Download Data (xlsx), Interactive Maps and Rankings

Social capital is almost surely an important factor driving many of our nation’s greatest successes and most serious challenges. Indeed, the withering of associational life is itself one of those challenges. Public policy solutions to such challenges are inherently elusive. But at present, policymakers and researchers lack the high-quality contemporary measures of social capital available at the state and local levels to even try proposing solutions that are attuned to associational life.

This report describes a new social capital index created to rectify this problem. It details the construction of the index, presents maps summarizing the geographic distribution of social capital, and establishes that the index is consistently—and often strongly—related to a range of economic, social, and demographic indicators. The report also presents the geographic distribution of several subcomponents of social capital, including family unity, family interaction, social support, community health, institutional health, collective efficacy, and philanthropic health.

The Social Capital Project is concurrently providing the state and county data underlying each index, as well as the indices and subindices themselves. It is our hope that the availability of the index will inspire researchers to focus more on social capital and its relationship to other features of economic and social life. And we hope it will aid policymakers as they seek to address the country’s needs.

Among the findings:

• The top fifth of states, in terms of social capital scores, are home to just nine percent of Americans, while 29 percent live in bottom-fifth states.

• We have social capital scores for 2,992 of 3,142 counties, containing 99.7 percent of the American population. Just eight percent of Americans live in the top fifth of these counties, while 39 percent of the population lives in the bottom fifth of counties. Nearly six in ten (59 percent) of Americans live in the bottom two fifths of counties, compared with 24 percent living in the top two fifths.

• The 12 states with the highest social capital scores are distributed across two continuous blocs: nine states running from Utah, through Wyoming and Colorado, across the Dakotas and Nebraska, and over to Iowa, Minnesota, and Wisconsin; and the three Northern New England states of Maine, New Hampshire, and Vermont. These states tend to rank highly across all seven subindices as well. Utah has the highest social capital score, followed by Minnesota and Wisconsin.

• Of the 11 states with the lowest levels of social capital, ten of them fall within a contiguous bloc of states running from Nevada, across the Southwest and South over to Georgia and Florida. New York is the only state in the bottom 11 that is outside this group. Louisiana has the lowest social capital score, followed by Nevada, Arizona, and New Mexico.

• Of the nine states ranked just above this bottom group, seven border and extend the southern bloc, filling out most of the rest of the South. The 17 southern states in the bottom 20 are home to 45 percent of Americans and 74 percent of Americans in bottom-fifth counties. Six in ten (59 percent) of people in the 17 states live in bottom-fifth counties. Only 17 of 1,338 counties in these states are in the top fifth.

• Our indices are not dominated by any single subindex, and our state and county indices appear to be approximating social capital in the same general way.

• Among the component variables underlying the state index, the strongest associations with the index itself across states were for the volunteer rate (0.86), heavy television watching by children (-0.81), the share of adults who made charitable contributions (0.80), the share with emotional and social support (0.80), heavy usage of electronics among children (-0.77), the share of adults that are married (0.75), the share of children living with a single parent (-0.72), and the share of births that were to unwed mothers (-0.71).

• At the county level, the highest correlates of social capital were violent crime (-0.73), the share of children with a single parent (-0.71), the share of adults currently married (0.69), voting rates (0.59), and nonprofits plus congregations (0.57).

• Despite the outsized role that religious communities have played in social capital investment, indicators of religious adherence and commitment were generally weakly (or even negatively) correlated with our social capital scores, both at the state and county levels. This may suggest that social capital organized around religion may be displaced by secular sources of social capital, that the availability of resources provided by secular social capital weakens religious commitment, or that people in distressed places turn to religious communities for the support that is missing in other parts of their lives. This question is a subject for future Social Capital Project research.

• Our social capital indices correlate strongly with earlier social capital indices across states and counties, and with other indices such as the Family Prosperity Institute’s Family Prosperity Index, Opportunity Nation’s Opportunity Index, and the Economic Innovation Group’s Distressed Communities Index.

• We show the correlations of our indices and subindices with 59 state-level and 50 county-level benchmarks reflecting a range of economic, social, demographic, educational, health, and other outcomes.

• Our index is a clear improvement on the Penn State index, based on this benchmarking, but remarkably, Robert Putnam’s state index from Bowling Alone, published nearly two decades ago, has slightly higher benchmark correlations than ours. Because our index captures the health of family life, and because it is based on up-to-date and freely available data (including at the county level), we still prefer it to the Putnam measure. The fact that the correlation between the two state-level indices is 0.81 reassuringly suggests that very different approaches to social capital measurement capture the same essential construct.

Subindices

We transformed the original values of each indicator to “standard scores,” by subtracting the mean value and dividing by the standard deviation of the indicator’s distribution (a measure of variation). The mean of each standard score is zero, and the standard deviation is one. Standardizing the scores puts them on a comparable scale, allowing us to combine multiple indicators despite their initially having different distributions (including different minimum and maximum values). We reversed the polarity of certain measures, such as the share of births to single mothers, so that a larger positive standard score always corresponded with “more” social capital.

Each subindex is comprised of a weighted sum of standard scores. Simply adding multiple standard scores gives them equal weight in contributing to the subindex, but we wanted indicators to be weighted more or less depending on how well they reflected the concept embodied in the subindex. The weights are selected through a statistical technique called “principal components analysis” (PCA). Specifically, the weights are estimated so that the resulting subindex accounts for the maximum possible “variance,” or variability, across the original scores. Some information in the original set of indicators is lost by using this “first principal component score” as the subindex, but the loss is minimized versus any other set of weights. It is analogous to finding the best angle from which to photograph a three-dimensional object so that the two-dimensional rendering retains the most information.

In the domain of family health, we created a “family unity” subindex and a “family interaction” subindex. The family unity subindex combines state-level data from the American Community Survey (2012-2016) on the share of births that are to unwed mothers (weight of 0.57), the percentage of children living in families headed by a single parent (0.60), and the percentage of women ages 35-44 who are married (and not separated) (0.57). (The weights could, in theory, range between -1.0 and 1.0, and they reflect the extent to which an indicator is correlated with the subindex itself. Ideally, the weights should be sizable and should all have the same sign.) ²⁴ The subindex accounts for 91 percent of the variability across the original three variables of which it is comprised.

The family interaction subindex combines data from the 2016 National Survey of Children’s Health on the share of children ages 0-5 read to every day by a family member (weight of 0.47), the share of children who watch television or videos or play video games at least four hours a day (0.65), and the percentage of children who use computers, cell phones, and other electronic devices for purposes other than schoolwork at least four hours a day (0.60). Our family interaction subindex accounts for 68 percent of the variability across the original three variables.

We created a social support subindex, comprised of several indicators from multiple sources. It includes the share of adults who sometimes, rarely, or never “get the social and emotional support [they] need,” taken from 2006 and 2010 Behavioral Risk Factor Surveillance System data (weight of 0.50). It also includes, from the 2013 Civic Engagement Supplement to the CPS, the share of adults who do favors for neighbors at least once a month (0.49) and the share who trust most or all of their neighbors (0.54). The last indicator is the average number of “close” friends that adults report having in the 2008 Civic Engagement Supplement to the CPS (0.47). The resulting index accounts for 70 percent of the original variability across the four variables that comprise it.

Our community health subindex incorporates information on the share of adults who reported volunteering for an organization in the past year (weight of 0.33), the share who attended a public meeting to discuss community affairs (0.38), and the share who worked with neighbors to improve the community (0.39), all from the 2015 Volunteer Supplement to the CPS. It also includes the share of adults who served on a committee or as an officer of a group in the past year, from the 2013 Civic Engagement Supplement to the CPS (0.38). From the 2008 Civic Engagement Supplement to the CPS, we include the share who attended a public meeting where political issues were discussed (0.39) and the share who participated in a march, protest, rally, or demonstration (0.29). Our community health subindex accounts for 65 percent of the variability across eight indicators.

Further, we estimate, from 2015 County Business Pattern data on establishments, membership organizations per capita (weight of 0.30). Finally, we include a measure of non-profit organizations per capita (weight of 0.36). This was created by summing registered non-religious not-for-profit organizations per capita and religious congregations per capita. The former is from the December 2015 Internal Revenue Service Business Master File (accessed through the Urban Institute’s National Center for Charitable Statistics). Most faith-based organizations, excepting colleges and health care organizations, are not required to register with the IRS. Only half of religious congregations do so, and the share varies by congregation. ²⁵ We therefore added congregations per capita from the 2010 Religious Congregations and Membership Study, conducted by the Association of Statisticians of American Religious Bodies and accessed through the Association of Religion Data Archives.

Finally, we created an institutional health subindex. This subindex combines the rate at which citizen adults of voting age cast ballots in the 2012 and 2016 presidential elections (averaged over the two years, weight of 0.38), the rate at which residents returned the 2010 decennial census questionnaire through the mail (0.44), and the share of adults with “great” or “some” confidence in corporations (0.49), the media (0.38), and public schools (0.53) to do what is right. The voting data is from the U.S. Election Assistance Commission annual reports, the census response rates are from the Census Bureau, and the confidence measures are from the 2013 Civic Engagement Supplement to the CPS. The institutional health subindex accounts for 48 percent of the variability across the original five indicators. The lower proportion that it explains relative to the other subindices may reflect the weaknesses in the presidential voting and census response indicators discussed above.

Stand-Alone Indicators

We included in our state-level social capital index two stand-alone indicators to represent two other dimensions of associational life. The violent crime rate was included to reflect the level of “collective efficacy” (or conversely, of social disorganization). The idea is that communities high in social capital are better positioned to informally police their own community and enforce pro-social norms, and their residents are less likely to do harm to one another. ²⁶ Violent crimes are better reported than crimes generally, which is why we do not use a broader measure. The source for this measure was the Federal Bureau of Investigation’s Uniform Crime Reporting Program.

