The Primacy of Race in the Geography of Income-BasedVoting: New
Evidence from Public Voting RecordsEitan Hersh and Clayton
Nall12First Draft: August 24, 2012This Draft: February 12, 2013,
8:09am1Eitan Hersh is Assistant Professor, Department of Political
Science, Yale University. Address: Institute forSocial and Policy
Studies, Yale University, 77 Prospect Street, P.O. Box 208209, New
Haven, CT 06520-8209;Email: [email protected]; Phone:
203-436-9061. Clayton Nall is Assistant Professor, Department of
PoliticalScience, Stanford University. Address: 616 Serra Street Rm
100, Stanford CA 94305; Email: [email protected];Phone:
(650)725-4076.2We thank David Broockman, Anthony Fowler, Andrew
Gelman, Justin Grimmer, Lauren Davenport, JacobHacker, David
Laitin, and participants in the Stanford Political Science Methods
Workshop for comments. Thanksto Yales Institution for Social and
Policy Studies and the Center for the Study of American Politics
for nancialsupport and to Jonathan Rodden for sharing a combined
version of the Harvard Election Data Archive
precinctdata.AbstractWhy does the relationship between income and
partisanship vary across U.S. regions? Some answershave focused on
economic context (in richer environments, economics is less
salient), while others havefocused on racial context (in diverse
environments, rich voters oppose support for poor minorities).
Using73 million geocoded registration records and 185,000 geocoded
precinct returns, we examine income-based voting across a diverse
set of local areas. We show the political geography of income-based
votingis inextricably tied to racial context,not economic context.
Within homogeneously non-Black areas,an areas wealth has no bearing
on the individual income-party relationship. The correlation
betweenincome and partisanship is strong in heavily black areas of
the Old South and weaker nearly everywhereelse, including urbanized
areas of the South. Differences between racially diverse local
areas of theSouth and comparable areas elsewhere account for
state-level differences in income-based voting.1 IntroductionWhy
are people in some places more likely to vote with their alleged
income-based class interests thanpeopleinotherplaces? What isit
about certainareasoftheUnitedStatesthat maketheincome-partisanship
relationship stronger or weaker? While the political consequences
of citizens alleged falseconsciousness has been a topic of intrigue
at least since Marx, recent work in behavioral social sciencehas
sought to explain deviations from class-based voting (typically
operationalized in terms of voterincome) across U.S. states and
across countries. Within American politics, these explanations fall
pri-marily under two schools of thought. The rst, exemplied by the
research of Gelman et al. (2008), hasfocused attention on how the
wealth or modernization of a state corresponds to the
income-partisanshiprelationship. The key nding is that within
wealthier state jurisdictions, the relationship between incomeand
partisanship is weaker than it is within poorer state
jurisdictions. The second school has focusedattention on race. For
example, Alesina and Glaeser (2004) suggest that much of the
regional differencein support for social welfare policies and
income redistribution (and support for the political party
advo-cating redistribution) may be related to the level of racial
heterogeneity in the region. The key ndingis that jurisdictions
that are more racially diverse exhibit stronger partisan
differences between rich andpoor.These two schools of thought offer
useful paradigms for understanding cross-regional patterns
inincome-based voting. The choice of paradigm yields different
hypotheses about the micro-level factorsthat affect the voting
patterns of rich and poor citizens. Under the paradigmfocused on
economic context,we might look to how post-materialistic
preferences, such as religion, affect voters in places that
arepredominantly afuent versus areas that are predominantly poor.
Under the racial context paradigm, wemight look to explanations of
racial threat, sorting, and historical migration as factors that
explain whyrich and poor vote the same in some places and
differently in others.In this paper, we sort out these paradigms by
showing that racial context trumps economic contextin structuring
the geographic patterns of income and partisanship in the United
States. Whatever
micro-levelprocessesareattherootofjurisdictionaldifferencesintheincome-partisanrelationship,
those1processes are closely tied to racial context, and they are
only tied to economic context to the extent thateconomics and race
are collinear. By narrowing the set of viable explanations for
regional
variationinincome-basedvotingtothosebearingonracialcontext,
wemovetheliteratureastepforwardinunderstanding geographic variation
in the political ssures between the wealthy and the poor.Building
on research on racial heterogeneity and redistribution, our efforts
here expand on the racial-contextual paradigm, focusing on
geographic variation in racial and political orders, and
capitalizingon new data sources that permit study of voter behavior
at much lower geographic levels than couldpreviously be achieved
using either survey data or aggregate-level election returns. Just
as opposition toredistribution has been predicted by white reaction
to the local prevalence of racial minorities, we
expectthatthestrengthoftherelationshipbetweenincomeandsupportforthepartythatfavorseconomicredistribution
will vary with local racial composition. We test this expectation
using full-populationdatasets of all 73 million voters who register
with one of the two parties in party registration states, aswell as
a geographic le containing data on the 185,000 precincts from all
49 states in which electionreturns are organized by precinct and
published. Once we account for the racial composition of
smallgeographic areas within states, the relationship between the
wealth of a jurisdiction and income-basedvoting vanishes.We show
that variation in the income-partisan relationship is tied,
foremost, to the racial compositionof local areas. In the
homogeneously white areas where most Americans live, the
relationship betweenincome and partisanship does not vary with the
wealth of the community nor the wealth of the state. Onlyin local
areas with substantial minority populations do regional differences
emerge. In the Northeast andMidwest, more afuent voters living in
districts with larger proportions of African-Americans are
onlyslightly more Republican than low-income voters in these areas.
