%20-%20https://sketch.com%20--%3e%3ctitle%3e3ECBBF50-FD08-412A-943D-2CFC5E025A16%3c/title%3e%3cdesc%3eCreated%20with%20sketchtool.%3c/desc%3e%3cg%20id='Stylesheet'%20stroke='none'%20stroke-width='1'%20fill='none'%20fill-rule='evenodd'%3e%3cg%20transform='translate(-926.000000,%20-668.000000)'%20id='BUTTONS'%3e%3cg%20transform='translate(100.000000,%20597.000000)'%3e%3cg%20id='BTN-/-RIGHTS'%20transform='translate(816.000000,%2065.000000)'%3e%3cg%20id='ICON-/-Rights'%20transform='translate(10.000000,%206.000000)'%3e%3cpath%20d='M13.5752841,0%20L1.34943182,0%20C0.59375,0%200,0.586666667%200,1.33333333%20L0,22.6666667%20C0,23.4133333%200.620738636,24%201.34943182,24%20L17.6505682,24%20C18.40625,24%2019,23.3866667%2019,22.6666667%20L19,5.36%20L13.5752841,0%20Z%20M13.5752841,1.89333333%20L17.0838068,5.36%20L13.5752841,5.36%20L13.5752841,1.89333333%20Z%20M1.34943182,22.6666667%20L1.34943182,1.33333333%20L12.1988636,1.33333333%20L12.1988636,5.36%20C12.1988636,6.10666667%2012.7926136,6.69333333%2013.5482955,6.69333333%20L17.6235795,6.69333333%20L17.6505682,22.6666667%20L1.34943182,22.6666667%20Z'%20id='Shape'%20fill='%23595959'%20fill-rule='nonzero'%3e%3c/path%3e%3ccircle%20id='Oval'%20stroke='%23595959'%20stroke-width='1.5'%20fill='%23595959'%20cx='9.5'%20cy='13.5355339'%20r='5.5'%3e%3c/circle%3e%3cpolyline%20id='Path-2'%20stroke='%23FFFFFF'%20stroke-width='1.5'%20transform='translate(9.500000,%2013.035534)%20rotate(45.000000)%20translate(-9.500000,%20-13.035534)%20'%20points='8%2016.0355339%2011%2016.0355339%2011%2010.0355339'%3e%3c/polyline%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/g%3e%3c/svg%3e)
In 2008, journalist Bill Bishop published a book with a novel and controversial hypothesis. The Big Sort: Why the Clustering of Like-Minded America is Tearing Us Apart argued that during the last few decades, American migration patterns have increasingly produced geographic clusters of like-minded individuals: liberals are increasingly more likely to be found in enclaves like San Francisco and New York City, while conservatives are clustering in places like Orange County and the suburbs of Dallas (Bishop Reference Bishop2009). According to Bishop, this geographic clustering along ideological lines imposes a cost on society: for most Americans in most American cities, the odds of encountering and interacting with someone of the opposite political persuasion have greatly diminished since the middle part of the previous century. Because exposure to contrary viewpoints has a moderating effect on our beliefs, Bishop argues, our reduced access to those viewpoints (as a result of the sorting process) has made our collective beliefs more extreme and has contributed—perhaps significantly—to polarization of the mass electorate.
In many ways, the sorting hypothesis extends one of the more pervasive political memes of the previous decade—the “blue state/red state” partitioning of the nation popularized by media coverage of the 2000 and 2004 election cycles. As such, it is not surprising that The Big Sort was a big hit: it was mentioned regularly during the 2008 presidential election cycle and received a ringing endorsement from former President Clinton (Abrams and Fiorina Reference Abrams and Fiorina2012).