The second stand-alone indicator reflects philanthropic health—the share of adults who gave more than $25 in the past year to “charitable or religious organizations.” By setting the threshold for giving low, this measure ensures that cross-state differences are not driven by income concentration at the top, where charitable giving is also somewhat concentrated. ²⁷ This measure comes from the 2015 Volunteer Supplement to the CPS.

Combining the Dimensions of Social Capital

The final step was to create the index itself. We standardized the five subindex scores and the two stand-alone indicators to put them all on a common scale. We then ran principal components analysis on these seven variables to create weights for each of them: family unity (0.38), family interaction (0.41), social support (0.45), community health (0.33), institutional health (0.36), collective efficacy (0.28), and philanthropic health (0.40).

Each state’s social capital index score was computed by taking the weighted sum of the seven standard scores and then standardizing this weighted sum. Index scores range from -2.2 to 2.1; a score of, say, 1.5 means that a state lies one-and-a-half standard deviations above the mean index score across states. Roughly, its social capital levels are higher than the average state’s social capital by an amount 1.5 times the typical gap between a state and the average.

There is an unavoidable element of arbitrariness in creating a one-dimensional index to reflect a concept as complex and diffuse as “social capital.” The usefulness of our index depends on its ability to represent a potentially important factor affecting a range of social, economic, and health outcomes. The index represents a “noisy” measure of a fuzzy concept. But it reflects those aspects of its constituent indicators that all measure the same “thing,” and lets them contribute to the measure insofar as they reflect that thing. Our social capital index accounts for 56 percent of the variability across the two stand-alone indicators and the five subindices (each of which accounts for 48 to 91 percent of the variability across its constituent indicators).

Constructing a New County-Level Social Capital Index

While the state index includes seven dimensions of social capital, our county index includes only four because fewer variables are available at the county level. These include three subindices—two of which contain their own subindex—and one stand-alone indicator. The construction of the county-level index is more complicated than for the state-level index. Table 2 shows the variables that go into the index. (See Appendix Table A1 for the full list of county-level variables we considered.)

Subindices

As when we created the state index, we transformed all original values to standard scores, though this time the mean and standard deviation applied to the distribution of values across counties rather than across states. As with the state-level subindices, our county-level subindices are weighted sums of standard scores, with the weights determined through PCA. These weights need not be the same as those produced from state-level data, where it is variation across states that is being analyzed.

We created the same family unity subindex as we did at the state level. The subindex combines county-level data from the American Community Survey (2007-2011²⁸ and 2012-2016) on the share of births that are to unwed mothers (weight of 0.52), the percentage of children living in families headed by a single parent (weight=0.62), and the percentage of women ages 35-44 who are married (and not separated) (weight=0.59). Reassuringly, these weights are very similar to those produced in the state-level analyses, suggesting that the state- and county-level subindices are measuring the same underlying construct. It does explain less of the variability in the original three variables than the state-level subindex does—73 percent instead of 91 percent.

We also created a community health subindex, though due to data availability issues, the county subindex incorporates less information than the corresponding state-level one. We were concerned that the available county-level indicators of community health did not fully capture the underlying concept. In particular, we lacked the CPS indicators of informal civil society and activities requiring a time commitment that were available at the state level—working together with neighbors, attending public meetings, serving on committees or as officers, volunteering, attending political meetings, and participating in demonstrations. We worried about this omission, in particular, because professionalized services offered through membership organizations and other nonprofit groups might be expected to crowd out informal and time-intensive volunteer activities, potentially leaving the stock of social capital thinner than it might have been. Inherently, formal organizations that serve members’ or clients’ interests allow people to “farm out” social capital activities. To include only a measure of the health of formal organizations would penalize places where community involvement is more informal.

To resolve this concern, we first went back to the state data and created a new subindex of “informal civil society” for each state. The subindex score was the first principal component score combining the six CPS variables above. ²⁹ We then assigned this subindex score to every county within a state. In other words, the only variation in the subindex score is between states, and all counties within a state get the same score.

Next, back in the county data, we created five different candidate subindices, using different combinations of the informal civil society subindex score, membership organizations per capita, non-religious non-profit organizations per capita, congregations per capita, and the combination of non-religious non-profits and congregations. We computed, for each candidate subindex, the population-weighted average subindex score across a state’s counties. Then we correlated each of these state averages with the state-level community health subindex. We selected the subindex out of the five candidates that produced the strongest correlation.

The final county-level community health index combines non-religious non-profits per capita (weight of 0.70), congregations per capita (0.48), and the informal civil society subindex (0.53). The population-weighted average of this subindex across a state’s counties correlated at 0.97 with the state-level community health subindex. For context, the correlation of the state-level subindex with the version of the county community health subindex we favored prior to adding in the informal civil society subindex was 0.75. The county-level community health subindex accounts for 55 percent of the variability in the three original variables that go into it.

Finally, we included an institutional health subindex. As with the community health subindex, we were concerned about the incomplete data we had at the county level. In this case, we lacked information about confidence in institutions. We took the same approach as for community health. In the state data, we created a confidence subindex that included the three institutional confidence variables. ³⁰ We assigned every county in a state the state’s subindex score. Then we created three versions of a county-level institutional health index, using different combinations of presidential voting rates, census response rates, and the confidence subindex.

As before, we created population-weighted state averages across a state’s counties and compared them to the state-level institutional health index. The version that correlated most strongly included presidential voting rates (weight of 0.63), census response rates (0.41), and the confidence subindex (0.66), accounting for 44 percent of the variability in those three measures. ³¹ The correlation of the population-weighted state average across counties with the state-level institutional health subindex was 0.98.

We did not attempt to create subindices at the county level for family interaction or social support, lacking data. ³²

Stand-Alone Indicator

The county-level social capital index includes one stand-alone indicator. As for the state-level index, the violent crime rate was included to reflect the level of collective efficacy in a county. It comes from the Federal Bureau of Investigation’s Uniform Crime Reporting Program.

The charitable giving measure from the CPS is not available at the county level, so it is not included as a stand-alone indicator.

Combining the Dimensions of Social Capital

Computing the county-level index was also a bit more involved than for the state index. We standardized the three subindex scores and the collective efficacy stand-alone indicator to put them all on a common scale. We then ran PCA on these four variables. The weights were 0.53 for family unity, 0.47 for community health, 0.49 for institutional health, and 0.51 for collective efficacy. We took the weighted sum of the four standard scores to get the first iteration of the index, which accounted for 51 percent of the variability in the original four constituent measures.

However, information on violent crime rates was missing for 178 counties (out of 3,142). We were able to compute scores for 103 of these counties by creating an alternative index that left out violent crime. (The weights were 0.55 for family unity, 0.56 for community health, and 0.62 for institutional health. The subindex accounted for 56 percent of the variability in the three original variables.) Where a county lacked a score using the original index, we gave it the score on the alternative index. These two indices were correlated with each other at 0.94, so where states ranked on one was largely where they ranked on the other.

The final county-level index scores range from -4.3 to 2.9, indicating greater dispersion than exists across states.

To assess how the county-level and state-level indices might differ from one another, we created another state-level index using only the three subindices and the stand-alone violent crime indicator that are in the county index. This index correlated with the fuller state-level index at 0.96. ³³ We also computed for each state the population-weighted average across counties of the county-level social capital index. The correlation between it and the state-level social capital index was 0.95, and the correlation between it and the state-level index based on the county-level methods was 0.98. Thus, the thinner county-level index likely ranks counties very similarly to the way in which the fuller state-level index would rank them.

In sum, our state index captures a fuller set of social capital indicators than any previous effort. We could not find a reliable measure of generalized trust at the state or county levels, but we believe we have covered most of the essential domains discussed by past social capital theorists. We considered including measures of segregation by race and income in our indices. The idea is that places where different types of people largely live apart are likely to be missing out on some benefits of social capital. However, research suggests that more diversity actually tends to reduce levels of social capital. ³⁴ Intuitively, it is more difficult to engage with people when they are not “like” us. In the end, we decided not to incorporate segregation into our indices. We view segregation as having an indeterminate effect on levels of the many dimensions of social capital. It seems more likely that segregation affects the distribution of social capital within a state or county. ³⁵

Findings

Table 3a lists the state social capital index ranks and the rankings on the individual subindices of the index. Table 3b lists the county social capital index and subindices as national percentiles. Figure 1 displays the state social capital scores in a map, and Figure 2 displays the county-level data. We have social capital scores for 2,992 of 3,142 counties, containing 99.7 percent of the American population. Before examining the places with the highest and lowest social capital scores, we provide some initial details about the distribution of social capital in America.

Table 3b. County-Level Index and Subindices as National Percentiles

^{Click a column to reorder by index or subindex county percentile (100th percentile is best, 0 is worst). For county rankings within states, click a county in Figure 2. Use the search box to jump to a specific county or state.}

The maps in Figures 1 and 2 display states and counties broken out into five (roughly) equally-sized groups—ten states per group and 598 counties. ³⁶ These groups do not contain the same number of people, however. The states with the lowest social capital include 29 percent of the nation’s population, while the top grouping is home to just nine percent of Americans. Over half the population (56 percent) is in the lowest two groups of states, while 21 percent is in the top two groups. At the county level, 39 percent of the population in non-missing counties lives in the bottom fifth, while just eight percent lives in the top fifth. Nearly six in ten (59 percent) of Americans live in the bottom two fifths of counties, compared with 24 percent living in the top two fifths.

Across states, the social capital scores are strongly correlated with each subindex. The correlations are 0.89 for social support subindex scores, 0.82 for family interaction, 0.80 for philanthropic health, 0.76 for family unity, 0.72 for institutional health, 0.65 for community health, and 0.55 for collective efficacy. The 21 correlations between the seven subindices are all positive, except that community health and collective efficacy, surprisingly, are correlated at -0.11. Otherwise, the correlations range from 0.17 (family unity and community health) to 0.74 (family interaction and community health).

