The Road to Division: Interstate Highways and
Geographic Polarization
Clayton Nall 1
June 18, 2012
1 Assistant
Professor, Department of Political Science, Stanford University, 616 Serra Street, Stanford CA
94305. Phone: 617–850–2062. Email: [email protected]. Web:http://scholar.harvard.edu/nall. Thanks to Nate
Baum-Snow for kindly sharing and answering questions about his carefully assembled and geocoded replication
archive of Federal Highway Administration data. Gary King, Dan Carpenter, Claudine Gay, Justin Grimmer, Karen
Jusko, Jonathan Rodden, Eitan Hersh, Ryan Enos, Ed Glaeser, and participants in the Harvard American Politics
and Political Behavior workshops provided valuable comments on earlier drafts, as did attendees at talks at Princeton, Stanford, Washington University in St. Louis, UC-San Diego, UC-Merced, UW-Madison, NYU-Wagner,
UC-Berkeley, UCLA, and Ohio State. Thanks to the Center for American Political Studies and the Taubman Center for State and Local Government at Harvard University for financial support. Upon publication, data and code
necessary for replication will appear at http://dvn.iq.harvard.edu/dvn/dv/claynall.
Abstract
What explains geographic polarization, the tendency of Americans to live near fellow partisans? Existing theories suggest that segregation of all kinds can be explained by the aggregate exercise of individual
preferences for homogeneity, or by discriminatory public policies that limit residential choice. While
both of these factors explain sorting, this article considers an alternative explanation: that public policies
that facilitate mobility make partisan geographic sorting easier and contribute to geographic polarization. I test this theory by examining the political consequences of the largest transportation program in
American history: the Interstate Highway System. Combining data from a federal highway construction
database with county-level presidential election results, I use matching and regression to demonstrate
that suburban counties with Interstates became more Republican than they would have been otherwise. I
then show that metropolitan areas with greater Interstate density became more geographically polarized
than comparable areas with fewer Interstates. The observed effects are especially strong in the South,
where highways opened rural areas to industrialization and suburbanization. These findings demonstrate
that policies can change politics not only by influencing individual behavior, but also by influencing
citizens’ geographic distribution.
The segregation of Americans into urban Democratic enclaves and suburban and rural Republican
enclaves is a defining and important feature of American politics, but remains, along with other forms of
geographic polarization, one of the unsolved “open questions” of research on the sorting and polarization
of American citizens (Fiorina and Abrams, 2008). While citizens of all political stripes have occupied
lower-density residential contexts over the last sixty years, between 1952 and 2008, the average gap in the
Democratic presidential vote between the central and peripheral counties of the 70 largest metropolitan
areas grew by more than 10.5 points. Compared to Democratic voters, Republican voters have become
increasingly prone to live in lower-density counties and low-density Congressional districts (?). In 1952,
33% of Republican and 28% of Democratic presidential voters lived in the central county containing
major cities. By 2008, 25% of Democrats and only 15% of Republicans lived in those central counties.
The population densities of counties in which members of the two parties live have similarly diverged. In
1952, the average county-level contextual population density experienced by the Democratic voters was
only 1.4 times as high as Republicans’, but is now 2.3 times as high. Republicans have become a party
representing small towns and suburbia, while Democrats owe their electoral chances to an urbanized
electorate. Democratic and Republican voters have sorted on an urban-suburban, high-density to lowdensity axis.
The urban-suburban partisan split is the result of a decades-long shift of population from the cities
to the suburbs in the postwar period. Yet, this population shift did not occur equally across different
metropolitan areas for the two parties. In some places the ascent of Republican suburbs happened more
rapidly than in others, and the partisan polarization of urban-suburban areas happened more rapidly in
some metropolitan areas. These differences were the result of individual-level residential decisions in
which Republican voters have been relatively more likely than Democrats to choose to live in low-density
and suburban areas, while Democrats have expressed their preference for more urban, high-density areas.
1
Yet existing sorting models–such as those introduced in Tiebout (1956) and Schelling (1971), offer an
insufficient explanation of city-by-city and county-by-county variation in geographic polarization that
resulted from this sorting process. This article presents evidence that federal policy towards urban areas
played a major role in enabling the geographic sorting of Democrats and Republicans. Specifically,
it examines the consequences of one of the most important of these policies for residential sorting:
the Federal-Aid Highway Act of 1956, colloquially known as the Interstate Highway Act. Highways
built under the act facilitated the expansion of American suburbs (Baum-Snow, 2007), but I show the
changes wrought by the act were political as much as they were demographic or economic. Highways
enabled Republicans (and voters likely to shift to the Republicans) to move away from cities and into
peripheral areas that highways made suitable for development. They also induced demographic and
socioeconomic changes in cities and suburban areas concurrently with partisan change, enabling the
growth of a suburban Republican constituency over multiple decades.
I present two sets of results consistent with this model of transportation-induced sorting. Using
county-level presidential election results as the key measure of partisanship, and adopting methods of
observational causal inference appropriate to the data (a combination of difference-in-difference, matching, and regression) (Ho et al., 2007; Iacus, King and Porro, 2009, 2011a,b; Diamond and Sekhon, 2005),
I first demonstrate that Interstate highways contributed to a decline in the Democratic share of the presidential vote in suburban counties where they were built. These suburban changes have been indirectly
related to a higher-scale change in metropolitan areas. In a second, related analysis that examines that
gap between urban counties and suburban counties (aggregated into couplets consisting of the urban and
suburban portions of each metropolitan area), I show that the highway-induced shift in the suburban
vote contributed to an increased urban-suburban partisan gap in presidential vote choice. Among other
findings, I show that these effects mostly appear in the South, where highways had more potential to
2
influence suburbanization and affected the distribution of Republican and Democratic voters during a
major realigning period. I suggest several causal mechanisms behind these changes, and show through
additional county-level data analyses that highways changed suburban correlates of a range of socioeconomic correlates of vote choice. I close by discussing these findings’ implications for partisan sorting
models and for scholarship on the study of politics and behavior.
1
How Transportation Infrastructure Facilitates Partisan Geographic
Sorting
Scholars have attempted to explain residential sorting from one of two perspectives, both of which are
crucial to understand how highways have facilitated residential sorting. The first approach, which I
call the agent-based approach, treats observed aggregate-level polarization and clustering as an emergent property of autonomous, individual decisions. The most widely referenced model of residential
sorting suggests that individuals choose their neighborhoods by aligning their preferences with the multidimensional features offered by autonomous communities that compete for residents (Tiebout, 1956).
Similarly, models of residential racial segregation have tended to explain residential sorting from one of
two perspectives. The Schelling sorting model Schelling (1971) suggests that stable racial segregation
can be explained by the aggregate level manifestation of individual preferences for homogeneity; even
in the absence of discriminatory policies, individuals with even slight preference for racial homogeneity
will, in the aggregate, produce stable segregation outcomes. As applied to the geographic sorting of partisans, either of these models suggests that even mild preferences to live near people who happen to share
the same partisanship can manifest themselves in severe partisan segregation. The opposing viewpoint,
which I call the structural approach, explains geographic sorting as a consequence of discriminatory
3
policies that constrain choice. Massey and Denton (1993), for example, point to the extensive system
of de jure and de facto housing discrimination, including against potential residents on the basis of income (Levine, 2006). An important claim of this research is that even policies with nondiscriminatory
intent can reinforce residential income and race segregation, and, even if the Schelling (1971) model is a
good one, public policies influence the residential choice set posed to each individual. While this debate
has largely occurred within the context of racial residential segregation, the theories and lessons are as
applicable to observed residential segregation of political partisans.
Both of these schools study sorting without accounting for how some places become subject to residential sorting while others have not. Of course, contentious local politics related to neighborhood homogeneity were routine in mid-century America, and residential composition commonly changed block
by block (Sugrue, 1995). But segregation and sorting at the metropolitan level required active state support for metropolitan transportation development. Imagine these residential choices from the perspective
of a Republican white male head of household living in a Rust Belt city in the early 1950s and commuting a short distance to the central business district. According to the earliest available random-sample
surveys on partisanship and residential preference, such an individual was more likely than a Democrat
with the same income, race, and education to want to move to the suburbs (Roper Organization, 1976;
Gallup, 1983). In the simplest versions of the Tiebout sorting model, this move would be frictionless:
our subject would pull up stakes and leave town. However, in practice, if this individual wanted to retain
employment options in the city, congested local roads in the postwar period would have made a daily
work commute untenable. Moreover, real estate developers would have been unlikely to build housing
tracts for these urban refugees knowing that few of them would tolerate the long commute.
This scenario would have played out differently after construction of an Interstate highway between
the suburbs and the central business district. The urban escapee’s daily commute to the central business
4
district (or to a job in another suburb) would now be tolerable. Moreover, with a new pool of potential
residents available, real estate developers would have identified the city’s outskirts as a prime location
for new residential subdivisions, offering potential residents more substantial housing options.
This vignette shows that transportation infrastructure is a sine qua non of large-scale partisan geographic sorting. In particular, construction of Interstate highways induced two responses, one from
potential movers and the other from those who provide them with residential opportunities. The first response came in the form of individual sorting behavior, by giving some citizens a larger radius over which
to search for housing and jobs. The second is that in response to this selective expansion of residential
opportunity, developers built new communities, the construction of which locked in persistent partisan
change. Improved transportation made construction in undeveloped rural “greenfield” areas more appealing to developers and community-builders, who had new access to a blank slate of open country
land in which to design communities for select interest and lifestyle groups (Burns, 1994, 35-37). By
influencing early community development, these infrastructure projects may have initiated “increasing
returns” processes (Krugman, 1992) in which initial settlement by homogeneous groups locked in longterm partisan stability. Being “literally cast in concrete” (Pierson, 1993, 609), infrastructure projects may
have helped lock in durable changes and constraints on future development that other policies–such as
the home mortgage interest deduction–could not achieve.
The connection of highways to partisan sorting may come in the form of local contextual effects.
Aggregate-level data presented here do not permit easy inference about contextual effects versus residential sorting. We know, however, that while ideological conversion in response to context happens
rarely (King, 1996), some partisan conversion took place during this period, especially in the South. Such
partisan (if not ideological) conversion may have occurred in response to changing local circumstances,
especially as suburban and rural whites throughout the country shifted their allegiance to the Republi5
can Party. Such partisan label-switching may have occurred more quickly in suburbanizing areas where
middle and upper-class whites realized the two parties’ issue sorting and adopted coherent conservative
ideologies rooted in economic conservatism (Shafer and Johnston, 2006).
2
Empirical Strategy
To ascertain the effect of the Interstate Highway System on partisan geographic sorting, I present two
analyses of highways and their effect on partisan geographic sorting, as manifested in county-level voting
in presidential elections between the 1950s and the present. The first analysis estimates the effect of
highways on the partisan makeup of suburban counties. The second, related analysis uses metropolitan
areas as the unit of analysis; it separately aggregates the outcome variables in suburban counties and
urban counties of each metropolitan area, then estimates the effect of overall density of the highway
network on the urban-suburban split in the Democratic vote between the urban and suburban regions of
each metropolitan area.
These analyses capitalize on knowledge of the assignment mechanism by which highways were built
in some places but not others (Rubin, 1991). To estimate highways’ effects on metropolitan political geography, it is necessary to explain the selection of highway routes across counties within major
metropolitan areas. Our knowledge of these decisions is based substantially on major highway planning documents, most notably the 1944 Interregional Highways report published by the Public Roads
Administration, a predecessor of the modern Federal Highway Administration (National Interregional
Highway Committee, 1944). The proposal laid out in the report encompasses most routes later included
in the Interstate Highway System. While the report proposes locations of major expressways without
accounting for county boundaries, the decision to build an Interstate highway through any given county,
or to build more Interstates in one metropolitan area over another, can be predicted using county-level
6
Census variables that correspond to route placement criteria discussed in the 1944 report.
3
Highways and Suburban Political Development
Using available data on highway placement and political development, I first present a set of empirical
analyses on highways’ localized effects on partisan change in suburban counties. These results are based
on highway placement data from a federal highway construction database covering 1956 to 1996 (encompassing almost all Interstate highway construction), combined with county-level Census and presidential
election return data for the years 1952-2008. Combining multiple identification approaches (nonparametric matching, difference-in-differences, and linear regression), I quantify the extent to which highways
reduced support for the Democratic Party in areas of the metropolitan periphery in which they were built.
3.1
Data and Methods
Defining the Suburban Sample: To capture highways’ effects on partisan sorting in suburban areas, it is
first important to identify a consistent sample across an extended period. Here, I define suburban counties
in the metropolitan periphery as the population of interest. Counties are ideally suited to this purpose.