Although popular with general audiences, Bishop's sorting hypothesis has received a mixed response from the academic community. The primary analysis used by Bishop to support his argument is a measured increase in the number of landslide counties over the last 30 to 40 years, where “landslide” is defined as one party winning the two-party vote in the presidential election by a margin of 20% or more. This approach has been scrutinized on a number of fronts, with the general criticism being that the data selected for analysis (county-level presidential election results over time) may not be appropriate for making valid inferences about the geographic sorting of voters (Abrams and Fiorina Reference Abrams and Fiorina2012; McGhee and Krimm Reference McGhee and Krimm2009). In particular, Abrams and Fiorina present three distinct caveats regarding Bishop's analysis: first, that for the purposes of detecting polarization in the electorate, presidential vote data are not an ideal measure because votes taken in different elections over time are inherently a function of candidate-specific attributes as well as of the construct for which measurement is desired (voter preferences); second, that Bishop's analytical strategy (totaling the number of landslide counties) is flawed primarily because results are sensitive to the arbitrary choice of what constitutes a “landslide”; and third, that data compiled at the county level are not sufficiently granular to allow for detection of a phenomenon (sorting) that purportedly occurs at the neighborhood level. With respect to this third criticism they issue an explicit call for future research: “The simple fact is it will take much more detailed research to settle questions about geographic sorting one way or the other. In particular, to examine the subject of residential polarization in a systematic manner requires data at a much lower level than the county level” (Abrams and Fiorina Reference Abrams and Fiorina2012).
Using both a new analytical strategy and a new data source, I assess the validity of Bishop's sorting hypothesis while addressing each of the criticisms raised by Abrams and Fiorina. In addition to presidential election returns, voter registration data are used as a measure of local ideological preferences. Instead of arbitrary classification of “landslide” areas, formal measures of geographic segregation are borrowed from the rich sociological literature in this area (Massey and Denton Reference Massey and Denton1988). Finally, in addition to county-level data, I examine microgeographic block group- and tract-level data for evidence of increasing clustering. As a result, this article addresses the debate about how geographic sorting along ideological lines should best be measured, while testing the hypothesis that such partisan sorting has increased during the last two decades in California.
DATA AND METHODS
This article uses voter registration and presidential vote data developed and maintained by the Statewide Database (SWDB) at the University of California (California Statewide Database 2012). The SWDB was originally created (and continues to serve) as the redistricting database for California. As part of their work to support the redistricting process, SWDB researchers examine the spatial distribution of voters and registrants in California by using statistical methods to reallocate precinct-level data to US Census geographies, including Census blocks and block groups (McCue Reference McCue2011). These data, which SWDB makes publicly available, present a unique opportunity to assess changes in the spatial distribution of political data over time and at an extremely granular level.
This broad approach computes measures of segregation for individual geographic units and for each year for which data are available and then uses statistical methods to ascertain whether those measures are increasing over time. Thus four underlying concepts in the analysis require explanation: measures of local ideology, computed measures of segregation, choice of geographic level of analysis, and statistical model specification.
Measures of Local Ideology
What is the best way to measure the ideological preferences of voters at the local level? An ideal survey dataset would be both deep, by asking a comprehensive battery of questions about voting behavior, party registration, and policy preferences from a representative sample of Americans (as does, for example, the American National Election Study (ANES) Time Series Survey), and be broad, by having robust coverage of every microgeographic unit within the study area. Unfortunately, the ANES study and others like it are representative only to large geographic areas such as individual states or the entire nation; as a result their relatively small sample size makes them inappropriate for local analysis (Johnston Reference Johnston2005).
Both party registration data (Abrams and Fiorina Reference Abrams and Fiorina2012; Jacobson Reference Jacobson2004; McGhee and Krimm Reference McGhee and Krimm2009) and presidential vote data (Bishop Reference Bishop2009; Einstein Reference Einstein2011; Kinsella Reference Kinsella2011; McDonald Reference McDonald2011) have been used as proxies for ideological preferences within local geographic areas. The validity of using presidential vote data has been challenged, however, primarily because changes in voting patterns across elections are partly a function of changing candidates rather than being wholly determined by changes in the underlying distribution of voter preferences (Abrams and Fiorina Reference Abrams and Fiorina2012; McGhee and Krimm Reference McGhee and Krimm2009); use of party registration statistics as a proxy for local preferences is also not ideal as registration rolls are often outdated (McDonald and Popkin Reference McDonald and Popkin2001; McGhee and Krimm Reference McGhee and Krimm2009).
Because of the lack of consensus in the literature on which of these methods is a better measure of local ideological preferences, the approach used here is to construct separate sets of ideological segregation measures: one set analyzes changes over time in the geographic segregation of Republican and Democratic voters in the presidential election, and the other set analyzes changes over time in the geographic segregation of registered Republicans and Democrats.