However, in rural areas with high con-centrations of minorities,
particularly in the Black Belt areas of the South and some
agricultural areasof the West such as Californias Central Valley,
the relationship between income and partisanship ismuch stronger
than would be expected otherwise. This is not only because poor
blackswho are over-whelmingly Democraticare a larger share of the
population and contribute to a stronger income-partyrelationship,
but because afuent non-blacks living in these areas tend to
register and vote for Repub-2licans at much higher rates than those
living among black voters elsewhere. We show that
alternativeexplanations for this pattern, including religiosity and
other indicators of social conservatism and liber-alism, provide
only minimal additional predictive value. Regional differences
between the preferencesof whites living in close proximity to
racial minorities largely accounts for geographic variation in
theincome-party relationship.2 Toblers Law, the Modiable Areal Unit
Problem, and the Geog-raphy of the LeftOur impetus for exploiting
variation in income-based voting at lowgeographic levels, and our
hypothesesresultant from that decision, are rooted in three
principles that inform us about the appropriate estimationof
geographic aggregate statistics, contextual effects, and their
inuence on redistribution preferences.Therstprinciple,
knownasToblersLaw, states, Everythingisrelatedtoeverythingelse,
butnearer things are more closely related than distant things
(Tobler, 1970). Toblers Law helps us tounderstanding where the
income-party relationship ought to be strongest:if racial or
economic contextcorresponds with variations in income-base voting,
we expect that variation to occur at the local level.Toblers Law is
consistent with much of the behavior contextual effects literature,
which has emphasizedthe importance of voter behavior within areas
that are typically much smaller than states (Gimpel andSchuknecht,
2004). While much of that literature relies on arbitrary
aggregate-level data (notably, metroareas and counties), it
constitutes nearly a century of scholarship that has consistently
demonstrated thatvoting behavior is geographically contingent and
local, even if it has not established that differences invoter
behavior are due to exogenous contextual effects.1As early as
Tingstens study of Swedish voting behavior (Tingsten, 1937) and
Keys scholarship oncounty-level variation in Democratic primary
voting across the South (Key, 1949), research foundedon local-level
election results and surveys has established that models that
account for neighborhood,district, or county-level social context
more accurately predict political behavior than analyses using1For
recent scholarship on the specic geographies that voters use as
their contexts, see Wong et al. (2012).3individual-level voter
predictors alone (e.g., Lazarsfeld 1948; Katz and Lazarsfeld 1955;
Huckfeldt andSprague 1995; Sampson, Raudenbush and Earls 1997;
Oliver 1999; Gay 2006; Putnam 2007). Recentresearch based on
geographically coded voter les suggests that Keys (1949) racial
threat hypothesisappears in low-level neighborhood contexts as well
(Enos, 2010). British political geographers of the1960s and 1970s
found that class-based voting varied according to the geographic
distribution of voters:working class voters living within more
working-class communities tended to exhibit stronger supportfor the
Labour Party than working class voters living elsewhere (Butler and
Stokes, 1969, 146). Researchon racial contextual effects commonly
identies effects at very local levels (Gay, 2002, 2006).
Similarly,major changes that may predict differences in income
effects, such as income segregation (Reardon andBischoff, 2011;
Saez, 2008) and partisan gaps between core and periphery, have been
observed withinmetropolitan areas, which are still substantially
smaller geographic regions (in most cases) than the stateswidely
used when relying on survey data (Chen and Rodden, 2011). All
together, this work suggests thatwhen studying the relationship
between geographic context and the income-party relationship, we
shouldinvestigate contextual correlates at levels well below state
lines.The second principle is a seemingly mundane regularity in the
study of geographic statistical method-ology known as the Modiable
Areal Unit Problem, or MAUP (Fotheringham and Wong, 1991). TheMAUP
tells us that parameters estimated on aggregate data are vulnerable
to how the data are groupedfor aggregation. Of special relevance to
our claims here, the MAUP tells us that geographic
correlationsbased on state-level data may obscure substantial
within-state variation in these correlations. The MAUPmeans that as
we dig beneath the state-level, we cannot take for granted that
prior results identied at thestate- or country-level will map on to
local areas.Models that assume the independent and identical
distribution of voter preferences within states (evenconditional on
demographics) may obscure key aspects of income-party relationship,
particularly if thisrelationship is geographically contingent (Cho
and Rudolph, 2008).2The caricature of the stereotypicalrich,
secular New York Democrat and her rich, religious Alabama
Republican counterpart may only
hold2Weusethetermincomeeffectsandincome-basedvotingasashorthandforthebivariaterelationshipbetweenpersonal
income and partisan outcomes, not to suggest that income has a
causal effect on party and vote choice, or even thatincome is the
primary consideration in voters minds.4within some subregions of
each of those states. For example, unlike the urban areas of the
South, theBlack Belta set of counties with a history of slave-based
cotton plantations and, later, sharecroppingunder Jim Crow lawshas
a high concentration of black poverty and some of the highest rates
of overallDemocratic voting. Similarly, the persistence of
Democratic voting among key white ethnic groups inthe Northeast
(Gimpel and Cho, 2004) has helped maintain the partys dominance in
those areas. Thesame can be said of areas dominated by ethnic and
religious groups in which the Republican Party holdssway, such as
the Mormon-dominated areas of Utah, parts of Idaho, and northern
Nevada (Gimpel andChinni, 2011). In short, there are persistent
political caricatures of voters in different regions of theU.S.,
but those caricatures have within-state, not just across-state,
variants.3Our interest in local context,then, stems not just from
the idea that local forces may be more potent in inuencing
political behaviorthan coarser geographic phenomena, but that
comparisons across large boundaries like states obscuresimportant
within-state differences.The rst two principles, Toblers Law and
the MAUP, justify our investigation of contextual effectsat local
levels. When we examine the income-party relationship within local
geographic areas,then,what do we expect to nd?Prior research, both
in studies of countries and of U.S. states, falls withintwo primary
schools of thought about the contextual correlates of income-based
voting. One school iscentered around the effects of economic
context. For example, Gelman et al. (2007), building on work
byHuber and Stanig (2007) and Inglehart (1990), suggest that
economic issues might well be more salientin poorer states (p. 74).
