This paper by John Landis, Professor Emeritus of City Planning at the Weitzman School of Design, examines income inequality in the United States, particularly at the urban level. While the U.S. is known for its economic opportunities, it also leads among the wealthiest nations in terms of income inequality. The paper explores the nuanced picture of income inequality across the 74 largest urban counties in the United States from 2000 to 2020. The study employs two measures of inequality: nominal income inequality and spatial income inequality. It reveals a trend of increasing nominal inequality alongside a decline in spatial inequality, offering a complex perspective on economic disparities in urban areas.
The findings present a mixed outlook: nominal income inequality has increased, particularly in counties with more productive economies, while spatial income inequality has generally decreased. However, the relationship between these changes and factors such as residential segregation and racialized economic disparities remains tenuous. This highlights the limited utility of income inequality measures in formulating anti-poverty policies and emphasizes the need for a more nuanced understanding of local economic dynamics.
The U.S. is heralded globally as a land of economic opportunity and justly so. American print and social media are full of testimonials from immigrants and native-borns alike who have pulled themselves up by their bootstraps to join the ranks of the wealthy and well-off. Yet the U.S. is also a land of deepening inequality. Based on Census Bureau income tabulations, the Gini coefficient for the U.S. currently stands at 48, up from 39 in 1968¹ (Kitov & Kitov 2013).2 Of the world's 10 wealthiest countries with at least ten million inhabitants, the U.S.'s Gini coefficient is a full five points higher than second-place Italy's.³ Incomes in the U.S. are especially unequal when compared by race and ethnicity, with the median income among Black households in 2021 standing at just 65% of that of white households. These same race and ethnicity-based disparities extend to wealth, homeownership, health care delivery, and school spending (Federal Reserve 2021 (wealth); Choi et al. 2019 (homeownership); Williams & Rucker 2000 (health care); Weathers et al. 2022 (school spending)).
Within the U.S., income inequality is surprisingly consistent among states. At .498, the average Gini coefficient for the five most unequal states (New York, Connecticut, Louisiana, Mississippi, and California) in 2019 was just five decimal points higher than the average for the five least unequal states (Utah, Idaho, Wyoming, South Dakota, and Alaska). 5 Compared by region, inequality is slightly greater among states in the Southeast, Southwest, New England, and Mid-Atlantic regions than among states in the Pacific, Mountain, and Midwest regions. Compared by state, income inequality is largely invariant concerning population size, population growth and growth rate, GDP per capita, urbanization share, and educational attainment (Appendix A).
What about inequality as measured at the metropolitan, county, or city scale-the places where people actually live? Different studies produce different results depending on the choice of geography, study period, and inequality indicator. Comparing changes in Gini coefficient values between 2006 and 2012, Florida (2017) found income inequality to have increased in 226 out of 356 U.S. metro areas. Leading Florida's list of the most unequal U.S. metro areas in 2012 were New York City-Newark-Long Island (with a 2012 Gini coefficient of .504), Miami-Ft. Lauderdale-Pompano Beach (.496), Los Angeles-Long Beach-Santa Ana (.485), Memphis (.482) and Houston (.479).
Using a different inequality measure, the 95/20 percentile ratio (the ratio of the 95th percentile income to the 20th percentile income), a 2016 Brookings study found income nominal income inequality to have increased significantly between 2007 and 2014 in 57 of the nation's 100 largest U.S. areas (Berube & Holmes 2016). The Brooking study found upsurges in inequality to be especially pronounced among central cities of the largest metro areas, with the largest significant increases occurring in mid-sized cities such as New Haven, New Orleans, Boise, Knoxville, Stockton, and Cincinnati. As with metro areas, city-level increases in numerical income inequality were driven more by those at the bottom end of the income distribution becoming poorer than by those at the top becoming richer.
Because households of similar incomes tend to group themselves spatially, researchers have focused as much on income segregation-the degree to which rich and poor households live apart from one another--as on numerical inequality. Among U.S. metropolitan areas, Jargowsky (1996) found income segregation as measured at the census tract level to have increased between 1970 and 1990 for both Black people and whites. Using a slightly different indicator, Mayer (2001) found census tract-level income segregation to have decreased during the 1970s but then to have jumped sharply upward during the 1980s. Focusing on poverty rather than income, Massey and Fischer (2003) found the spatial incidence of poverty within U.S. metro areas to have increased during the 1980s, but then to have declined during the 1990s. Studying families rather than households, Watson (2006a) found income segregation among U.S. metro areas to have increased steadily between 1970 and 1990, but then moderated between 1990 and 2000. In a separate paper, Watson (2006b) observed a positive relationship between rising numerical income inequality and rising income segregation; that inequality at the top of the income distribution had a bigger effect on income segregation in rapidly growing metro areas than in slowly growing ones; and that among slowly-growing metro areas, large increases in income segregation were coupled with higher-than-expected rates of new home construction.
