Jump to ContentJump to Main Navigation
Altered StatesChanging Populations, Changing Parties, and the Transformation of the American Political Landscape$

Thomas M. Holbrook

Print publication date: 2016

Print ISBN-13: 9780190269128

Published to Oxford Scholarship Online: August 2016

DOI: 10.1093/acprof:oso/9780190269128.001.0001

Show Summary Details
Page of

PRINTED FROM OXFORD SCHOLARSHIP ONLINE (www.oxfordscholarship.com). (c) Copyright Oxford University Press, 2019. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a monograph in OSO for personal use (for details see www.oxfordscholarship.com/page/privacy-policy). Subscriber: null; date: 25 June 2019

Compositional Change and Political Change

Compositional Change and Political Change

Chapter:
(p.79) 4 Compositional Change and Political Change
Source:
Altered States
Author(s):

Thomas M. Holbrook

Publisher:
Oxford University Press
DOI:10.1093/acprof:oso/9780190269128.003.0004

Abstract and Keywords

This chapter wades through the many conceptualizations of how group membership is likely to affect political support, offers some individual-level evidence from the 2012 election, and puts together a parsimonious model in which changes in party support at the state level are explained by changes in key state demographic characteristics. It considers some of the commonly cited sources of political change, such as changes in the racial and ethnic composition of the state electorates, as well as changes in socioeconomic and occupational status, indicators of cultural values, and state party identification and ideology. It identifies three important sources of political change in states: changes in the percentage of the state citizen voting-age population (CVAP) who are foreign born, changes in the pattern of internal migration among the CVAP, and changes in the percentage of state CVAP who are educated beyond a four-year college degree.

Keywords:   political change, political party, party support, group membership, voting pattern, state population, voting age population

Most accounts of changing geographic patterns of political support include references to the changing demography of state electorates. Although many Republicans focus explicitly on changes in the racial and ethnic makeup of key states (Gluek 2014; Ladd 2014; Murphy and Wisecup 2013), others have focused on other potential sources of change, including changes in the size of the well-educated professional class (Hood and McKee 2010; Judis and Teixeira 2004; Manza and Brooks 1999) and migration patterns (Bishop 2009; Gimpel and Schuknecht 2001; Jurjevich and Plane 2012; Robinson and Noriega 2010). There are many ways to slice the demographic pie—many more than just race, ethnicity, and place of origin—and many of the pieces overlap significantly with each other. The goal of this chapter is to wade through the many conceptualizations of how group membership is likely to affect political support, offer some individual-level evidence from the 2012 election, and put together a parsimonious model in which changes in party support at the state level are explained by changes in key state demographic characteristics.

(p.80) Groups and Party Support

The idea of a group basis for voting behavior has always played an important role in studies of U.S. elections. Two foundational studies of modern electoral studies were among the first to identify and quantify the group basis of voting behavior and public opinion. These studies—Lazarsfeld et al.’s (1944) The People’s Choice and Berelson et al.’s (1954) Voting—are often referred to as representing the Columbia school of electoral studies, based on the authors’ affiliations with Columbia University. Both studies used relatively elaborate panel survey designs to analyze voter decision making during campaigns, Lazarsfeld et al. in Erie County, Ohio, during the 1940 presidential election, and Berelson et al. in Elmira, New York, during the 1948 election. In the end, however, it turned out that exposure to campaign events and discussions had little impact on voters. Instead Lazarsfeld et al. uncovered a social group dimension to the vote—religious, class, and geographic identities lined up squarely behind the parties—that was so strong that whatever individual-level movement in vote intention there was during the campaign tended mostly to be from misplaced voters returning to vote with their dominant social group. This pattern of group-based voting led to the creation of the Index of Political Predisposition (IPP), on which voters were scaled according to the Republican or Democratic predispositions of their social groups. The strength of the empirical connection between the IPP and votes was so strong that Lazarsfeld et al. (1944, 27) concluded, “A person thinks politically as he is socially. Social characteristics determine political preferences.” Berelson et al. (1954) found much the same process at work in the 1948 election: very few people changed their vote intention during the campaign, and those who did tended to fall in line with the voting patterns of their social group.