Computed Measures of Segregation
According to Erbe, two distinct elements need to be considered when measuring segregation: first, the extent to which populations within a geographic area deviate from an equal distribution, and second, the degree to which the residential contiguity of populations influences the likely rate of contact between different groups (Erbe Reference Erbe1975).
In their seminal article on racial/ethnic segregation, Douglas Massey and Nancy Denton describe five distinct dimensions of residential segregation: evenness and exposure (which correspond to the elements described by Erbe) as well as concentration, centralization, and clustering. I use two of the measures described by Massey and Denton (Reference Massey and Denton1988): the Index of Dissimilarity (D)—a measure of evenness—and Eta 2—which measures the dimension of exposure.
The Index of Dissimilarity is the most widely used measure of racial segregation (Alba, Logan, and Stults Reference Alba, Logan and Stults2000; Massey and Denton Reference Massey and Denton1988). It has also been used to model gender segregation across occupations (Jacobs Reference Jacobs1989) and to measure the ideological segregation of individual media outlet users across the liberal-conservative spectrum (e.g., Fox News versus the Huffington Post) (Gentzkow and Shapiro Reference Gentzkow and Shapiro2011). Some precedent exists for its use as a measure of segregation between Democrats and Republicans in the United States: Glaeser and Ward generated a global D statistic for the United States (computed at the county level) for presidential elections between 1840 and 2004 (Glaeser and Ward Reference Glaeser and Ward2006), whereas Einstein used it to compare precinct-level segregation across different metropolitan statistical areas (MSAs), again using presidential election data (Einstein Reference Einstein2011). It can be computed for any binary population and, in our case, is calculated as follows:
$$
D = {1 \over 2}\sum\limits_I^N {\left| {{{dem_i } \over {DEM}} - {{rep_i } \over {REP}}} \right|}
$$
where dem i and rep i are the number of Democrats and Republicans, respectively, in the ith microgeographic unit (block group, tract, precinct, etc.), and DEM and REP are the total number of Democrats and Republicans in the macrogeographic unit for which the statistic is being calculated (county, state, nation, etc.).
D is bounded by [0,1]; a value of 0 indicates perfect integration (the ratio of Democrats to Republicans in each microdistrict is identical to that of the macrodistrict), whereas a value of 1 indicates perfect segregation. D has the desirable property of being independent of the cross-group proportion within the macrogeographic unit. In other words, the fact that Orange County has a higher proportion of Republicans than Los Angeles County will not confound comparison of the D statistics calculated for each. D may be thought of as measuring “the percentage of a group's population that would have to change residence for each neighborhood to have the same percentage of that group as the metropolitan area overall” (Iceland, Weinberg, and Steinmetz Reference Iceland, Weinberg and Steinmetz2002). In this sense, D reports the evenness of the distribution of the two groups by measuring the relative similarity of the microlevel proportions to the macrolevel proportions.
Figure 1 portrays two hypothetical scenarios that would generate high and low values of D. In the high segregation scenario, each subgeographic unit is either 100% Democratic or 100% Republican, whereas the overall proportion in the macrogeographic unit is 50/50. This scenario would yield a calculated D of 1—the maximum value. In the low segregation scenario, each microgeographic unit has a Democrat/Republican ratio of either 4/5 or 5/4, very close to the overall proportion in the macro area. This scenario yields a calculated D value of 0.111, close to the index's minimum value of 0.
High- and Low-Segregation Neighborhood Configurations for the Dissimilarity Index
Note: Adapted from Iceland, Weinberg, and Steinmetz (2002). (Color online.)
By contrast, the dimension of exposure refers to the “degree of potential contact, or possibility of interaction, between minority and majority group members” (Massey and Denton Reference Massey and Denton1988). This dimension was included because it mirrors one of the concepts underpinning the hypothesis of The Big Sort—diminished levels of interaction between Democrats and Republicans contribute to an escalation of polarization along a liberal-conservative axis (Bishop Reference Bishop2009).
