Chapter 4 The Neighborhood Stabilization Program and Voting

4.1 Relationships between statistical geographies

Figure 4.1: Organizational hierarchy of different geographies.

Census geography runs as follows. The basic unit, blocks, were originally designed for Census takers to walk, and so were constrained by physical boundaries difficult or illogical to cross. However, since the Bureau undertook to cover the entire United States with census blocks, there are millions of census blocks with no inhabitants, covering lakes, parks, and barren landscapes; this fact will become important shortly. Blocks are then aggregated into census tracts containing between 1,200 and 8,000 people. Tracts are the smallest unit for which American Community Survey data—a continuously-run telephone survey that collects additional demographic data such as median household income and educational attainment—are published. Census tracts sit inside counties and then inside states; unlike tracts and blocks, county boundaries are also legal jurisdictions and therefore rarely change. See Figure 4.1 for a graphical representation.160160 Christopher Lemery, “United States Census Information @ Pitt: Understanding Census Geography,” LibGuides at University of Pittsburgh (https://pitt.libguides.com/uscensus/understandinggeography, March 2020).

Figure 4.2: Relationships between different statistical entities

While tracts and blocks were designed to be semi-permanent, acting as the basis for other, non-Census areas, in practice their shapes shift with population changes. These shifts mean that: (a) not all 2000 tracts and 2010 tracts are congruent (though many are); and (b) non-Census areas (such as state-designed voting precincts) can cut census blocks. Figure 4.2 shows the Census Bureau’s graphical representation of geographical relationships. Since these Census Bureau does not make public the responses of individual persons or homes until several decades after, placing each record in its appropriate area is impossible. Rather, ecological methods approximate the distribution of persons, votes, or homes among geographies. Ecological methods assume that persons, votes, or homes are equally distributed throughout an area: ecological methods would assume, for instance, that since six percent of Americans are millionaires, then six out of every hundred persons sampled—no matter their neighborhood—would also be millionaires. Strictly speaking, ecological methods are logical fallacies of de-composition, of taking the whole for the part. However, guided by the right principles, ecological methods can be perfectly valid statistical tools.

The first principle is that smaller geographies are better. Take for example tract 12099005947, a nearly-optimal tract of 4,406 people in Palm Beach County. The owner occupation rate is 97%, and 88% of the residents are aged 65 or older. However, Palm Beach County entire contains tracts with as low as 4% 65-and-up, and owner occupancy rates around 33%. Taking the county for the tract would poorly represent the diversity within Palm Beach, smoothing out the nuances that differentiate one neighborhood’s politics from the next. Zooming in on census tracts improves the accuracy of ecological methods, allowing for a fuller picture.

The second principle is that errors should be distributed evenly. Errors here refer mainly to one party’s votes that are misplaced among another party’s voters, a result of the process by which voter tabulation districts must be decomposed and recomposed in order to translate non–2010 Census geography into 2010 Census geography. This assumption may not always hold in my data: the history of American urban planning has shown that physical barriers—the same used to bound census blocks—often bind communities in more ways than one. But since census tracts are also (if more weakly) constrained by physical boundaries, the aggregation of census blocks into tracts ought to be roughly consistent with the physical construction of communities by planners. Additionally, most voting precincts do not cut tracts. This means that the nightmare problem of gerrymandering deliberately undermining the assumption of evenly-distributed errors should be a moot point. Gerrymandering divides districts into a biased collection of precincts, it does not redraw precinct lines. Since my analysis deals with already-made and locally-created voting precincts, it skirts this particular problem potentially befalling analyses that begin at the county or Congressional-district level.