Covering the same 1980-2000 period and using a more robust measure of income segregation but limiting their analysis to the 100 largest U.S. metro areas, Reardon and Bischoff (2011) found there to be a strong and robust relationship between within-race income inequality and within-race income segregation; and that increased income segregation was primarily driven by the housing choices of upper rather than lower-income households. In a later paper updating their analysis through 2009, Bischoff and Reardon (2014) found income segregation to be strongly and positively associated with income inequality, metro area population size, the proportion of children in the population, and average educational attainment, but to have no relationship with unemployment or the share of manufacturing jobs.
Florida (2017) provides more recent estimates of income segregation. Using dissimilarity indices to compare the spatial distributions of poor and rich residents (based on the number of households in census tracts in which the 2010 median household income either fell below the poverty line or exceeded $200,000) Florida identified Cleveland, Detroit, Memphis, Milwaukee and Columbus (Ohio), Philadelphia, Phoenix, Buffalo, Kansas City and Nashville as the ten most income segregated U.S. metro areas. Overall, Florida found income segregation levels to be strongly associated with metro area location (higher in the Rustbelt), population size and density (increasing with both), as well as the presence of high technology industries and a well-educated workforce. Confirming Watson's and Bischoff and Reardon's earlier findings, Florida found overall patterns of income segregation to be driven by the exclusionary housing preferences of the wealthy and near wealthy rather than by the lack of housing choices available to the poor. Coupling his income inequality and segregation findings with the place-based economic mobility findings produced by Chetty et al. (2014), Florida observed an inverse relationship between the two, noting that many of the metros with the highest rates of upward economic mobility are also the metros with the highest levels of income inequality and segregation. The results of these and other key income inequality and segregation studies are summarized in Table 1. For a fuller discussion of the biases and limitations of the various income segregation measures, see Reardon et al. (2018).
This working paper draws on recent U.S. Census data to update and clarify the local income inequality story. First, using census tract and block group-level medians, it generates two complementary income inequality indicators for the nation's seventy-three largest urban counties for the years 2000 and 2020. The first is based on 90/10 percentile ratios and provides a summary measure of county-level income inequality regardless of census tract or block group location. The second makes use of Moran's I, a statistical measure of spatial autocorrelation, to measure the level of income segregation; that is, the degree to which wealthy neighborhoods cluster far away from poor ones. Next, it builds a series of regression models that explain changes in county-level income inequality and income segregation during the 2000-2020 period. Third, using a series of scatterplots and trendlines, it takes on the "so what" question of whether changes in income inequality and segregation when measured at the county level are associated with widening racial income, poverty, and homeownership gaps. A fourth section looks at the relationships between local income inequality and sprawl, gentrification, and local land use regulations. Reflecting on these empirical findings, a final section reviews the usefulness of local income inequality measures to policymakers considering local economic development and anti-poverty policies.
The upshot of these county-level comparisons is that: (i) while nominal income inequality worsened between 2000 and 2020, the level of income segregation declined; (ii) nominal income inequality worsened more in counties with more productive economies (as measured by GDP per worker) while income segregation declined more in counties with rising median incomes; (iii) when measured at the county level, worsening residential segregation and widening racialized income, poverty, and homeownership gaps are largely unrelated to changes in nominal or spatial income inequality; and (iv) except for increased gentrification activity, which is associated with an decrease in income segregation, recent changes in county-level income inequality are unrelated to where population growth occurs and to whether it is overly impacted by restrictive land use regulations.
It is important when reviewing these findings to continually remind oneself of three things. The first is they are based on county-level aggregations of census tract and census block group data and not on individual or household observations. The fact that income inequality may not be associated with economic and racial disparities when measured at the county level in no way means that residents of poor and minority neighborhoods have economic and housing opportunities comparable to those available to residents of wealthy and majority neighborhoods. The second is that they refer to income inequality and not wealth inequality. Whereas income counts economic gains realized in a single year, wealth is cumulative across years and generations. This is one reason the incidence of wealth inequality is greater than the incidence of income inequality (Horowitz et al., Pew Research Center 2020). Finally, and in a related vein, although they are related, income inequality is not the same thing as poverty. Poverty is about absolute economic hardship whereas income inequality is about comparative access to economic resources. Even when they live in middle-income and wealthy neighborhoods, poor Americans still face systemic shortages of jobs, housing, health care, and schooling opportunities compared to their wealthier neighbors. If the goal is to reduce the incidence and burdens of poverty, policies designed to reduce income inequality can help but they will be far from sufficient.