The theory offered is that in areas that are more modernized, or
that have higheraverage living standards, or that have higher
incomes, economics becomes a less salient part of a voterspolitical
calculus. Post-materialist concerns are not just a function of ones
own economic conditionbut, under this view, they arise depending on
the local economic conditions one sees on a day-to-daybasis. Beyond
this line of reasoning,there are many other theories focused on the
role of economiccontext in driving support for redistributive
policies and/or support for parties that advocate or
opposeredistribution. For example, research on income inequality
(Melzer and Richard, 1981; Romer, 1975),3Wong et al. (2012) address
the implications of choice of context for individual-level effects,
assessing how individu-als dene the geographic scope of their
neighborhoods. Their ndings are an extension of MAUP reasoning to
individualpsychology, while our results are more focused on the
basic statistical problem.5natural resource abundance (Auty and
Gelb, 2000), and class bias in the electorate (Hill and
Leighley,1992) test variation in political support by measures of
economic context. Whether economic contextis measured by GDP,
average income, inequality, or another measure, these theories
share in common afocus on the economic conditions of a jurisdiction
affecting individual-level attitudes or voting behavior.The main
alternative school of thought, especially pertinent to American
politics, is that preferencesfor and attitudes towards economic
redistribution are inuenced by response to out-group threat or
out-group dislike (e.g. Alesina and Glaeser 2004; Luttmer 2001;
Gilens 1999). Hypotheses under this schoolof thought include that
people are predisposed to feel sympathetic to those who look like
them but notto those who look different, and that politicians
capitalize on racial differences by pitting groups againstone
another for electoral gain. Whatever the micro-level foundations,
Alesina and Glaeser (2004) haveshown that racial heterogeneity in a
jurisdiction explains much more of the variation in economic
prefer-ences than do economic factors like inequality. As our third
motivating principle, we apply Alesina andGlaesers result about the
preeminent role of racial context, previously found to correlate
with redistribu-tive preferences, to geographic variation in the
income-party relationship. We argue that racial context,not
economic context, ought to explain why in some places rich and poor
vote alike and in other placesthey do not. Local areas are
typically racially homogenous. In the absence of racial diversity
in theselocal areas, we do not expect to see wide gaps between rich
and poor. If nearly all voters in a local areaare white, for
example, we expect to see small differences in average support for
the Republican Partyacross income levels.We expect that most
variation in income-based voting occurs in racially diverse
localities. In suchareas, the poor end of the income spectrum is
often overwhelmingly black, and black voters of all in-come levels
everywhere are overwhelmingly Democratic (Dawson, 1995). In diverse
local areas wherewealthier voters living near blacks (and,
increasingly, other racial minorities) are politically liberal,
therelationship between income and partisanship will be weak. But
in areas where wealthy voters livingnear blacks are politically
conservative, the relationship between income and partisanship will
be quitestrong. The former condition is likely to be especially
common urban areas with cosmopolitan voters (asGelman et al
argues), whereas the latter condition rarely exists in central
cities but may be more likely6in suburban and rural areas of the
U.S. South. To the extent that the relationship between income
andpartisanship varies across local jurisdictions in the United
States, then, we expect that variation to becontained in racially
diverse areas.Our contribution, then, is to test three empirical
hypotheses about the role of local context in thestrength of
income-based voting: First, after accounting for racial context, we
do not expect income-base voting to vary with area-level income.
Second, in homogenous white areas, we expect rich and poorvoters to
have minimal differences in their political afliations and vote
choices. Third, in heterogenousplaces, we expect to see regional
variation in the income effect. In some parts of the country,
poorminorities live in close proximity to rich, white,
conservatives. In other parts, they live near rich, white,liberals.
We expect strong income-base voting patterns only in areas that are
both racially diverse andwhere the rich people are historically
racially conservative, namely the South. Only when both
theseconditions are met (diverse districts with racially
conservative Whites) do we expect sizeable partisangaps between
rich and poor.3 Data and MethodsWe test these hypotheses using data
sources that have been unavailable to previous researchers.
Pastworkonthepoliticalgeographyofincome-basedvotinghasreliedprimarilyonsurveymeasuresofself-reported
income, party identication, and vote choice, or on geographically
coarse aggregate-levelelection data. Because nationally
representative surveys such as the American National Election
Studyand even the National Annenberg Election Study have small
sampleseach NES release is sample ofonly about 2,000 respondentsthe
studies commonly sidestep questions of sub-state geographic
variation(GelmanandLittle, 1997;Park, GelmanandBafumi,
2004;LaxandPhillips, 2009;Vigdor, 2006).In some cases, scholars
supplement these data with analysis of aggregate correlations in
counties andother similarly sized units below the state level
(Gelman et al., 2008).4This paper expands on these4Among recent
contributions to this literature, recent work has bridged across
multiple surveys Tausanovitch and Warshaw(2011) to provide similar
estimates within congressional districts. Other work has provided
estimates of the voting behaviorof racial and economic subgroups
within states (Ghitza and Gelman, 2012), but like many existing
models, these have beenlimited to estimation of statewide effects
within non-geographic subgroups.7research ndings using 73 million
party-registration records and presidential election results from
morethan 185,000 precincts. These data, which even a decade ago
were not available due to limitations intechnology and data
quality, now permit us to explore the relationship between
geographic context andincome-based voting with extremely high
resolution.The rst of our data sources is a voter le containing 73
million voter records from states that requirepartyregistration.