The group basis for vote choice continued to play an important part in the development of voting studies, assuming the role of antecedent factor in the development of the concept of party identification in The American Voter (Campbell et al. 1960; Miller and Shanks 1996) and also assuming (p.81) a prominent role in many conceptions of political realignment (Key 1959; Mayhew 2002; Petrocik 1981; Pomper 1967).1

Before examining some of the important group-based voting cleavages that have generated attention in American politics, it is useful to reiterate the theoretical connection among groups, parties, and candidates. As described in chapter 2, some group connections make a great deal of sense based on a material or “benefits” perspective: groups align with parties that pursue policies that have some direct, tangible benefit to the group. Racial, occupational, and class cleavages can be understood in this context, to some extent. Racial minorities, union members, and low-income voters could reasonably see a more direct connection between their interests and the success of the Democratic Party. Higher income voters and business people should have an easy time connecting their interests to the Republican Party. But interests are only part of the story. Some groups are not tied to parties so much by shared, tangible interests as they are by shared preferences; for these groups the connection to political parties may be based as much or more on preference as on interests (Petrocik 1981). Consider, for example, the connection between religiosity and party support, whereby support for the Republican Party is strongly and positively related to indicators of religious commitment (Olson and Green 2006). It is not so much the case that the deeply religious derive clear, tangible benefits from Republican policies; rather the Republican Party, more than the Democratic Party, supports a traditional and conservative social agenda that dovetails nicely with the preferences of deeply religious voters (Bishop 2009; Olson and Green 2006). Similarly, as pointed out in chapter 2, it is difficult to imagine that the Democratic Party provides direct tangible benefits to voters with advanced degrees who, after all, tend to be white, are unlikely to belong to unions, and have relatively high levels of income. Instead the appeal of the Democratic Party for the highly educated and professional class is more likely to stem from shared ideological perspectives, especially on social and postmaterial issues, such as gay rights, reproductive issues, and environmentalism (Feldman and Johnston 2014; Houtman 2001; Kahn 2002).

(p.82) Most of my analysis of group-based change in the states comes from the study of aggregate data on group size in the states. However, the vast majority of group-based analysis of voting behavior comes from individual-level studies in which the demographic characteristics of survey respondents are correlated with their vote choice, producing evidence of gaps in group preferences (Green and Olson 2009). Before moving on to the core of the aggregate analysis, I review some of the existing evidence of group-based voting, using survey data from the 2012 election to illustrate the utility of those gaps as potential explanations of contemporary voting patterns.

Race and Ethnicity

Historically one of the largest gaps in voting behavior occurs along racial and ethnic lines. Although the contemporary focus is on differences in voting behavior between whites and nonwhites at the presidential level, the idea of racial or ethnic political solidarity is not new to American politics and historically has even encompassed ethnic divides within the white community (Lorinskas et al. 1969; Pomper 1966). The most notable change in the dynamics of racial alignment is reflected in the gulf between white and black voters, which grew in magnitude beginning in the mid-1960s in the wake of Democratic support for civil rights, the Goldwater candidacy, and the Republican embrace of the “southern strategy” (Black and Black 2009; Carmines and Stimson 1989; McClerking 2009). However, black voters were not alone in moving toward the Democratic Party. The Latino vote has also moved somewhat steadily to the Democratic column, and the increased size and cohesion of this group is of significant enough magnitude that it has become pivotal in several states, perhaps providing the margin necessary for Democrats to win at the national level (Barreto et al. 2010). The source of Latino support for the Democratic Party likely stems from the overall level of liberal attitudes of Latinos (Segura 2012) and from Democratic appeals on issues related to immigration policy (Collingwood et al. 2014). In addition, though less work has been done on Asian American political attitudes, recent elections have (p.83) seen steep increases in support for Democratic candidates among Asian Americans and Pacific Islanders (Ramakrishnan 2014, 2015).

Socioeconomic Status

One of the earliest identified voting cleavages was social class, with blue-collar, unionized, and low-income voters throwing their support behind Democratic candidates, and white-collar and high-income voters supporting Republican candidates (Berelson et al. 1954; Campbell et al. 1960; Lazarsfeld et al. 1944). The relationship between socioeconomic status and party support has changed substantially and is not nearly as strong or linear as it once was (Brewer and Stonecash 2006; Manza and Brooks 1999; Ortiz and Stonecash 2009). Among other factors that have altered this relationship is the fact that many socioeconomic variables are also related to overall levels of cultural conservatism, with low-income and low-education voters tending to be culturally conservative, and high-income voters holding moderate to liberal social views. This tends to cross-pressure socially conservative low-income voters, whose loyalty to the Democratic Party can erode as a result, as well as high-income voters, who hold more liberal views on social issues and whose support for the Republican Party can erode as well (Ortiz and Stonecash 2009). In fact Ortiz and Stonecash maintain that the increasingly constrained relationship between income and vote choice may end up as a net loss to the Republican Party; they argue that the changing relationship “has not occurred because Republicans have been able to attract the less affluent, but because they are alienating the more affluent.”