While their political function varies by state, they are often the basis for organizing school districts and
public services that provide the basis of Tiebout sorting. In addition, counties have the advantage of
retaining constant boundaries over most of the study period, permitting over-time comparisons that are
not possible with the limited fine-grained data that can be obtained over this period. Suburban counties
are defined as those with geographic centroids between 20 and 100 kilometers from the center of the
100 largest cities in 1950. (The approximate location of the central business district is defined using the
point location of the current Cities layer in StreetMap USA (ESRI, 2008a).) Under this rule, the national
sample contains n = 988 counties, which appear in the map in Figure 1. Counties in which an Interstate
7
highway was built at any time before 1996 appear in gray, while counties without Interstate highways
appear in black. To exclude urbanized counties that may have fallen within the catchment area of large
cities (e.g., Kings County, New York or Middlesex, Massachusetts), the sample excludes counties with
a 1950 Census population greater than 300,000. Summary statistics for the suburban counties appear in
Table 1.1
Defining County-Level Partisanship: To estimate the causal effect of an Interstate highway on a
county’s partisan composition, the outcome of interest is defined as the difference in the county-level
Democratic presidential vote share between 1952 and election year t, t ∈ 1960, 1964, . . . , 2008.2 Presidential election returns through 1988 are drawn from the ICPSR county-level presidential election returns
data (Clubb, Flanigan and Zingale, 2006; Inter-university Consortium for Political and Social Research,
1995). Data for 1992 to the present, and corrections to errors in ICPSR data sets, were drawn from CQ
Press (2010) and Leip (2010).3
Defining Treated Counties: Each suburban county is assigned a binary treatment variable indicating
whether an Interstate was open to travel in the county. This variable is coded so that, for each election year t and county i, a binary treatment variable Zit indicates whether an Interstate highway passed
through county i by year t − 4. This variable is constructed using data from the Federal Highway Administration PR-511 database, the official record of Interstate highway construction progress (Baum-Snow,
2007; Michaels, 2008; Chandra and Thompson, 2000), which runs through 1996 and covers almost all of
1
Appendix Section A.1 includes tests of the sensitivity of these findings to various coding rules,
including the use of different inner and outer radii and the adoption of density-based measures.
2
Recent research suggests that across most recent elections, the presidential vote has been a suitable
proxy for an area’s latent partisanship (Levendusky, Pope and Jackman, 2008).
3
Among other coding decisions that merited correction, votes cast for the Minnesota Democratic
Farmer Labor Party presidential candidate were coded as votes in the “other” party category.
8
the expansion of the Interstate Highway System. This model assumes that any county with an Interstate
highway in place for at least 4 years received the same treatment dosage, a plausible assumption considering that most counties became “treated” with an Interstate highway in the relatively narrow window
between 1956 and 1970.
Figure 1: Map of the suburban county sample. Counties that contained an Interstate highway at any point
through the final year of the FHWA database (1996) appear in gray, while counties without Interstate
highways appear in black.
Estimation: Interstate highways’ effect on suburban county-level Democratic presidential vote share
is estimated by least squares regression, independently for each election year t between 1960 and 2008:
Yt − Y1952 = β0t + βzt zt + β1t x1 + . . . + βkt xk + (1)
where zt is the vector of treatment indicators constructed for each election year and x1 . . . xk are included
covariates. The linear model is assumed to correctly capture the relationship among the variables. For βzt
9
Variable
Mean
SD
Min.
Max.
Earliest Interstate Opening, (Treated Counties)
1965
6.717
1942
1992
Interstate Highway Built At Any Time
0.554
0.497
0
1
Republican Presidential Vote, 1948
0.413
0.192
0.006
0.841
Republican Presidential Vote, 1952
0.548
0.152
0.081
0.930
Republican Presidential Vote, 1956
0.556
0.154
0.095
0.926
Median Family Income, $, 1949 (9-category)
5.16
1.524
2
9
Proportion Non-White, 1950
0.110
0.164
0
0.818
Log(Persons Per Acre, 1950)
-2.54
0.872
-7.26
0.2825
Proportion in Urban Residence, 1950 (10-category)
2.60
2.231
0
9
Crop Value Per Capita, 1950 ($)
112
133
0
1619
Mfg. Establishments, 1939
36
51
0
457
On Strategic Route, 1941
0.657
0.475
0
1
Proportion Non-Resident One Year Earlier, 1950
0.013
0.13
0.000
0.192
Region=South
0.42
0.49
0
1
Treatment Variables
Covariates
Table 1: Summary statistics for suburban county sample (n = 988).
10
to capture the causal effect of the treatment of interest, the model assumes that the units in the treatment
group undergo the same average potential change in presidential vote choice over time as units in the
control group, conditional on included covariates.
To reduce model dependence before use of linear regression, the same covariates are used to generate a balanced sample at each election year t using coarsened exact matching (CEM), a nonparametric
alternative to propensity-score matching (Iacus, King and Porro, 2011a; Ho et al., 2007). This matching
method places observations in multidimensional bins using coarsened covariate values, then assembles
a sample from the treated and control observations that appear in the same bin.4 The default procedure
in CEM is to trim both treated and untreated observations from the sample, possibly generating a more
balanced sample when covariate overlap is minimal than methods that discard only control-group or
treated-group units. CEM is applied to the sample for each election year. While imbalance between the
treated and control groups is only partially reduced in the early years of the Interstate program, likely
as a result of the small and unusual set of treated units, CEM removes almost all of the imbalance on
included covariates in the later years of the program. The standardized imbalance for the treated and
untreated groups–the normalized difference in means for each covariate between the treated and control
groups–appears in Appendix Figure A-6.
4
5
Unlike propensity-score methods, CEM allows for matching on covariates that are not normally or
t-distributed, which avoids the introduction of bias (Rubin, 1976).
5
Along with the matching procedure, the sample is truncated to exclude “early-adopter” states that
built Interstate-quality highways before passage of the 1956 Highway Act. For example, the Massachusetts Turnpike was built in the 1930s and was later incorporated into the Interstate system. Such
toll roads did not draw down the generous 90% federal funding granted to states for free highways, and
they benefited only indirectly from the Interstate Highway Act.
11
Covariates: Because we are interested in the total effect of Interstate highways on election outcomes
over time, we want to control for factors that contributed to the decision to build a highway in a county
that could be related to the potential outcomes, and that were measured prior to the placement of highways in a county. Most of this decision was made in a period before passage of the Interstate Highway
Act in 1956, the primary function of which was to provide sufficient funding to execute the plans, so the
analysis includes only covariates measured up to 1956. Controlling for post-treatment variables (e.g.,
racial composition of counties years after highway construction) has the potential to introduce bias and
would change the outcome of interest. Therefore, covariates used in modeling include Census, political, and other planning-related variables that predict placement of an Interstate highway in a county and
were measured before passage of the Federal-Aid Highway Act of 1956. The most important of these are
covariates are based on the Interregional Highways planning document, which lists factors used to decide the geographic scope of the Interstate System. These include Log population density in 1950, Crop
value per capita, 1950, and a categorical variable for the Percentage urban, a measure of preexisting
development and another way to pick up population density. The Number of manufacturing establishments in 1939 provides the latest available pre-treatment indicator of county manufacturing production,
another factor mentioned in the report (Fitch and Ruggles, 2003). Finally, to capture a county’s military importance, a dummy variable was included indicating whether a county was on a 1941 military
strategic route (National Interregional Highway Committee, 1944, 33). Many of these strategic routes
could have been designated as part of the Interstate program had it covered more highway mileage. To
account for suburban growth at baseline, which may have factored in highway placement decisions, the
regressions include the Percentage of households that lived outside the county in 1949, an indirect measure of the pre-existing suburbanization rate. Three additional controls included to account for possible
partisan interference in highway planning (and to account for major regional differences in the partisan
12
realignment) are the Republican presidential vote share in 1948, 1952, and 1956. Median family income in 1950 captures a correlate of county automobile ownership, which was accounted for in the 1944
planning document, while Percentage nonwhite in 1950 may capture the possible confounding influence
of county racial composition on highway route placement. Lastly, a dummy variable for the South is
interacted with the treatment variable to account for possible pro-Southern bias in highway construction
and to pick up regional heterogeneity in the treatment effects.
To construct accurate confidence intervals and reduce sensitivity of findings to election-specific variation in the presidential vote, a non-parametric smoothing procedure is followed. The sample is bootstrapped and estimates generated for all years on each bootstrapped sample. A lowess curve (Cleveland,
1979; Cleveland and Devlin, 1988) is drawn through each set of estimates, and the appropriate quantiles of the smoothed values used to construct the 95% and 80% confidence intervals.6 This procedure
borrows information across multiple elections, generating an estimate of highways’ impact on election
results in a handful of adjoining election years. This nonparametric approach remains close to the data
and avoids functional form assumptions of multi-level models.7
3.2
Results
Results based on the combination of regression and matching demonstrate that suburban counties with
Interstate highways became less Democratic than they would have been otherwise. The estimates on
6
The confidence intervals used throughout the paper are quantile-based confidence intervals (Keele,
2008, 181).
7
The lowess function uses a smoothing span incorporating one third of the data points. At the middle
of the study period, this incorporates approximately two elections on either side of the chosen election,
using the default tricube kernel-based weighting within the smoothing window. Each lowess fit uses
three “robustifying iterations” (Becker, Chambers and Wilks, 1988).
13
the national matched sample suggest modest overall effects. Predicted values simulated from a model
interacting the treatment and the South dummy variable suggest that the overall result owes to substantial
regional heterogeneity in effects, with most of highways’ effects taking place in Southern states.
The bootstrapped, lowess-smoothed estimates on the matched sample for each election year from
1960 to 2008 appear in Figure 2, and, across nearly the entire time period, are in the expected direction.
In the matched sample, the presence of an Interstate highway in a county reduced the Democratic presidential vote by between 2 and 3 points over a period between 1976 and 2004. These effects appear to
have declined in subsequent elections, though the direction of the estimates remains negative. Highways’
effect on partisanship in Southern counties has been remarkably stable over time. Interstate highways
made suburban counties in the matched sample in the South about 5 points less Democratic than they
would have been without an Interstate highway between the years 1972 to 2008. Effects outside the
South have been smaller and less persistent, though between 1980 and 1996 the presence of an Interstate
highway in a county facilitated a drop in the Democratic vote of between 1 to 2 points, on average, in
counties outside the South.
3.3
Responses to Data and Modeling Concerns
In this section, I address and respond to several concerns about the suburban county analysis relating, broadly, to the regional heterogeneity of treatment effects and the robustness of data and modeling
choices.
3.3.1
Regional Effect Heterogeneity
One objection that could be raised about this result is that it appears to depend substantially on the
South. On its face, this is no reason to discount the results: the South is not only part of the country,
but a region that continues to grow and assert its primacy in American politics–all the more reason to
14
Highway Effect on Change in County Dem Vote Share (Points), 1952−[Year]
All Suburban Counties
8
7
6
5
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
−9
● ● ●
●
1964
●
1976
●
● ●
● ●
1988
●
2000
South
●
8
7
6
5
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
−9
Non−South
●
●
●
1964
●
●
1976
Year
●
●
● ●
● ● ●
1988
Year
2000
8
7
6
5
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
−9
●
● ● ● ● ●
● ● ● ●
1964
1976
1988
●
●
2000
Year
Figure 2: Interstate highways facilitated drop in the Democratic vote share in suburban counties.
Smoothed OLS estimates of the effect of construction of an Interstate highway in a county by year t − 4
on the difference in the Democratic presidential vote between 1952 and year t, using the CEM-matched
sample and excluding early-adopter states. 80% (thick line) and 95% (thin line) confidence intervals
accompany each estimate.
15
be concerned with the role of federal public policy in bringing about partisan sorting in the region. A
substantive reason for regional effect heterogeneity is that earlier construction of freeways and in nonSouthern areas, especially in the Northeast, produced residential segregation at an earlier stage than in
the Southern sample, so that suburban areas in the fixed catchment area used in this study began their life
cycles decades before comparable neighborhoods comparably geographically situated in the South. It is
unsurprising that large effects would be observed in a region that depended substantially on highways
and other forms of federal investment for its postwar economic growth (Schulman, 1994). As highways
connected a poor rural labor force to the outside labor market, Interstate highway crossroads became
suburban and exurban boomtowns (Schulman, 1994, 116) that often supported the Republican Party.
The overall poverty of rural areas, low land costs, and good weather attracted migrants and made these
areas affordable targets for new residential development.
Simultaneously, highways provided white Southerners with easier use of the exit option to circumvent
desegregation efforts: white flight in the South fed more rapid growth of suburban Republicanism in
these areas as white voters left urban school districts and cities when the federal government intervened
on voting rights and school segregation (Kruse, 2005; Lassiter, 2006), and the organization of school
districts by county in some Southern states (e.g., North Carolina, an epicenter of the school desegregation
fight) meant that white residents looking to avoid desegregating school systems had particularly strong
reasons to move across county boundaries. The substantial realignment of the Solid South and large
swings in partisan allegiance during this period may have moderated highways’ effects, as changes in
residential patterns induced by highways coincided with changing party affiliation among individuals in
the electorate.
Metropolitan areas dominated by older housing–a group dominated by cities in the Northeast and
Midwest–were already largely developed, and may have been less susceptible to infrastructure-induced
16
change. This result is consistent with the model of suburban development described in Mieszkowski and
Mills (1993). This does not suggest that the findings about infrastructure’s effects are not generalizable
outside the South, but rather that major infrastructure policies can have a larger effect on the political
makeup of a metropolitan region when the rural areas of the region are a blank slate for development. In
the 1950s, present-day major cities of the South were less suburbanized compared to the major Northern
industrial centers. The larger effects in the South, where more greenfield development opportunities
existed within the urban catchment area, are unsurprising.
There is also a plausible methodological reason for this result. The coarsened exact matching procedure trims the sample of both treated and control observations. If the resulting sample is balanced, the
estimate of the average treatment effect on the treated (ATT) will provide an unbiased estimate of effects
on the set of treated units retained in the sample, which are disproportionately on the periphery of the
catchment area used to define the suburban sample.8
3.3.2
Responding to Concerns About Data and Models
Like all methods for observational causal inference (except true natural experiments), the potential for
bias always looms over the results, but a series of robustness checks, primarily presented in the appendix,
address many of the concerns.