Weapplyacombinationofnonparametriccomparisonsandmulti-levelmodelingto
voter registration data from Catalist, a leading Democratic data
vendor that compiles, cleans, andkeeps current registration records
from every state.5Our analysis using these data is restricted to
the29 states that invite registrants to register with a party (or
as independent or unafliated voters). Thesecond data set, compiled
through the Harvard Election Data Archive is a record of the 2008
presidentialelection returns from 185,000 precincts, which allows
us to estimate the relationship between precinct-level income and
the 2008 Republican presidential vote in the 49 states that have
published and madeavailable their data (Ansolabehere and Rodden,
2011).The Catalist voter database permits us to make
individual-level inferences about partisan identica-tion by linking
party registration records to block-group level demographic
characteristics. The versionof the Catalist database made available
to academic researchers links individual registrants to the Cen-sus
block-group of their residential addresses. The median household
income of each registrants blockgroup in the 2000 Census is used as
a proxy for their income class. These data are represented in
theCatalist data as categorical variables that place each voter
into a block-group income category at $20,000intervals, ranging
from $20,000 or less to $200,000 income (the top-coded value
reported in the Censusdata). Each voter i also associated with the
racial composition of his or her block group.6To study the link
between local racial and economic context and voter partisanship,
we analyze theCatalist data at the state house district level. We
have several reasons to study these relationships withinstate house
districts. Foremost, these are the bodies that elect
representatives to the lower chamber of5These data are explained in
more detail in Ansolabehere and Hersh (2012).6Individual-level race
data is only available in some Southern states. While we would
prefer to have individual-leveldemographic information, given the
level of racial segregation in the United States, using block-group
level race data isunlikely to induce sufcient ecological bias to
reverse our ndings. We discuss the merits of using block-group
income as ameasure of economic class in the Online Appendix.8the
state legislatures, making them the lowest-level geographic areas
that elect politicians responsiblefor major policymaking. These
legislators set state welfare policy and are responsible for
state-to-localtransfers that result in both income redistribution
and geographic redistribution. So, the relationshipbetween district
income and partisanship within these districts has important
implications for each state.State house districts are also useful
to study because they are approximately comparable across statesand
there are nearly 5,000 of them across the country. Because state
house districts are numerous andtypically equal in population, they
are especially useful in the study of metropolitan areas, which
aregenerally contained within just a few populous counties but are
almost always divided into many morestate legislative districts.7To
measure district income, we take the unweighted mean of the median
household block groupincome of every registered voter in each
district. We similarly calculate the black population percentagein
the district by averaging the black population percentage in
registrants block groups in each district.8Thus, we study the
relationship between a registered voters individual-level
partisanship, the incomeof their block group,and the income and
race of their district. For registrants in either of the
majorparties, we have measures of party, income, racial and
economic context, and geography for over 73million voters. Note
that at the end of the manuscript and in a detailed online
appendix, we offer adetailed justication for the use of party
registration as a measure of partisanship, block group incomeas a
measure of income, and state house districts as a measure of local
area.Though recent scholarship has justied use of party
registration states by arguing that they are repre-sentative of the
nation at large (see Abrams and Fiorina 2012; McGhee and Krimm
2009), we supplementour analysis by using additional data from the
Harvard Election Data Archive (hereafter, HEDA) (An-solabehere and
Rodden, 2011), a GIS database containing 2008 presidential election
results from 49states.9We merge the precinct-level data with
block-group level income and race data from the 20007Unlike states
and counties, the boundaries of state house districts are drawn
strategically, subject to legal requirementsabout population size,
racial composition, contiguity, compactness, and other criteria
imposed by courts. That these districtsare drawn purposefully is
not a weakness in our research design, as many so-called
gerrymanders act to combine communi-ties of interest (Forest,
2004).8This will provide a registered-voter-weighted average of the
the local population gures, which will differ slightly fromthe
black proportion of the district population.9The omitted state,
Oregon, conducts its elections by mail and does not release
precinct-level voting returns.9Census using a spatial join in
ArcGIS.1011After these Census data were aggregated by precinct,
eachprecinct was linked the Census shapele of 2006 state house
districts and the ESRI county shapele byway of a spatial join (of
the Census, 2006). This was achieved by rst dening the precincts as
points,then assigning them to the county or district in which they
fell. The resulting precinct-level data setcontains 185,002
precincts distributed across districts and counties in all 49
states. Summary statisticsfor these data appear in Appendix Table
A-3. Table 1 summarizes the two data sources.Table 1: Summary of
Data SourcesCatalist Data HEDA DataIndividual 73,170,970 D. and R.
Registrants 185,002 precinctsUnits with block group (BG) income
measure with BG income measureAggregate State House Districts
Counties and State House DistrictsUnits with aggregated BG inc. and
race with aggregated BG inc. and raceCoverage 29 Party Registration
States 49 States with Published DataTo test the relationship
between racial context and income-based voting, we present a series
of resultsthat vary in the strength of their assumptions and
progressively bore below state boundaries. We adopttwo approaches
in our analysis of the Catalist party registration data. First, we
present nonparametricsummaries of Republican registration within
different income groups, taking advantage of the abundantCatalist
data. Then, adopting stronger modeling assumptions, we estimate a
series of hierarchical linearmodels in which both the group level
intercept and income coefcients are allowed to vary (Gelmanand
Hill, 2007). We estimate these district-specic effects of income on
Republican registration
amongtwo-partyregistrantswithinstatehousedistrictsinthe29party-registrationstates,
andforprecinctswithin districts (and in the appendix, within
counties) for the 49 precinct-data states. We present themodel
results using cartograms, maps in which the area of each geographic
unit is proportional to itspopulation.Finally, we present a set of
regressions in which district-level random-effect estimates of the
district-level deviation in the rich-poor partisan gap are dened as
the outcome variable, and various social and10Race and population
data were obtained fromthe ESRI block group layer (ESRI, 2008).
Aggregate and median householdincome for the block group were
obtained from the National Historical GIS (Fitch and Ruggles,
2003). Block groups wereconverted to points, and data from all
points within each precinct were averaged to generate the
precinct-level data.11Block groups from states that do not appear
in the HEDA data were excluded from the spatial matching
procedure.10cultural variables are used as the explanatory. The
purpose of this analysis is to test the marginal explana-tory value
of competing social explanations commonly offered for income-based
voting: religiosity, ruralculture, education, and family
structure.4 Nonparametric Tests of the Local Contextual Income
EffectsTobuildintuitionaboutthestrengthoftheincome-partyrelationshipwithindistricts,
ourempiricalinvestigation begins with scatter plots based on
aggregate-level regressions of district-level partisanshipagainst
district-level income. Using the full dataset of registered
Democrats and Republicans in 29 states,we calculate, in Figure 1,
the percent Republican of each state house district and the average
income levelof each district. We organize the party registration
states into four approximately even groups, based onthe states
income. We calculate state income by aggregating the block group
median income value foreach voter in the jurisdiction. Similar to
state income, the district income is calculated by averagingeach
registered voters median block group income within a district. We
plot the proportion of two-partyRepublican registration against the
logged mean district income, and then t a lowess curve to
eachstates set of state house districts.We display Figure 1 because
it draws immediate attention to four facets of local political
geographythat might be obscured by state-level analysis. First, the
gure indicates the large variance in district-level income. State
house districts in the U.S. differ in average income levels by as
much as $130,000.By comparison, the richest and poorest states
differ in average income by a much narrower $35,000.Second, in all
but two of the 29 states plotted (Wyoming and Oregon), on a
state-wise basis richer districtsare more Republican districts.