At the county level, social capital scores are also strongly correlated with all four subindex scores. The correlations are 0.76 for the family unity subindex, 0.73 for collective efficacy, 0.71 for institutional health, and 0.65 for community health. The fact that these correlations are all fairly strong means that our state and county indices do not simply reflect a single dimension driving the results. The correlations between the four subindices range from 0.24 (family unity and community health) to 0.47 (family unity and collective efficacy).

Diving deeper into the components of the indices at the state level, the indicators with the strongest correlation to social capital were the volunteer rate (0.86), heavy television watching by children (-0.81), the share of adults who made charitable contributions (0.80), the share with emotional and social support (0.80), heavy usage of electronics among children (-0.77), the share married (0.75), the share of children living with a single parent (-0.72), and the share of births that were to unwed mothers (-0.71). While not included in the index, the share who trust most of their neighbors was correlated at 0.86 with it. At the county level, the highest correlates of social capital were violent crime (-0.73), the share of children with a single parent (-0.71), the share of adults currently married (0.69), voting rates (0.59), and nonprofits plus congregations (0.57). The importance of the absence of many of the key state-level variables at the county level is evident.

A few state-level indicators had low correlations with the index, including membership organizations per capita (0.07), confidence in the media to do what is right (0.20), having participated in a march or demonstration (0.21), and non-religious non-profits and congregations per capita (0.29). Three of these indicators go into the community health subindex, which may explain why it is less strongly correlated with social capital scores than most of the other subindices. Relatedly, the Penn State social capital index relies on variants of the membership organization and non-profits indicators. Our replication of the Penn State index correlates only at 0.37 with our index, as we will see below. At the county level, census response rates—one of the four Penn State components—was correlated with our index at only 0.26, but the correlation between the 2014 Penn State index and our county index was 0.56. We view this as evidence that the relatively thin county-level indices do not measure social capital as strongly as our richer state-level index does.

Figure 1. Social Capital Index and Subindex Scores by State

^{Coming Soon: Social Capital Index data at the metropolitan/micropolitan, commuting zone, and congressional district levels. Click here to open in a new window}

Figure 2. Social Capital Index and Subindex Scores by County

^{Coming Soon: Social Capital Index data at the metropolitan/micropolitan, commuting zone, and congressional district levels. Click here to open in a new window}

Table 4. Benchmarking the State Social Capital Index (Bivariate Correlations)

^{Source: Social Capital Project. Download Data (xlsx)}

The first group of indicators includes three indices of social capital. Our index is highly correlated with the Putnam and Alesina/La Ferrara indices, as well as the Family Prosperity Institute’s Family Prosperity Index. It is more strongly related to these than the Penn State index is. It has essentially the same strength of relationship with the Alesina/La Ferrara index as does Putnam’s index, and because Putnam’s index does not incorporate family unity, our index is more strongly related to the Family Prosperity Index than Putnam’s is.

The next set of indicators relates to employment. For all four benchmarks, the correlation with our index is greater than 0.5 (or less than -0.5), and for all four the correlation with our index is larger than the correlation with the Penn State index. Two of the four benchmarks are more strongly associated with our index than with Putnam’s.

The five benchmarks in the next group are indicators of income and poverty. Three of five correlations with our index are below -0.5 or above 0.5. For each benchmark, the correlation with our index is stronger than the correlation with the Penn State index, and that is true of the correlation with the Putnam index for four out of five benchmarks.

Next, we show nine benchmarks related to segregation by income, inequality, and economic mobility. For five of these benchmarks, the correlation is above 0.5 (in absolute value), including the share of the state’s population in ZIP codes deemed “economically distressed” by the Economic Innovation Group, and the opportunity score assigned to a state by Opportunity Nation. States with more inequality have lower social capital scores. States with more social capital have lower relative intergenerational immobility—the relative positions of children are more scrambled in adulthood given where they started. However, absolute mobility (exceeding one’s parents’ income) is only slightly higher in states with higher social capital scores. States where poor residents are concentrated have less social capital, but concentration of rich people is not related to social capital.

The correlations between inequality benchmarks and the Penn State index have the opposite sign as those for our and Putnam’s indices. Only for absolute mobility is the Penn State correlation stronger than the correlation using our index. The Putnam index is related to all nine benchmarks in the same way that our index is. In fact, it more strongly correlates with five of the nine measures than our index does (though the difference is sizable for only three benchmarks).

The next three benchmarks are related to education. Our index is strongly correlated with the share of a state’s population that graduated from high school, but less strongly correlated with obtaining a bachelor’s degree or with on-time high school graduation rates. Both the Putnam and the Penn State indices are more strongly correlated with graduating from college than is our index, but the Penn State index is much less strongly correlated than ours is with graduating from high school.

We show nine benchmarks related to health. Our index is strongly correlated with being in only fair or poor health, dying prematurely, having diabetes, and giving birth to a low-birthweight baby. Mortality from “deaths of despair”—involving alcoholism, drug overdoses, or suicide—has no relationship with our index. Our index is more strongly correlated in the expected direction with six of the nine health benchmarks than is the Penn State index, and the Penn State index has the opposite sign we would expect for premature mortality and smoking. Putnam’s index generally has somewhat higher correlations with health outcomes than ours.

The next group of benchmarks reveals that age has little to do with social capital. The exception is that states with fewer children have higher Penn State social capital scores, which is counter to the near-zero correlations shown by our index and Putnam’s.

In the next grouping, we cross the Census Bureau’s racial categories with its Hispanic ethnicity question. States with more non-Hispanic whites have higher social capital, and states with more African Americans and Hispanics have lower social capital. The Putnam correlations are generally consistent with ours and are stronger for four of the eight groups. The Penn State index has the opposite sign relative to our and Putnam’s indices for the share of the population that is African American.

Of note, in preliminary research, we also found that states with a large share of residents who identified their ancestry as “American” also have substantially lower social capital scores (on par with the correlation between social capital scores and the Hispanic share). Self-identified Americans are overwhelmingly comprised of southern whites. ⁴⁰

The reasons for correlations between ethnic shares and social capital are surely complicated. Historical oppression and ongoing discrimination could weaken social capital investment and institutions, particularly to the extent that it leaves a group with high poverty rates and concentrated residentially. Immigrants may bring to the United States a mix of values that reflects the history and culture of their countries of origin or that reflects the unique values of the self-selected group of people that left their homeland for a new life—values that might strengthen or weaken social capital. Alternatively, communities with many newcomers to the country might be in a state of flux, as those newcomers assimilate and as institutions such as schools and churches experience shifts in composition. That could weaken social capital. Indeed, diversity itself may weaken some dimensions of social capital, as some research suggests, simply by creating barriers to easy social cooperation.

That places with larger black populations have lower social capital may reflect the deleterious consequences of racial segregation. The next grouping in Table 4 shows that states with greater segregation between blacks and non-Hispanic whites have lower social capital. Similarly, the lower social capital in states with bigger Hispanic populations may reflect the unique challenges of immigrant communities. The share of the population comprised of foreign-born residents is associated with lower state social capital levels. ⁴¹ Our social capital index reflects these correlations somewhat more strongly than do the other two indices, except that the Penn State index indicates a stronger correlation for the share foreign born.

The final grouping is a grab bag of 13 indicators. More social capital is strongly associated with more internet subscribers (counter to the notion that technology hurts social capital), lower average temperatures, and being further from the equator. The latter two are obviously related, and they are unsurprising given the pattern of northern states having high social capital levels and southern states having low levels.

The other correlations are weaker, but social capital is higher in states that are less dense, more rural, and higher above sea level, and in states with more homeowners, newer housing, and shorter commutes. It is mostly unrelated to greater state and local government spending (counter to the notion that greater independence from the federal government yields greater social capital) or to net migration (counter the notion that more-rooted communities have stronger social capital). For eight of the 13 indicators, the correlations are stronger using the Putnam index than ours. For six of the 13, the correlations using the Penn State index are stronger than using ours. For another five, the Penn State correlations are in the opposite direction versus our and Putnam’s indices.

Of the 59 benchmarks in Table 4, the correlation with our social capital index was over 0.5 for 25 of them. That was also true of 24 correlations of benchmarks with our family unity subindex, and the social support subindex was close behind at 22. The correlation was over 0.5 for 20 of the 59 benchmarks when comparing them with the philanthropic health subindex, and that was true in 18 cases looking at the family interaction subindex.

Of the 25 benchmark correlations with our index that were over 0.5 (or below -0.5), 12 were stronger using the index than using any of the seven subindices. Eight times, the family unity subindex is more strongly correlated with a benchmark than the index, and that is true three times for the social support subindex, two times for the philanthropic health and collective efficacy subindices, one time for the family interaction and community health subindices, and never for the institutional health subindex. The fact that the index usually appears to provide more predictive power than the individual subindices is another indication that the index measures something corresponding with social capital in a way that is an improvement on the individual subindices. It is also reassuring that the correlations of the index with benchmarks do not appear to be driven by one or two of the subindices.

Of the 18 benchmarks where the correlation with the Penn State index is greater than 0.30 (or less than -0.30), our index has a stronger correlation with the benchmark in nine instances. Of the 18 benchmarks where the correlation with the Putnam index is greater than 0.60 (or less than -0.60), our index has a stronger correlation in just seven instances. Our index appears to improve on the Penn State methodology, but if the association with benchmark variables is the sole criterion, our index is not an obvious improvement on the Putnam index.