The first of these is a placebo test, which runs all the same analyses conducted in the main text, but
8
Estimation on a sample generated using genetic matching (Diamond and Sekhon, 2005) and entropy
balancing (Hainmueller, 2012) yield substantively similar estimates but smaller point estimates in the
South, and regional estimates are more homogeneous. However, in neither of these cases are results
directly comparable to those presented here, and slightly more imbalance appears when doing matching
with the same covariates, using default parameters.
17
uses the average of the 1948 through 1956 Democratic presidential vote as the outcome variable for each
of the annual estimates. (This test includes both the Eisenhower elections, which were somewhat exceptional for the support the Republican elicited among Democratic voters, and the 1944 and 1948 elections,
which presented different partisan voting patterns.) Under such placebo tests, failure to reject the null
for the baseline effect is held to support unbiasedness. Placebo regressions on the CEM-matched sample
yield point estimates of around zero to one-half points, as desired, and the 80% and 95% confidence
intervals, while not uniformly overlapping with zero, are small in magnitude relative to the estimated
treatment effects across most of the study period. This suggests that even if there is a small amount of
omitted variable bias, it is not nearly enough to affect the results. (See Appendix Figure A-7.)
Additional placebo tests discount the potential risks associated with the difference-in-difference analysis being conducted over an unusually long (fifty-six year) time period. It could be argued that this
explains the attenuation of results in later years. One plausible explanation for this is that the relative impact of highways may diminish over time as a consequence of the life cycle of neighborhoods.
Neighborhoods could take random divergent paths over time after their initial formation, and this could
undermine the plausibility of the difference-in-difference assumption over time, though it might simply
push the causal effect of the highways downward without introducing bias. But the quality of matches
across later time periods is actually quite strong. While matching does not eliminate all imbalance on
the included covariates at all time periods (Appendix Figure A-6), most imbalance on included covariates exists in the earliest years of the program, during which few units had been treated. By the 1970s,
by which time most units had been treated, coarsened exact matching eliminates nearly all imbalance
on all but a few variables. If the difference-in-difference assumption were failing in later time periods,
we might expect placebo tests for a wide range of pre-treatment variables to provide some evidence of
imbalance, potentially in the form of higher variance in estimated placebo effects. In addition to the
18
placebo estimates conducted using presidential election results (Appendix Figure A-7), an additional set
of placebo tests on an array of outcomes measured before the Interstate Highway Act, including variables from the 1930 and 1940 Census, yield a set of estimates that, with only a few exceptions, suggest
that balance has been achieved. To present the results on a common scale, these results are presented as
t-statistics (Appendix Figure A-8).
Another array of concerns speak to the potential confounding effects of a range of events that might
have occurred after the main period of Interstate highway construction, including urban school desegregation, state-specific candidate effects, and construction of additional highways.. Nearly all of these
concerns can be dismissed on the grounds that each occurred simultaneously with or after the construction of Interstates and controlling for them could introduce post-treatment bias. For example, if school
busing in metropolitan areas was made possible by Interstate highways, controlling for the existence of
a school desegregation plan in a metropolitan area would lead to a biased estimate of highways’ total
effect. Nevertheless, Appendix Figure A-13 presents the same estimates as presented in the text, after accounting for the presence of a desegregation order in the metropolitan area (Coleman, Kelly and
Moore, 1975). Similarly, one could claim that there are common shocks to state-level presidential voting
that are not accounted for here; the largest of these is the home-state candidate effect. After indicator
variables for Democratic and Republican home-state candidates are included, the results are not meaningfully affected (Appendix Figure A-9). Finally, we could be concerned that construction of highways
in one county could lead to additional non-Interstate construction in adjoining counties. If so, this could
violate the non-interference assumption of causal inference (Rubin, 1986). Point estimates in the first
few decades after Interstate construction are unlikely to be affected meaningfully by this issue, while
estimates in later years may be affected. However, it is most likely that the effect of such interference
would work against the findings presented here and bias estimated effects towards zero.
19
4
Highways and Urban-Suburban Geographic Polarization
If highways are making suburban counties in which they were built more Republican, a related question
is whether metropolitan areas with more highways have more Republican suburban regions and are
more polarized as a result. In this section, I present a second set of analyses that aggregate county
data into urban and suburban regions for each metropolitan area, then assess the extent to which the
highway density in a metropolitan area influences residential sorting on the urban-suburban axis. While
the previous section demonstrated that Interstate highways made suburban counties less Democratic,
these analyses demonstrate that an increase in highway density across a metropolitan area facilitated
the partisan split between the center and periphery of metropolitan areas. Defining a new metric of
highways’ spatial impact on a metropolitan region–Interstate exit density–I show that metropolitan areas
with Interstate exit density above the median in recent years became more polarized by almost four points
relative to comparable metropolitan areas with highway exit density below the median.
4.1
Data and Methods
To measure urban-suburban polarization, it is first necessary to devise a new unit of analysis specifically
intended to capture the partisan gap between city and periphery areas. This unit, the urban-suburban
couplet, contains, for each metropolitan area, the central (i.e., urban) county (or counties) containing the
Top 100 U.S. cities by 1950 population, and the aggregation of suburban counties within 80 kilometers
of the central city (cities).9 To address the concern that some urban areas are multicentric, urban counties
within the catchment area of a larger metropolitan area are merged to define a combined metropolitan
9
A summary of the robustness of these findings to the choice of inner and outer radius appears in
Appendix Figure A-17.
20
area.10 The included counties are aggregated and weighted averages of the urban and suburban outcome
variables are respectively calculated. After the metropolitan areas are combined, the resulting sample
contains 70 urban-suburban couplets. Summary statistics for these variables appear in Table 2.
Defining urban-suburban polarization: Within these couplets, outcome, treatment, and covariate
values are defined. To capture urban-suburban polarization, the outcome is defined as the difference
between election year t and 1952 in the gap in the Democratic presidential vote between the urban and
suburban portions of each couplet. Let D̄ist represent the Democratic presidential vote share in the
suburban region of each metro area i in year t, and D̄iut represent the Democratic presidential vote share
in the urban portion of the metropolitan area i in year t. The outcome of interest is then:
∆i,t = (D̄iut − D̄iu,1952 ) − (D̄ist − D̄is,1952 )
(2)
A plot summarizing the change in this quantity between the years 1952 and 2008 appears in the Appendix Figure A-14. As with the suburban county-level analysis, defining the outcome variable in this
way permits us to adopt the assumptions of difference-in-difference estimation.
Highway density as treatment: Because a binary treatment variable no longer suffices when studying
the effect of highways on metropolitan areas, I introduce a different measure of the Interstate highway
treatment for metropolitan areas, Interstate highway exits per square mile of land area. This captures the
extent to which the area is not just traversed by Interstate highways, but also connected to and dependent on them for transportation within the metropolitan area. Extending the widely used cardiovascular
metaphor of highway “arteries,” a greater number of exits suggests that highways are used not only as
10
These combinations will be fully defined in the replication code. Among other differences with
the suburban-county analysis, early-adopter states are not excluded. Doing so would cut out parts of
multi-state metropolitan areas.
21
Variable
Mean
SD
Min.
Max.
Exits Open Per Square Mile, 1956
0.005
0.008
0
0.050
Exits Open Per Square Mile, 1964
0.023
0.027
0
0.155
Exits Open Per Square Mile, 1968
0.032
0.033
0
0.162
Exits Open Per Square Mile, 1972
0.038
0.038
0
0.197
Exits Open Per Square Mile, 1976
0.041
0.040
0.004
0.214
Exits Open Per Square Mile, 1980
0.044
0.042
0.004
0.214
Exits Open Per Square Mile, 1984
0.045
0.042
0.004
0.214
Exits Open Per Square Mile, 1988
0.046
0.042
0.004
0.217
Exits Open Per Square Mile, 1992
0.047
0.043
0.006
0.220
Exits Open Per Square Mile, 1996
0.049
0.043
0.006
0.220
Urb-Sub Diff. in Dem. Pres. Vote, 1948
0.034
0.086
-0.201
0.210
Urb-Sub Diff. in Dem. Pres. Vote, 1952
0.033
0.093
-0.169
0.197
Urb-Sub Diff. in Dem. Pres. Vote, 1956
0.024
0.092
-0.202
0.226
Urb-Sub Diff. in Proportion Nonwhite, 1950
0.026
0.056
-0.132
0.185
Urb-Sub Mean, Proportion Nonwhite, 1950
0.098
0.112
0.003
0.440
Urb-Sub Mean, Mfg. Estabs., 1939
801
1804
0
14320
Urb-Sub Mean, Co. on Strategic Route
0.83
0.19
0.1
1
Log Persons/Sq. Mi., 1950
5.10
0.93
3.74
8.013
South
0.275
0.449
0
1
Table 2: Summary statistics for treatment variables and covariates used to construct urban-suburban
couplets (n=70).
22
a conduit of long-range traffic, but as connections to local street network at the “capillary” level. The
exit location data come from the 2008 StreetMapUSA highway exits layer (ESRI, 2008b) and California CalTrans shapefiles. To create a panel data set capturing exit construction over time, a cartographic
shapefile containing contemporary exits was merged with the Federal Highway Administration PR-511
master file, a chronological record of the start and completion dates of local Interstate segments. To
establish the opening date for each exit, these geocoded points were linked to the 2008 exit shapefile,
producing a database of actual exit locations and their imputed year of construction. The count of exits
is then divided by the total land area in the urban-suburban metropolitan couplet. The resulting variable
captures Interstate highways’ overall spatial presence in each metro area each year between 1956 and
1996, taking into account the physical geography of the area.11 As an example, the treatment variable is
displayed in the context of a typical urban-suburban couplet, the Atlanta metro area (Figure 3). An 80kilometer buffer centered on the city of Atlanta delimits the scope of land area used to calculate Interstate
exit density. Interstate highway exits appear as points.
Estimation: Multiple modeling assumptions are adopted to obtain an estimate of the highways’ causal
effect on urban-suburban polarization. Least-squares regression is used to estimate the growth in polarization in each metro area after 1952 as a linear function of exit density. Setting ∆t as the vector of ∆i,t
values (the difference in the urban-suburban gap in the Democratic presidential vote in couplet i between
year t and 1952), the causal effect of the treatment variable zt (exit density or log exit density at election
year t), βzt , is estimated along with other parameters according to the following linear model:
∆t = β0 + βzt zt + β2 x1 + . . . + βk xk + 11
(3)
I examine presidential election results through 2008, but the PR-511 data (and thus, the exit data)
are updated through 1996, by which point the Interstate Highway System was substantially complete.
23
¯
!
(
Legend
!
(
Central City
80-Kilometer Radius
Interstate Highway Exits
Fulton County (Urban)
0
Suburban Counties
10
20
40 Kilometers
Other Counties
Figure 3: Map of the Atlanta metro area, illustrating the construction of an urban-suburban couplet and
exits used to construct the exit-density measure.
24
where βzt represents an estimate of the causal effect of a one exit-per-square-mile increase in exit density,
∆t is the difference in between 1952 and year t in the urban-suburban difference in the Democratic
proportion of the presidential vote, and x1 , . . . , xk represent the included covariates. Following the
same bootstrapping procedure adopted in the county-level analysis, I estimate βzt on the matched and
unmatched samples and again lowess-smooth the βzt estimates (Cleveland, 1979).
A second analysis, presented here primarily to check the robustness of findings to model selection, dichotomizes the treatment variable around the sample median, then generates matched samples of treated
and untreated units to estimate the effect of highway construction. Urban-suburban couplets were considered “treated” if the density of exits or the number of rays was above the full-sample median.12 Using
genetic matching (Diamond and Sekhon, 2005), I create a matched sample of treated and similar untreated units. This method often yields better improvements in observed balance in small samples than
other, propensity-score based matching methods, while tolerating some additional increase in imbalance
in a subset of covariates.13 The genetic matching algorithm employs the same set of covariates that
appear in the linear regression models applied to the matched sample.
Couplet-level covariates: Including couplet-level versions of all variables used in the suburbancounty analysis is infeasible given the limited sample size. All of these covariates are calculated by
12
Dichotomizing the treatment variable introduces the assumption that the “dose” of the treatment of
interest is identical for units above the median and for units below the median.
13
All matching methods that assume that data are multivariate t- or normal-distributed (Rubin, 1976)
face this problem. When sufficient observations are available, variants of exact matching, such as coarsened exact matching, may be preferable, as they are not vulnerable to this assumption but may, under varied coarsening assumptions adopted in practice, discard too many observations (Iacus, King and Porro,
2011a).
25
taking the population-weighted average of the county averages within the urban and suburban portions
of each metropolitan couplet. The analysis presented here includes a set of three variables predictive of
metropolitan highway density. The most important of these measures is Log population density of the
metropolitan area in 1950; this is one of the primary factors listed in the Interregional Highways report
and Other variables are coded in two ways: the urban minus suburban (within-couplet) difference and the
unweighted mean of the urban and suburban averages within each couplet. Two of the most important
of these are the within-couplet Mean number of manufacturing establishments in 1939 and the Mean
proportion of counties included in the military strategic route map. To account for the possible role of
partisan politics as a factor in highway construction, the models include the Lagged urban-suburban
difference in the Democratic presidential vote share for 1948, 1952, and 1956. Census variables include
both the Urban percentage non-white minus the suburban percentage non-white and the Unweighted
mean of urban and suburban percentage non-white. One reason to include these racial variables is that
urban highways may have been easier to build in places with more poor black neighborhoods in which
property takings were less costly, and this construction could have been correlated with white flight and
eventual partisan changes that coincided with it (Mohl, 2004). An indicator variable for cities in the
South captures differences in the pace of freeway construction in that region: older, non-Southern cities
were more likely to build turnpikes and urban freeways before the passage of the Interstate Highway Act,
and voting behavior and voting rights across the region differed from the rest of the country during the
first few decades of the 1952-2008 study period.