According to Massey and Denton, one of the primary measures of exposure is the isolation index, which estimates the degree to which group members within a macropopulation are exposed only to one another, and not to members of the other group. For our purposes, the isolation index, which was originally denoted x P x * by Lieberson (Lieberson and Carter Reference Lieberson and Carter1982), is calculated for the joint cases of Democratic and Republican isolation as follows:
$$
_{dem} P_{dem}^* = \sum\limits_i^N {\left( {{{dem_i } \over {DEM}}} \right)\left( {{{dem_i } \over {tot_i }}} \right)} {\rm\ }_{rep} P_{rep}^* = \sum\limits_i^N {\left( {{{rep_i } \over {REP}}} \right)\left( {{{rep_i } \over {tot_i }}} \right)}
$$
where dem i , rep i , and tot i are the total number of Democrats, Republicans, and individuals in the ith microgeographic unit, and DEM and REP are the total number of Democrats and Republicans in the macrogeographic unit for which the statistic is calculated.
The isolation index is also bounded by the interval [0,1], and has a similar interpretation to D in the sense that everything else being equal, higher values imply greater levels of segregation. However, the isolation index differs from D in that its calculated value is affected by the macrolevel proportions of the groups being considered: ceteris paribus, the estimated isolation of Democrats from Republicans in Orange County will differ from that in Los Angeles County simply as a result of the differing percentages of Democrats and Republicans in those places. This characteristic becomes problematic if cross-geographic or intertemporal comparisons of isolation index value are attempted: relative differences in macro proportions across time and/or space will confound attempts to make inferences about increases or decreases in the cross-group level of exposure.
The isolation index can be adjusted to account for differences in macroarea population proportions to allow for valid comparisons across time and space (Massey and Denton Reference Massey and Denton1988; Iceland, Weinberg, and Steinmetz Reference Iceland, Weinberg and Steinmetz2002). This adjusted isolation index, also known as the Correlation Ratio or Eta 2, is calculated as follows:
$$
Eta^2 = [(_x P_x^* - P)/(1 - P)]
$$
where x P x * is the isolation index described above (for X = Democrat or Republican), and P is the macrolevel proportion of group X. After this adjustment, the measure of exposure becomes symmetric, that is, Eta dem 2 = Eta rep 2. Eta 2 is similar to D in that it ranges on [0,1], and also in that higher values are associated with higher levels of segregation. It is the second measure of segregation used in this article.
Figure 2 illustrates the conceptual difference between high and low segregation as measured by Eta 2. In both scenarios, the number of Democrats (21), out of a total population of 144, is identical. In the high segregation scenario, however, the Democrats are clustered in a way that reduces the probability of cross-group exposure, whereas in the low segregation scenario, Democrats are dispersed throughout the macrogeographic area more evenly; this increases the probability that a random Republican will encounter (and interact with) a Democrat, and vice versa. The high segregation scenario yields an Eta 2 value of 0.604, whereas the low scenario yields a value of 0.046.
High- and Low-Segregation Neighborhood Configurations for Eta2
Note: Adapted from Iceland, Weinberg, and Steinmetz (2002). (Color online.)
Choice of Geographic Level of Analysis
Computing the measures described in the previous section requires selection of micro and macro units of geography. However, neither the (robust) literature on segregation by race/ethnicity nor the (relatively sparse) literature on segregation by political ideology presents consistent guidance on the optimal geographic units of analysis. The measures described previously have been used to assess segregation, variously, at the level of precincts within MSAs (Einstein Reference Einstein2011); block groups within cities (Collins and Williams Reference Collins and Williams1999); tracts within counties and tracts within states (Massey and Hajnal Reference Massey and Hajnal1995); counties within the United States (Glaeser and Ward Reference Glaeser and Ward2006); and states within the country (Massey, Rothwell, and Domina Reference Massey, Rothwell and Domina2009), to name a few of the combinations extant in the literature. Because of this diversity of approaches, I analyze segregation within California across multiple levels of geography. Specifically, the two measures described are computed and analyzed in three different scenarios that use block groups, tracts, and counties, respectively, as the microgeographic unit, with the entire state as the macrogeographic unit: for each year in the observation period, a single statistic is calculated for the entire state. However, the practice of using county-level data (that is, the use of county as the microlevel unit of analysis) to make inferences about a phenomenon—the geographic sorting of voters—that purportedly occurs at the neighborhood level has been cautioned against (Abrams and Fiorina Reference Abrams and Fiorina2012). As Kinsella argues, doing this may lead to commission of the ecological fallacy: analysis of trends within units of observation that have large areas and/or population sizes (counties) may lead to incorrect conclusions about patterns at more microgeographic levels (Kinsella Reference Kinsella2011).