Unlike states, rich localities are more Republican, and this holds
in richliberal states like Connecticut as much as it holds in poor
conservative states like Louisiana. Third, noticethat while there
is state-by-state variation in the nature of the slope lines that
are plotted, that variationappears to be completely unrelated to
state-level income. Each quadrant of Figure 1 contains stateswith
very steep curves, gradual curves, and exponential-like curves.
Fourth, and perhaps most importantforourargument,
whiletheconnectionbetweenstate-levelincomeandthebivariaterelationshipofdistrict
income and district partisanship appears to be tenuous in Figure 1,
observers familiar with the11Figure 1: Regardless of State Income,
Rich Districts are Republican DistrictsKYKYKYKYKYKYKYKYKYKYKY
KYKYKYKYKYKYKYKY KYKYKYKYKYKYKYKYKYKY
KYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKY
KYKYKYKYKYKYKYKYKYKYKYKYKYKYKYKY
KYKYKYKYKYKYKYLALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALA
LALALALALALALALALALALALALALALALALALALALALALALALALALALA
LALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALALAOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOK
OKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOK
OKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOK
OKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKOKSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDSDWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWVWV
WVWVWVWVWVWVWVWVWVWVWVMEMEMEME MEMEMEMEMEMEME ME MEME
MEMEMEMEMEMEMEMEMEMEMEMEMEMEMEME MEME
MEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEME
MEMEMEMEMEMEMEMEMEMEME MEMEMEMEMEMEMEMEMEMEME
MEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEMEME
MEMEMENMNMNMNMNMNMNMNM NMNMNMNMNMNMNMNM
NMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMNMWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWYWY
WYWYWYWYWYWYWYWYWYWYWYWY0.2.4.6.81Pr. Republican of Two-Party
Reg.Ln($12) Ln($20) Ln($33) Ln($55) Ln($90) Ln($150)Log of Mean
District Income in $10KKYLAMENMOKSDWVWYPoorest
StatesIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIA
IAIAIAIAIAIAIAIAIAIAIAIAIAIAIAIA IAIAIAIAIAIA IAIAIAIAIANC
NCNCNCNCNCNC NCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNC
NCNCNCNCNCNCNCNCNCNC
NCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNC
NCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCNCFLFLFLFLFLFLFLFLFLFLFLFLFLFL
FLFL FLFLFLFLFLFLFLFLFLFLFLFLFLFLFL FLFLFLFL FLFLFL FLFLFLFLFLFLFL
FL FLFL FLFLFLFLFLFLFLFLFLFLFLFLFLFL
FLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFL FLFLFLFLFLFLFLFL
FLFLFLFLFLFLFLFLFLFL
FLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLFLNENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENENE
NENENENENENENENENENENE NENEOROROROROROROROROROROROROROR
ORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORORPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA PAPAPAPAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPA PAPAPAPAPAPAPAPAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA PAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA
PAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPAPA PAPAPAPAPAPAPAPAPAPAPA
PAPAPAPAPAPA PAPAPAPAPAPAPAPAPAPAPAPAPARIRIRI
RIRIRIRIRIRIRIRIRIRIRIRI RIRIRI RIRI RIRIRIRIRI
RIRIRIRIRIRIRIRIRIRIRIRIRIRIRI RIRIRIRIRIRI RIRIRIRIRIRIRIRI RI
RIRIRIRIRIRIRIRIRIRI RIRIRIRI RIRIRIRIRIRI0.2.4.6.81Pr. Republican
of Two-Party Reg.Ln($12) Ln($20) Ln($33) Ln($55) Ln($90)
Ln($150)Log of Mean District Income in
$10KFLIANCNEORPARIMiddle-Poor StatesAZAZAZAZAZ
AZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZAZDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDEDE
DEDE DEDEDEDEDEDEDEDEDEDEKSKSKSKSKSKSKSKSKS KSKSKSKSKSKSKS
KSKSKSKSKSKSKSKSKSKS KSKSKSKSKSKS KS KSKSKS KSKSKSKSKSKSKSKSKSKS
KSKSKSKSKSKSKSKSKSKSKSKSKS KSKS
KSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKS
KSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSKSNVNVNVNVNVNV
NVNVNVNVNVNVNVNV NVNVNVNV
NVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNVNV NVNVNVNYNYNY
NYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNY
NYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNY NYNYNYNYNYNYNYNYNYNYNYNYNY
NYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNYNY NYNYNYNYNYNYNY
NYNYNY NYNYNYNYNYNYNYNYNYNYNYNYNY NYNYNYNY
NYNYNYNYNYNYNYNYNYNYNYNYNY NY NYNYNYNYNYNYNYNYNY NYNY NYNY
NYNYNYUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUT
UTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUTUT UTUTUTUTUTUTUTUTUTUTUTUTUT
UTUTUTUTUTUTUTUT UT UTUTUTUTUTUTUTUTUTUTUTUTUT
UTUTUTNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNH
NHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNH NHNHNHNHNHNHNHNHNHNHNHNHNH
NH NHNHNH NHNHNH NHNH NHNH NHNH
NHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNHNH NH0.2.4.6.81Pr.