4.2
Results
The first set of results presents the effect of the number of exits per square mile in each metropolitan area
via least squares regression on the full sample, using the untransformed and log-transformed versions
26
of the exits variable. The first differences presented are those associated with a shift between the 25th
and 75th percentile of highway density in each year. The left panel of Figure 4 displays results using
the untransformed highway density variable. While the point estimates are positive starting in 1972,
they are estimated with substantial uncertainty and approach significance at standard levels only by the
2000s. The right panel of Figure 4 presents results using the a log-transformed version of the exit density
variable. Effect sizes reach 4 points by the 1990s, with estimates nearing statistical significance at the
α= 5% significance level across most of the study period.14
Exits Per Sq. Mi.
Log(Exits Per Sq. Mi.)
5
8
4
7
3
6
●
1
●
●
●
●
●
●
●
●
First Difference, Points
First Difference, Points
2
●
0
●
−1
●
−2
●
5
●
4
●
●
●
3
●
●
●
2
−3
●
●
−4
1
●
−5
●
0
●
●
−6
−1
−7
1960
1968
1976
1984
1992
2000
2008
1960
Year
1968
1976
1984
1992
2000
2008
Year
Figure 4: Simulated estimate of the first difference in urban-suburban party polarization, comparing
across the interquartile range of exits per square mile in each year, using the untransformed (left panel)
and log-transformed versions of the variable. Bootstrapped 80% and 95% confidence intervals accompany the point estimates.
Matching “treated” observations (those with a highway exit density greater than the sample median)
14
The interquartile range of the number of exits per square mile was [0.003, 0.1] in 1960, rising to
[0.19, 0.56] in 1996.
27
to “untreated” observations (those with highway exit density below the sample median) and applying
linear regression to the matched sample substantively similar, though less precise estimates. Across all
covariates across all years, matching reduced the average standardized imbalance between treatment and
control groups to 0.25 standard deviations of the control-group values in the pre-matched population.
Applying a linear regression model to reduce the consequences of the remaining imbalance (and accepting remaining model dependence) produces estimates of the average treatment effect on the treated
(ATT). The estimated treatment effect reaches more than 4 points by the 2008 election, with overall effect size is statistically significant at approximately the α = 0.2 level (two sided test) from 1996 to the
present (Figure 5). A graph of the absolute average standardized balance obtained on covariates after
matching appears in Appendix Figure A-16.
4.3
Responses to Data and Modeling Concerns
Of the results presented here, the most perplexing is that the effects of highways appear to be greater in
later periods, compared to estimated effects that are larger in the 1970s and 1980s in the suburban-county
analysis. But these seemingly divergent results are consistent. First, the inferential target presented here
is different: analyses in the suburban-county analysis were based on comparisons between suburban
counties with and without highways, while this analysis compares polarization within entire metropolitan areas. Because of the modifiable areal unit problem (Openshaw, 1984; Fotheringham and Wong,
1991), estimates on the suburban counties would not be expected to appear at the same times, or with the
same magnitude, as units that are assembled using weighted averages of county-level data. Second, the
treatment is defined differently, and varies by election year, with later years entailing a larger-magnitude
treatment because the density of Interstate highways, and therefore both the sample median and the sample quartiles, increased over time. We would, of course, expect a metropolitan area with more Interstate
28
OLS Estimates of ATT
Exits Per Sq. Mi. > Median
11
10
9
Polarization Effect, Points
8
7
6
5
●
4
●
3
●
2
●
1
0
●
●
●
●
●
−1
●
●
●
−2
●
−3
−4
−5
1960
1968
1976
1984
1992
2000
2008
Year
Figure 5: Effect of highways on the urban-suburban Democratic presidential vote gap, estimated by least
squares regression on the matched sample. The estimate captures the difference in the urban-suburban
Democratic presidential vote difference, shifting the number of built exits from below the median in
year t to above the median. Bootstrapped 95% and 80% confidence intervals are generated from the
lowess-smoothed curves to incorporate information from proximate observations.
29
highways to also have more suburban counties with highways, and the diffusion of residential development into suburban counties without highways would be expected to be more prevalent in metropolitan
areas with greater highway density. These spillover effects would explain attenuation of the causal effect estimated on the suburban county sample and the apparent concurrent spike in polarization in more
recent years.
While geographic analyses are often subject to arbitrary coding decisions, the couplets assembled
from counties within a metropolitan area are defensible on several grounds. Why not, one might ask, rely
on Census metropolitan statistical areas, a widely accepted definition of a metropolitan area? While this
may be preferred for contemporary cross-sectional analyses, such official boundaries may themselves be
a consequence of highway-induced suburban growth. In addition, metropolitan statistical areas, like the
couplets defined here, are composed of sets of counties. The adoption of a uniform radius is similarly
defensible on grounds that one could not know, in 1950, whether metropolitan population would expand
into a broader area, so the 80-kilometer outer radius was chosen to capture a roughly one-hour commuting radius from the central city. While the results are somewhat sensitive to this choice of outer radius,
the results are substantially similar to those presented in the main text over a range between 40 and 100
kilometers (Appendix Figure A-17).
As with all analyses using geographic data, there are modest sources of measurement error, most of
which are not subject to serious concerns. One is whether current exit placement is a suitable measure
of freeways’ overall impact on an urban area. The response to this concern can be found in historical
documentation. During early stages of implementation, the Bureau of Public Roads was deliberate in approving a limited number of exits to maintain the advantages of building expressways. Many exits were
built to connect to pre-existing highways, and approval of new exits on the Interstate system required
approval of the Secretary of Commerce (who at the time oversaw the Bureau of Public Roads) (Brag30
don, 1959). While exceptions to this fixed number of exits have since arisen, these amendments have
been small relative to the number of exits connected to the existing road network at the time of initial
construction. The density of exits in metropolitan areas was, to a much larger degree, a consequence of
local population density, which is included as a control in the analysis presented here.
5
Highway-Induced Changes in the Correlates of Partisanship
Partisan change in suburban counties and across metropolitan areas did not occur in a controlled environment. Highways had multiple effects, including changing the demographic and economic composition of
communities, and, while we cannot completely disentangle sorting and compositional effects, we would
expect highway-induced influences on both the political and demographic makeup of communities. In
the case of the suburban-county vote, highways may have aided the sorting of ex ante Democratic and
Republican identifiers, or placed voters in a context in which they are more likely to vote for one of
the major parties. We could attempt causal mediation analysis to engage these demographic changes’
relative influence over the total effect of highways on partisanship, but even in an experimental setting
such an analysis would require strong assumptions (Imai et al., 2011). As an alternative, we can estimate
highways’ effect on a range of correlates of partisanship to demonstrate the aggregate-level outcomes
that were most affected by the presence of highways. Available evidence from both the suburban counties and urban-suburban couplets during time period in question (using county-level Census data from
1970 to 2000) suggests that in the suburban comparative analysis, highways’ most pronounced effect
was to aid construction of wealthier commuter neighborhoods in suburban counties in which they were
built. Within urban-suburban couplets, it appears that racial change was a more important consequence
of highway construction.
The analyses presented here present compositional changes brought about by highway construction.
31
Estimated effects on five correlates of partisan voting behavior are presented, using the same models
used to estimate highways’ effect on cross-suburban and metropolitan area differences. Four of these are
socioeconomic correlates of partisanship, measured using data from the National Historical GIS (Fitch
and Ruggles, 2003): Average home value (2010 dollars), Per capita income (2010 dollars), Percentage
of workers over 16 working outside the county (a measure of a county’s commuter status), and the
Percentage of homes built in the previous decade. To test for highways’ effect on racial compositional
change, I estimate the effect on the Nonwhite percentage of the population.
The first of these findings applies the methods for the suburban-county analysis as presented in Section 3. Results of this analysis (Figure 6) suggest that an explanation of the partisan differences between suburban counties may stem from highway-induced changes in counties’ economic composition.
Specifically, counties with Interstate highways have had higher income, higher home values, and a higher
out-of-county commuter rate than comparable counties without them. These results are consistent with
the model of residential sorting laid out earlier: highways enable creation of new, upper-middle-class
and wealthy commuter suburbs in previously rural areas. They offer little evidence that highways have
brought about differences in racial composition between suburban counties, though this is unsurprising
given the overwhelmingly white population in the suburban sample.
One may speculate about the degree to which urban-suburban partisan polarization is similarly tied
to correlates of partisanship influenced by highway-induced residential sorting. Once again, we would
not expect the mechanisms at work in a comparison of suburban counties to carry over to an urbansuburban analysis, and, as with the analysis of the urban-suburban Democratic gap, estimates are likely
to be imprecise. Here, we can estimate the effect of highway density on the urban-suburban gap in per
capita income, age of housing stock, home values, percent non-white, and the percentage of residents
working outside the county. Lowess-smoothed estimates of highway density’s effects on urban-suburban
32
Per Capita Income (2010 $)
% Homes Built Last Decade
Avg Home Value (2010 $)
15000
2000
●
1500
●
4
●
2
●
●
5000
1000
0
500
●
1990
2000
●
●
1970
1980
1980
1990
2000
1970
1980
1990
2000
% Work Outside County
0.5
●
●
−10000
1970
% Non−White
1.0
●
●
●
−2
0.0
●
0
●
1980
●
−5000
0
1970
10000
−0.5
7
6
5
4
3
2
1
●
●
●
●
−1.0
1990
2000
1970
1980
1990
2000
Figure 6: Interstate highways’ impact on economic and racial correlates of partisanship in suburban
counties (average treatment effect on the treated in the population represented by the matched sample).
33
differences in selected Census variables are presented in Figure 7. Estimated on a small matched sample
for each election year, fairly large confidence intervals accompany each of these estimates. While these
imprecise estimates are by no means decisive, they suggest that geographic partisan change is linked to
highway-induced changes in key socioeconomic correlates of partisanship.
6
Conclusion
Public policies that change spatial relationships among citizens change politics. This article has considered how one such policy–the Interstate Highway System–changed the political composition of communities. Transportation networks can act as a catalyst for partisan sorting in the presence of partisan
homophily or when the public selects residential locations using factors tied to partisanship. While residential choice is often discussed in terms of relatively frictionless agent-based models, highway-induced
changes in the spatial distribution of Democrats and Republicans are not momentary changes; transportation networks indirectly subsidize the construction of new physical communities. In the case of the
Interstate Highway System, highways facilitated the growth of new suburban communities that attracted
and retained residents who either were (or became) more Republican than residents elsewhere.
This article has presented multiple findings consistent with the important role of transportation policy
in residential sorting and geographic polarization. Across a national sample, highways reduced the
Democratic vote by about two points in the counties where they were built, and these effects were
most pronounced exactly where highways had the most potential to have an effect: in the South, where
highways were responsible for an increase in the Democratic vote of as much as 5 points. Subsequent
analyses of the urban-suburban gap suggest a link between these results and urban-suburban polarization.
Metro areas with greater Interstate highway density (at the 75th percentile) have observed up to 4 points
more urban-suburban polarization in the Democratic presidential vote than comparable metropolitan
34
Per Capita Income (2010 $)
% Homes Built Last Decade
0.02
0
●
Polarization Effect, Points (ATT)
Polarization Effect, Points (ATT)
●
●
−1000
●
−2000
●
0.00
●
−0.02
●
●
−0.04
−3000
1970
1980
1990
2000
1970
1980
Year
1990
2000
Year
Avg Home Value (2010 $)
% Non−White
0.06
5000
0.04
0
●
Polarization Effect, Points (ATT)
Polarization Effect, Points (ATT)
●
−5000
●
−10000
●
−15000
●
●
0.02
●
0.00
−0.02
●
−20000
−0.04
−25000
1970
1980
1990
2000
1970
Year
1980
1990
2000
Year
% Work Outside County
Polarization Effect, Points (ATT)
0.05
0.00
●
●
●
●
−0.05
−0.10
1970
1980
1990
2000
Year
Figure 7: Highways and urban-suburban polarization on correlates of partisanship. Lowess-smoothed
estimate of the first difference in urban-suburban polarization in five key Census variables, using the
matched exit-density treatment variable. Bootstrapped 80% and 95% confidence intervals accompany
the point estimates.
35
areas with highway density at the 25th percentile.
These results suggest that substantial opportunities exist to expand studies of public policies’ effects
to incorporate political geography. The still new field of research on the political consequences of public
policy has focused on social policies’ impact on individual-level behavior (Campbell, 2003; Soss, 2000;
Mettler, 2002), but these studies have rarely taken into account how some public policies change political
behavior by shaping individuals’ geographic context (Enos, 2010; Gay, 2008). Such studies have rarely
engaged how place-making policies (Glaeser and Gottlieb, 2008) shape the American political map.