This finding ties in nicely with research on occupational status and vote choice. As discussed earlier, one of the key findings in this area is that the once very Republican professional class has moved significantly in the direction of the Democratic Party over time (Brooks and Manza 1997a, b; Hout et al. 1999; Judis and Teixeira 2004; Manza and Brooks 1999). Manza and Brooks offer two different explanations of this change: that the professional class is increasingly made up of public employees and people who work in the nonprofit sector and therefore have a vested interest in (p.84) Democratic success, or the professional class has become increasingly liberal on social issues. Manza and Brooks find that increased social liberalism among professionals provides the best empirical explanation. This fits nicely with the suggestion that increased attention to cultural issues on the part of political elites has helped activate ideology and moved professionals toward the Democratic Party (Judis and Teixeira 2004). One aspect of professionals that helps explain this pattern is that they have a high level of education compared to other occupational groups. While high level of education tends to be connected to greater economic conservatism, it is also true that the highly educated tend to be relatively liberal on social issues (Feldman and Johnston 2014; Houtman 2001) and on issues such as environmental regulation (Kahn 2002).

The Gender Gap

If the racial gap stands out as one of the most profound gaps in terms of magnitude, the gender gap surely stands out in terms of the attention it gets. Often hailed as decisive by political commentators (Kornacki 2012), the gender gap in voting has ranged in magnitude (female percentage voting Democratic minus male percentage voting Democratic) from 4 to 11 points in presidential elections from 1980 to 2012.2 Like the relationship between race and vote choice, the nature of the gender gap has evolved over time. Kaufmann and Petrocik (1999) were among the first to track the dynamics of the gender gap, finding that women were actually slightly more Republican than men (in both voting patterns and party identification) in the 1950s and early 1960s, but that the gap was reversed and grew throughout the subsequent decades. Notably the gap is not a result of women flocking to the Democratic Party but of men increasingly embracing the Republican Party over time (Box-Steffensmeier et al. 2004; Kaufmann and Petrocik 1999). Despite conventional wisdom that the gender gap is the result of differences of opinion on “women’s” issues, such as abortion, there have not been substantial changes in the opinion gaps between men and women (p.85) across these and other issues (Kaufmann 2009). Instead the primary explanation for the emergence and continuation of the gender gap is changes in the emphasis men and women attach to groups of issues; with increasing numbers of single women in the electorate, as well as changes in the traditional family structure (Box-Steffensmeier et al. 2004), social welfare and cultural issues (reproductive rights, gay rights, women’s equality) have emerged as prominent sources of the gender gap (Kaufmann 2002, 2009).3

Religion

Religion also plays an important role in shaping party identification and voting behavior, though our understanding of this relationship has changed over time. Religious denomination used to help define the social basis for party support; Catholics traditionally were strong supporters of the Democratic Party and Protestants were strong supporters of the Republican Party (Berelson et al. 1954; Campbell et al. 1960; Lazarsfeld et al. 1944). Today, however, religious denomination is not viewed as nearly as important as the strength of religious commitment (Leege and Kellstedt 1993; Olson and Green 2006). Although measures of religious commitment vary across studies, one measure that is commonly used is frequency of attendance at religious services, and those who attend frequently are much more likely to vote Republican than those who attend less frequently or not at all. Part of the explanation for this relationship is that regular attenders are generally more conservative, more traditional, and place more emphasis on “moral” issues than do infrequent attenders (Olson and Green 2006). It is generally assumed that this relationship grew stronger in the 1980s and 1990s due to the Republican Party’s courtship of political elites in the Christian conservative movement (Bishop 2009; Judis and Teixeira 2004; Milkis et al. 2013). A rival hypothesis is that religious conservatives were pushed to the Republican Party by the “irreverent left,” although some evidence suggests that any hostility toward religion by the Left was a result of religious alignment rather than a cause (Pieper 2011).