Gini coefficients have long been the metric of choice when comparing income inequality across nations and states, but because of the way the Census Bureau keeps track of survey respondent locations, they are harder to construct at the local level. In their place, most local inequality studies make use of percentile ratios, which compare average or median incomes among census tracts in the upper-most percentile of the local income distribution to those in the bottom-most. The Brookings study referenced above compared average incomes in the 95th percentile of census tracts (in each metropolitan area) to those in the 20th percentile. A different state-level inequality study by the Economic Policy Institute (Sommeiller et al. 2016) was more extreme in its approach to measuring percentile income inequality. It compared incomes in the 99th percentile of the census tract income distribution to those in the first percentile.
One limitation of percentile ratios is that they are indifferent to questions of space and location. That is, they do not measure the level of spatial separation between high- and low-income households. This type of spatial income inequality, also known as income segregation, can be easily measured using the spatial autocorrelation statistic known as Moran's I. Spatial autocorrelation refers to the fact that there are many characteristics, for example, household income, for which we observe adjacent or nearby observations to have similar values, and those further away to have different values. Moran's I values vary between 0 and 1, with higher values indicating a high level of spatial autocorrelation or clustering, and lower values indicating spatial autocorrelation or dispersion. When used to compare income segregation levels among places like cities or counties, higher values of Moran's I (e.g., values above 0.4) indicate that upper and lower-income residents each cluster among themselves in far-apart locations. Lower Moran's I values (e.g., values below 0.20) indicate that upper and lower-income households are spatially intermixed. Unlike other segregation measures such as dissimilarity or the entropy indices, Moran's I is non-linear, meaning that a Moran's I value of 0.5 indicates a clustering level that is more than twice the level associated with a Moran's I value of .25.
For this analysis, I use 90/10 median income percentile ratios to measure non-spatial income inequality and Moran's I values to measure income segregation among the nation's seventy-three largest urban counties. To explore the effects of geographic resolution on the two measures, I did so at two spatial scales, census tracts and census block groups. Census tracts are small, relatively permanent county subdivisions typically ranging in size from 1,200 to 8,000 inhabitants. Census block groups are statistical divisions of census tracts containing between 600 and 3,000 residents and are the smallest spatial scale at which median income data is available. The typical census tract contains between three and five block groups.
When calculating percentile ratios, I sorted the census tracts and block groups within each county from low to high based on their median household income and then calculated the ratio of the 90th percentile tract (or block group) median income to the 10th percentile tract (or block group) median income. Provided that there are enough census tracts to form a meaningful distribution-fifty is an appropriate minimum-the 90/10 percentile ratio hits the sweet spot between being overly sensitive to extreme median income values and not sensitive enough. I also used tract- and block group-level median household incomes to calculate Moran's I values. For the sake of brevity, I will henceforth refer to tract-based median income Moran's I values in the text as MIMIs.
I had three reasons for using counties as my aggregation unit rather than cities or metropolitan areas. The first is that counties, unlike cities and metro areas, have permanently fixed boundaries. Cities, particularly those in the western U.S., are constantly expanding via annexation or incremental incorporation. And while metro area boundary changes occur less frequently than city ones, they do happen. Second, depending on the state, individual census tracts may include multiple municipalities. This adds unnecessary complications when trying to calculate or compare city-level percentile ratios or MIMI values. Except in Texas, and even then, only rarely, census tracts do not cross county boundaries. Lastly, when it comes to calculating income percentiles, cities with fewer than 300,000 residents typically have too few census tracts to counteract the biasing effects of extreme values. Large metro areas have the opposite problem when calculating MIMI values. To the degree that different income groups cluster in more than a handful of locations, MIMIs may understate the degree of county-level income segregation.