While numerous other long-term factors, including the geographic persistence of ethnic voting patterns
(Gimpel and Cho, 2004) can explain American partisans’ spatial distribution, the findings here point
clearly to public policy’s central role in these changes. The results presented here, triangulated with
limited national survey data (Roper Organization, 1976; Gallup, 1983; Belden, Russonello & Stewart,
2011), point to highways catalyzing individual-level sorting around a mix of partisan, economic, and
racial considerations. But substantial opportunities exist to extend this research using more fine-grained
geographic data, public databases, and custom surveys.15
The Interstate Highway System’s effects on both local and metro-level political geography call
into question studies that explain metropolitan development exclusively in terms of agent-based models founded on individual choice. To previous findings on local governments’ limited control over the
factors that attract residents and businesses (Peterson, 1981), we can add that local places (and the people who would like to live in them) may be at the mercy of infrastructure programs implemented by the
federal and state governments. Even in those areas in which municipalities have residential gatekeeping
power–zoning, policing, and other housing regulations–their ability to control the influx of new residents
15
For examples of research that uses more fine-grained data than is available from most national
surveys, see Gimpel, Cho and Hui (2009) and McDonald (2006).
36
is often at the mercy of transportation policies dictated from above.
Metropolitan regions with higher levels of geographic polarization are that way not only because of
individual sorting decisions, but because of the intervention of public policy. To be sure, the deindustrialization of American cities, the suburbanization of industry, urban renewal, racial segregation, the relative
quality of suburban and rural school districts, and a plethora of other factors contributed to growth of
the Republican vote in suburban areas. As surely as such changes reshaped metropolitan areas, however, highways acted as a catalyst, enabling residents to select into communities consistent with their
preferences. Policies intended to bring Americans closer together instead allowed them to live apart. An
unintended consequence for American politics was the creation of new suburban Republican enclaves
and a larger partisan split between major cities and their hinterlands.
References
Baum-Snow, Nathaniel. 2007. “Did Highways Cause Suburbanization?” The Quarterly Journal of
Economics 122(2):775–805.
Becker, Richard A, John M Chambers and Allan R Wilks. 1988. The New S. language. New York:
Wadsworth.
Belden, Russonello & Stewart. 2011. “2011 National Community Preference Survey.” Washington, DC:
.
Blackwell, Matthew, James Honaker and Gary King. 2011. “Multiple Overimputation: A Unified Approach to Measurement Error and Missing Data.”.
URL: http://gking.harvard.edu/gking/files/measure.pdf
37
Bragdon, John. 1959. “Bertram Tallamy to John Bragdon.” Letter, National Archives II, Bureau of Public
Roads, Box 67, “Interstate Highways 15 1958-59”.
Burns, Nancy. 1994. The Formation of American Local Governments. Oxford: Oxford.
Campbell, Andrea L. 2003. How Policies Make Citizens: Senior Political Activism and the American
Welfare State. Princeton: Princeton University Press.
Chandra, Amitabh and Eric Thompson. 2000. “Does Public Infrastructure Affect Economic Activity? Evidence from the Rural Interstate Highway System.” Regional Science and Urban Economics
30(4):457–490.
Cleveland, William S. 1979. “Robust Locally Weighted Regression and Scatterplots.” Journal of the
American Statistical Association 74(368):829–836.
Cleveland, William S. and Susan J. Devlin. 1988. “Locally Weighted Regression: An Approach to
Regression Analysis by Local Fitting.” Journal of the American Statistical Association 83:596–610.
Clubb, Jerome, William H. Flanigan and Nancy H. Zingale. 2006. Electoral Data for Counties in the
United States: Presidential and Congressional Races, 1840-1972. Ann Arbor, MI: Inter-university
Consortium for Political and Social Research.
Coleman, James S, Sara D Kelly and John A Moore. 1975. Trends in School Segregation, 1968-1973.
Technical report Urban Institute.
CQ Press. 2010. “CQ Press Voting and Elections Collection.” Online.
Diamond, Alexis and Jasjeet Sekhon. 2005. “Genetic Matching for Estimating Causal Effects: A New
Method of Achieving Balance in Observational Studies.” http://sekhon.berkeley.edu/.
38
Enos, Ryan D. 2010. “What Tearing Down Public Housing Projects Teaches Us About the Effect of
Racial Threat on Political Participation.” Unpublished.
ESRI. 2008a. Major Cities. In StreetMap North America. ESRI.
ESRI. 2008b. US and Canada Exits. In StreetMap North America. ESRI.
Fiorina, Morris P. and Samuel J. Abrams. 2008. “Political Polarization in the American Public.” Annual
Reviews in Political Science 11:563–588.
Fitch, Catherine A and Steven Ruggles. 2003. “Building the National Historical Geographic Information
System.” Historical Methods 36(1):41–51.
Fotheringham, A.S. and D.W.S. Wong. 1991. “The Modifiable Areal Unit Problem in Multivariate
Statistical Analysis.” Environment and Planning A 23(7):1025–1044.
Gallup. 1983. “Gallup Poll #1207G [computer file].”. Roper Center for Public Opinion Research.
Gay, Claudine. 2008. “Moving Out, Moving Up: Housing Mobility and the Political Participation of the
Poor.”.
Gimpel, James G. and Wendy K. Tam Cho. 2004. “The Persistence of White Ethnicity in New England
Politics.” Political Geography 23:987–1008.
Gimpel, James, Wendy Tam Cho and Iris Hui. 2009. “Regional Migration Flows and the Partisan Sorting of the American Electorate.” Paper presented at the American Politics Workshop, University of
Buffalo.
Glaeser, Edward L. and Joshua Gottlieb. 2008. “The Economics of Place-Making Policies.” Unpublished.
39
Hainmueller, Jens. 2012. “Entropy Balancing for Causal Effects: A Multivariate Reweighting Method
to Produce Balanced Samples in Observational Studies.” Political Analysis 20:25–46.
Ho, Daniel, Kosuke Imai, Gary King and Elizabeth Stuart. 2007. “Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference.” Political Analysis 15:199–
236.
Iacus, Stefano, Gary King and Giuseppe Porro. 2009. “CEM: Software for Coarsened Exact Matching.”
Journal of Statistical Software 30(9).
Iacus, Stefano M., Gary King and Giuseppe Porro. 2011a. “Causal Inference Without Balance Checking:
Coarsened Exact Matching.” Political Analysis .
Iacus, Stefano M., Gary King and Giuseppe Porro. 2011b. “Multivariate Matching Methods That are
Monotonic Imbalance Bounding.” Journal of the American Statistical Association 106:345–361.
Imai, Kosuke, Luke Keele, Dustin Tingley and Teppei Yamamoto. 2011. “Unpacking the Black Box
of Causality: Learning about Causal Mechanisms from Experimental and Observational Studies.”
American Political Science Review Forthcoming.
Inter-university Consortium for Political and Social Research. 1995. “General Election Data for the
United States, 1950-1990.” Online.
Keele, Luke. 2008. Semiparametric Regression for the Social Sciences. Chichester, UK: John Wiley &
Sons, Ltd.
King, Gary. 1996. “Why Context Should Not Count.” Political Geography 15(2):159–164.
40
King, Gary and Bradley Palmquist. 1998. “The Record of American Democracy, 1984-1990.” Sociological Methods and Research 26(3):424–427.
Krugman, Paul. 1992. Geography and Trade. Cambridge, MA: MIT Press.
Kruse, Kevin M. 2005. White Flight: Atlanta and the Making of Modern Conservatism. Princeton, NJ:
Princeton University Press.
Lassiter, Matthew D. 2006. The Silent Majority: Suburban Politics in the Sunbelt South. Princeton, NJ:
Princeton University Press.
Leip, Dave. 2010. “Dave Leip’s Atlas of U.S. Presidential Elections.” Online.
Levendusky, Matthew, Jeremy Pope and Simon Jackman. 2008. “Measuring District Level Partisanship
with Implications for the Analysis of U.S. Elections.” Journal of Politics 70(3):736–753.
Levine, Jonathan. 2006. Zoned Out: Regulation, Markets, and Choices in Transportation and Metropolitan Land-Use. Washington, D.C.: Resources for the Future Press.
Massey, Douglas and Nancy Denton. 1993. American Apartheid: Segregation and the Making of the
American Underclass. Cambridge, MA: Harvard University Press.
McDonald, Ian. 2006. “Migration and Sorting in the American Electorate: Evidence from the 2006
Cooperative Congressional Election Study.” American Politics Research 39(3):512–533.
Mettler, Suzanne. 2002. “Bringing the State Back In to Civic Engagement: Policy Feedback Effects of
the GI Bill for World War II Veterans.” American Political Science Review 96(2):351–365.
Michaels, Guy. 2008. “The Effect of Trade on the Demand for Skill: Evidence from the Interstate
Highway System.” The Review of Economics and Statistics 90(4):683–701.
41
Mieszkowski, Peter and Edwin S. Mills. 1993. “The Causes of Metropolitan Suburbanization.” Journal
of Economic Perspectives 7(3):135–147.
Mohl, Raymond A. 2004. “Stop the Road: Freeway Revolts in American Cities.” Journal of Urban
History 30(5):674.
National Interregional Highway Committee. 1944. “Interregional Highways.” Washington, DC: .
Openshaw, Stan. 1984. The Modifiable Areal Unit Problem. Norwich: Geo Books.
Peterson, Paul E. 1981. City Limits. Chicago: Chicago.
Pierson, Paul. 1993. “When Effect Becomes Cause: Policy Feedback and Political Change.” World
Politics pp. 595–628.
Roper Organization. 1976. “Roper Reports Poll # 1976-02.” Roper Center for Public Opinion Research.
Rubin, Donald B. 1976. “Multivariate Matching Methods that are Equally Percent Bias Reducing, II:
Maximums on Bias Reduction for Fixed Sampled Sizes.” Biometrics 32:121–132.
Rubin, Donald B. 1986. “Comment: Which If’s Have Causal Answers.” Journal of the American Statistical Association 81(396):961–962.
Rubin, Donald B. 1991. “Practical Implications of Modes of Statistical Inference for Causal Effects and
the Critical Role of the Assignment Mechanism.” Biometrics 47(4):1213–1234.
Schelling, Thomas. 1971. “Dynamic Models of Segregation.” The Journal of Mathematical Sociology
1(2):143–186.
Schulman, Bruce J. 1994. From Cotton Belt to Sunbelt: Federal Policy, Economic Development, and the
Transformation of the South, 1938-1980. Durham, NC: Duke University Press.
42
Shafer, Byron E and Richard Johnston. 2006. The End of Southern Exceptionalism: Class, Race and
Partisan Change in the Postwar South. Cambridge, MA: Harvard University Press.
Soss, Joe. 2000. Unwanted Claims: The Politics of Participation in the US Welfare System. Ann Arbor:
University of Michigan Press.
Spilerman, Seymour. 1973. “Strategic Considerations in Analyzing the Distribution of Racial Disturbances.” American Sociological Review 38(493-499).
Sugrue, Thomas. 1995. “Crabgrass-Roots Politics: Race, Rights, and the Reaction against Liberalism in
the Urban North, 1940-1964.” Journal of American History 82(2):551–578.
Tiebout, Charles M. 1956. “A Pure Theory of Local Expenditures.” The Journal of Political Economy
64(5):416.
43
Online Appendix
This online methodological appendix contains a range of supplemental analyses and robustness checks,
divided among the paper’s two main empirical sections.
A
County-Level Analysis
A.1
Suburban-County Sample
A.1.1
Defining the Suburban County Sample
This section addresses one of the basic concerns that arises in the study of American metropolitan areas:
the potential sensitivity of findings to rule-based sample definitions. The results presented in the body of
the paper define suburban counties by including only counties more than 20 kilometers from the central
city and less than 100 kilometers from the central city. Here, I demonstrate that these findings are robust
to both selection of the inner and outer radius and to a sampling frame based on population density. I
also show that the results are robust to the population cap used for suburban counties.
First, I demonstrate the robustness of the findings to selection of inner and outer radii. Results for
the full suburban county sample, excluding pre-1956 freeway adopters and applying coarsened exact
matching, appear in Figures A-1 and A-2. To begin, the full sample, excluding early-adopter states,
is subsampled using a grid-search procedure. This grid search covers three different values of the inner radius (10, 20, and 30 kilometers) and six values of outer radius (40, 50, 60, 70, 80, 90, and 100
kilometers). This yields a total of 21 different subsamples. On each of these samples, coarsened exact
matching and least squares regression are used to estimate the average treatment effect of highways on
the suburban county vote, following the procedure described in the text. For all samples with an outer
44
radius greater than 50 kilometers, the results corroborate those presented in the main text.
Figure A-3 displays a similar grid search over sample definition rules. Instead of defining the suburban sample for each metropolitan county in terms of defines the sample in terms of county population
density. Each sample is defined by selecting different minimum and maximum county population densities, defined in terms of sample quantiles. This grid search applies the same matched sampling and
regression methods to samples defined by shrinking the density window progressively. Samples that
include more low-population-density counties yield results closer to those obtained. The findings are not
robust to including only counties near median population density.
Finally, Figure A-4 displays the consequences of changing the maximum county population (from the
1950 Census) permitted in the suburban county sample. This population cap removes from the sample
large counties that are outside the central city of a metropolitan area but may nevertheless be heavily
urbanized. Varying the maximum permissible population from 100,000 to 600,000 has no meaningful
bearing on the results.
A.2
Simple Difference in Means
One may be concerned that the estimates presented here are a result of specification hunting. One
intuitive way to allay these concerns is to present a simple difference-in-means estimate. While there is
no clear reason to expect that the results must run in the same direction as the results presented in the
body of the paper, this is a helpful way to present the extent to which model choices may be responsible
for the point estimates presented in the paper. This figure Figure A-5 presents the difference in the mean
Democratic vote (in percentage points) for these three sets of estimates.