(p.86) Marital Status

The potential list of demographic characteristics is almost endless, and there is no good reason to simply segment demography in any way possible looking for influences on vote choice. However, one other potentially important voting gap is the marriage gap. Evidence of a marriage gap first emerged in the 1972 election (Weisberg 1987) and has persisted over time, growing in magnitude over the past several election cycles (Gerskoff 2009). Findings in this area of research show that married people have a substantially higher probability of voting for the Republican candidate than do unmarried people (Gerskoff 2009; Plutzer and McBurnett 1991; Weisberg 1987). In some analyses this pattern persists even when controlling for variables that might explain the gap (Plutzer and McBurnett 1991). While the precise source of the marriage gap remains elusive, Gerskoff (2009) offers three potential explanations: financial insecurity (unmarried people are more likely to be financially insecure and hence could be attracted to the Democratic Party based on its traditional support for programs aimed at helping the poor); values (marriage is a traditional institution, a choice people make, and people who are married are more likely to be culturally conservative and find Republican cultural appeals attractive); and mobilization (Democrats are more likely to reach out to single voters, and Republicans are more likely to reach out to married voters). The values-based explanation offers interesting possibilities, especially for examining changes in the relationship between marital status and vote over time, as the parties diverged on social and cultural issues.

Demographics and the 2012 Presidential Election

Data from the 2012 presidential election provide insights into the relative importance of a set of demographic characteristics on vote choice. The web-based module of the American National Election Study is used to demonstrate the effects of many of the variables just discussed.4 These (p.87) effects are summarized in Figure 4.1. The probability of voting for the Democratic incumbent, Barack Obama, is estimated for people with discrete demographic characteristics, using logistic regression. These are simple estimates, representing the effects of each category of characteristics without controlling for overlapping group memberships. Perhaps the starkest differences in party support occur along racial and ethnic lines. Support for Obama ranges from a low of .44 among non-Latino white respondents to .56 among non-Latino “other” racial and ethnic groups (this includes Native Americans, Asian Americans, and others who are neither Latino nor black), .71 among Latino respondents, and .96 among black respondents. This pattern of differences is statistically significant (chi-square = 90.31, p = .00) and substantively important and is very much in keeping with findings from recent elections (McClerking 2009). As demonstrated in chapter 3, migrant status is also tied to race and ethnicity; the relationship between the size of the foreign-born population and changes in Democratic support shown earlier is supported in this figure, where we see that foreign-born respondents were about 15 points more likely to vote Democratic than native-born respondents (chi-square = 8.86, p = .003).

The results for indicators of socioeconomic status generally comport with expectations, with a few wrinkles. Income (chi-square = 42.36, p = .000), occupation (chi-square = 12.65, p = .006), and education (chi-square = 21.12, p=.009) are all significantly related to vote choice. President Obama enjoyed his highest vote margin among those with relatively low and high incomes, those with the lowest and highest levels of education, and those with a professional occupation; he received the least support from middle-income groups, middle-education groups, and people in management occupations.5 Interestingly there are no clear differences between union and nonunion voters in the 2012 ANES survey data (chi-square = 1.26, p = .262). This finding is significantly at odds with previous decades of ANES surveys, though exit poll surveys show some drop-off in union voting beginning in 2004.6

The gender gap also appears to have been alive and well during the 2012 election, with female voters somewhat more likely (.08) than male voters (p.88)

Compositional Change and Political Change

Figure 4.1 Demographic Characteristics and Presidential Vote Choice in 2012

Note: All entries are the estimated probability of voting for Obama in 2012 for each group characteristic, based on simple logistic regressions, with dichotomous independent variables representing individual group characteristics. Because the measure of occupational status used here was not available for the face-to-face sample, data from the web-based sample of the 2012 ANES were used to estimate the probabilities. Web-based weights were used in all analyses.

(p.89) to vote Democratic. This effect is relatively small in comparison to many others in Figure 4.1 but is statistically significant (chi-square = 10.02, p = .002) and is on par with recent gender gaps. The marital gap also shows a fairly pronounced division, with single respondents (defined as those who have never been married) about 24 points more likely to vote Democratic than those who are currently married, and about 8 points more likely than those who had been but are no longer married (chi-square = 37.58, p = .000).