Model Specification
Given two different measures of local ideology (presidential vote and voter registration data), two different computed measures of segregation (the Dissimilarity Index and Eta 2), and three different microgeographical levels, 12 sets of results are examined. My approach is to compute a segregation statistic for each combination of local ideology measure, segregation measure, and microgeography, for each year in the data set, and to examine each of those time series for evidence of a statistically significant positive trend. The baseline year is 1992, the first year for which SWDB data are available; therefore I look for evidence of increasing partisan segregation during 1992–2008 (for measures generated from presidential vote data) or 1992–2010 (for measures generated from party registration data).
Whether each of the 12 segregation indices is growing over time—and if so, whether those increases are statistically significant—is determined by this two-step process: First, I examine whether each series is growing over time by visually inspecting the calculated values of each index for each year. To ensure that any perceived increases are not artifacts of stochastic variation in the data, I also generate a bootstrapped distribution (1,000 repetitions) for each index for each year and examine the mean values from those distributions for evidence of an upward trend over time. Next, to evaluate whether any observed temporal trends are statistically significant, I examine whether the differences between the baseline year (1992) and nonbaseline year index values are statistically different from zero. This evaluation is accomplished by performing a second bootstrapping procedure (again using 1,000 repetitions) to obtain a distribution for the inter-year difference statistic. In this framework, statistical significance is assessed by letting α equal the percentage of the 1,000 replications of this difference statistic that are greater than zero; then (1 − α) is equivalent to the p-value for rejecting the null hypothesis when H a is that the difference is greater than zero. This “difference” approach is preferred because it retains the stochastic element (i.e., the error distribution) of the estimates of segregation index values for both the baseline and the current year.
Because the numbers of block groups and tracts (both within individual counties and within the state as a whole) vary across time, year-specific segregation index values were produced from generated datasets in which observations were randomly resampled to ensure comparable numbers of block groups and tracts across years.
RESULTS
This analysis generally supports the hypothesis that partisan segregation—as measured within California and across the period 1992–2010—is increasing over time. For each of the 12 measures, two conditions should jointly hold to conclude that evidence of rising partisan sorting has been found: first, somewhat obviously, the year-specific calculated segregation index values should increase over time; second, evidence that these trends are statistically significant and not just due to random variation in the data should be shown.
The first condition is assessed through visual inspection of the bootstrapped mean values of the different indices of segregation (table 1). A general upward trend is present for all of the statistics; although not dramatic, the magnitude of these growth paths also is not trivial. For example, when calculated at the level of block groups and tracts within the state, the party registration D measure rises by 9.3% and 7.4% between 1992 and 2010, respectively; for Eta 2 the figures are 10.7% and 2.1%. (The relative magnitude of these figures should be considered in light of the fact that because of limited data availability, the interval of analysis begins in 1992, roughly 16 years after Bishop suggests that sorting first began: if Bishop is correct, the growth estimates in table 1 will understate the overall increase in partisan sorting by a significant amount.)
Segregation Index Values by Ideology Measure, Segregation Measure, Geographical Combination, and Year
(1) Both D and Eta 2 are scored on the interval [0,1], with higher values representing higher levels of segregation.
(2) Bootstrapped difference from calculated value for 1992:
* significant at 0.05.
† significant at 0.01.
‡ significant at 0.001.
Although an upward trend is present for all 12 measures across the time period, the statistics generated using presidential election data demonstrate a significant drop between 2004 and 2008. No such drop is present in the segregation measures generated from party-registration data. In summary, the first condition required for support of the sorting hypothesis appears to be met—weakly for segregation measures generated from presidential election results (largely because of the decline between 2004 and 2008) and strongly for measures generated from party registration data. The second condition for concluding support for the sorting hypothesis—statistical significance of any observed growth trends in the segregation statistics—is also indicated in table 1.
Results testing the second condition are mixed but generally support the sorting hypothesis. Critically, all segregation measures generated from party registration data strongly support the hypothesis of increased sorting over time: at the block group, tract, and county microlevels, the positive difference between later-year estimates of D and of Eta 2 and the values from 1992 are highly statistically significant, with the null hypothesis of no difference rejected at p = 0.01 or less in all cases.