45
−0.2
−0.3
0.00
2.0
1.5
1.0
0.5
0.0
−0.5
−1.0
−1.5
1988
●
● ● ● ● ● ● ● ● ● ●
1964
Outer radius: 40 km
Inner radius: 30 km
Obs: 64
0.1
0.0
−0.1
● ● ● ●
●
● ● ● ●
● ●
−0.2
−0.3
1988
1964
1988
Year
Year
Year
Outer radius: 50 km
Inner radius: 10 km
Obs: 168
Outer radius: 50 km
Inner radius: 20 km
Obs: 147
Outer radius: 50 km
Inner radius: 30 km
Obs: 119
●
●
●
●
●
−0.05
●
●
●
●
● ●
−0.10
1964
0.00
●
●
●
●
−0.05
●
●
●
●
●
−0.10
●
●
−0.15
1988
1964
Effect of Hwy in Cnty
Effect of Hwy in Cnty
1964
Outer radius: 40 km
Inner radius: 20 km
Obs: 92
Effect of Hwy in Cnty
● ●
● ● ●
● ● ● ●
● ●
−0.1
Effect of Hwy in Cnty
0.0
Effect of Hwy in Cnty
Effect of Hwy in Cnty
Outer radius: 40 km
Inner radius: 10 km
Obs: 113
● ●
0.4
●
0.2
●
0.0
● ● ●
−0.2
●
●
−0.4
1988
1964
1988
●
Year
Year
Year
Outer radius: 60 km
Inner radius: 10 km
Obs: 255
Outer radius: 60 km
Inner radius: 20 km
Obs: 234
Outer radius: 60 km
Inner radius: 30 km
Obs: 206
●
−0.02
●
●
●
−0.06
●
●
●
−0.08
1964
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
1988
●
●
●
●
●
●
●
●
●
1964
● ●
0.04
0.02
0.00
−0.02
−0.04
−0.06
−0.08
1988
●
● ●
●
● ●
● ●
● ● ●
1964
1988
Year
Year
Year
Outer radius: 70 km
Inner radius: 10 km
Obs: 347
Outer radius: 70 km
Inner radius: 20 km
Obs: 326
Outer radius: 70 km
Inner radius: 30 km
Obs: 298
●
●
● ●
● ●
●
●
●
●
●
1964
1988
Year
0.02
●
0.00
−0.02
●
● ●
●
−0.04
●
●
●
●
● ●
−0.06
1964
1988
Year
Effect of Hwy in Cnty
Effect of Hwy in Cnty
●
●
●
−0.04
0.02
0.00
−0.02
−0.04
−0.06
−0.08
−0.10
Effect of Hwy in Cnty
●
Effect of Hwy in Cnty
0.00
Effect of Hwy in Cnty
Effect of Hwy in Cnty
●
0.02
0.00
−0.02
●
●
−0.04
●
●
●
●
● ●
● ●
●
−0.06
1964
1988
Year
Figure A-1: Robustness of findings to definition of a suburban county sample based on the inner and outer
radius defining rings drawn around the metropolitan central city. Outer radii of 40 to 70 kilometers.
46
●
●
●
−0.05
−0.06
●
●
●
●
●
●
●
●
Outer radius: 80 km
Inner radius: 30 km
Obs: 393
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
1988
●
●
●
●
● ●
●
1964
●
●
●
●
1988
Year
Year
Outer radius: 90 km
Inner radius: 10 km
Obs: 528
Outer radius: 90 km
Inner radius: 20 km
Obs: 507
Outer radius: 90 km
Inner radius: 30 km
Obs: 479
● ●
●
●
●
●
●
●
●
●
●
−0.04
1964
−0.01
−0.02
●
●
● ●
●
−0.03
●
●
−0.04
●
●
●
●
−0.05
−0.06
1988
1964
Effect of Hwy in Cnty
Year
−0.02
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
1988
●
● ● ●
●
●
●
● ●
1964
●
●
1988
Year
Year
Outer radius: 100 km
Inner radius: 10 km
Obs: 615
Outer radius: 100 km
Inner radius: 20 km
Obs: 594
Outer radius: 100 km
Inner radius: 30 km
Obs: 566
●
−0.01
● ●
−0.02
●
●
●
● ●
●
−0.04
1964
1988
Year
●
●
0.00
●
−0.01
−0.02
● ●
●
●
●
●
●
●
●
●
−0.03
−0.04
1964
1988
Year
Effect of Hwy in Cnty
Year
0.00
−0.03
●
● ●
1964
−0.01
−0.03
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
1988
Effect of Hwy in Cnty
Effect of Hwy in Cnty
● ●
●
1964
Effect of Hwy in Cnty
●
●
−0.03
−0.04
● ●
Outer radius: 80 km
Inner radius: 20 km
Obs: 421
Effect of Hwy in Cnty
●
−0.02
Effect of Hwy in Cnty
−0.01
Effect of Hwy in Cnty
Effect of Hwy in Cnty
Outer radius: 80 km
Inner radius: 10 km
Obs: 442
0.01
0.00
−0.01
−0.02
−0.03
−0.04
−0.05
●
●
●
1964
●
● ●
●
●
● ●
●
1988
Year
Figure A-2: Robustness of findings to definition of a suburban county sample based on the inner and outer
radius defining rings drawn around the metropolitan central city. Outer radii of 80 to 100 kilometers.
47
0.6
0.4
0.2
●
●
●
●
●
●
●
●
●
1976
1988
0.04
0.02
●
0.00
●
●
●
●
●
●
●
●
●
●
−0.02
−0.04
2000
1964
1976
1988
2000
Year
Log Density Quantile Interval: [0.3,0.7]
Obs:303
Log Density Quantile Interval: [0.25,0.75]
Obs:356
Effect of Hwy in Cnty
Year
0.04
0.02
0.00
●
●
●
●
●
●
●
●
●
●
−0.02
●
−0.04
1976
1988
0.02
●
0.00
●
●
●
●
●
●
−0.02
●
●
●
●
−0.04
2000
1964
1976
1988
2000
Year
Year
Log Density Quantile Interval: [0.2,0.8]
Obs:413
Log Density Quantile Interval: [0.15,0.85]
Obs:471
Effect of Hwy in Cnty
Effect of Hwy in Cnty
●
−0.4
1964
Effect of Hwy in Cnty
●
−0.2
1964
0.02
●
0.00
●
−0.02
●
●
●
●
●
●
●
●
●
−0.04
1964
Effect of Hwy in Cnty
Effect of Hwy in Cnty
0.8
0.0
Log Density Quantile Interval: [0.35,0.65]
Obs:226
1976
1988
0.02
0.00
●
●
−0.02
●
●
●
●
●
●
●
●
−0.04
●
−0.06
2000
1964
1976
1988
2000
Year
Year
Log Density Quantile Interval: [0.1,0.9]
Obs:520
Log Density Quantile Interval: [0.05,0.95]
Obs:566
0.02
0.01
0.00
−0.01
−0.02
−0.03
−0.04
●
●
1964
●
●
●
1976
●
●
●
●
1988
●
●
Effect of Hwy in Cnty
Effect of Hwy in Cnty
Log Density Quantile Interval: [0.4,0.6]
Obs:149
2000
0.01
0.00
−0.01
−0.02
−0.03
−0.04
●
●
●
1964
Year
●
●
1976
●
●
●
●
1988
●
●
2000
Year
Figure A-3: Robustness of findings to definition of suburban county sample by population density.
48
Population cutoff: 1e+05
Obs: 567
Population cutoff: 2e+05
Obs: 591
0.00
0.00
●
−0.01
●
●
−0.02
●
●
●
●
●
●
−0.03
●
●
−0.04
Effect of Hwy in Cnty
Effect of Hwy in Cnty
0.01
−0.01
●
●
●
●
●
●
●
●
−0.02
●
●
●
−0.03
−0.04
−0.05
1964
1976
1988
2000
1964
1976
1988
2000
Year
Year
Population cutoff: 3e+05
Obs: 594
Population cutoff: 4e+05
Obs: 597
●
−0.01
●
●
●
−0.02
●
●
●
●
●
●
●
−0.03
Effect of Hwy in Cnty
Effect of Hwy in Cnty
0.01
0.00
●
−0.01
●
●
●
●
●
●
●
●
−0.02
●
●
−0.03
−0.04
−0.04
1964
1976
1988
2000
1964
1976
1988
2000
Year
Year
Population cutoff: 5e+05
Obs: 597
Population cutoff: 6e+05
Obs: 598
0.00
●
−0.01
●
●
●
●
●
●
●
●
−0.02
●
●
−0.03
Effect of Hwy in Cnty
Effect of Hwy in Cnty
0.00
0.00
●
−0.01
●
●
−0.02
●
●
●
●
●
●
●
●
−0.03
−0.04
−0.04
1964
1976
1988
2000
1964
Year
1976
1988
2000
Year
Figure A-4: Robustness of findings to exclusion of high-population counties with varying cutoffs.
49
Difference in Democratic Vote, (Treated−Untreated)
Simple Difference in Means
Suburban Counties
10
●
●
0
●
●
●
●
●
●
●
●
●
●
●
−10
−20
1960
1968
1976
1984
1992
2000
2008
Year
Figure A-5: Difference in mean Democratic vote share between counties with an Interstate highway and
counties without one. Counties in early-adopter states excluded.
A.3
Balance Checks and Placebo Tests
A.3.1
Coarsened Exact Matching
I present the standardized imbalance for the covariates included in each matched sample, covering election years between 1964 and 2008. With a few isolated exceptions, across all years, the matched sample
has substantially less imbalance on the included covariates. Note that the initial imbalance on most
covariates is surprisingly modest, which may help to explain why even the simple difference in means
estimate of the causal effect does not deviate much from the estimates presented elsewhere in the paper.
These results appear in Figure A-6, which displays the standardized imbalance (obtained by dividing
the difference in means between the treated and control group by the standard deviation of the control
group), for both the full sample and sample matched using coarsened exact matching.
50
Standardized Imbalance
Log Population Density, 1950
1.0
0.5
0.0
F
F F F F F F F F F F F
M M
Standardized Imbalance
F
0.5
0.0
F
M M
F F
F F F F F F F
M M M M M M
M M M M
0.5
0.0
−0.5
−1.0
−1.0
−1.0
1980
1996
1964
Year
1952 GOP Pres. Vote
1.0
1980
1996
F
F F F
F F F F F F F
M M M M M M M M M M M
1980
1996
Year
Percentage Non−White, 1950
1.0
0.5
F
M M F
F F F F F F F F F
M M
M M M M M M M M
M
1964
Year
1948 GOP Pres. Vote
1.0
F
0.5
0.0
F
1956 GOP Pres. Vote
1.0
−0.5
0.5
M
0.0
F
F
F F
M M M
F M
M M
F M
F M
F M
F M
F M
F
0.0
M M
M M M
F M
F M
F M
F M
F
M M M
F F F
F
F F
−0.5
−0.5
−0.5
−1.0
−1.0
−1.0
1964
Standardized Imbalance
F
−0.5
1964
1980
1996
1964
Year
Pct. Urban, 1950
1.0
1980
1996
1964
Year
Crop Value Per Capita, 1950
1980
1996
Year
Pct. Incoming Residents, 1950
1.0
1.0
0.5
0.5
F
F F F F F F F F F F F
0.5
0.0
M M
M M M M M M M M M M
0.0
M M M M M M M M M M M
F F F F F
M
F F F F F
F F
0.0
−0.5
−0.5
−0.5
−1.0
−1.0
−1.0
1964
Standardized Imbalance
M M M M M M M
M M M
F
Median Family Income
1.0
1980
1996
1964
Year
South
1.0
1996
M M M M M M M
M M M M M
F F F F F F F
F
F F
1964
Year
Mfg. Estab., 1939
1.0
0.5
0.0
1980
F F F F F F F F
M M M F M
M M M M M M M
F F M
F
F
F
1.0
0.5
F
0.0
M M M M M M M M M M M M
F F F F F F F F F F F
1980
F
1996
Year
Mil Need
F
F F F F F F F F
0.5
F
0.0
M M M M M M M M M M M M
F
−0.5
−1.0
F
1964
1980
1996
−0.5
−0.5
−1.0
−1.0
1964
1980
1996
1964
1980
1996
Figure A-6: Standardized imbalance of included covariates (difference in means between the treated and
control group, divided by the standard deviation in the original treated group) under CEM using the timevarying version of the Interstate highway treatment variable and coarsened exact matching. Standardized
imbalance in the [M]atched and [F]ull (original) sample are presented for the matched sample for each
election year, 1960-2008.
51
A.3.2
Placebo Test
I present a placebo test of the effect of Interstate highway construction on the pre-treatment Democratic
presidential vote and on a range of pre-treatment variables in the suburban-county sample, including
several variables from earlier Censuses that were not included in the matching procedure. These tests
use the same methods used to generate the main set of estimates, with the exception that the differencein-difference analysis is not applicable in this setting. The first of these placebo tests, in Figure A-7,
presents an estimate of the presence of an Interstate highway in each county in each election year on
the average Democratic vote share in the 1944 thru 1956 presidential elections. Figure A-8 presents tstatistics from placebo tests for a range of pre-1956 measures under the 1930, 1940, and 1950 Censuses,
as well as the presidential vote results included in the above placebo test. Note that the placebo tests
in Figure A-7 and in Figure A-8 are done without exploiting difference-in-difference estimation (i.e.,
without subtracting out the 1950 or 1952 values). This makes them a difficult test for balance, because in
practice with difference-in-difference estimation, some imbalance in the potential outcomes in election
year t is acceptable as long as the potential trends between 1952 and year t in the average presidential
vote are balanced between treatment and control groups, conditioning on the covariates. This makes this
particular placebo test a difficult one.