4.2 Data collection and aggregation

I built my data almost exclusively from official statistics and estimates. Given that most of my sources would eventually lead back to the Census Bureau, which covers the entire United States, I knew that the factors limiting my coverage would lie outside the Census Bureau. Collecting and translating precinct-level voting data to census tracts proved to be the most limited and time-consuming aspect of this process. I began by downloading precinct-level voting returns from the Harvard Election Data Archive (HEDA).161161 Stephen Ansolabehere, Maxwell Palmer, and Amanda Lee, “Precinct-Level Election Data, 2002-2012” (Harvard Dataverse, 2014). Started in 2008 by political scientists, this database compiled voting returns from most states and reported results for state and national races in a standardized format.

These figures were then wedded to files specifying the geographic boundaries of each voting precinct within a state. This stage saw the complete loss of several states’ data as some systems used to identify precincts at the state-level were inconsistent with Census Bureau naming conventions. For each state, I looked for an isomorphic mapping from state-level naming to the standardized naming of the Census Bureau or Harvard Election Data Archive, and only for links where the match was clear were the data used. Case in point was Maryland: the voting returns data clearly contained every single specified tract in Maryland, but matching algorithms could only identify a few hundred of the several thousand voting districts. Geographic files came primarily from the Census Bureau’s TIGER/Line redistricting project, the official database for all Census geography,162162 “2010 TIGER/Line Shapefiles” (US Census Bureau, 2012). via the Election-Geodata mapping project, an open-source repository for redistricting data,163163 Nathaniel V. Kelso, “Nvkelso/Election-Geodata,” March 2020. though, as alluded to above, some states’ geographic files were also compiled by HEDA.

Figure 4.3: Voting precincts and census tracts in Palm Beach County, Florida without census tracts…

Figure 4.4: …and with census tracts.

Translating the voting precincts to Census geography was a complicated process of decomposition and recomposition. Since tracts often straddle two voting precincts, distributing votes to particular tracts relies on ecological methods, essentially peeling off the part of a precinct lying in one tract and stitching it together with parts of the other precincts that also coincide with a tract. This process generates Frankenstein tracts, where 20% of one precinct’s votes are summed with 100% of a second precinct’s votes and 3% of a third precinct’s votes, depending on how much of a precinct coincides with a tract. Since the voting records of these tracts are purely statistical constructions, it is perfectly fine that one have fractions of votes. Take again, for example, our tract 12099005947 in Palm Beach County Florida: it recorded 732.00 Republican votes out of 3114.00 total votes in the 2010 Congressional race. But it’s neighbor, tract 12099005949, recorded 461.85 Republican votes against 1414.25 total votes.

There is, however, more than one way to cut a census tract. The obvious way to slice census tracts is by geography: the spatial intersection forms the basis for dividing up votes. The better way to slice census tracts is by voting-age population: the intersection of eligible people forms the basis for dividing up votes. This method better leverages even more granular data that the Census Bureau publishes to partially overcome ecological approximations and account for differing population densities within tracts. The way tracts in exurban and coastal areas are created necessitates this approach. Ideal census tracts are fairly homogeneous in terms of density, as it is the goal of the Census Bureau to preserve neighborhoods and actual street blocks where possible. In contrast, tracts containing blocks where population density begins to shift can be quite irregular. Since tract size is constrained by population, tracts not only can subsume those zero-population blocks I mentioned earlier, but must subsume some uninhabited blocks.

Public Law 94-171 (PL 94-171) ensures that the Congressional districting mandate of equally-populated districts be met with small-area population counts. In addition, requirements of the Voting Rights Act pressed states to consider the racial makeup of their Congressional districts. Since these are governmental tabulations, the population counts were required to be public. I drew from these data tables, comprising every block in the United States, in order to divide up census tracts. The advantage from data granularity is large: half of the 11,166,336 census blocks in 2010 were smaller than a tenth of a square mile164164 “FCC Form 477: More About Census Blocks” (Federal Communications Commission, March 2015), 1.—Charlottesville itself contains 803 census blocks, averaging 54.14 persons.165165 “2010 Census Summary File 1” (US Census Bureau, 2011).

