025). PLACES Measure Definitions: Health-Related Social Needs. CDC PLACES: Local Data for Better Health, web page updated December 7, 2025. www.cdc.gov/places/measure-definitions/health-related-social-needs.html
Centers for Disease Control and Prevention (2025). PLACES Methodology. CDC PLACES: Local Data for Better Health, web page last reviewed December 7, 2025. www.cdc.gov/places/methodology/index.html
Zhang, Xingyou; Holt, James B.; Lu, Hua; Wheaton, Anne G.; Ford, Earl S.; Greenlund, Kurt J.; Croft, Janet B. (2014). Multilevel regression and poststratification for small-area estimation of population health outcomes: a case study of chronic obstructive pulmonary disease prevalence using the Behavioral Risk Factor Surveillance System. American Journal of Epidemiology 179(8): 1025-1033. doi:10.1093/aje/kwu018

How to cite

@misc{singh2026civiclifelonelinessandpartnership,
  author = {Singh, Vishal},
  title = {Civic life, loneliness and partnership},
  year = {2026},
  url = {https://vishalsingh.org/zipcode/civic-life-loneliness-and-partnership},
  note = {In ZIP Code Destiny. Published 9 October 2026.}
}

Reading paths

Outcomes that do not follow the gradient

Participation falls with county deprivation

Voting tracks mobility less than family structure and friendship do

The partner premium in the ladder score

The premium does not shrink where county social capital is higher

Modeled loneliness follows who lives in a ZCTA

What these data cannot show

Data sources

Methods

County deprivation, civic measures and intervals

The Gallup sample is unweighted and its years differ by question coverage

Adjusted comparisons and their standard errors

County social capital and the interaction models

Modeled loneliness and depression across ZCTAs

References

How to cite

Reading paths