A.4
Use of Presidential Outcomes as a Measure of Partisanship
One potential concern is that by adopting the presidential vote as the outcome of interest, the analysis
is not fairly capturing partisanship. This section offers two responses to this concern. The first is that
the presidential vote is, perhaps, the most important manifestation of partisan voting, so it should be
an outcome of interest in its own right. Second, lowess-smoothing of the point estimates minimizes
the consequences of unusual election years–such as 1964 and 1976–in which partisan voting behavior
52
All Suburban Counties
2
1
0
●
●
●
●
●
●
●
●
●
●
●
●
−1
Avg Effect on Dem %
(ATT, Matched Sample)
−2
−3
−4
−5
1964
1972
1980
1988
1996
2004
Year
Figure A-7: Placebo test: effect of Interstate highways in each year on the average Democratic presidential vote, 1944-1956.
shifted suddenly in response to different candidates. Finally, Levendusky, Pope and Jackman (2008),
who develop Bayesian methods for estimating latent partisanship, find that the presidential vote correlates highly with their measure of latent partisanship, and “provide reassurance to researchers who have
used district-level presidential vote as a proxy for district-level partisanship” (Levendusky, Pope and
Jackman, 2008, 750). The discussion in Levendusky, Pope and Jackman (2008) also suggests that a large
proportion of the difference between the two measures can be attributed to home-state “favorite son” effects. To demonstrate robustness to this aspect of presidential returns, a robustness check presented here
includes two dummy variables: one for home-state Democrats, the other for home-state Republicans.
Including these variables does not meaningfully shift the estimates presented in the text (Figure A-9).
These results do little to shift the overall substantive interpretation of the results. If anything, accounting
for favorite-son candidates appears to lessen some of the regional heterogeneity in effect sizes.
53
4
Placebo Test t Statistic
2
% Urban, 1940
Value
PerCapita,
Capita,1940
1940
Crop Added
Value Per
% Non−White 1930
% Urban,
1930 1940
Mfg Wages
Per Capita,
% Non−White, 1940
%
Negro,
1940
%Family
Negro,
1930 1950
Median
1950
Crop%Value
PerIncome,
Capita,
Non−White,
1950
%
Urban,
1950
0
Mfg Ests Per Capita, 1940
Mfg Ests Per Capita, 1930
Value Added Per Capita, 1930
Mfg Wages Per Capita, 1930
−2
% Homes Owned, 1930
−4
1964
1972
1980
1988
1996
2004
Year
Figure A-8: T-statistics of placebo tests on a range of pre-treatment Census and political covariates.
Abbreviated labels for each variable appear at right. Absolute values of these t-statistics are mostly less
than two, suggesting that conditional ignorability has been satisfied.
54
● ●
● ●
● ●
● ●
1964
1976
1988
● ●
●
●
2000
0.03
0.02
0.01
0
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
South
● ●
●
●
●
●
● ●
●
1964
1976
Year
1988
Year
●
●
●
2000
Diff. in Change in Dem. Vote Share, 1952−(Year)
All Suburban Counties
Diff. in Change in Dem. Vote Share, 1952−(Year)
Diff. in Change in Dem. Vote Share, 1952−(Year)
0.03
0.02
0.01
0
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
0.03
0.02
0.01
0
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
Non−South
● ●
● ●
● ●
● ●
●
1964
1976
●
1988
●
●
2000
Year
Figure A-9: Robustness check using Democratic “favorite son” and Republican “favorite son” dummy
variables. Lowess-smoothed, bootstrapped OLS estimates of the effect of construction of an Interstate
highway in a county by year t − 4 on the difference in the Democratic vote between 1952 and year t,
applied to CEM-matched sample. Results apply to the full suburban county sample, excluding early
adopters. 95% (solid line) and 80% (dashed line) confidence envelopes are generated from lowess
smoothing of point estimates. Note that the additional covariate for “favorite son” candidates is included
only at the analysis stage, not in the matching procedure.
55
A.5
Accounting for Uniform Swings
The outcome is also recoded to account for year-to-year national changes in elections. The outcome
variable is defined by subtracting the national Democratic presidential vote from the county level vote in
each election year. This step allows slightly more direct comparability of the presidential election results
across elections.
The results derived from this approach vary slightly from those that appear in the body of the paper.
On the full sample, the smoothed estimates of the ATT on the county-level Democratic vote (after accounting for national swings) reach the level of -2 percentage points, holding steady through the 2008
election (Figure A-10). An analysis excluding early adopter units yields slightly smaller, and sometimes
imprecise, estimates, reaching 1 percentage point. Several of these smoothed estimates fail to reach statistical significance at the customary 95% confidence level (Figure A-10). Once again, the effect sizes
appear to be slightly larger in the South, but accounting for the national swing in the vote may slightly
lesson interregional differences.
A.6
Urban Riots
The construction of highways through predominantly black neighborhoods was a major urban grievance
(among many others) that motivated urban unrest. As such, it is one of many causal pathways by which
highways may have contributed to the developments presented here. At the same time, many urban
disturbances may have developed in response to a broad array of other grievances and due to varying
local demographics. Thus, accounting for the occurrence of disturbances may anticipate one instance of
omitted variable bias not anticipated . For each metropolitan area in the dataset, a “riots” dummy variable
was coded with the number of years in the 1961-1968 period for which Spilerman (1973) reported a
disturbance. Again, including this covariate has a minimal effect on the point estimates. An analysis
56
●
●
●
●
●
● ●
●
● ● ●
1964
1976
1988
2000
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
South
●
●
●
●
● ●
●
● ●
●
●
1964
1976
Year
1988
Year
2000
●
Diff. in Change in Dem. Vote Share, 1952−(Year)
●
Diff. in Change in Dem. Vote Share, 1952−(Year)
Diff. in Change in Dem. Vote Share, 1952−(Year)
All Suburban Counties
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
Non−South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
●
●
● ●
●
●
1964
● ●
1976
●
● ●
1988
2000
Year
Figure A-10: Robustness check recoding the outcome as the county-level difference from the Democratic
percentage of the national presidential vote. This captures major national swings in the Democratic vote.
Lowess-smoothed, bootstrapped OLS estimates of the effect of construction of a Interstate highway in
a county by year t − 4 on the difference in the Democratic vote between 1952 and year t, applied to
CEM-matched sample. 95% (solid line) and 80% (dashed line) confidence envelopes are generated from
lowess smoothing of point estimates.
57
●
● ●
● ●
●
1964
● ●
● ● ●
1976
1988
2000
South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
●
●
●
●
●
● ●
● ●
● ●
1964
1976
Year
1988
2000
Diff. in Change in Dem. Vote Share, 1952−(Year)
●
Diff. in Change in Dem. Vote Share, 1952−(Year)
Diff. in Change in Dem. Vote Share, 1952−(Year)
All Suburban Counties
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
Year
Non−South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
● ●
●
●
1964
●
● ●
● ● ● ●
1976
1988
●
2000
Year
Figure A-11: Robustness check including a covariate for the number of years between 1961 and 1968
during which a racial disturbance occurred in a metropolitan area (Spilerman, 1973). Lowess-smoothed,
bootstrapped OLS estimates of the effect of construction of a Interstate highway in a county by year
t − 4 on the difference in the Democratic vote between 1952 and year t, applied to CEM-matched
sample. Results apply to the full suburban county sample, excluding early adopters. 95% (solid line) and
80% (dashed line) confidence envelopes are generated from lowess smoothing of point estimates.
using this alternative explanation is excluded from the body text because recorded urban unrest occurred
during the post-treatment period in almost all cases. These results, applied to the full sample (with
matching and excluding early-adopter states), appear in Figure A-11.
58
The next analysis aims to account for sensitivity of the findings to the use of baseline election year.
One could assert that the results presented here are a result of the selection of 1952 as the baseline year
in the difference-in-difference analysis. This year could have been vulnerable to particular year-specific
effects, including the fact that it was Eisenhower’s first presidency.
A.7
The Influence of Desegregation Orders
Another robustness check examines the effect of school desegregation orders and agreements on the
growth of Republican suburbs. Metropolitan areas with such orders may, according to hypotheses presented in the late 1970s, lead to more white flight. Of course, such orders may be at least partially enabled
by the construction of the highways themselves. Busing is a more practical solution in a metropolitan
area with a highly developed metropolitan highway system. Nevertheless, we can observe how robust
findings are to the coding of these desegregation orders. Using the dataset assembled in Coleman, Kelly
and Moore (1975), a variable is created for each election year to indicate whether a desegregation program had been adopted in the central city (or cities) of a metropolitan area at least four years before each
election year.16 Results of the main regressions, with the added desegregation variable included, appear
in Figure A-13.
B
Polarization Analysis
B.1
Defining Urban-Suburban Couplets
To account for large metropolitan areas in which a top-100 city falls within the suburban zone of another
major city (e.g., Jersey City and New York; Dallas and Fort Worth), such paired cities were merged into
a single large metropolitan area, and their suburban catchment areas combined into a single larger catch16
No desegregation orders had been adopted as of the 1960 election.
59
ment area. For cases in which a smaller central city was clearly subordinate to a larger city (e.g., New
York and Jersey City), the county containing the smaller central city was redefined as “suburban.” However, in the case of clear multi-city metropolises such as Dallas-Fort Worth and Minneapolis-St. Paul,
each of the counties containing the central cities were classified as “urban” and their data aggregated.
B.2
Limitations of Precinct-Level Data
I explored the use of municipal-level election returns, which would have provided finer spatial resolution. This included calculations based on data from the Record of American Democracy (ROAD)
project (King and Palmquist, 1998), which provides solid precinct-level coverage at least for the country’s northeastern quadrant. However, many states do not retain election records below the county level,
and in other cases historical Census data to match electoral geography are not readily available. The
absence of pre-treatment Census data at low aggregation levels (Fitch and Ruggles, 2003) prevents the
inclusion of suitable covariates in model-based analyses, making causal analysis considerably more difficult. Spatial mismatches between the ROAD data and Census data similarly required substantial missing
data imputation, adding noise to included covariates and potentially biasing point estimates (Blackwell,
Honaker and King, 2011). Recent systematic projects to collect precinct-level data can do little to solve
this historical data mismatch problem. With these limitations in mind, the analysis presented here is
limited to county-level data.
B.3
The Secular Change in Urban-Suburban Polarization
A plot of ∆i,2008 , the change in the urban-suburban difference in the Democratic presidential vote between 1952 and year t, appears in Figure A-14. Taking the unweighted average of all metro areas in the
sample, this shift was about 10.5 percentage points, but, like the results from the suburban county-level
analysis, the increase in polarization varies by region. While metro areas across the country have be60
come increasingly polarized, the shift has been particularly pronounced in the South, where five major
metropolitan areas experienced the largest urban-suburban shift. Memphis (Shelby County) and New
Orleans (Orleans Parish), both of which contained large, newly enfranchised Black Democratic populations, underwent the largest absolute increase in urban-suburban polarization, presumably as a result of
enfranchisement of and political mobilization of large Black populations under the Voting Rights Act.17
Other, fast-growing areas of the New South from Austin to Atlanta similarly became polarized, with their
central counties becoming 30 points more Democratic relative to their suburbs than they were in 1952.
Across most of the rest of the country, the difference in the growth of the urban Democratic advantage
and the growth of the suburban Democratic advantage was at least ten points in most metropolitan areas.
Urban-suburban polarization increased almost everywhere, with the few exceptions appearing in large
Western counties that encompass entire metropolitan areas (thus introducing error in the measurement of
the urban-suburban gap) and a handful of depopulating, deindustrializing, usually smaller metropolitan
areas mostly located in the Rust Belt. In the metropolitan areas whose centers became more Republican
than their peripheries, the size of the shift was typically not more than 10 points.
B.4
Bivariate Plots
One reasonable concern is that the results presented here are a consequence of specification hunting, a
concern that follows all observational studies. One approach to verify the robustness of the findings to
model specification is to present bivariate relationships without controls. If the bivariate relationship is
in the same direction as the causal estimate, it does not establish whether results were biased as a result
17
Remarkably, in the case of Memphis, even the presence of the nearby Mississippi River Black Belt
did not prevent a massive increase in urban-suburban polarization in the Memphis area over the past 56
years.
61
of model specification hunting, but it does suggest that the results are not as dependent on the model
and correct choice of covariates for regression and matching. Figure A-15 presents a series of bivariate
regressions for each election year after the election. Each point in these graphs represents a metropolitan
area, defined by the 80-kilometer catchment area around the central city (cities). Across the board during
the post-1956 period, the point estimate of the effect of highway density on urban-suburban polarization
is positive.
B.5
Matching Quality
A summary of balance statistics obtained for each year appear in Figure A-16. This presents the average of the absolute standardized imbalance for covariates included in matching in each year, using the
dichotomized treatment variable obtained by dividing the sample at the median.
B.6
Placebo Test
B.7
The Baum-Snow Data
Baum-Snow (2007) assigned geographic locations to the nearest mile, interpolating between the start and
end points of each freeway project segment. When the Bureau of Public Roads and the Federal Highway Administration monitored state governments’ highway construction progress, data on each highway
segment were collected and organized on “strip maps,” long sheets displaying an abstract version of the
highway segment annotated with information on the construction start and end dates. The segments that
appeared on each map were as short as a fraction of a mile near the center of urban areas, permitting a
great deal of geographic precision. However, in the most rural areas, these construction areas typically
run from one end of a county to another. This is fine enough detail that in metropolitan areas, the PR-511
data are sufficient to identify whether a county was “treated” with an Interstate highway or not.