Both religious affiliation and frequency of religious service attendance have important effects on vote choice. Respondents who report weekly attendance at religious services have a .34 probability of voting Democratic, while those who never attend religious services have a .63 probability. The pattern of support across attendance categories in not perfectly monotonic, but the group differences are very strong (chi-square = 87.14, p = .000). Patterns of religious affiliation fit with expectations from previous research: non-Catholic Christians are the least likely to vote Democratic, followed by Catholics, then “other” religions, and “not religious” (chi-square = 103.04, p = .000).7 The differences across affiliation groups may overstate the relationship a bit, since only 2.5% of the sample affiliated with a non-Christian (“other”) religion. Affiliation and attendance patterns overlap a bit (Cramer’s V = .35), with non-Catholic Christians most likely to attend religious services, followed by Catholics, then other religions, and “not religious.”

All in all, the patterns in Figure 4.1 reflect expectations from previous research. As noted, however, many of these variables overlap significantly with other variables. For example, race and ethnicity overlap with religious affiliation, religious attendance, and all measures of socioeconomic status; income, education, and occupation overlap with each other. This begs the question of which characteristics remain significantly related to vote choice when controlling for all other characteristics. When all variables are entered into a single model, many of the patterns presented in Figure 4.1 emerge again, though all of the relationships are weaker and a couple of them are no longer significant. The most substantial changes are among socioeconomic indicators: income differences are no longer (p.90) statistically significant as a group; education is still significant as a group, but this is driven by advanced-degree recipients, who stand out as significantly different from all groups other than those without a high school diploma; and occupational status is also significant as a group, though there is no significant difference between management and “other” occupations, while professionals stand out as giving significantly more support to Obama than either of the other two occupation categories. Interestingly nativity loses significance in the combined model, and union household becomes significant, though the probability of voting for Obama is only .06 greater for people living in union households.

The relationships presented in Figure 4.1 represent a set of demographic characteristics typically included when estimating models of vote choice. It is possible to partition demography into even finer distinctions (white males, single females, white Protestants, etc.); however, the point here is not to account for every possible permutation but to illustrate how people with different characteristics often line up behind different parties and candidates.

Moving to Aggregate Patterns

This evidence is illustrative and helps build the case for examining compositional changes in the states as a determinant of changes in patterns of presidential support at the state level. However, the data presented in Figure 4.1 are based on individuals at one point in time, and it is important to bear in mind that the process of aggregation, or confounding relationships with other variables, could lead to different patterns at the aggregate level. This is typically a greater concern when trying to make inferences about individuals based on aggregate patterns (Robinson 1950), but it is also sometimes the case that relationships found at the individual level are not manifested, or even present perverse findings, in aggregate analyses. Perhaps one of the best-known examples of this is found in the relationship between income level and voting behavior. While it is common to find that Democrats enjoy their highest level of (p.91) support among lower-income voters, the exact opposite is true at the state level: Democrats typically fare best in the relatively wealthy New England and Mid-Atlantic states and fare worst in the relatively poor states of the South and Plains. This paradox is explored in fine detail by Gelman et al. (2008), who find that the phenomenon is fairly recent, perhaps due in part to the combined effects of the emergence of cultural and social issues and the tendency for poor internal migrants, who are generally socially conservative, to move to “Red” states, and wealthy internal migrants, who are generally socially liberal, to move to “Blue” states.

My own sense is that this type of problem is more likely to affect findings in static research designs in which the focus is on levels of variables rather than changes in the values of variables. The remainder of this chapter focuses on changes over time in the levels of key demographic and political variables, aggregated at the state level. Intuitively this is different from looking at levels of the independent and dependent variables at a fixed point in time. Consider race, for example. We know from Figure 4.1 that nonwhite voters overwhelmingly support the Democratic Party. However, we also know that states in the conservative South have relatively high concentrations of nonwhite voters,8 which leads to fairly weak or, in some cases, perverse cross-sectional correlations between the percentages of nonwhite and Democratic vote share in any given year. Because this and other variables examined in this chapter are expressed in terms of increases or decreases over time, the hypothesis is not that Democrats should do better in states with a substantial nonwhite population than in other states at any given point in time but that—regardless of the historical size of the nonwhite population—Democrats should see their vote share increase more in states that have seen the greatest increases over time in their nonwhite population than in other states. This does not mean Democrats are expected to win in states with substantial increases in nonwhite populations, just that they should improve their margin more in those states than in states with smaller increases.

