The Changing Shape of Spatial Income Disparities in the United States

Abstract

Spatial income disparities have increased in the US since 1980, a pattern linked to major social, economic, and political challenges. Yet, today’s spatial inequality, and how it relates to the past, remains insufficiently well understood. The primary contribution of this article is to demonstrate a deep polarization in the American spatial system—yet one whose character differs from that commonly reported on in the literature. The increase in spatial inequality since 1980 is almost entirely driven by a small number of populous, economically important, and resiliently high-income superstar city-regions. But we also show that the rest of the system exhibits a long-run pattern of income convergence over the study period. A secondary contribution is historical: today’s superstars have sat durably atop the urban hierarchy since at least 1940. Third, we describe six distinctive pathways of development that regions follow between 1940 and 2019, with certain locations catching up, falling behind, and surging ahead. We explore the role played by initial endowments in driving locations down these pathways, finding population, education, industrial structure, and immigrant attraction to be key distinguishing features. These insights are enabled by a fourth contribution: methodologically, we use group-based trajectory modeling—an approach new to the field that integrates top-down and bottom-up views of the evolving national spatial system. We conclude by exploring implications for the mid-twenty-first century.

Key words:

• inequality
• geography
• cities
• convergence
• economic history

Acknowledgments

The authors wish to thank Max Nathan and Dylan Connor for early conversations, Sergio Petralia for sharing patent data, and Jan Helmdag and Valeria Lima-Passos for help with GBTM model selection. This article benefited from feedback from attendees at meetings of the American Association of Geographers, the Global Conference on Economic Geography, GEOINNO, as well as seminar participants at the Gran Sasso Science Institute, LSE, Birkbeck, and the Paris School of Economics. For comments and suggestions that helped improve the article, we wish to also thank anonymous reviewers and Andrés Rodríguez-Pose at Economic Geography.

Notes

1 There is no single perfect measure that captures all the relevant inequalities of places within a spatial-urban system; in this research, we rely on average income as a robust, if not perfect, indicator of the underlying quality of the economic development that is occurring in a place, and the distribution of such quality of regional development across the system. See Appendix C in the online material for wider discussion of this topic.

2 Though framed differently, recent work by Connor, Uhl et al. (2022), Connor, Berg et al. (forthcoming), and Houlden et al. (2022) can be thought of as exceptions.

3 Four CZ locations cannot be created in the 1940 data; these small locations are omitted from the analysis. We benefit from 1960 mappings made available by Evan K. Rose. See https://ekrose.github.io/resources/.

4 We also explore variation in total pretax income from all sources, which additionally includes income earned from business, welfare, social security, retirement, and other sources. Household-level aggregations are also analyzed.

5 If $Y$ is unadjusted nominal household income for region $j$ in time $t$, we obtain deflated real incomes $R$ using the following formula: $R_{\mathrm{jt}}=Y_{\mathrm{jt}}/\mathrm{LCP}{I}_{\mathrm{jt}}$. The local consumer price index, $\mathrm{LCPI},$ is calculated as $\mathrm{LCP}{I}_{\mathrm{jt}}=(1\ast w_{\mathrm{nt}})+\left[{\displaystyle \frac{\mathrm{REN}{T}_{\mathrm{jt}}}{\mathrm{REN}{T}_t}\ast w_h}\right]$, where $w_{n}h$ and $w_h$ represent expenditure shares derived from the CPI-U.

6 For raw data, see https://www.census.gov/quickfacts/fact/note/US/LND110210.

7 These data are available at the USPTO’s PatentsView website: https://patentsview.org. Thanks to Sergio Petralia for sharing the cleaned data.

8 A correlation table for key variables is in Appendix A in the online material.

9 Results are consistent when we exclude locations in the South, defined as South Atlantic, East South Central, and West South Central Census Regions. They remain consistent if we generate separate estimates for men and women, and if we restrict the sample to workers who are full-time, full-year employed. The shape of results is also unchanged when we estimate $\beta$-convergence using Baumol regressions (Baumol 1986). All of these alternative estimates are available upon request.

10 See Appendix B in the online material for fit statistics and other details on model selection.

11 This reduces the number of regions from 722 to 297. Consult Appendix B in the online material for a report on this and a battery of other diagnostics and robustness checks.

12 Appendix Figure B3 in the online material depicts these national trends.

13 Exploring this, Figure C1, in Appendix C in the online material, shows trends in population shares for each trajectory, going all the way back to 1900. It is striking that the two groups with high or improving income ranks (superstars and pulling ahead) are the places whose shares of national population have grown considerably over the study period, in contrast to all the other groups. Income, at first glance, seems a powerful attraction for people, although causality may be due to the attributes of those who have sorted into these places.

14 Coefficients and standard errors are presented in Appendix B in the online material.

15 While in a linear regression, generating standardized $z$-scores might be a sensible approach to maximizing comparability, such normalization techniques reduce predictive accuracy in logistic regression (Turing n.d.), while also reducing the comprehensibility of odds ratios.