62
B.8
Instrumental Variables Analysis
The instrumental variables regressions presented in this chapter act primarily to bolster findings obtained
using OLS estimation and matching, though the technique can require questionable assumptions. Instrumental variables estimation must satisfy both conditional ignorability of the instrument and the rarely
plausible exclusion restriction. Moreover, though OLS and 2SLS yield different estimates, they often
yield similar substantive conclusions. Baum-Snow (2007) reaches the same substantive conclusions
about the direction of freeways’ effect on suburban growth using OLS and 2SLS estimation. Though
the estimates are not directly comparable, in both cases the estimated effect was positive and statistically significant. Finally, the substantive assumptions used in instrumental variables estimation are often
questionable. For example, the same reports from which highway-plan instruments have been drawn
anticipated that highway construction would stimulate suburban growth. If this knowledge had changed
behavior among policy makers in areas other than highway construction, or if this map was in any way
related to the outcomes of interest by means other than highway construction, then the highway instrument would violate the exclusion restriction, potentially adding bias to causal estimation rather than
reducing it, and with no way of assessing the potential direction of the bias.18 Instrumental variables
18
While Baum-Snow (2007) repeats the widely held belief that the Interstate Highway System was
designed with military prerogatives in mind, such claims typically overstate the War and Defense Departments’ involvement in highway planning. While the military offered input on the location of freeways
relative to military bases, its involvement was at most peripheral: though some local highway routes were
planned with military imperatives in mind, the facilitation of interstate commerce and easing of urban
traffic congestion were the primary guiding principles of highway system planning, and defense officials
never adopted an integral role in highway planning.
63
regression is suitable as another estimation method, but is not inherently less biased than combinations
of matching, difference-in-difference, and linear-model-based estimation.
B.9
Varying the Metropolitan-Area Radius
I test the robustness of the results to selection of the outer radius of each metropolitan area. The first
sensitivity analysis, which varies the outer radius from 40 to 100 kilometers for the matching-based
estimates of the effects of exit density, confirms that the results are highly robust to the choice of outer
radius (Figure A-17). Varying the urban outer radius for the rays-based analysis yields point estimates
that are typically in the correct direction, but are not as robust to the selection of an outer radius. The
already marginal OLS results using the rays variable are substantially less robust to the selection of outer
radius (Figure A-17).
B.10
Highways’ Effect on Urban and Suburban Portions of Couplets
Turning to the mechanisms behind the urban-suburban split, it is important to ascertain whether highways facilitate this split by concentrating Democrats in cities, decreasing the Democratic share of the
suburbs, or some combination of the two. If highways induce localized Republican growth in the suburbs, we would expect suburban portions of metropolitan regions with more highways to become more
Republican than comparable suburban portions of other regions. To consider these differences, I estimate the difference in the Democratic vote share between year t and 1952 separately for the urban and
suburban components of each couplet.
This analysis suggests that highways influence on the two-party vote occurs through its effect on
suburban areas. Figure A-18 displays the relevant lowess-smoothed estimates. The left panel displays
the effect within the urban portions. These estimates provide little evidence that highways have induced
polarization by increasing the urban Democratic vote share. The right panel, however, suggests that high64
ways stimulated the Republican ascent of suburbs. Over a period from 1976 on, the suburban portions
of metropolitan areas with exit density above the median became more Republican than comparable untreated areas. This effect reached a peak of more than four points in 1984, remaining relatively stable
since then, albeit with larger standard errors over the last decade.19
19
These estimates represent average effects in each group estimated separately, so addition of these
two estimates will not equal the total average effect on urban-suburban polarization.
65
●
●
●
●
● ●
●
●
● ● ●
1964
1976
1988
2000
2
0
−2
−4
−6
−8
−10
−12
−14
−16
● ●
●
●
●
● ●
1964
1976
All Suburban Counties
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
●
●
●
●
●
● ●
●
● ● ●
1964
1976
1988
2000
Non−South
2
0
−2
−4
−6
−8
−10
−12
−14
−16
●
● ● ● ● ● ● ●
2000
●
●
●
● ●
●
● ●
●
●
1964
1976
Year
1988
●
●
1964
1976
1988
2000
Year
South
●
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
●
Year
Diff. in Change in Dem. Vote Share, 1956−(Year)
Diff. in Change in Dem. Vote Share, 1956−(Year)
Year
1988
● ● ●
●
Diff. in Change in Dem. Vote Share, 1948−(Year)
●
South
●
2000
●
Diff. in Change in Dem. Vote Share, 1956−(Year)
2
0
−2
−4
−6
−8
−10
−12
−14
−16
Diff. in Change in Dem. Vote Share, 1948−(Year)
Diff. in Change in Dem. Vote Share, 1948−(Year)
All Suburban Counties
Non−South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
●
● ●
●
●
● ●
●
1964
Year
1976
●
● ●
1988
2000
Year
Figure A-12: Sensitivity of suburban-county analysis to choice of baseline year in calculating outcome
variable. Top panel: Baseline year of 1948. Bottom panel: Baseline year of 1956.
66
●
●
●
●
1964
1976
● ●
1988
●
●
2000
●
South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
●
● ●
●
●
●
●
● ●
●
1964
1976
Year
1988
Year
●
●
2000
Diff. in Change in Dem. Vote Share, 1952−(Year)
● ●
●
Diff. in Change in Dem. Vote Share, 1952−(Year)
Diff. in Change in Dem. Vote Share, 1952−(Year)
All Suburban Counties
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
Non−South
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
● ●
● ●
●
●
● ● ●
1964
1976
1988
●
●
●
2000
Year
Figure A-13: Results, with addition of variable indicating whether a desegregation order had been put in
place in a metropolitan area at least four years earlier.
67
●
●
●
HARTFORD
SEATTLE
NEW BEDFORD
−0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
●
●
●
●
●
●
●
0.2
●
●
NEW ORLEANS
MEMPHIS
●
●
●
●
●
●
●
●
●
RICHMOND
SHREVEPORT
DALLAS
NASHVILLE
ATLANTA
SAVANNAH
AUSTIN
CHARLOTTE
CHATTANOOGA
SALT LAKE CITY
CINCINNATI
HOUSTON
JACKSONVILLE
BATON ROUGE
LOUISVILLE
INDIANAPOLIS
BALTIMORE
PITTSBURGH
PORTLAND
SAN ANTONIO
KANSAS CITY
COLUMBUS
ST. LOUIS
TULSA
MILWAUKEE
OKLAHOMA CITY
EVANSVILLE
SPOKANE
ALLENTOWN
CHICAGO
NEW YORK
FORT WAYNE
YONKERS
EL PASO
NORFOLK
MINNEAPOLIS
PHILADELPHIA
CLEVELAND
DAYTON
CANTON
BRIDGEPORT
KNOXVILLE
SYRACUSE
WILMINGTON
BUFFALO
WICHITA
CORPUS CHRISTI
PEORIA
ERIE
SPRINGFIELD
TRENTON
SCRANTON
WORCESTER
TOLEDO
DETROIT
YOUNGSTOWN
ROCHESTER
BOSTON
READING
FALL RIVER
GRAND RAPIDS
PROVIDENCE
AKRON
DES MOINES
FLINT
OMAHA
ALBANY
SOUTH BEND
MOBILE
TAMPA
0.4
●
0.6
Change in Urban−Suburban Difference in Dem Pres Vote, 1952−2008
Figure A-14: Shift in the urban-suburban gap in the Democratic presidential vote in leading metropolitan
areas, 1952 to 2008. Couplets are constructed from counties inside an 80 kilometer outer radius.
68
0.30
●
0.00
0.15
0.30
0.2 0.4
1968
−0.2
0.2 0.4
−0.2
●
●●
●
●●
●
● ●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
Urban−Suburban D Vote Gap
0.15
Urban−Suburban D Vote Gap
0.2 0.4
−0.2
Urban−Suburban D Vote Gap
0.00
● ●
● ●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●
0.00
●
0.15
0.30
1984
0.15
0.30
0.00
●
0.15
0.30
●
●
● ● ●●
●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●
●
0.00
0.15
0.30
−0.2
●
●●●● ● ●
●
●● ●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
−0.2
●
−0.2
●
●
●
●
●
●
●
● ●●●
●
●
●
●●
●●●● ●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●●
●
●
●
0.2 0.4
1980
Urban−Suburban D Vote Gap
1976
0.2 0.4
1972
Urban−Suburban D Vote Gap
Exits Per Sq. Mi.
0.2 0.4
Exits Per Sq. Mi.
Urban−Suburban D Vote Gap
Exits Per Sq. Mi.
●
●
●●
●●
●
●
●
●●●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●●
●
●●
●
●
●● ●
0.00
0.15
0.30
1996
2000
0.00
0.15
0.30
●
●
0.00
0.15
0.30
Exits Per Sq. Mi.
Exits Per Sq. Mi.
2004
2008
●
●
●
●●●●● ●
●●
●
●
●
●●
●
●
●●●● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
● ●
●
●
0.00
0.15
0.30
Exits Per Sq. Mi.
●
●
● ●●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●●
●
●●
●●
●
●
●
●
●
0.00
0.15
0.30
Exits Per Sq. Mi.
●
●
● ●
● ●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●●
●
●
−0.2
●
●● ●
●
●
● ●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●● ● ●
●
●
●
●
●
●
●●● ●
●
●
●
●●●
●
●
●●
●
●
●●
●
−0.2
●
●
−0.2
●
●
●
● ●●
●
●
●
●
●● ● ●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●● ●
●●
0.2 0.4
1992
Urban−Suburban D Vote Gap
1988
0.2 0.4
Exits Per Sq. Mi.
Urban−Suburban D Vote Gap
Exits Per Sq. Mi.
0.2 0.4
Exits Per Sq. Mi.
Urban−Suburban D Vote Gap
Exits Per Sq. Mi.
0.00
0.15
0.30
Exits Per Sq. Mi.
●
0.2 0.4
0.2 0.4
−0.2
0.30
●
●
●
●● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
1964
Exits Per Sq. Mi.
0.00
0.2 0.4
0.15
1960
−0.2
−0.2
0.2 0.4
0.00
Urban−Suburban D Vote Gap
0.2 0.4
−0.2
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−0.2
Urban−Suburban D Vote Gap
Urban−Suburban D Vote Gap
Urban−Suburban D Vote Gap
Urban−Suburban D Vote Gap
1956
●
●
●
●
●
●●
●
●
●●
●●●
●● ●
●
●
● ● ●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
● ●
0.00
●
●
0.15
0.30
Exits Per Sq. Mi.
Figure A-15: Bivariate plots of the urban-suburban gap in the Democratic vote against exit density per
square mile at year t − 4. The sample correlation of exit density and urban-suburban polarization is
uniformly positive across all years in the study.
69
0.8
0.6
Standardized Imbalance
(Mean for All Covariates)
Full Sample
0.4
0.2
Matched Sample
0.0
1960
1968
1976
1984
1992
2000
Year
Figure A-16: Average of standardized imbalance across all covariates for genetic matching, using the
dichotomized exit-density treatment variable.
70
OLS Estimates of ATT
Exits Per Sq. Mi. > Median
OLS Estimates of ATT
Exits Per Sq. Mi. > Median
14
7
13
12
6
11
Polarization Effect, Points
Polarization Effect, Points
10
9
8
●
7
●
●
6
●
5
●
4
●
●
3
●
2
●
●
1
0
−1
●
5
4
3
●
2
●
●
●
●
●
●
●
●
●
●
1
●
●
●
0
●
−2
−3
−1
−4
1960
1968
1976
1984
1992
2000
2008
1960
1968
1976
1984
Year
1992
2000
2008
Year
OLS Estimates of ATT
Exits Per Sq. Mi. > Median
OLS Estimates of ATT
Exits Per Sq. Mi. > Median
10
11
9
9
8
8
7
7
Polarization Effect, Points
Polarization Effect, Points
10
6
5
●
4
●
3
●
2
●
1
0
●
●
●
●
●
−1
●
●
●
−2
●
6
5
4
●
3
●
●
2
●
1
0
●
●
●
●
●
●
●
●
●
−1
−3
−2
−4
−3
−5
1960
1968
1976
1984
1992
2000
2008
1960
Year
1968
1976
1984
1992
2000
2008
Year
Figure A-17: Sensitivity of urban-suburban polarization results to choice of outer radius for metropolitan
area, exit density analysis. The county containing the central city (or cities) is defined as the urban county
in each analysis, while the set of suburban counties used to construct each urban-suburban couplet vary
with the outer radius. Top row (left to right): 40 and 60 kilometer outer radius. Bottom row (left to
right): 80 and 100 kilometer outer radius.
71
Suburban
6
5
4
3
●
Effect on Dem Vote, Points (ATT)
Effect on Dem Vote, Points (ATT)
Urban
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
−9
−10
−11
●
●
●
●
●
●
●
●
●
●
2
●
●
1
●
0
●
−1
−2
●
●
−3
●
●
−4
●
●
●
●
●
−5
−6
−7
●
−8
●
−9
−10
−11
1960
1968
1976
1984
1992
2000
2008
1960
Year
1968
1976
1984
1992
2000
2008
Year
Figure A-18: Lowess-smoothed estimate of the first-difference in urban-suburban polarization in the
Democratic vote share, generated using coefficients from least squares regression model applied to the
matched data. Interstate highway exit density on the Democratic vote share in the urban (left) and
suburban (right) portions of the constructed urban-suburban couplets. Bootstrapped 95% confidence
intervals accompany the point estimates.
72