The Cartography of Opportunity: Spatial Data Science for Equitable Urban Policy

Abstract

As evidence of the contextual effects of place upon individual outcomes has become increasingly solid over time, so too have urban policies and programs designed to connect underserved people with access to spatial opportunity. To this end, many attempts have been made to quantify the geography of opportunity and quite literally plot it on a map by combining evidence from studies on neighborhood effects with spatial data resources and geographic information systems (GIS) technology. Recently, these opportunity maps have not only become increasingly common but their preparation has been encouraged and facilitated by the U.S. Department of Housing and Urban Development. A closer look at the foundations and methods that underlie these exercises offers important lessons I examine the practice of opportunity mapping from both theoretical and methodological perspectives, highlighting several weaknesses of the common methods. Following this, I outline a theoretical framework based on Galster’s categorization of the mechanisms of neighborhood effects. Using data from the Baltimore metropolitan region, I use confirmatory factor analysis to specify a measurement model that verifies the validity of the proposed theoretical framework. The model provides estimates of four latent variables conceived as the essential dimensions of spatial opportunity: social-interactive, environmental, geographic, and institutional. Finally, I develop a neighborhood typology using unsupervised machine learning applied to the four dimensions of opportunity. Results suggest that opportunity mapping can be improved substantially through a better connection to the empirical literature on neighborhood effects, a multivariate statistical framework, and more direct relevance to public policy interventions.

Keywords:

• Opportunity mapping
• neighborhood effects
• community development
• ecometrics
• fair housing

Acknowledgments

The author wishes to thank George Galster, Karen Chapple, and Ingrid Gould-Ellen for helpful conversations during their visits to the University of Maryland, the contents of which helped shape this research. In addition, Casey Dawkins, George Galster, Rolf Pendall, Kris Marsh, Willow Lung-Amam, Zack Patton, and three anonymous reviewers provided comments on earlier drafts for which I am also grateful.

Notes

1. In their seminal article introducing ecometrics, Raudenbush and Sampson (1999) advocated the use of methods borrowed from psychometrics, namely factor analysis and item-response theory, in conjunction with novel data collection methods such as systematic social observation. Their approach was designed to move beyond the limitations of traditional administrative data such as that collected by census. Thus, a more conservative view might reject the label of ecometrics as applied to the present analysis, given its liberal use of Census data and lack of item-response models. The first critique is less damning as O’Brien, Sampson, and Winship (2013) have successfully applied ecometrics to large-scale administrative data, although the second has considerable merit. The liberal interpretation of ecometrics might argue that the label remains accurate because of the use of factor analysis, which attempts to capture the latent ecological aspects of neighborhoods. Regardless of interpretation, the ecometric label in the current use is open for debate.

2. To be classified as a highly qualified teacher, Maryland State Department of Education (MSDE) requires that instructors in core academic subject areas must: hold at least a bachelor’s degree from a regionally accredited institution of higher education; hold a valid Standard Professional Certificate, Advanced Professional Certificate or Resident Teacher Certificate in the subject area they are teaching; and satisfy additional requirements associated with specific teaching levels and experience. For additional information, see http://www.marylandpublicschools.org/msde/programs/esea/docs/TQ_Regulations/general_definition.htm.

3. A middle school factor is not estimated because it only includes two distinct data points.

4. It should be noted that the 60-min threshold applied here is a subjective decision, reinforcing the notion that any opportunity analysis is not a purely technical exercise. Transportation accessibility studies typically define accessibility measures somewhere between 15 and 90 min, depending on the destination and the type of study, but there is very little consensus on which threshold is the most general or appropriate for most uses (Anderson, Levinson, & Parthasarathi, 2013; El-Geneidy & Levinson, 2006; Knaap, Ding, Niu, & Mishra, 2015). In this case, 60 min is chosen as it represents a reasonable commute time that is able to capture substantial intraregional variation. Future work is necessary to validate this threshold and determine whether different thresholds should be applied to different destinations.

5. In future work, researchers could take advantage of the Environmental Justice Screening tool (EJSCREEN) recently released by the EPA, which helps overcome some of the estimation issues outlined here. More details on the EJSCREEN tool can be found at https://www.epa.gov/ejscreen

6. Note that because all three variables in the environmental dimension are negative, it should be assumed that this represents the inverse of opportunity. In other words, crime, toxic exposure, and vacancy are allowed to load positively on the environmental factor, and the inverse of the environmental factor is taken to be a measure of opportunity.

7. Note that because the latent variables used in the cluster analysis follow distributions similar to z scores, their interpretation follows similarly; positive numbers represent higher than average scores, negative numbers represent lower than average scores, and scores near zero are close to the average.