For all these reasons, this analysis is limited to U.S. urban counties with populations in 2020 of one million or more, or to counties with 700,000 or more inhabitants having a core city with at least 300,000 inhabitants. Applying these criteria limits the number of counties and city/county combinations included in this analysis to seventy-three and the number of states to twenty-five. Four states have four or more counties meeting the minimum size criteria: California (12 counties), New York (8 counties), Texas (7 counties) and Florida (6 counties). Ten counties are included that lack a central city with at least 300,000 inhabitants; they are Palm Beach County (Florida), Gwinnett County (Georgia), Montgomery County (Maryland), Middlesex County (Massachusetts), Oakland County (Michigan), Nassau and Suffolk Counties (New York), Collin County (Texas), and Fairfax and Arlington Counties (Virginia). Two wraparound counties, Baltimore County and St. Louis County are combined with their core cities, as are Fairfax and Arlington Counties in northern Virginia. Among the large cities not included in this analysis are Omaha, Wichita, New Orleans, Stockton, St. Paul, Greensboro, Lincoln, Durham, Madison, Toledo, and St. Petersburg. Collectively, the seventy-three counties included in this analysis accounted for 38 percent of the U.S. population in 2020.
When comparing local income inequality between places, the scale at which the data is collected matters. For the counties included in this analysis, using census block groups as the unit of observation instead of census tracts raises the average 90/10 percentile ratio in 2020 from 3.32-- meaning that the average 90th percentile census tract median income in 2020 was 3.32 times the average 10th percentile median income-to 3.80 and the average MIMI value from 0.45 to 0.57.8 These differences are partly due to some block group median income estimates being coded as "missing" for confidentiality reasons. Because of this, I will henceforth use tract-based observations for all comparisons.
As noted earlier, 90/10 percentile ratios and MIMI values measure inequality in different ways. 90/10 percentile ratios measure inequality in purely nominal terms without regard to census tract location. By contrast, MIMI values measure inequality spatially and are higher when census tracts with similar income levels are located adjacent to one another. The complementary nature of the two indicators is evident from Figure 1, which compares county-level 90/10 percentile ratios and MIMI values for 2000 and 2020. While both measures track together generally, the correlation coefficient between the two in 2000 was just 0.29. By 2020, it had fallen to 0.19.
Figure 2 makes this same point in another way, by mapping 2020 median income quintiles by census tract for four different combinations of nominal and spatial income inequality. The top left panel of Figure 2 depicts 2020 median household incomes by census tract for Shelby County (Tennessee), which is the city of Memphis' home county. Note the variations in shading (which indicate different median income quintiles) and the way similarly shaded census tracts are clustered. These patterns are indicative of high levels of both nominal and spatial inequality. The top-right panel of Figure 2 presents a comparable quintile map for Westchester County (New York) in which there is a large numerical gap between high-income and low-income tracts, but the two are integrated spatially. This is indicative of places with high levels of nominal inequality but low levels of spatial inequality. The bottom-left panel of Figure 2, which is for King County in Washington (the home county to Seattle), depicts a situation in which residents of high-income and low-income tracts have not-so-dissimilar incomes but are sharply divided along spatial lines. The lower-right panel of Figure 2 is for Queens County in New York City and presents a situation in which the gap between high-income and low-income neighborhoods is relatively small and the two types of neighborhoods are spatially interspersed in checkerboard fashion. Appendix B includes the full set of tract-based and block group-based 90/10 percentile ratios and Moran's I values for 2000 and 2020 for all seventy-three counties.
The most important thing to know about income inequality in America is how widely it varies between places. In terms of nominal inequality, the average 90/10 percentile ratio for the five most unequal counties in our 73-county sample was 2.3 times the ratio for the five least unequal counties. In terms of spatial inequality, the average MIMI value for the five most income-segregated counties was 3.4 times the value for the five least income-segregated counties or income segregation.
Table 2 adds further specifics to these comparisons. As listed in the left-hand panel of Table 2, among the most nominally unequal counties in 2020 were Essex (the home county of Newark with a 90/10 percentile ratio of 5.6), Fulton (Atlanta: 5.3), Shelby (Memphis: 5.1), the District of Columbia (4.8), Suffolk (Boston: 4.2), Hamilton (Cincinnati: 4.2) and Cuyahoga (Cleveland: 4.1). The sources of these inequities varied by place. In Essex and Shelby Counties, there were many more low-income tracts than high-income ones. The opposite was the case in Suffolk and the District of Columbia in which high-income neighborhoods far outnumbered low-income neighborhoods. In Fulton, Hamilton, and Cuyahoga Counties, there were numerous high-income and low-income neighborhoods but few middle-income ones. Among the counties with the lowest levels of nominal inequality in 2020 were Suffolk and Queens in New York State, Salt Lake, Middlesex (Cambridge, MA), and Gwinnett, just north of Atlanta. What these counties have in common is that they mostly consist of middle-income census tracts.