⊕\[\begin{equation} votes_{t} = \sum_{p=1}^{P \cap t} votes_p \cdot \frac{\sum_{b=1}^{B \cap t} VAP_b}{\sum_{b=1}^{B} VAP_b} \end{equation}\] Where t represents a particular tract, p a precinct from the set P of precincts intersecting with that particular tract, and b a block from the set B of blocks within that particular precinct:

Then, to distribute portions of voting precincts to tracts, I counted how many census blocks from each particular tract lay within each precinct. For computational reasons, the relevant statistic here was the block’s spatial center rather than its nuanced boundaries. Since blocks are so small, and voting precincts are supposed to be fashioned from census blocks (known exceptions exist in a handful of states), using the entire boundaries of a block would likely have generated errors that had more to do with the resolution of downloadable mapping files than actual, in-practice boundaries. Summing up the voting-age population (the census does not ask about citizenship status or felonious histories) of those blocks that lay within a particular precinct, and then comparing that figure to the total voting-age population of said precinct, allocates the percentage of votes from a precinct to a tract. This method is represented formally to the right. To recap, this process first decomposes precincts into blocks, and then recomposes precincts from tracts (where the block-tract connection is trivial by construction). Each tract is allocated votes according to its share of the total precinct population.

Each tract in this dataset was allocated votes by this method, with the exception of Californian tracts. The Redistricting Database for the State of California, hosted at the University of California Berkeley School of Law, had already performed a more precise version of this process using voting registration records.166166 “2010 General Election Precinct Data” (University of California, 2011). Voter registration records allow researchers to locate each address—and thus each resident voter—precisely within tracts and precincts. A more careful version of my analysis would implement such methods, but the time required would have exceeded my timeframe.

A similar process of de/recomposition was used for the Department of Housing and Urban Development’s foreclosure data,167167 Shawn J Bucholtz, “Neighborhood Level Foreclosure Data” (Department of Housing & Urban Development Office of Policy Development and Research, September 2008); Mirian Cody, “Neighborhood Stabilization Program Data,” Research Portal, HUD User (https://www.huduser.gov/portal/datasets/NSP.html, October 2010); Todd Richardson, “NSP2 Data and Methodology,” Research Portal, HUD User (https://www.huduser.gov/portal/NSP2datadesc.html, December 2009). which was tabulated at the 2000 tract- and block group–level. While most of these tracts are the same, and almost all 2010 tracts only subdivide a single 2000 tract, updates to geographic surveying, changes in the landforms, and population fluctuations necessitate updated tracts. The Census Bureau publishes relationship files that precisely place the percentage of housing units within each tract and block group in their 2010 equivalents. Housing units were used here instead of persons since the relevant statistics concern properties, not people. I first decomposed 2000 tracts by the percentage of housing units intersecting with a particular 2010 tract, and then recomposed those 2010 tract figures by allocating foreclosure starts and total housing units according to the percentage each 2010 tract contained.

Lastly, data from the 2010 US Religion Census168168 Clifford Grammich et al., “U.S. Religion Census Religious Congregations and Membership Study, 2010 (County File)” (Open Science Framework, 2018). and the Bureau of Labor Statistics169169 “Local Area Unemployment Statistics” (Bureau of Labor Statistics, March 2020). were collected at the county-level. Since tract geography follows county geography, I simply joined those county figures to each tract, so for each tract the county average was used. All other data used were native to the 2010 tract level, without translation. They came primarily from the 2010 decennial census,170170 “SF1”; Kyle Walker and Bob Rudis, “Tigris: Load Census TIGER/Line Shapefiles,” February 2020. with some additional information from the 2005-2009 American Community Survey.171171 “2006-2010 American Community Survey 5-Year Summary File” (US Census Bureau, 2011); Kyle Walker, Kris Eberwein, and Matt Herman, “Tidycensus: Load US Census Boundary and Attribute Data as ’Tidyverse’ and ’Sf’-Ready Data Frames,” January 2020. American Community Survey data comes attached with margins of error, but, as no other variables had these margins of error, they were entirely ignored so as not to complicate calculations. In the interest of time and data quantity, I took the Census Bureau at their word that errors were distributed evenly.