Republican of Two-Party Reg.Ln($12) Ln($20) Ln($33) Ln($55) Ln($90)
Ln($150)Log of Mean District Income in
$10KAZDEKSNHNVNYUTMiddle-Rich StatesAKAKAKAKAK AK
AKAKAKAKAKAKAKAKAKAKAKAKAKAKAKAKAKAK AKAKAKAKAKAKAKAKAKAKAKAKAKAKAK
AKCACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACA
CACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACACTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCT
CTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCTCT
CTCTCTCTCTCTCTCTCTCTCTCT CTCTCT CTCTCTCOCOCOCOCOCOCOCOCO
COCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCO
COCOCOCOCOCOCOCOCOCOCO COCOCOCOCO
COCOCOCOCOCOCOCOCOCOCOCOCOCOCOCOCO MAMAMAMAMAMAMAMAMAMAMAMAMAMA
MAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMA
MAMA MAMAMAMAMAMA
MAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMA
MAMAMAMAMAMAMAMAMAMA MAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMA
MAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMAMA
MAMAMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMD MD
MDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMDMD
MDMDMDMDMDMDMDMDMDMDMDMDMDNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJNJ
NJNJNJNJNJNJNJNJNJNJNJNJNJNJ0.2.4.6.81Pr. Republican of Two-Party
Reg.Ln($12) Ln($20) Ln($33) Ln($55) Ln($90) Ln($150)Log of Mean
District Income in $10KAKCACTCOMAMDNJRichest StatesNote: Figures
are derived from the full population of registered Democrats and
Republicans in party registration states. N=73,170,970. Average
median household income is the unweighted average of the median
household income of block groupsin the district; partisan
registration is obtained from individual-level voter le
records.American political landscape will notice a connection
between the geography of race and the district-level relationship.
The states that have very Republican wealthy districts and very
Democratic poordistricts are states that have pockets of
high-poverty minority areas. These include rich states like
NewJersey and Connecticut as well as poor states like Louisiana and
Oklahoma. States that have a relativelyat relationship include both
rich and poor states, including Wyoming, Kentucky, Iowa, Oregon,
UtahandNewHampshire. Allofthesestatesarelessthan2%African-American,
withtheexceptionofKentucky,
whichinthe2010CensushadthelowestAfrican-Americanpopulationshare(7.71%)ofany
state south of the Mason-Dixon Line. Other states have not a linear
relationship but a curvilinearrelationship. InSouthDakota,
Nebraska, Delaware, andCalifornia,
pocketsofpoorracial-minority12districts appear to explain the
relatively at relationship between income and party in all but the
poorestparts of the state.4.1 Local Racial Composition Matters to
the Income-Party RelationshipHaving viewed the data in the
aggregate, we now examine the rate of Republican registration
amongvoters falling within each block-group income category within
state house districts and within states. Westratify the state house
districts based on their average income levels and on their racial
compositions.Our rst empirical strategy is to exploit our massive
data sets to adopt as few modeling assumptions aspossible and
estimate the difference in partisan registration across income
groups in different geographicsettings.
Thelargenumberofobservationsinthevoterlegoesfarinoverwhelmingthecurseofdimensionality
problem by brute force, allowing us to plot conditional means,
without any models orsemiparametric smoothing techniques such as
locally weighted regression (Cleveland, 1979).Figure 2 presents a
nonparametric tests in the form of a set of graphs (lattice plots)
that demon-strate the role of racial and economic context in the
relationship between block-group median
householdincome(expressedin$20,000incomeintervalsstartingatincomelevelsof$20,000andbelowandcapped
at $100,000+) and the Republican proportion of two-party
registrants within each income cate-gory.12These proportions are
calculated and plotted for registered voters in block groups that
fall withinthree state-house-district income categories: those with
average block-group median income of less than$40,000, $40,000 to
$60,000, and more than $60,000. Each of these categories is
represented by a dif-ferent line style on the graph. Two sets of
such lines are presented: red lines represent voters in
thewealthier 14 party-registration states in our sample, while
purple lines represent voters in states in the 15poorer
party-registration states. Each of these estimates is further
subdivided into graphs based on thedistricts black population
percentage: 0 to 5 percent, 5 to 10 percent, 10 to 25 percent, and
25 percent ormore black. Block-group median household income is on
the horizontal axis. To cite one example fromthe gure, the solid
red line in the upper left plot shows the relationship between
block-group income12These bins are not selected arbitrarily; rather
they are how the data are grouped by Catalist. Because there are so
fewblock groups with median household income above $100,000, income
groups above this level were collapsed into the topcategory.Such
plots summarize bivariate relationships within researcher-dened
strata (Becker, Cleveland and Shyu, 1996).13and partisanship in
districts that are poor, less than 5% black, and located in the
rich states.13If there was a tendency of voters in poor states to
vote their class interests more than voters in richstates, we would
expect to see the purple lines (representing poor states) to have
steeper slopes than theset of red lines (representing rich states).
Likewise, if there was a tendency for voters in poor districts
tovote with their class interests, we would see different slopes in
the richest districts (shown with dottedlines) and the poorest
districts (shown with solid lines). As the graph demonstrates,
these patterns donotgenerallyhold.
ThekeyexceptionisindistrictswithlargerproportionsofAfrican-Americans.In
districts with more African Americans, we nally see a divergence
between the richest half of thestates and the poorest half. While
we are not attributing these ndings to a causal effect of
contextualracial composition, the ndings presented here show that
economic explanations are, at a minimum,descriptively explained by
behavior among rich voters who live among black voters in rich
states, versusthose who live among black voters in poor
states.State-by-state differences in the income-party relationship
appear to rest almost entirely on
differ-encesinbehaviorinareaswithsizeableblackpopulations.