The right-hand panel of Table 2 presents a comparable listing of most and least income-segregated counties. The counties in which rich and poor neighborhoods were the most segregated in 2020 include Jefferson County in Kentucky (with a MIMI value of 0.67), St. Louis City and County (0.66), Shelby County in Tennessee (0.65) and Cook County in Illinois (0.64). Jefferson, Shelby and Cook counties are the home counties to Louisville, Memphis, and Chicago, respectively. The counties in which high-income and low-income neighborhoods were the least segregated in 2020 include Suffolk and Nassau Counties in New York State and Montgomery County in Maryland.
Based on the simple correlation coefficients shown in Table 3, when compared with other county-level demographic and economic characteristics, incomes in 2020 were nominally more unequal in counties with larger shares of Black residents, higher levels of Black-white housing segregation, higher homeownership costs (relative to rents), higher poverty rates, and lower rates of population growth. In terms of spatial inequality, higher levels of income segregation in 2020 were associated with lower median household incomes and higher poverty rates (especially among Black people and Hispanics), lower monthly homeownership and rental housing costs, lower housing values, fewer foreign-born residents, and lower levels of economic productivity per worker. These results are consistent with those of earlier studies by Jargowsky (1996), Massey and Fischer (2000), and Bischoff and Reardon (2014). As the old saying goes, correlation is not causation, so when making causal inferences about whether inequality drives other disparities or the reverse is true, these results should be regarded with care. For that, readers should consult the next section which concerns patterns of inequality change over time.
As per the title of this paper, the income inequality picture among America's largest urban counties is a mixture of good and bad news. On the bad news side, since 2000, nominal income inequality mostly increased. Between 2000 and 2004, the average 90/10 percentile ratio for the counties included in this analysis increased from 3.08 to 3.32 while the average MIMI value declined from 0.53 to 0.45. As indicated in Table 4, among the counties in which nominal income inequality worsened most were Hudson (the home county to Jersey City whose 90/10 percentile ratio increased by 1.38), Suffolk (Boston: +1.26), King (Seattle: +1.15), and the District of Columbia (+.77). Among the counties in which nominal income inequality improved most were Davidson (Nashville: -.50), Los Angeles (-.42), Travis (Austin: -.38) and Mecklenburg (Charlotte: -.35).
The exact nature of these changes varies by place. As indicated in the left-hand panel of Figure 3, which tracks median incomes by census tract in 2000 and 2020 for Hudson County, in situations where nominal income inequality increased, it was generally because the incomes of the residents of upper-income tracts rose faster than those of lower-income tracts. This is evident from the fact that the cumulative share curve for 2020 is everywhere below the cumulative share curve for 2000. By contrast, in the equity-improving Los Angeles County case, the 2020 cumulative share curve is for most of its run above the 2000 cumulative share curve. This is indicative of the fact that relative income gains between 2000 and 2020 in Los Angeles' poorer neighborhoods exceeded the gains in its wealthier ones.
On the good news side of the ledger, spatial income inequality (a.k.a. income segregation) mostly declined with the average MIMI value (based on census tract median income) falling from 0.53 to 0.45. Among the counties in which income segregation was most reduced were Contra Costa County, California (whose MIMII value fell by .37), Santa Clara County (San Jose: -.24), Montgomery County, Maryland (-.23), and Fulton County (Atlanta: -.23). On the other side of the ledger, there were ten counties in which income segregation worsened between 2000 and 2020, led by Hudson County (Jersey City: +.10), San Bernardino County (+.08), and Kern County (Bakersfield: +.06). When interpreting these results, readers should recall that because Moran's I is non-linear, a change in MIMI value of +.1 is more than twice as big as a change of +.05.
What's behind these changes? To find out, I tested three linear regression models comparing 2000-to-2020 changes in nominal and spatial income inequality with various county-based demographic, income, and housing conditions and trends. The regression results are presented in Table 5 and identified as Model A: 2000-to-2020 changes in 90/10 median income percentile ratios; Model B: 2000-to-2020 percentage changes in 2000-to-2020 90/10 median income percentile ratios; and Model C: 2000-to-2020 changes in median income MIMI values. Because Moran's I is non-linear, there is no MIMI percentage change regression model. All the variables are measured and reported at the county level.