Interpretation of findings for statistics generated from presidential vote data is less straightforward. If the terminal year of this analysis had been 2004, both of the conditions for concluding support for the sorting hypothesis would have been universally met: first, both D and Eta 2 are increasing (almost, but not quite, monotonically) at all levels of microgeography on the interval [1992, 2004], and second, for all measures, the inter-year differences between 2004 and 1992 are highly significant. However, the terminal year of this analysis of presidential election data is not 2004 but 2008. As such—consistent with the pattern of a 2004–2008 drop presented in table 1—only four of the six estimates of differences between presidential vote-based segregation statistics calculated in 2008 and 1992 achieve significance. Indeed, in one case (Eta 2 at the tract level) the 2008 point estimate is actually lower than the 1992 value (although this negative difference does not achieve statistical significance). This pattern could be a statistical “blip”—perhaps estimates generated from the 2012 election will show a reversion to the general trend of increased segregation seen between 1992 and 2004. This pattern could also be an artifact of the previously discussed point made by Abrams and Fiorina: presidential vote data are an inherently flawed measure of local ideology, subject to disturbance by candidate-specific effects. In any event, the analysis of segregation measures generated from presidential vote data is still sufficient to conclude limited support for the hypothesis of increased ideological sorting within California.
DISCUSSION
By examining microgeographic party registration and presidential vote data from California, this article attempts to determine whether partisan geographic segregation has been rising in the United States. Reasonably strong evidence supports the claim that it has. When presidential election returns are used to calculate statewide segregation statistics, findings are mixed but generally supportive of the sorting hypothesis; when party registration data are used results provide substantial support for the sorting hypothesis, with highly statistically significant upward trends present for indices calculated at all geographic levels.
The results presented here are subject to several important limitations. As previously mentioned, SWDB data are available only as far back as 1992; as a result, trends in geographic segregation for earlier periods, including roughly half of the period (1976 to present) in which Bishop argued that sorting has occurred, cannot be assessed. Results are generated from a California-specific data set, and, as such, external validity to the rest of the country cannot be claimed. However, Bishop describes physical relocation as the primary mechanism through which ideological sorting occurs; if Americans are moving as a result of local ideological preferences (or at least allowing those preferences to influence their decisions as they move for career or other reasons), there is no a priori reason to believe that this phenomenon is occurring only in California.
In addition, the restriction to geographies within California creates a different kind of limitation: consider the case of a liberal living in a conservative rural tract of Fresno County who moves elsewhere to be among fellow liberals. Several different cases are possible: he or she might move to a more liberal part of Fresno County (for example, the urban core of the city of Fresno), or to a liberal city in a different county in California, like Berkeley or Los Angeles, or, finally, to a liberal area in a different state, like Seattle or New York City. Ceteris paribus, the first two cases would increase calculated statewide measures of partisan segregation. Because it involves an exit from the system, the third case would also increase measures of segregation (because the rural tract would become proportionally more conservative following the migrant's departure), but with a smaller marginal effect than in the first two cases. If migration is one of the driving factors behind rising partisan geographic sorting, then measures calculated for individual states (such as those in this article) will not fully account for the effect of partisan migration across state borders. Despite these limitations, this analysis provides fairly compelling evidence that between 1992 and 2010, Republicans and Democrats in California became increasing geographically isolated from one another; while by no means dispositive, these results are consistent with Bishop's original claim of a gradual geographic realignment of liberals and conservatives in the United States.
Whether or not partisan geographic segregation is broadly rising in the United States is not merely an academic question. If the spatial distribution of liberals and conservatives is moving from a relatively diffuse structure to a more clustered one, an important consideration is the impact this might have on polarization in our legislative bodies, and, by extension, on their ability to generate meaningful policy solutions. Scholars generally agree that the voting behavior of legislators is partly a function of local constituency preferences (Brewer, Mariani, and Stonecash Reference Brewer, Mariani and Stonecash2002; Gerber and Lewis Reference Gerber and Lewis2004; Jackson and Kingdon Reference Jackson and Kingdon1992). If Bishop's sorting argument is valid, then the widely noted rise in congressional polarization may be a function of the fact that legislative districts have become increasingly ideologically homogenous over the last several decades.