OncestatehousedistrictswithsizeableAfrican American populations are
excluded from the analysis, the link between income and
partisanshipdoes not vary meaningfully with respect to either
district-level or state-level contextual income. This factis
important, because districts that are less than 5% black contain
over half the U.S. population. Theseare typically white voters who
live in rural and suburban areas and small cities and have minimal
socialexposure to African-American citizens. The substantial
overlap among block groups in both rich andpoor districts and rich
and poor states indicates that homogeneously white districts are
not the source ofmajor differences between so-called rich and poor
states. A second, related nding is that the rich-poorgap is only
meaningfully correlated with state income when we limit the
analysis to state house districtsthat have sizable black
populations. Unsurprisingly, places with a larger proportion of
black voters havepoor voters who are much more Democratic than poor
voters in other settings. This is unsurprisingbecause black voters
everywhere register with the Democratic Party at rates of about 90
percent, andpoor voters in districts with large black minorities
(or black majorities) are disproportionately black. In13In the
Online Appendix (Table A-1), we conduct a version of this analysis
using regression techniques.14Figure 2: The Income-Party
Relationship is Closely Linked to Racial
Context0.1.2.3.4.5.6Proportion Republican$0-20K Med. Household
Income $100K+State House Districts 0-5%
Black0.1.2.3.4.5.6Proportion Republican$0-20K Med. Household Income
$100K+State House Districts 5-10% Black0.1.2.3.4.5.6Proportion
Republican$0-20K Med. Household Income $100K+State House Districts
10-25% Black0.1.2.3.4.5.6Proportion Republican$0-20K Med. Household
Income $100K+State House Districts 25%+ Blackin Varying Racial and
Economic ContextsPartisanship by Block Group
Income0.1.2.3.4.5.6Proportion Republican$0-20KMed. Household
Income$100K+Dist. Income $60KRich StatesPoor StatesState House
Districts 0-5% BlackNote: Figures are derived from the full
population of registered Democrats and Republicans in party
registration states. N=73,170,970.rich states, the poorest voters
in the most black districts afliate with the Republicans at rates
below tenpercent, a number that is only slightly higher in black
areas in poor states.Here appear the sub-state origins of the
red-state, blue-state paradox (Gelman et al., 2008)
and,potentially, the local origins of opposition to redistributive
policy in poor states: a huge gap existsbetween rich voters who
live in more heterogeneous racial contexts in the poor states and
rich voters wholive in more heterogeneous contexts in richer
states. Voters with block group incomes over $100,000 inheavily
black districts in rich states afliate with the Republicans, on
average, at just over 20 percent,while voters in the same income
group in poorer states afliate with Republicans at twice that
rate.Several plausible explanations exist for these differences
other than the legacies of local race relations15and white response
to perceived racial threat, but man of these can be dismissed. One
might think
thatthesedifferencescanbeattributedtolargermiddleclassblackpopulationsinblackdistrictsinlesspoorstateswhomaybeonlyslightlymorelikelytobeRepublicanthanpoorblackvoters.
Thisisunlikely, both because the black middle and upper class are
increasingly geographically segregated
frompoorblacks(ReardonandBischoff, 2011), andtheirpartisanvoting,
whichremainsalmostentirelyDemocratic (Dawson, 1995), probably
explains very little of the partisan difference between rich
andpoor areas within such districts. Instead, it is much more
likely that differences at the upper end of theincome spectrum can
be explained by voter behavior among whites consistent with Keys
racial threathypothesis. For our purposes, it is irrelevant whether
this is a result of individual-level contextual effects,or is due
to the persistence of partisanship arising from social interactions
in these areas over multipledecades.14In the Appendix (Figures A-6
and A-7) we replicate Figure 2 but restrict the data to blockgroups
that are less than 10% black and more than 90% non-Hispanic white.
While the income-partyintercept is somewhat higher (due to the
removal of most black voters in such districts) we nd that
thedivergence in voting behavior in high-minority districts varies
even when we limit the analysis to votersin block groups that are
90-100% white, allowing us to much more safely assume that the
voters beingstudied are white voters. These results uphold the
strength of the income-party relationship.15Another plausible
reaction to the results presented in this section is that perhaps
they are attributabledifferencesinahandfulofstates.
Weaddressthisconcerninseveralways. First, intheAppendix(Figure
A-5), we present a graphical summary of the party registration gap
between rich and poor
votersindistrictsofdifferentracialcompositionsbystate.
Theresultsconrmthedivergenceinbehavioramong voters in rich and poor
states, but the differences are concentrated in districts that are
more than10% black. Second, we present the analysis in Figure 2
separately for each state, displaying estimates fornearly 2,000
conditional means. These graphs provide a transparent summary of
the partisan registration14For an excellent discussion of the false
choice of contextual versus compositional effects, see Sampson
2012, Ch. 1.15Another feature of Figure 2 establishes the primacy
of race:in poor states, no state house districts with average
block-group median incomes above $60,000 per year are more than 25%
African-American. While this could be seen as a weaknessof the
nonparametric analysis, in fact it reinforces our ndings that race
and income are inseparable in the American context,and income
effects should not be discussed without accounting for the
geographic distribution of black poverty and its role inpartisan
identication.16of voters in each race-income stratum in each group
of state house districts for all party-registrationstates. Third,
in the Appendix (Figure A-8), we replicate Figure 2 using precinct
data from 49 states, andnd the same pattern. Finally, in the next
section we adopt slightly stronger modeling assumptions byusing the
Catalist data and presidential election results data in multi-level
models.5 Testing Sub-State Variation Using Linear Mixed-Effect
ModelsWe now continue to concentrate on local areas to demonstrate
geographic clustering in the income-partyrelationship. For each
state house district, the Catalist data provides counts of voters
in each incomecategory, as well as state house district-level
racial and income variables. We use these frequencies ina
hierarchical linear model in which the Republican registration
proportions in each income bin in eachdistrict are weighted
according to the total number of two-party registrants in each
district income bin.We then estimate a model in which mean
Republican registration is the dependent variable and interceptand
the effect of the categorical income variable are allowed to vary
within district. The six incomecategories are included as a set of
dummy variables in the six income categories, using as the
omittedbase category block-group income of $0 to $20,000. Both the
district level intercept, and district-levelcoefcients on the
categorical variables are allowed to vary. This model can be
written as follows:yij= + 2x2i + . . . + 6x6i + (2jx2ij + . . . +
6x6ij) + (2jx2i[j] + . . . + 6jx6i[j]) + ijwhere yij represents the
Republican share of the two-party vote in for unit i situated in
district j, repre-sents an overall intercept term, and 2 . . . 6
represent xed effects for each income category
controllingfortheestimateddistrict-levelrandomeffects.