The independent variables in the three regression models are organized into three groups. The first group consists of the intercept term and initial (Year 2000) 90/10 percentile ratio and the 2000-to-2020 percentage changes in 2000-to-2020 90/10 median income percentile ratios; and Model C: 2000-to-2020 changes in median income MIMI values. These are meant as trending variables and are included to determine whether an acceleration or deceleration trend is occurring. In an acceleration trend, high and low values move further to the extremes. In a deceleration trend, high values decline faster than low values. A rising intercept value would indicate that income inequality is worsening sample wide. A falling intercept value would indicate a sample-wide improvement in income inequality.
A second group of seven independent variables consists of initial-year anchoring variables. These are included to determine whether subsequent changes in nominal or spatial income inequality follow from initial values of various demographic and economic variables of interest. Except as noted, all are based on decennial census and ACS county-level tabulations. The anchoring variables include:
A third set of six independent variables consists of potential change drivers. These are included to identify the effects of county-level demographic and economic trends on concomitant changes in nominal and spatial income inequality. They include:
Finally, to determine whether rising nominal inequality is contributing to increased spatial inequality, I also included the change in the 90/10 percentile ratio as a change driver variable, but only in Model C, the spatial inequality change regression model. Except where noted, each independent variable is included in all three regression models. That said, I expect some to be more strongly associated with changes in nominal income inequality and others to be more strongly associated with changes in income segregation.
As with any statistical model that purports to explain change over time, one must be wary of potential endogeneity effects. By this, I mean the possibility that the dependent variable exerts a direct or feedback effect on one or more independent variables. This is unlikely to be an issue in this case. Both the 90/10 percentile ratio and MIMI change variables are statistical constructs, meaning that they do not exist outside of the analysis and thus cannot directly influence the decisions or behavior of Census Bureau respondents. Both are also summary measures describing the aggregate behaviors of hundreds of thousands of individual actors who are unlikely to have prior knowledge of their magnitudes or effects. Finally, in no case is there any behavioral conduit through which potential feedback effects might occur. Descriptive statistics and source documentation for the full set of dependent and independent variables are included in Appendix C.
In terms of how well the regression models explain the data, the results are so-so, with r-squared values of 0.50 for all three models. These values are acceptable for this type of cross-sectional change model, but they also suggest that there is still much about recent county-level changes in income inequality left to be explained. The intercept estimates are positive and significant in the two nominal inequality change models (Models A and B), indicating that there was a sample-wide increase in nominal income inequality between 2000 and 2020 equivalent to a 69 percent rise in the 90/10 percentile ratio. In the case of the spatial inequality change model (Model C), the intercept value is not statistically significant.
The coefficients of the two trending variables, 90/10_PR and MI-Value are negative and statistically significant in all three models. What this means is that income inequality declined more among counties where it was initially higher. For every unit difference between counties in 90/10 percentile ratios in 2000, there was an additional 0.29 ratio point decline and a 13-percentage point decline in 90/10 percentile ratios by 2020. This is a significant decline, but in most counties, it was not large enough to overcome the sample-wide increase in nominal income inequality indicated by the constant term. For every unit difference among counties in 2000 MIMI values, there was a 0.3 decline in MIMI values by 2020.
Of the anchoring and change driver variables, only a few are statistically significant. Whether a particular county initially had a higher proportion of Black people, was more segregated, was home to more forward-looking industries, or had a higher-status workforce made no difference in terms of income inequality rising or falling between 2000 and 2020. Nor did changes in demographic or workforce composition. These results differ from Richard Florida's (2017), who argued that there was a strong correlation between an increasingly bifurcated workforce and rising income inequality.
The exceptions to this no-difference finding are for GDP per worker (GDP/job) and housing affordability (MedValu/MedInc) in the two nominal inequality change models (Models A & B) and for initial median household income (Medinc) in the income segregation change model (Model C). Holding other factors constant, nominal income inequality increased more between 2000 and 2020 in counties that produced more high-value-added goods and services. For every $10,000 difference among counties in GDP per employee in 2000, the 90/10 percentile ratio increased by an additional .09 ratio points (Model A) and four percentage points (Model B) by 2020. By contrast, spatial income inequality (Model C) declined more in counties that were initially wealthier. For every $1,000 difference in county median household income in 2000, the MIMI value fell by.003 by 2020. All three differences are small. Nominal income inequality also declined more in counties in which housing was initially less affordable as measured by the ratio of median home value to median household income in 2020. This last finding is a bit of a surprise in that it suggests that the initial presence of more high-value homes contributes to rather than detracts from increasing income equality. There are several mechanisms through which this effect could occur, including the home mortgage interest deduction (which disproportionately benefits wealthier homeowners) as well as the fact that wealthier homeowners generally find it easier to extract equity from their homes.