However, we prefer our index for several reasons. First, substantively, our index, unlike the Bowling Alone index, reflects the health of family life—a dimension of social capital that has been overlooked in past research on social capital per se. We equate high levels of social capital with the health of our associational life—our “middle layers” between the individual and the state. As Yuval Levin has described, the middle layers

begin in loving family attachments. They spread outward to interpersonal relationships in neighborhoods, schools, workplaces, religious communities, fraternal bodies, civic associations, economic enterprises, activist groups, and the work of local governments. They reach further outward toward broader social, political, and professional affiliations, state institutions, and regional affinities. And they conclude in a national identity that among its foremost attributes is dedicated to the principle of the equality of the entire human race. ⁴²

Whatever social capital is, it is difficult to justify leaving out the part that exists and is maintained in families. Practically, our index is also based on up-to-date data that is freely available (and, therefore, readily updateable). All this said, it is striking that Putnam’s landmark attempt nearly two decades ago to measure social capital holds up so well that it correlates with contemporary benchmarks at least as well as our measure does, and often better. And the fact that our and Putnam’s different approaches to measuring social capital produced similar results—the correlation between the two measures is 0.81—is reassuring; both indices apparently capture the same underlying construct.

Unfortunately, the Penn State county-level index does not appear so robust.

County Validation

In Table 5, we show bivariate correlations for county-level variables, comparing our index, the Penn State index, and our subindices to 50 different benchmarks. The Penn State index is the 2014 version of the index available on the website of the Northeast Regional Center for Rural Development. ⁴³

The first row shows the correlation of our index and subindices with the Penn State index. The correlation of the two indices is 0.56, and our community health index has the strongest association with the Penn State index. That reflects the fact that both are built, in part, on IRS data on nonprofit organizations. Similarly, there is a moderate correlation between our institutional health subindex and the Penn State index, since both are partly built on presidential voting rates and census response rates. The family unity and collective efficacy subindices are only weakly related to the Penn State index, since the latter does not attempt to measure those dimensions of social capital.

Table 5. Benchmarking the County Social Capital Index (Bivariate Correlations)

^{Source: Social Capital Project. Download Data (xlsx)}

The correlation of our index with the benchmarks is greater than 0.5 (or less than -0.5) for 17 of the 50 benchmarks, while the Penn State correlations are never above that threshold. For 38 of the benchmarks, our index has a stronger correlation than the Penn State index. The Penn State correlations are stronger for nine benchmarks, and the two indices’ correlations are signed in opposite directions for three benchmarks. Of the 12 benchmarks that are correlated with the Penn State index at a level greater than 0.30 (or less than -0.30), the correlation of the benchmark with our index is stronger in nine instances.

Our social capital index is more strongly correlated with all three of our employment benchmarks than is the Penn State index and with all five of our income and poverty benchmarks. For five of the eight, the correlation with our index is greater than 0.5 (or less than -0.5).

Next, Table 5 shows eight benchmarks related to segregation by income, inequality, and economic mobility. The correlation between our index and the benchmarks is below -0.5 for four of the eight. There is little relationship between the extent to which a county’s rich residents are concentrated together and its social capital score. The three education benchmarks are all correlated with our index in the expected direction, and the association with high school graduate shares is particularly large.

Three of nine health benchmarks are correlated with our index at less than -0.5. All nine correlations are in the expected direction. As in our state validation analyses, our county index is not strongly correlated with age. However, the Penn State index is moderately correlated with the three age benchmarks.

The race/ethnicity correlations are uniformly consistent with the state-level validation results, as are the black-white segregation and foreign-born correlations. The eight “grab-bag” benchmarks all show similar correlations with our county index as they do with our state index.

Across the 50 county-level benchmarks, the correlations with our subindices were generally lower than was the case at the state level, possibly reflecting the relative dearth of county-level measures available related to social capital. The family unity subindex was correlated with 16 benchmarks at a level of more than 0.50 (or less than -0.50). Just five correlations were that strong using the institutional health subindex, only two using the community health subindex, and none using the collective efficacy subindex.

Looking at the 17 benchmarks where the correlation with our social capital index is above 0.5 (or below -0.5), our social capital index is more strongly correlated with the benchmark than any of the four subindices are for 11 of them. For six benchmarks, at least one subindex correlates more strongly than does our overall index. In five instances, it is the family unity subindex that is more strongly correlated. The community health subindex is more strongly correlated with one benchmark than the overall index is. Once again, the index appears to provide more predictive power than the individual subindices.

To sum up, while our county-level social capital index is an unambiguous improvement on the Penn State index, it does seem to measure social capital levels less well than our state-level index does. It would be better to have more county-level data related to social capital than currently exists.

Conclusion

Despite the theoretical importance of social capital for understanding our national challenges and for crafting effective public policies, past efforts to measure the concept have suffered from a number of problems. These range from overly narrow or broad conceptualizations of “social capital,” to data unavailability at the state or county levels, to out-of-date data, to the inaccessibility of non-public data. Our state and county social capital indices rectify these problems to a large degree. It is our hope that by making our data publicly available, researchers may be inspired to relate social capital to any number of other aspects of American life and to policy-relevant outcomes.

Social capital is markedly unequally distributed across the United States. A clear “north-south” divide is apparent, and the clustering of states into similar contiguous blocs suggests that geographic differences may have deep-seated roots in historical immigration and internal migration patterns, regional culture, and perhaps even features of climate and topography.

While our county-level index is a clear improvement on the only other county measure available, from Penn State University, the evidence we have accumulated suggests that a lack of data at the county level on indicators related to social capital reduces the accuracy of local estimates. Adding more social capital measures to surveys such as the American Community Survey or the Current Population Survey could greatly improve county estimates.

Nevertheless, we are reassured by a number of findings detailed in this report. When we mimicked the county-level approach at the state level, the resulting state index and our preferred state index were correlated at 0.96. When we averaged county-level scores in each state, weighting by the county population, the resulting state averages correlated with our preferred state index at 0.95, and the correlation between the state estimates mimicking the county approach and the state averages of county estimates was 0.98. The fact that Robert Putnam obtained similar results nearly 20 years ago using very different data and measures also suggests that it is possible to measure something meaningful that corresponds with social capital

Having constructed these indices, the Social Capital Project, in future work, will attempt to explain the geographic patterns identified here and to explore in greater depth the relationship between social capital and a variety of outcomes. There is clearly much to learn, and just as clearly, the regional inequalities we have uncovered demand that policymakers and researchers better understand the distribution of social capital in America.

Appendix

In this appendix, we provide greater detail about the process used to select the measures that go into our social capital indices and subindices. Our objective is to be as transparent as possible about the process. We also detail the source data for our benchmarks. Finally, we provide additional maps of social capital and its components at the state and county levels.

State Analyses

We began with the list of indicators shown in Appendix Table A1, below. We standardized all variables and reversed the polarity (multiplying by -1) for 21 of them so that higher standard scores always indicated more “social capital.”

We started with some initial analyses estimating Cronbach’s alpha and using principal components analysis, using both county- and state-level analyses. These gave us a general sense of the domains of social capital that appeared using inductive methods. We then attempted to determine how to best measure the underlying concept reflected in these domains.

Family Health

We constructed four measures of “adverse childhood experiences,” [⁴⁴] from the National Survey of Children’s Health (NSCH). ⁴⁵ One indicated the average number out of nine items experienced by a child, one dropped three items not related to experiences in a child’s household and indicated the average number of the remaining six experienced, one measured the share experiencing at least one of the six experiences, and a fourth measured the average number of items from four related to parental behavioral problems. They all correlated with each other at above 0.90. We retained the one including just four problems, which excludes parental divorce (already addressed in other variables), parental death, poverty (only indirectly about family interactions), neighborhood violence, and discrimination (both involving outside-the-home experiences).

The share of families with children with a single parent and the share of children in families headed by a single parent, both from the Census Bureau’s American Community Survey (ACS), were correlated at 0.99. ⁴⁶ We dropped the former.

Looking at the percent of births to unwed mothers, percent of women married, and percent of children in single-parent families, all from the ACS, the alpha was 0.95. Adding the percent of women never married (ACS) and the percent of children experiencing divorce (NSCH) lowered it to 0.88. Adding only the percent never married lowered it to 0.93. Adding both plus the share eating dinner with their family (from the September 2013 Volunteer Supplement to the Current Population Survey ⁴⁷) lowered it to 0.84. Adding only the nonmarital fertility rate (from the ACS) lowered it to 0.87.

Looking at the percent of births to unwed mothers, the nonmarital fertility rate, and the percent of children in single-parent families, the alpha was 0.83. It rose to 0.85 if the percent of children experiencing divorce was added. It rose to 0.87 if percent married was added instead. It fell if the percent in one-person households, percent never married, or having dinner with one’s family was added. It rose to 0.95 if the nonmarital fertility rate was dropped. The percent of births that were to unmarried women correlated much more strongly with other family indicators than the nonmarital fertility rate.

Appendix Table A1a. Social Capital Indicators Considered (State and County-Level)

^{Source: Social Capital Project.}

Appendix Table A1b. Social Capital Indicators Considered (State-Level Only)

^{Source: Social Capital Project.}

Dinner with family was correlated with percent married and percent never married, but not with other family indicators. A second indicator involving eating meals with family (from the NSCH) also was not well correlated with other variables and was often signed the wrong way.

Looking at adverse childhood experiences, reading to children every day (NSCH), heavy exposure to TV/videos/video games (NSCH), heavy exposure to electronic devices (NSCH), and dinner with family, the alpha was only 0.61. Dropping adverse childhood experiences and reading, the alpha rose to 0.71. The alpha using only reading, TV, and electronic devices was 0.76.

Ultimately, we decided to create a family unity index from the percent of births to unmarried women, the percent of children in single-parent families, and the percent of women who were married (alpha=0.95). These variables are all available at the county level. We also created a second family unity index using the weights on the three variables produced in the county-level principal components analysis. These weights were very similar to those using PCA at the state level, and the two indices correlated at 0.9999.