Therandomintercept, j, istheinterceptcoefcientestimated by
partially pooling the mean of observations i located in district j
to the global mean inter-cept. Our primary substantive interest,
though, is in the coefcients on the categorical income
variables.These similarly vary by district and are represented by
the term kj for each income category k = 2 . . . 6,where the base
category is the bottom income category. These random effect
estimates in each commu-nity lend themselves to easy
interpretation. The coefcient on the top income category (for
voters in17block groups with income of $100,000 or above) can be
interpreted as the district-specic differencein the gap between the
richest and poorest voters. Higher positive (negative) values
indicate a higher(lower) income-party gradient between the poorest
and richest voters than would be expected otherwise.Finally, the
error term, ijis independently and identically distributed in each
districtj. The randomeffects and errors are each assumed to be
normally distributed with constant mean and variance in eachincome
category k (Gelman and Hill, 2007, 258):kj N(k, 2k)kj N(k, 2k)
ij N(j, 2j)This model is represented using the R function lmer
applied to panel data containing two-party regis-tration data and
voter numbers in each district-income bin. (For additional details
on the implementationof these models in the lme4 package (Bates,
Maechler and Bolker, 2011), see the Online Appendix.)We apply a
similar varying-intercept, varying-slope model to estimate the
income-party relationshipusing HEDA precinct data clustered within
state house districts, dening the McCain share of the two-party
vote as the outcome variable and dening the income variable as the
average of the block-grouplevel median household income values in
each precinct. This model can then be estimated using thesame
methods, allowing the relationship between precinct-level income
and the two-party vote to varywithin state house districts. The
hierarchical model for the income-party relationship for precinct i
withindistrict j is thenyij= + xi + jxi[j] + jxi[j] + ijwhere is a
general intercept term, is a general coefcient on xi, the
precinct-level median householdincomevariableexpressedintensof
thousandsofdollars(afteraccountingfordistrict-levelrandomeffects),
andjandjare, respectively, the random intercept and income
coefcients for precincts ineach district j. This model is estimated
using similar assumptions using lmer (see Appendix).18We are most
interested in the random effect for the income slope in each state
house district, whichis analogous to an interaction term between
income variable and the district or county dummy variable.Rather
than presenting such estimates using the usual dot plots or
condence interval plots, we integratethis with geographic knowledge
by presenting our results in maps, but with one more added
changetotraditionalapproaches.
Whilebasicchoroplethmapshavebeenusedtopresentmodelestimates,the
weakness of maps that preserve areas or distance is that the area
of the units presented is rarelyproportional in area to the amount
of data used. As a result, most maps of political quantities in
theUnited States devote a majority of their area to unpopulated
areas such as wheat elds of Kansas andthe barren Nevada and Utah
high desert, while high-density population centers such as New York
Cityand Long Island are nearly imperceptible to the naked eye. We
avoid this problem by presenting ourresults using cartograms, maps
in which geographic units have been distorted so that their areas
areapproximately proportional to their population, using Gastner
and Newmans diffusion-based cartogrammethod (Gastner and Newman,
2004) as implemented in ArcGIS (Gross, 2009).16The size of the
districtpresented on each of these maps will be approximately
proportional to the 2007 population of the countyor counties that
it overlaps. For each model, the j random income effects estimated
for each groupinggeography j are presented as a heat map
constituting twelve equal intervals. Areas in which the
marginaldistrict-level difference is below expectations appear at
the green end of the scale, while areas with ahigher than expected
income-party relationship appear at the redder end of the scale.5.1
Income-Based Partisanship and Voting Within State House DistrictsA
cartogram of state-house-district level random income effects in
Figure 3 visualizes interregional dif-ferences in partisan voting
and highlights the importance of sub-state context. Here, we
present 6,the random effect coefcient on the indicator for
registered voters in block groups with incomes above$100,000 per
year, relative to a base category of households earning $20,000 per
year or less.
Thus,thismapcapturesthemagnitudeoftherich-poorgaprelativetoanationalbaseline.
Theserandomeffects on the Republican proportion of two-party
registration between voters in the richest versus
poor-16County-level population data was used to dene the cartogram
transformation.19est block groups range from -0.34 to 0.45. Large
swaths of the three party-registration states of theold
ConfederacyNorth Carolina, Florida, and Louisianahave a larger
income effect than would other-wise be expected, most notably in
the rural Black Belt areas of those states. Similarly, the
ethnicallydiverse and economically polarized areas of Californias
Central Valley region, an area dominated bypoor Latino agricultural
workers, have a strong relationship between income and party. The
relation-ship between income and Republican partisanship is lower
than would otherwise be predicted in mostmetropolitan areas,
especially in districts of the Northeast megalopolis running from
DC to Boston thathave been the subject of blue-state voter
stereotyping. However, overall there is little evidence thatmajor
differences in voting behavior are explained well by state-level
contexts.What are the sources of these major differences? Based on
the nonparametric analyses, we againexamine the explanatory value
of the racial composition of places. We replicate the racial
straticationin the earlier analysis by subsetting the random
effects estimates to display only districts with the highestand
lowest African-American populations. In Figure 4 we show districts
that are at least 10% black andless than 5% black. This
demonstrates the link between the racial composition of places and
income-based voting. The variation in voting behavior in
homogeneously white areas varies substantially withinstates, but
does not differ dramatically across states (top panel). The
regional divide in the income-basedparty afliation shows up,
however, in the map of black-dominated state house districts
(bottom panel).Except in a few urbanized areas of the South where
urban blacks live among cosmopolitan whites, suchdistricts are in
rural Black Belt areas where partisan afliation is strongly related
to income. Note, inparticular, the large-magnitude effect in the
districts along the Mississippi River in Louisiana and innorthern
and eastern counties of North Carolina. Both of these areas have a
legacy of slavery and blacksharecropping, particularly relative to
the Appalachian areas of those states.These ndings are not merely a
consequence of an accounting identity related to the nearly
universalDemocratic afliation of black voters who fall near the
bottom end of the economic spectrum. If
weisolatetheresultstoincludedatafromblockgroupsthatarenotmorethan10%African-American(emulating
the results that appear in Figure A-7), areas identied as having
the strongest income-partydivide continue to stand out (see Figure
A-9 in the appendix). Voters fromblock groups with
fewAfrican-20Figure 3: Income Effects are Stronger than Expected in
Districts with Rural Minority PovertyWithin-District Random Effect
of Income on Republican RegistrationShift from 10%
BlackWithin-District Random Effect of Income on Republican
RegistrationShift from 5% BlackWithin-District Random Effect of
Income on Republican RegistrationShift from