Note that this discussion is couched in terms of changes in relative standing over time. This is because for many of the variables considered here, there are strong secular trends over time such that all states move (p.92) in the same direction, so we may see gains (or losses) in every state. For instance, from the early 1970s to 2012 all states saw a decline in the percentage of their white population. Some states saw relatively steep declines, while others saw relatively modest declines. But we don’t expect to see Democratic gains in all states just because all states lost white population. Instead the expectation is that Democratic gains (Republican losses) should be greatest in states with steep declines in white population, and Democratic losses (Republican gains) should be the greatest in states with the slightest declines. In effect, when measuring change in population characteristics over time we are measuring changes in the relative levels over time, since all states are measured at the same points in time.9

Aggregate Patterns of Change

I contrast aggregate patterns of changes in key demographic and political characteristics with changes in state-level election outcomes to gain a simple, bivariate impression of how changes in state characteristics are related to changes in support for the parties in presidential elections. My strategy is similar to that used in my analysis of migration effects in chapter 3: change in average estimated centered Democratic vote from 1972–80 to 2004–12 is the dependent variable, and change in state characteristics during the same time periods is the independent variable. It bears repeating that the change in vote measure is based on the trend in party support, estimated in Figures 1.3 and 1.4, which provides a measure of candidate support purged of home-state effects as well as the southern region effects particular to the Carter and Clinton candidacies.

Race and Ethnicity

As noted earlier, changes in the racial and ethnic composition of the states are frequently cited as an important source of political change, with a primary focus on the increased size of the minority (nonwhite) population in key states (Gluek 2014; Judis and Teixeira 2004; Murphy and Wisecup (p.93) 2013). My approach is to examine the overall relationship between change in the nonwhite CVAP and change in party support, and then examine changes in the separate components of the nonwhite CVAP—changes in the percentage of Latino, the percentage of non-Latino black, and the percentage of non-Latino other (mostly Native American and Asian and Pacific Islander). These relationships are presented in Figure 4.2, where the joint outcome of change in population characteristic and change in votes for each state is represented by the state abbreviation, and the solid line is a bivariate regression slope that summarizes the linear trend in the data. The upper left cell shows the relationship between changes in the nonwhite population and changes in votes. As mentioned earlier, there are no states in which the nonwhite population decreased during this time period: the average increase across all states was 8 percentage points, ranging from a low of 1.6 points in West Virginia to a high of 25.1 points in California. There is a general trend in the data toward increasing Democratic vote share in response to increases in the nonwhite population (r =.43). Generally speaking, states with the greatest increases in the nonwhite population are also states with relatively large swings toward the Democratic Party. On average Democrats gained 4.49 points in states with above-average gains in their nonwhite population and lost 2.15 points in states with below-average growth in the nonwhite population. States that fit this pattern most clearly are California, Florida, Maryland, New Jersey, Nevada, and New York. A couple of states clearly don’t fit the trend: Texas and Oklahoma have relative large increases in nonwhite population but are also states in which Democrats have lost ground over the forty years studied here. Also notable here (and in all three other plots in Figure 4.2) is that there is very little differentiation in political support at relatively low levels of population change. For instance, among those states with the smallest growth in the nonwhite population (less than 2.5 percentage points) are Kentucky and West Virginia, both places with steep declines in Democratic support; South Carolina, where party support barely changes; and Maine, New Hampshire, and Vermont, all states where Democrats made impressive gains. This is similar to the patterns found for the effects of foreign-born population and, to a lesser extent, (p.94)

Compositional Change and Political Change

Figure 4.2 Changes in Racial and Ethnic Composition of States and Changes in Party Support in Presidential Elections, 1972–80 to 2004–12

Note: The dependent variable in all scatter plots is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4). The independent variables are measured as the change in their average value from 1972–80 to 2004–12.

(p.95) weighted internal migration (Figure 3.2). As in the case of foreign-born migrations, this pattern probably indicates that changes in population need to be nontrivial before they register a systematic impact on changes in political support.