It is also worth noting that a couple of variables, MedValu/Medinc in Model B and ch-90/10_PR in Model C, are just on the cusp of statistical significance. When more compact versions of these models were estimated by removing variables that were not statistically significant to increase the available degrees of freedom, the MedValu/MedInc variable became significant in a more compact version of Model B but the ch-90/10_PR variable did not become significant in a compact version of Model C.
Taken together, the regression results tell four stories about the changing incidence of income inequality in America's largest urban counties. The first is that there has been a broad scale increase in nominal income inequality since 2000 but no comparable increase in spatial income inequality. The second is that in counties in which income inequality was most severe in 2000, it has since trended downward. A third story is that the principal factor driving increased inequality is differences in economic productivity, with nominal income inequality having increased more since 2000 in counties whose economies are dominated by high-value-added producers. A final story concerns the things that, contrary to conventional wisdom, do not seem to be important explainers of recently rising or falling income inequality. The first of these is race and the second is employment status. Despite higher levels of spatial inequality being associated with higher levels of racial segregation as measured in single years (i.e., 2000 and 2020), as a group, counties in which Black-white segregation was more severe in 2000 did not experience either a subsequent improvement or worsening in nominal or spatial income inequality. Nor did income inequality consistently rise or fall among counties with rising numbers of college graduates or high-status jobs and occupations.
These findings are subject to several caveats, the first being that the county-level inequality indicators that serve as dependent variables in the regressions are all calculated from census all the variables are based on county-level aggregates, the regressions cannot identify significant sub-county or sub-population associations. To see why this could be important, consider the case in which a county enacts a basic income subsidy program targeted toward households with annual incomes less than $30,000. While such a program would have clear equity benefits in aggregate, to the extent that the program recipients are randomly or uniformly distributed spatially, those benefits are likely to be undercounted by tract-based measures like the ones used in this analysis. Third, some of the effects I treat as continuous (such as population growth rate and segregation level) might better be treated as categorical and included in the regressions as either fixed effects or stratifying variables. Finally, I note that these findings are limited to trends occurring in large urban counties and may not apply in the same manner to smaller counties or cities and towns.
Is income inequality automatically a bad thing? Whereas there is some evidence connecting wealth and income inequality to reduced economic and social mobility when compared at the national level, support for the view that income inequality (as distinct from poverty) is associated with negative outcomes when measured at the local level is a good deal weaker. To bring more clarity to this issue, I compared county-level changes in nominal and spatial income inequality between 2000 and 2020 to simultaneous county-level changes in residential segregation levels and to racialized poverty, income, and homeownership disparities. These comparisons are presented in graphic form in Figure 4. To the extent that rising or falling nominal or spatial income inequality is associated with rising or falling segregation or racialized economic disparities, then the points indicating counties should cluster along trendlines with rising or falling slopes.
Before getting to the results, several cautions are in order. The first is that the comparisons presented in Figure 4 apply only to counties and not necessarily to smaller spatial units such as neighborhoods or census tracts. Indeed, it is not only possible but likely that the residents of neighborhoods in which incomes are unequally distributed will face vastly different education, housing, health, and economic opportunities than the inhabitants of neighborhoods comprised of uniformly rich or poor residents. Second, I do not expect any observed associations to necessarily be causal. The conditions required to establish causality require a much more robust and disaggregated series of comparisons than the one presented here. Third, as with the regression results presented previously, I do not automatically expect to observe comparable segregation and disparity associations for nominal income inequality which are presented on the left-hand side of Figure 4, as for spatial inequality, which are presented on the right-hand side. For an association to be considered statistically significant, the probability value associated with the slope estimate must be less than 0.10.
Overall, these comparisons paint a picture in which recent increases in nominal income inequality disadvantaged Black people more than Whites in terms of changes in family income and poverty but did not affect residential segregation levels or Black-White homeownership disparities. Nor was it the case that falling spatial income inequality helped close longstanding racialized income, poverty, and homeownership gaps or lessen the severity of local residential segregation.
Starting in the mid-1990s, after years of losing population, America's oldest central cities began gaining them back. The trend began slowly, gradually picked up steam, and by 2010, was firmly established. America's twelve largest cities in 1970, having collectively lost 1.4 million residents between 1970 and 1990, managed to gain back 1.1 million new inhabitants by 2010. Known colloquially as gentrification, in some places, central city population growth occurred alongside suburban population growth. Elsewhere, the renewal in core city population growth was accompanied by a slowdown in suburban housing construction. This homebuilding slowdown was not solely the result of changing housing location preferences. In many communities, it was augmented by increasingly restrictive land use and environmental regulations (Been et al. 2014).