We also created a “family interaction” index including reading to children, TV viewing, and electronic device viewing. These variables are not available at the county level.

We considered including adverse childhood experiences as its own index, but it was less strongly correlated with the other subindices, and it is unavailable at the county level.

Social Support

The average number of friends variable from the 2008 Civic Engagement Supplement to the Current Population Survey ⁴⁸ was correlated at 0.91 with the variable indicating the share with at least five friends (from the same survey), and it was more strongly correlated than the latter with the share having daily contact with family and friends (from the September 2013 Volunteer Supplement). We dropped the variable giving the share with at least five friends. The share of adults with emotional support (from the 2006 and 2010 Behavioral Risk Factor Surveillance System data) ⁴⁹ and the parenting support measure in the NSCH were correlated at 0.62. We dropped the latter.

Looking at emotional support, trust in neighbors, talking to neighbors, doing favors for neighbors, average number of friends, and contact with family and friends, the alpha was 0.83. ⁵⁰ Adding working with neighbors, from the September 2015 Volunteer Supplement, raised it to 0.84. ⁵¹ Adding dinner with family members lowered it to 0.82, while adding both left it at 0.83. Adding both plus the violent crime rate (from the Federal Bureau of Investigation Uniform Crime Reports) ⁵² raised it only to 0.84. Dropping average number of friends, doing favors for neighbors, or trusting neighbors lowered it below 0.80. Adding the percent of women married raised the alpha to 0.84. Further adding the percent of women never married did not change it. Adding the percent in one-person households lowered it.

Looking at emotional support, trust in neighbors, talking to neighbors, doing favors for neighbors, average number of friends, and contact with family and friends, the alpha was 0.83. It rose to 0.85 if talking to neighbors and contact with family and friends were dropped.

We created a social support index including emotional support, trust in neighbors, doing favors for neighbors, and average number of friends (alpha=0.85). Only emotional support is available at the county level, but it is missing for several hundred counties, and it comes from a survey that is not necessarily representative of every county.

Community Health, Religious Health, Civic Engagement

We created four variables related to participation in groups, using the September 2013 Supplement to the Current Population Survey. They ended up highly correlated with one another. We dropped the measure that excluded participation in recreation groups like golf and tennis clubs. Instead of using the separate variables we created for participation in religious groups and in nonreligious (and non-recreation) groups, we decided to use the variable for participation in all groups.

Church attendance and participation in prayer groups, both from the 2014 Pew Religious Landscape Study, were highly correlated (0.92). ⁵³ We dropped participation in prayer groups. Church attendance, religious adherence rates (from the 2010 U.S. Religion Census: Religious Congregations and Membership Study) ⁵⁴, and participation in religious groups were all correlated at above 0.50 with each other, but only participation in religious groups had any moderate positive correlation with other civic engagement variables. Looking at participation in religious organizations, church attendance, and religious adherence rates, the alpha was 0.79, and it fell if any of the three were excluded. These findings will be the subject of future Social Capital Project analyses. We used none of these measures in the end.

We looked at the Penn State measure of associations per capita, but we also created two of our own, splitting recreation and leisure establishments (like bowling centers and golf clubs) from membership organizations. (All from the Census Bureau’s 2015 County Business Patterns data. ⁵⁵) All three of these measures were correlated with the non-religious civic engagement measures. The Penn State measure and our measure for membership organizations correlated at 0.90.

The civic engagement measures from the Current Population Survey supplements were all strongly correlated with each other, and the registered non-religious nonprofit measure (from IRS data) also correlated well with the other non-religious civic engagement measures. ⁵⁶ The measure adding congregations to non-religious nonprofits generally was a bit more strongly correlated with these other measures, so we used that measure instead. Using that measure also tended to produce larger alphas with other variables than using non-religious nonprofits and congregations separately.

Looking at group participation, membership organizations per capita, recreation and leisure establishments per capita, non-religious nonprofits and congregations, serving on a committee or as an officer, attending a public meeting, volunteering, and working with neighbors on a project, the alpha was 0.91. Dropping recreation and leisure establishments, the alpha remained at 0.91. Swapping in the Penn St. measure for our membership organization measure (or for both it and our recreation and leisure establishment measure) also left the alpha unchanged. Adding voting in local elections (from the September 2013 Voting Supplement) lowered the alpha. Dropping group participation slightly lowered the alpha.

Adding measures of political participation to other civic engagement variables did not alter the alpha much. Of the six measures we considered (from the November 2008 Civic Engagement Supplement), we retained two—attending political meetings and participating in a demonstration—because of the degree of engagement involved. We dropped discussing politics (less obviously related to civic engagement and more related to interests and knowledge), boycotting companies (too private an act), and supporting a candidate (too imprecisely defined to include low-investment and –involvement “support”).

The share making charitable contributions of at least $25 (from the 2015 Volunteer Supplement) was negatively correlated with two IRS measures on charitable contributions (from 2014 IRS Statistics of Income data). ⁵⁷ The Current Population Survey measure was positively correlated with non-religious civic engagement variables; the IRS measures were negatively correlated. We dropped the IRS measures. Adding the Current Population Survey measure to the other community health indicators left the alpha unchanged, however.

We created a community health index that included membership organizations per capita, nonreligious not-for-profits and congregations per capita, serving on a committee or as an officer, attending a public meeting, volunteering, working with neighbors on a project, attending a political meeting, and participating in a demonstration (alpha=0.92). We excluded voting in local elections and charitable contributions, in part, because they are unavailable at the county level. We excluded group participation because it did not increase the alpha by much.

Institutional Health

Voting in local elections was negatively correlated with voting in presidential elections (from the Election Administration and Voting Survey) and census response rates (from the Census Bureau). ⁵⁸ It also correlated less strongly with other variables in our database and is unavailable at the county level. We dropped it.

The generalized trust variable from the American National Election Study (ANES) was correlated with the CPS variable relating to trust in neighbors at only 0.33. ⁵⁹ It has very small correlations with the CPS confidence in institutions variables (and two of them are negative). Since the ANES variable is only available (with sufficient sample sizes) for half the states, and since it is unavailable at the county level, we dropped it from consideration.

Presidential voting rates, census response rates, and our three confidence-in-institutions measures (from the September 2013 Volunteer Supplement) were all positively correlated. Looking at them together, the alpha was 0.72. Adding the three religion variables lowered it to 0.64. Dropping the confidence variables lowered it to 0.66.

We created an institutional health index from presidential voting rates, census response rates, and the three institutional confidence variables (alpha=0.72). The alpha using the two variables available at the county level was 0.66, and the two subindices are correlated at 0.73, but unfortunately at the county level, census response rates and voting rates are not strongly correlated, so we did not use the two-variable subindex.

Social Capital Index

We computed the preferred index from family unity, family interaction, community health, institutional health, social support, collective efficacy (violent crime rate), and philanthropic health (percent giving at least $25 to charity). The alpha was 0.86.

We also computed an index using the methodology used for the county level. This version correlated with the state index at 0.96.

County Analyses

We began with the list of indicators shown in Appendix Table A1, above. The county-level data required cleaning before it could be analyzed. One problem was that a small number of county definitions changed over the time period covered by our measures. ⁶⁰ A second was that some counties had one or more extreme values, often for smaller counties where an extreme value can reflect small samples. We inspected the distributions of each variable and recoded extreme values on a case-by-case basis. For some variables, we either bottom-coded or top-coded (often at the 99.75 percentile) the values. For others, we dropped percentages equal to 0 or to 100. For two variables, we recoded some rates that should not have exceeded 100 percent to 100 percent. Values of 0 or 100 were recoded to missing if a county had four or more variables at one of those extreme values.

After this cleaning, we standardized all variables and reversed the polarity (multiplying by -1) for eight of them so that higher standard scores always indicated more “social capital.”

Family Health

The share of families with children with a single parent and the share of children in families headed by a single parent—both from the same ACS data used at the state level—were correlated at 0.95. We dropped the former.

Looking at the percent of births to unwed mothers, percent of women married, and percent of children in single-parent families, the alpha was 0.81. Adding the percent never married raised it, but only to 0.82. Adding only the nonmarital fertility rate lowered it to 0.78. (All of these variables were from the ACS.)

Looking at the percent of births to unwed mothers, the nonmarital fertility rate, and the percent of children in single-parent families, the alpha was 0.75. It rose to 0.78 if the percent of women married was also added. It fell if the percent in one-person households or the percent of women never married was added. It fell to 0.73 if the nonmarital fertility rate was dropped. The percent of births that were to unmarried women correlated more strongly with other family indicators than the nonmarital fertility rate.

In the end, we created the same family unity subindex as at the state level, using percent of births to unwed mothers, the percent of women married, and percent of children in single-parent households (alpha=0.81). We created two versions, one that used the weights from PCA analyses at the county level, and one that used the weights from PCA analyses at the state level. They correlated with each other at 0.9997; the weights on the three variables were very similar at the state and county levels.

There were no family interaction variables available at the county level, so we were unable to create a subindex for this dimension.

Social Support

The only social support variable available at the county level is having emotional support, ⁶¹ but it is missing for several hundred counties, and it comes from a survey that is not necessarily representative of every county. We chose not to use it, and to thereby forego having a county-level social support subindex.

Community Health, Religious Health, Civic Engagement

Our two variables related to charitable contributions, from IRS data, ⁶² generally had low or negative correlations with the other indicators. For this reason, and because they were dropped from the state index, we dropped them here too. The Penn State social capital measure and our measure for membership organizations correlated at 0.95. ⁶³ We dropped the Penn State measure.