Outside the Census Bureau, the Federal Housing Finance Agency located addresses within their 2010 tracts, requiring no further pre-processing. Those data were constructed from tens of millions of federally-financed home loans. By virtue of that source, the indices are only composed of prime mortgages. However, two factors work in my favor here. First, significantly more prime mortgages than subprime mortgages defaulted during the foreclosure crisis.172172 Ferreira and Gyourko, “A New Look at the U.S. Foreclosure Crisis.” Second, the FHFA’s index only considers properties with at least 100 sales (and purchases).173173 Alexander Bogin, William Doerner, and William Larson, “Local House Price Dynamics: New Indices and Stylized Facts,” Real Estate Economics 47, no. 2 (June 2019): 365–98, https://doi.org/10.1111/1540-6229.12233. This feature matters because housing prices are strongly dependent spatially: nearby homes with nonconforming mortgages will likely sell for similar amounts as homes with conforming mortgages. Still, this limitation means that some tracts with high levels of nonconforming loans were omitted. There are many neighborhoods, for instance, with complete homogeneity of conforming or nonconforming loans, enclaves connected by mortgage brokers who handled large numbers of subprime loans or specific developments.174174 Dayen, Chain of Title. The political behavior of these developments is interesting due to their uniformity, but beside the aims of this thesis.

4.3 Data Analysis

My approach is simple: I will test my assumption that the Neighborhood Stabilization Program increased housing prices; then, I will model the effect of the NSP on voting with and without home price controls. If the NSP did raise home prices, then its presence ought to be associated with an increase in Republican voting over the model that controls for home prices, where no change in voting should be associated with the NSP. Additionally, this section allows for the possibility that impacts on voting were sharpened where housing was politically stigmatized by Tea Party politics, as Chapter 2 suggested. To evaluate those localized effects, trials are run that only include tracts in which a Tea Party candidate ran for office.

4.4 Conclusion

In sum, 2010 prices were negatively associated with the Neighborhood Stabilization Program’s first round of funding. Perhaps this result was due to unobserved variable bias, but, given the normalcy of regression diagnostics, and the results of later NSP rounds, it may instead result from the NSP’s timeline. Due to the legislative mandate that grantees purchase homes at a slight discount, there may not have been enough time to clear, repair, and re-appraise (or fully resell) properties by the 2010 midterms. My results indicate that the NSP reduced prices between 1.4% and 2.0%, a small reduction keeping with size of the mandated discount. Further, by NSP1’s maturation in 2013, the amount of funds spent on acquisitions dropped, while new construction and rehabilitation increased slightly compared to its 2010 progress.186186 18568, “HUD NSP-1 Reporting: Sept 2010 Program-Wide Report” (Department of Housing and Urban Development, November 2010); HUD, “HUD NSP1 Report: National Drawdowns by Activity (Q3 2013)” (Department of Housing and Urban Development, n.d.).

The labor risk–equity model was evaluated in spite of this fact, and evidence suggested the theorized effects were present in the 2010 midterms, though oddly not when only Tea Party tracts were considered. Economic anxiety, composed of low equity and risky employment magnified the conservatizing effects of increased home prices. These results were statistically significant and robust to several covariates. Still to be explained are the mechanisms by which labor risk propagates through mortgage markets, HELOCs, and political behavior. This variable proved “liberalizing”, in tension with prior literature on the subject.

Limitations of the measurement and statistical methods were explored, with strong evidence that further work is needed to extricate the impacts studied from spatial dependence and bounded response assumptions. Beta distribution regression was employed to correct for the latter issue, with slightly more ideal measures of model fitness and similarly strong evidence for the labor risk–equity model’s capacity to explain voter behavior in the full sample as well as the Tea Party tracts. These results build towards consensus with empirical literature, girding Ansell (2014) with voting results to provide a crucial middle step in his argument.