These trends were both caused and affected by shifts in local incomes. In terms of causes, the arrival of well-paid creative-class workers in many older urban neighborhoods caused incomes and home values in those neighborhoods to soar. In terms of effects, this same dynamic often resulted in long-time and poorer residents being displaced. How did these trends affect income inequality in America's largest urban counties? To find out, I compared changes in 90/10 percentile ratios and MIMI values among the counties included in this analysis to measures of core city population growth, suburban vs. core city population growth, gentrification activity, and local land use regulation restrictiveness. These relationships are presented graphically in Figure 5 and discussed below. In no case do I assert that increased core city population growth and gentrification activity, or more restrictive land use regulations are causing greater or lesser income inequality. As noted above, these relationships operate in both directions, often at the same time. Rather, the intent of Figure 5 is to establish whether there are consistent patterns that connect housing construction activity and regulation to nominal and spatial income inequality trends when measured at the county level.
Except for increased gentrification activity, which is associated with increased spatial income equality, recent changes in county-level income inequality are largely unrelated to where population growth occurs and to whether it is limited by local land use restrictions. This is not to say that the openness of particular cities and counties to additional growth does not affect who can buy or rent a home in which neighborhoods, but rather that these variations in growth rates and regulations are not enough to affect whether county-level income inequality rises or falls.
However helpful concepts and indicators of income inequality may be for comparative purposes, this research reveals their utility as policy-making measures to be more limited. This is so for four reasons. The first is that when compared locally, nominal and spatial measures of income inequality do not necessarily track together. When measured at the county level, nominal inequality indicators like Gini coefficients and 90/10 percentile ratios have been rising, while spatial inequality measures like Moran's I have been falling. Whether urban income equality in America is improving or getting worse depends on whether one cares more about equality regardless of where people live, in which case, inequality is getting worse; or whether low-income and high-income families have opportunities to live in the same neighborhoods and communities, in which case it is improving. Even in circumstances where they agree, summary inequality measures like 90/10 ratios and Moran's I may obscure important intra-metropolitan or intra-county distributional differences. Second, in situations where there are tradeoffs between nominal and spatial inequality, such as with gentrification, the nature of those tradeoffs may be unclear or vary from place to place. Third, nominal and spatial inequality measures appear to have different drivers. Whether nominal income inequality is rising or falling depends on whether aggregate business and worker productivity (as measured by GDP per employee) is rising or falling. By contrast, changes in income segregation have less to do with what is occurring in the local economy and more to do with where poor households can find affordable housing. Fourth, when compared at the county scale, it appears that recent nominal and spatial income inequality trends are only tangentially associated with widening or narrowing residential segregation disparities, or with racialized income, poverty, and homeownership gaps.
What this means for policymakers is that they are better off focusing on non-distributional measures like poverty rates and average wages when considering alternative economic development, anti-poverty, and income support policies. Programs that raise the local minimum wage or provide for targeted job training may have a salutary effect on local poverty rates, but, depending on how frequently and where households move, may have only a minimal effect on nominal or spatial income inequality. Alternately, programs that build low-income housing across many different neighborhoods may be beneficial in terms of reducing spatial inequality while doing little to address the deeper challenge of lifting low-income residents out of poverty. In this regard, rather than thinking in terms of equalizing income outcomes, policymakers should focus on equalizing opportunities.
At the same time, and especially in places where income inequality is already extreme (e.g., cities or counties with 90/10 percentile ratios greater than 4.0 or MIMI values larger than 0.5) policymakers should think twice about policies and programs that further exacerbate the effects of income inequality like across-the-board property tax reductions or cutbacks in public transit service.
Finally, it is important to note that the Census Bureau has updated its online data reporting formats in recent years that make it much easier to identify income and housing disparities at the neighborhood and municipal levels. These include reporting average as well as median incomes, rents, and housing values by census tract (making it easier to understand the effects of extreme values) and distinguishing between white and non-white Hispanics. When coupled with powerful statistical tools like Moran's I and data visualization tools like PolicyMap (https://www.policymap.com/) or the mapping tools included with the Census Bureau's data retrieval website (https://data.census.gov/map), these make it easier than ever for city governments and non-profit and civic organizations to identify the severity and extent of demographic, economic and spatial disparities.
Providing expert commentary on urban topics and highlighting Penn IUR's research in the context of pressing urban issues.