We had six remaining variables related to community health: non-religious non-profit organizations (IRS), non-religious non-profit organizations plus religious congregations, religious congregations, religious adherence (both from the 2010 U.S. Religion Census: Religious Congregations and Membership Study), membership organizations, and recreation and leisure establishments (both from County Business Patterns data). ⁶⁴ However, we were concerned that we lacked any indicators of informal civil society and activities requiring a time commitment. At the state level, several such measures are available from the Current Population Survey and included in our community health subindex—working together with neighbors, attending public meetings, serving on committees or as officers, volunteering, attending political meetings, and participating in demonstrations.

To resolve this concern, we first went back to the state data and created a new subindex of “informal civil society” for each state. The subindex score was the first principal component score combining the six CPS variables above. We then assigned this subindex score to every county within a state. In other words, the only variation in the subindex score is between states, and all counties within a state get the same score.

Next, back in the county data, we created five different candidate subindices, using different combinations of the informal civil society subindex score, membership organizations per capita, non-religious non-profit organizations per capita, congregations per capita, and the combination of non-religious non-profits and congregations. (Religious adherence was not strongly correlated with the others, and because of our concerns about the establishment data—noted in the report—we were wary of including recreation and leisure establishments when we already were using membership organizations.) These subindices were estimated using PCA.

Next, we computed, for each of the five candidate subindices, the population-weighted average subindex score across a state’s counties. Then we correlated each of these state averages with the state-level community health subindex. We selected the subindex, out of the five candidates, that produced the strongest correlation.

The final county-level community health index combines non-religious non-profits per capita, congregations per capita, and the informal civil society subindex.

Institutional Health

Looking at voting rates in presidential elections (2012 and 2016 Election Administration and Voting Surveys), 2010 census response rates (from the Census Bureau), and religious adherence, none were strongly correlated with one another, and the alphas were very low using any combination of the three. Census response rates generally correlated poorly with the other social capital indicators. ⁶⁵

As with the community health subindex, we were concerned about the incomplete data we had at the county level. In this case, we lacked information about confidence in institutions. We took the same approach as for community health. In the state data, we created a confidence subindex that included the three institutional confidence variables. We assigned every county in a state the state’s subindex score. Then we created three versions of a county-level institutional health index, using different combinations of presidential voting rates, census response rates, and the confidence subindex.

As before, we created population-weighted state averages across a state’s counties and compared them to the state-level institutional health index. The version that correlated most strongly included presidential voting rates, census response rates, and the confidence subindex.

Social Capital Index

We computed an initial index from family unity, community health, institutional health, and collective efficacy (violent crime rate). ⁶⁶ The alpha was 0.66. We also computed an alternative index, dropping violent crime, which was missing for 178 counties (out of 3,142). We were able to compute scores for 103 of these counties by creating an alternative index that left out violent crime. The alpha fell to 0.58, but it correlated at 0.94 with the initial index. Finally, we modified the (standardized) initial index by replacing any missing values on the index with values from the (standardized) alternative index. This is our final county social capital index.

Benchmarking Data

^{Appendix Tables A2a and A2b provide information on the benchmarks against which we compare our social capital indices and subindices.}

Appendix Table A2a. Benchmark Indicators (State and County-Level)

^{Source: Social Capital Project.}

Appendix Table A2b. Benchmark Indicators (State-Level Only)

^{Source: Social Capital Project.}

Additional Maps

In Appendix Figures A1 and A2, we provide state-level social capital maps using alternative (inferior) estimates. In Appendix Figure A1, we show the results when we construct state-level indices that mimic the approach to constructing county-level indices. In Appendix Figure A2, the state estimates are population-weighted average county social capital scores.

Appendix Figures A3-A9 display present each of the seven state subindices. Appendix Figures A10-A13 map the four county subindices.

Appendix Figures A1 through A13

^{Source: Social Capital Project. Maps powered by Leaflet.Download Data (xlsx)}

References

Alesina, Alberto and Eliana La Ferrara. “Participation in Heterogeneous Communities.” The Quarterly Journal of Economics 115, no. 3 (2000): 847–904.

Arrow, Kenneth J. “Observations on Social Capital.” In Social Capital: A Multifaceted Perspective, edited by Partha Dasgupta and Ismail Serageldin, 3-6. Washington, D.C.: The World Bank, 2000.

Bowles, Samuel and Herbert Gintis. “Social Capital and Community Governance.” The Economic Journal 112 (2002): F419–36.

Burt, Ronald S. “Structural Holes and Good Ideas.” American Journal of Sociology 110, no. 2 (September, 2004): 349-399.

Case, Anne and Angus Deaton. “Mortality and morbidity in the 21st century.” Brookings Papers on Economic Activity (2017). 
https://www.brookings.edu/wp-content/uploads/2017/08/casetextsp17bpea.pdf.

Coleman, James S. “Social Capital in the Creation of Human Capital.” American Journal of Sociology 94 (1988): S95-120.

Coyle, Diane. GDP: A Brief but Affectionate History. Princeton, New Jersey: Princeton University Press, 2014.

Felitti, Vincent J., Robert F. Anda, Dale Nordenberg, David F. Williamson, Alison M. Spitz, Valerie Edwards, Mary P. Koss, James S. Marks. “Relationship of Childhood Abuse and Household Dysfunction to Many of the Leading Causes of Death in Adults. The Adverse Childhood Experiences (ACE) Study.” American Journal of Preventive Medicine 14, no. 4 (May 1998): 245-58.

Gronbjerg, Kirsten A. and Laurie Paarlberg. “Extent and Nature of Overlap Between Listings of IRS Tax-Exempt Registration and Nonprofit Incorporation: The Case of Indiana.” Nonprofit and Voluntary Sector Quarterly 31, no. 4 (December 2002): 565-594.

Kim, Daniel, S.V. Subramanian, Steven L. Gortmaker, and Ichiro Kawachi. “US State- and County-level Social Capital in Relation to Obesity and Physical Inactivity: A Multilevel, Multivariable Analysis.” Social Science & Medicine 63, no. 4 (August 2006): 1045-1059.

Lee, Chul-joo and Daniel Kim. “A Comparative Analysis of the Validity of US State- and County-Level Social Capital Measures and Their Associations with Population Health.” Social Indicators Research 111, no. 1 (March 2013): 307-326, 
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4285351/.

Levin, Yuval. The Fractured Republic: Renewing America’s Social Contract in the Age of Individualism. New York: Basic Books, 2016.

McKeever, Brice S., Nathan E. Dietz, and Saunji D. Fyffe. The Nonprofit Almanac: The Essential Facts and Figures for Managers, Researchers, and Volunteers. Lanham, Maryland: Rowman & Littlefield, 2016.

Podolny, Joel M. and James N. Baron. “Resources and Relationships: Social Networks and Mobility in the Workplace.” American Sociological Review 62, no. 5 (October 1997): 673-693.

Prewitt, Kenneth, Christopher D. Mackie, and Hermann Habermann, eds. Civic Engagement and Social Cohesion: Measuring Dimensions of Social Capital to Inform Policy. Washington, DC: The National Academies Press, 2014.

Putnam, Robert D. Bowling Alone: The Collapse and Revival of American Community. New York: Simon & Schuster, 2000.

______ and David E. Campbell. American Grace: How Religion Divides and Unites Us. New York: Simon and Schuster, 2010.

Rotolo, Thomas and John Wilson. “State-level Differences in Volunteerism in the United States.” Research Based on Demographic, Institutional, and Cultural Macrolevel Theories 41, no. 3 (June 2011): 452-473, 
http://journals.sagepub.com/doi/abs/10.1177/0899764011412383.

Rupasingha, Anil, Stephan J. Goetz, and David Freshwater. “The Production of Social Capital in US Counties.” Journal of Socio-Economics 35 (2006, with updates): 83-101.

Sampson, Robert. Great American City: Chicago and the Enduring Neighborhood Effect. Chicago: The University of Chicago Press, 2012.

______, Stephen W. Raudenbush, and Felton Earls. “Neighborhoods and Violent Crime: A Multilevel Study of Collective Efficacy.” Science 277, no. 5328 (August 1997), 918-924.

Scheitle, Christopher P. The Congregation: The World of Christian Nonprofits. New York: Oxford University Press, 2010.

______, Erica J. Dollhopf, and John D. McCarthy. “Exploring Religious Congregations’ Registration with the IRS.” Nonprofit and Voluntary Sector Quarterly 45, no. 2 (July 2015): 397-408.

Scrivens, Katherine and Conal Smith. “Four Interpretations of Social Capital: An Agenda for Measurement.” Working Paper No. 55, OECD Statistics, 2013.

Solow, Robert M. “Notes on Social Capital and Economic Performance.” In Social Capital: A Multifaceted Perspective, ed. Partha Dasgupta and Ismail Serageldin. Washington, D.C.: The World Bank, 2000.

van der Meer, Tom and Jochem Tolsma. “Ethnic Diversity and Its Effects on Social Cohesion.” Annual Review of Sociology 40 (2014), 459-478.

U.S. Congress, Joint Economic Committee, Social Capital Project. “What We Do Together: The State of Associational Life in America.” Report prepared by the Vice Chairman’s staff, 115th Cong., 1st Sess. (May 2017), https://www.lee.senate.gov/public/_cache/files/b5f224ce-98f7-
40f6-a814-8602696714d8/what-we-do-together.pdf.

End Notes

1 Diane Coyle, GDP: A Brief but Affectionate History (Princeton, New Jersey: Princeton University Press, 2014).

2 U.S. Congress, Joint Economic Committee, Social Capital Project, “What We Do Together: The State of Associational Life in America,” report prepared by the Vice Chairman’s staff, 115th Cong., 1st Sess. (May 2017), https://www.lee.senate.gov/public/_cache/files/b5f224c
e-98f7-40f6-a814-8602696714d8/what-we-do-together.pdf.

3 But for dissenting views on using the “capital” metaphor, see, for example, Kenneth J. Arrow, “Observations on Social Capital,” in Social Capital: A Multifaceted Perspective, ed. Partha Dasgupta and Ismail Serageldin (Washington, D.C.: The World Bank, 2000), 3-6; and Robert M. Solow, “Notes on Social Capital and Economic Performance,” in Social Capital: A Multifaceted Perspective, ed. Partha Dasgupta and Ismail Serageldin (Washington, D.C.: The World Bank, 2000), 6-10.

4 See, for example, Robert D. Putnam, Bowling Alone: The Collapse and Revival of American Community (New York: Simon & Schuster, 2000); Samuel Bowles and Herbert Gintis, “Social Capital and Community Governance,” The Economic Journal 112 (2002): F419–36.

5 See, for example, James S. Coleman, “Social Capital in the Creation of Human Capital,” American Journal of Sociology 94 (1988): S95-120; Joel M. Podolny and James N. Baron, “Resources and Relationships: Social Networks and Mobility in the Workplace,” American Sociological Review 62, no. 5 (October 1997): 673-693; and Ronald S. Burt, “Structural Holes and Good Ideas,” American Journal of Sociology 110, no. 2 (September, 2004): 349-399.

6 Robert D. Putnam, Bowling Alone.

7 Alberto Alesina and Eliana La Ferrara, “Participation in Heterogeneous Communities,” The Quarterly Journal of Economics 115, no. 3 (2000): 847–904.

8 Daniel Kim, S.V. Subramanian, Steven L. Gortmaker, Ichiro Kawachi, “US State- and County-level Social Capital in Relation to Obesity and Physical Inactivity: A Multilevel, Multivariable Analysis,” Social Science & Medicine 63, no. 4 (August 2006): 1045-1059.

9 Chul-joo Lee and Daniel Kim, “A Comparative Analysis of the Validity of US State- and County-Level Social Capital Measures and Their Associations with Population Health,” Social Indicators Research 111, no. 1 (March 2013): 307-326, 
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4285351/.

10 National Conference on Citizenship, Broken Engagement: America’s Civic Health Index, September 2006, accessed March 27, 2018, 
https://www.ncoc.org/wp-content/uplo
ads/2015/04/2006BrokenEngagementCHI.pdf.

11 Legatum Institute Foundation, The Legatum Prosperity Index™ 2017, accessed March 6, 2018, 
http://www.prosperity.com/about/resources.

12 Anil Rupasingha, Stephan J. Goetz, and David Freshwater, “The Production of Social Capital in US Counties,” Journal of Socio-Economics 35 (2006, with updates): 83-101.

13 See, for example, “Guide to Using NCCS Data,” National Center for Charitable Statistics, 4-5, accessed March 27, 2018, http://nccs-data.urban.org/NCCS-data-guide.pdf, 4-5; Christopher P. Scheitle, The Congregation: The World of Christian Nonprofits (New York: Oxford University Press, 2010), 13-16; and Brice S. McKeever, Nathan E. Dietz, and Saunji D. Fyffe, The Nonprofit Almanac: The Essential Facts and Figures for Managers, Researchers, and Volunteers (Lanham, Maryland: Rowman & Littlefield, 2016), 1-24.

14 Christopher P. Scheitle, Erica J. Dollhopf, and John D. McCarthy, “Exploring Religious Congregations’ Registration with the IRS,” Nonprofit and Voluntary Sector Quarterly 45, no. 2 (July 2015): 397-408.

15 Kirsten A. Gronbjerg and Laurie Paarlberg, “Extent and Nature of Overlap Between Listings of IRS Tax-Exempt Registration and Nonprofit Incorporation: The Case of Indiana,” Nonprofit and Voluntary Sector Quarterly 31, no. 4 (December 2002): 565-594, 
http://journals.sagepub.com/doi/pdf/10.1177/0899764002238101.

16 The official website of the LDS Church indicates 5,069 Mormon congregations. See “Facts and Statistics,” The Church of Jesus Christ of Latter-day Saints, accessed March 27, 2018, 
https://www.mormonnewsroom.org/facts-and-statistics/country/united-states/state/utah.

17 For Penn State’s methods, see “Social Capital Variables, 2014,” accessed March 27, 2018, http://aese.psu.edu/nercrd/community/social-capital-r
esources/social-capital-variables-for-2014/data-dictionary-social-capital-variables.

18 We used the 2014 version of the Penn State data. State-level estimates for associations and nonprofits were created by multiplying associations per capita and nonprofits per capita by the 2014 population included in the data. We computed statewide census response rates by taking the 2014-population-weighted average of county response rates within a state. We computed statewide presidential voting rates by a similar population-weighted average, except that we used county estimates of the civilian population at least 18 years old from the 5-year 2016 American Community Survey data (covering 2012-2016, Table B05003).

19 “2017 Family Prosperity Index,” Family Prosperity Initiative, accessed March 6, 2018, http://familyprosperity.org/application/files
/7314/9161/3959/FPI-FullReport2017-WEB.pdf.

20 “Association of Statisticians of American Religious Bodies Projects,” Association of Statisticians of American Religious Bodies, accessed March 6, 2018, 
http://www.asarb.org/products.html.

21 Thomas Rotolo and John Wilson, “State-level Differences in Volunteerism in the United States,” Research Based on Demographic, Institutional, and Cultural Macrolevel Theories 41, no. 3 (June 2011): 452-473, 
http://journals.sagepub.com/doi/abs/10.1177/0899764011412383.

22 It finishes third to worst if the principal components analysis is conducted excluding Washington, D.C.

23 See, especially, Kenneth Prewitt, Christopher D. Mackie, and Hermann Habermann, eds., Civic Engagement and Social Cohesion: Measuring Dimensions of Social Capital to Inform Policy (Washington, DC: The National Academies Press, 2014); Katherine Scrivens and Conal Smith, “Four Interpretations of Social Capital: An Agenda for Measurement” (Working Paper No. 55, OECD Statistics, 2013), accessed March 7, 2018, 
http://www.oecd.org/officialdocuments/publicdisplaydocumentpdf/?cote=STD/DOC(2013)6&docLanguage=En; and Robert D. Putnam, Bowling Alone.

24 We round the weights to two decimal places in the report, but the actual weights are rounded to four decimal places.

25 See “Frequently Asked Questions,” National Center for Charitable Statistics, accessed March 6, 2018, 
http://nccs.urban.org/frequently-asked-questions.

26 See, for example, Robert J. Sampson, Stephen W. Raudenbush, and Felton Earls, “Neighborhoods and Violent Crime: A Multilevel Study of Collective Efficacy,” Science 277, no. 5328 (August 1997), 918-924; Robert Sampson, Great American City: Chicago and the Enduring Neighborhood Effect (Chicago: The University of Chicago Press, 2012).

27 “Who Gives Most to Charity,” Philanthropy Roundtable, The Almanac of American Philanthropy, accessed March 7, 2018, 
http://www.philanthropyroundtable.org/
almanac/who_gives_most_to_charity/.

28 After realizing that the counties including Austin and Dallas, Texas—two of the nation’s largest cities—was missing for this ACS variable, we attempted to pull estimates from earlier releases of ACS 5-year averages. Unfortunately, the 2011 data—averaging 2007 to 2011—was the most recent release that included data for the counties. We used this data to fill in missing data for Austin and Dallas but also for 25 other counties in and outside Texas.

29 Weights were as follows: volunteering, 0.39; attending public meetings, 0.44; working with neighbors, 0.45; serving on a committee or as an officer, 0.41; attending political meetings, 0.42; and participating in demonstrations, 0.31. The subindex accounted for 71 percent of the variability in the original six variables.

30 The weights were 0.59 for corporations, 0.53 for media, and 0.62 for schools. The subindex accounted for 65 percent of the variability in the original three variables.

31 County-level presidential voting rates were computed by taking votes from the U.S. Election Assistance Commission and dividing them by ACS estimates of the civilian population aged 18 or older. We do this for 2012 and 2016 (in both cases using the 5-year 2012-2016 population in the denominator), and then average them. Alaska votes are unavailable, so each county is assigned the statewide voting rate in each year.

32 Measures of emotional support at the county level are available through County Health Rankings (taken from multiple years of the Behavioral Risk Factor Surveillance System survey). However, this survey was never designed to be representative at the county level.

33 The thinner state-level index explained 62 percent of the variance across the four variables, higher than the 56 percent of variance across seven variables explained by the fuller index.

34 Tom van der Meer and Jochem Tolsma, “Ethnic Diversity and Its Effects on Social Cohesion,” Annual Review of Sociology 40 (2014), 459-478.

35 Segregation is clearly related to “bridging social capital”—social capital that allows groups to work together or that develops from such cooperation.

36 The highest social capital grouping has eleven states, and the two lowest social capital groupings have 599 counties. There are 150 counties for which we could not compute social capital scores.

37 Robert D. Putnam, Bowling Alone; Robert D. Putnam and David E. Campbell, American Grace: How Religion Divides and Unites Us (New York: Simon and Schuster, 2010).

38 We have developed some initial hypotheses to explain the low correlations. First, it may be that people who live in communities that have low social capital and substantial social and economic problems are drawn to religion as a source of support and a way of making sense of the world. Second, it may be that communities that are low in social capital for historical or demographic reasons may also be more religious for historical or demographic reasons. For instance, the institution of slavery may have had a lasting effect on the social capital levels and religiosity of African Americans. Third, secular social capital may “crowd out” religious social capital, so that places with robust community life that does not revolve around religious institutions find religious participation less valuable. Fourth, it may be that religious adherents are withdrawing from the broader civil society as it becomes more secularized (though it seems less likely that they would withdraw from their own families). Fifth, it may be that within a state or county, places with high religiosity and with low social capital are largely separate. That separateness might get obscured in aggregating up to the state or county level. Sixth, perhaps in some low-social-capital communities there is high religiosity within faith traditions that emphasize a personal relationship with God rather and that do not produce institutions of civil society to as great an extent.

39 See the discussion in the “Past Efforts to Develop a Social Capital Index” section for how we constructed the state-level version of the Penn State index. Putnam’s index scores are available at 
http://bowlingalone.com/?page_id=7.

40 U.S. Census Bureau, Ancestry: 2000, Census 2000 Brief, by Angela Brittingham and G. Patricia de la Cruz, June 2004, "Large Ancestry: 2000," Figure 3, accessed March 27, 2018, https://www.census.gov/prod/2004pubs/c2kbr-35.pdf; Anthony Daniel Perez and Charles Hirschman, "The Changing Racial and Ethnic Composition of the US Population: Emerging American Identities," Population and Development Review 35, no. 1(March 2009), 1-51, Table 5, accessed March 27, 2018, 
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2882688/.

41 Reinforcing this hypothesis, in preliminary research on ancestry we have conducted, states and counties with many people of Chinese and Asian Indian descent also have lower social capital, despite those groups having average household incomes that are 30 percent and 78 percent above the overall American average (according to the American Community Survey).

42 Yuval Levin, The Fractured Republic: Renewing America’s Social Contract in the Age of Individualism (New York: Basic Books, 2016).

43 Anil Rupasingha, Stephan J. Goetz, and David Freshwater, “The Production of Social Capital in US Counties.”

44 See, for example, Vincent J. Felitti et al., “Relationship of Childhood Abuse and Household Dysfunction to Many of the Leading Causes of Death in Adults. The Adverse Childhood Experiences (ACE) Study,” American Journal of Preventive Medicine 14, no. 4 (May 1998): 245-58, accessed March 27, 2018, 
https://www.ncbi.nlm.nih.gov/pubmed/9635069.

45 U.S. Census Bureau, National Survey of Children’s Health (NSCH), 2016 NSCH Data Release, accessed March 27, 2018, 
https://www.census.gov/programs-surveys/nsch/data/nsch2016.html.

46 U.S. Census Bureau, American Community Survey, accessed March 27, 2018, 
https://factfinder.census.gov/faces/nav/jsf/pages/index.xhtml.

47 See U.S. Census Bureau, Current Population Survey, September 2013: Volunteer Supplement, accessed March 27, 2018, 
https://www.icpsr.umich.edu/icpsrweb/ICPSR/studies/36035#.

48 See Sarah Flood, et al., Integrated Public Use Microdata Series, Current Population Survey: Version 5.0. [dataset], (Minneapolis: University of Minnesota, 2017), accessed March 27, 2018, 
https://doi.org/10.18128/D030.V5.0.

49 See U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, Behavioral Risk Factor Surveillance System, accessed March 27, 2018, 
https://www.cdc.gov/brfss/annual_data/annual_data.htm.

50 The variables related to neighbors and contact with family and friends are from the September 2013 Volunteer Supplement.

51 See U.S. Census Bureau, Current Population Survey, September 2015: Volunteer Supplement, accessed March 27, 2018, 
https://www.icpsr.umich.edu/icpsrweb/ICPSR/studies/36411.

52 See U.S. Department of Justice, Federal Bureau of Investigation, Uniform Crime Reporting Statistics, accessed March 27, 2018, 
https://www.ucrdatatool.gov/Search/Crime/State/TrendsInOneVar.cfm.

53 “Attendance at religious services by state,” Pew Research Center, 2014, accessed March 27, 2018, 
http://www.pewforum.org/religious-landscape-
study/compare/attendance-at-religious-services/by/state/; “Frequency of participation in prayer, scripture study or religious education groups by state,” Pew Research Center, 2014, accessed March 27, 2018, 
http://www.pewforum.org/religious-landscape-
study/compare/frequency-of-participation-in-prayer-
scripture-study-or-religious-education-groups/by/state/.

54 See Association of Religion Data Archives, U.S. Religion Census: Religious Congregations and Membership Study, 2010 (State File), accessed March 27, 2018, 
http://www.thearda.com/Archive/Files/Descriptions/RCMSST10.asp.

55 See U.S. Census Bureau, American Community Survey.

56 The IRS data is available at The Urban Institute, NCCS Data Archive, accessed March 27, 2018, http://nccs-data.urban.org/data.php?ds=bmf. The CPS civic engagement variables are from the September 2013 and 2015 Volunteer Supplements.

57 See Internal Revenue Service, SOI Tax Stats County Data 2014, accessed March 27, 2018, 
https://www.irs.gov/statistics/soi-tax-stats-county-data-2014.

58 For voting rates, see U.S. Election Assistance Commission, 2012 Election Administration and Voting Survey: A Summary of Key Findings, September 2013, “Turnout rates for voter participation using different bases,” Table 29, accessed March 27, 2018, https://www.eac.gov/assets/1/6/2012
ElectionAdministrationandVoterSurvey.pdf; U.S. Election Assistance Commission, The Election Administration and Voting Survey: 2016 Comprehensive Report, “Voter Turnout,” Overview Table 1, accessed March 27, 2018, 
https://www.eac.gov/assets/1/6/2016_EAVS_Comprehensive_Report.pdf. For census response rates, see U.S. Census Bureau, “2010 Census Response Rates for the 25 Most Populous Cities” April 26, 2010, accessed March 27, 2018, 
https://www.census.gov/2010census/news/xls/mail_part_rate_4-27-10.xls.

59 The data is available at ANES, 2012 Time Series Study, accessed March 27, 2018, 
http://www.electionstudies.org/studypages
/anes_timeseries_2012/anes_timeseries_2012.htm and ANES, 2016 Time Series Study, 
http://www.electionstudies.org/studypages
/anes_timeseries_2016/anes_timeseries_2016.htm, March 27, 2018. We used only the web samples, which are intended to be representative at the state level.

60 In 2007, the Skagway-Hoonah-Angoon Census Area was split into Skagway Municipality and the Honnah-Angoon Census Area. In 2008, the Wrangell-Petersburg Census Area and the Prince of Wales-Outer Ketchikan Census Area were divided into Wrangell City and Borough, Petersburg Census Area, and the Prince of Wales-Hyder Census Area, with part of Prince of Wales-Outer Ketchikan annexed by Ketchikan Gateway Borough. Subsequently, in 2013, part of the Petersburg Census Area and part of the Hoonah-Angoon Census Area were combined into Petersburg Borough, and the rest of the Petersburg Census Area was added to the Prince of Wales-Hyder Census Area. We dropped any data for Prince of Wales-Outer Ketchikan, Skagway-Hoonah-Angoon, or Wrangell-Petersburg. In 2015, the Wade Hampton Census Area became the Kusilvak Census Area and got a new FIPS code.

Outside Alaska, Shannon County, South Dakota became Oglala Lakota County in 2015 and got a new FIPS code. In 2013, the independent city of Bedford, Virginia was added to Bedford County. We added the totals for these two FIPS codes for variables measured in years before 2013.

61 See Robert Wood Johnson Foundation, County Health Rankings and Roadmaps, accessed March 27, 2018, http://www.countyhealthrankings.org/sites/default/
files/2014%20County%20Health%20Rankings%20Data%20-%20v6.xls.

62 See Internal Revenue Service, SOI Tax Stats County Data 2014.

63 Anil Rupasingha, Stephan J. Goetz, and David Freshwater, “The Production of Social Capital in US Counties.”

64 For non-religious nonprofit organizations, see The Urban Institute, NCCS Data Archive. For religious congregations and religious adherence, see Association of Religion Data Archives, U.S. Religion Census: Religious Congregations and Membership Study, 2010 (State File), accessed March 27, 2018, http://www.thearda.com/Archive/Files/Downloads/RCMSCY10_DL2.asp. For membership organizations and recreation and leisure establishments, see U.S. Census Bureau, American Community Survey.

65 For presidential voting rates, see U.S. Election Assistance Commission, Datasets, Codebooks and Surveys, accessed March 27, 2018, https://www.eac.gov/research-and-data/datasets-codebooks-and-surveys/. We had to add totals for a number of cities and towns into their respective counties in Illinois, Maine, Missouri, and Wisconsin. Alaska counties all have the statewide voting rate. Denominators are the civilian voting-age population, from the 2012-2016 ACS. For census response rates, see Michigan Population Studies Center, Institute for Social Research, Final Mail Participation Rates for All Counties [Census 2010], accessed March 27, 2018, 
https://www.psc.isr.umich.edu/dis/census/
Features/participation/allcounties_oct2010.html.

66 For violent crime rates, we used County Health Rankings data tabulating results from the FBI’s Uniform Crime Reports. We used 2012-14 estimates where available, otherwise 2010-12 estimates, otherwise 2009-11 estimates, or otherwise 2008-10 estimates. See Robert Wood Johnson Foundation, 2017 County Health Rankings Data, accessed March 27, 2018, http://www.countyhealthrankings.org/sites
/default/files/2017CountyHealthRankingsData.xls; Robert Wood Johnson Foundation, 2015 County Health Rankings Data, accessed March 27, 2018, http://www.countyhealthrankings.org/sites/default/files/2015%20Cou
nty%20Health%20Rankings%20Data%20-%20v3.xls; Robert Wood Johnson Foundation, 2014 County Health Rankings Data, accessed March 27, 2018, 
http://www.countyhealthrankings.org/sites/default/files/2014%
20County%20Health%20Rankings%20Data%20-%20v6.xls; and Robert Wood Johnson Foundation, 2013 County Health Rankings Data, accessed March 27, 2018, 
http://www.countyhealthrankings.org/sites/default/files/
2013CountyHealthRankingsNationalData.xls.