Decomposing the change in nonwhite population reveals a few differences across subgroups. Looking first at changes in the size of the black population (upper right quadrant), there simply isn’t much of a connection to changes in party support. One potential explanation for this is that most states did not see substantial changes in the size of their black population. Interestingly four states actually saw slight decreases in blacks as a percentage of CVAP (California, Idaho, South Carolina, and West Virginia), and in one state (Alaska) there was no change. Overall forty states had increases in the black CVAP of 2 or fewer percentage points. Among the states with the greatest increases there is a slight tendency toward improved Democratic position, but the trend is not strong enough to constitute much of a relationship (r =.21, p =.14). The pattern for change in Latino CVAP is strikingly similar to that for change in nonwhite population overall, reflecting the fact that the greatest source of change in nonwhite population is linked to the growth in the Latino population. The range of change in Latino population is much greater than that for the black population, from less than a single percentage point for several states to just over 14 points for California. The familiar pattern of low differentiation in political change at low levels of population change, coupled with generally high levels of increased Democratic support among the states with the greatest increase in the Latino population, persists. Overall there is a moderate correlation between change in Latino population and change in Democratic votes (r = .37). The same pattern also emerges when considering changes in other nonwhite populations (lower right quadrant), once again led by California, with an increase of just over 11 points, and an overall correlation with changes in Democratic votes of .31. A couple of states stand out across these figures as being somewhat impermeable to the political effects of changes in nonwhite population: Texas, Oklahoma, and, to a lesser extent, Alaska are states that experienced relatively substantial increases in their nonwhite populations (Texas via (p.96) increased Latino population; Oklahoma and Alaska via increased “other” populations) but grew increasingly Republican over this forty-year time span. So much for any Democratic hope that Texas will turn “blue” due to increases in its Latino population.

All told, changes in racial and ethnic composition have some impact on changes in state support for presidential candidates, but this impact is somewhat limited, at least in this simple bivariate analysis. In particular the effects of changing composition seem to be felt primarily in those states where truly substantial changes have taken place. With the exception of a couple of already noted cases, states with above-average increases in the nonwhite population also had increases in Democratic vote share from the early 1970s to the 2010s. Political change was much more heterogeneous among states with below-average increases in the nonwhite population, indicating that political change in some of those states must be explained by some other process.

Socioeconomic and Occupational Status

Other demographic characteristics with important links to voting behavior are socioeconomic and occupational status. Four different indicators are used to measure these concepts at the state level: change in the percentage of CVAP whose income falls below the poverty rate, change in the percentage of the workforce who are unionized,10 change in the percentage of the CVAP workforce who are employed as professionals, and percentage of the CVAP with college education beyond a standard four-year degree. Figure 4.3 presents the bivariate relationships between each indicator of socioeconomic and occupational status and changes in party support for presidential candidates. Here we see a lot of variation in the strength of relationships. For both changes in poverty rate and changes in unionization rates (upper left and upper right panels) there are slightly positive relationships that are barely significant using a one-tailed test (p =.04 and p =.09, respectively).11 States with below-average changes in poverty moved very slightly Republican (–.87 change in centered Democratic (p.97) vote), while states with above-average changes in poverty moved 1.7 points in the Democratic direction, on average. The relationship for change in union coverage is of roughly the same magnitude: Democrats lost 1.1 points in centered Democratic vote among states with greater than average declines in union coverage and gained 1.9 points among states with smaller than average union losses. Note, though, that these average gains and losses reflect central tendencies, and there is a lot of variation in outcomes.

The relationships for both occupational status and educational attainment are appreciably stronger. There is a moderately strong positive relationship (r =.46) between change in the percentage of the CVAP workforce with professional occupations and change in centered Democratic vote: Democrats gained 3.7 points in states with above-average increases in professionals, and on average lost 2.01 points in states with below-average growth in professionals. This difference in average outcome (5.7 points) is substantial, especially compared to the differences found for poverty and unionization, and is also larger than that found for change in nonwhite population in Figure 4.2. The pattern is even stronger (r =.65) for the relationship between change in percentage of CVAP with education beyond a four-year college degree and change in centered Democratic vote (lower right quadrant): Democrats gained an average of 5.8 points in states with above-average growth in advanced degrees and lost an average of 2.4 points in states with below-average growth in advanced degrees, for an overall difference between the two groups of 8.2 points. The relationship between education and political outcomes is stronger than that for professional occupations, but the two variables overlap appreciably. This overlap is revealed somewhat by a close inspection of the scatter plots in Figure 4.3, where states with the greatest growth in both professionals and higher education tend to be in the Northeast and Mid-Atlantic region, and states with low levels of both tend to be in the Plains and the South. The connection is further confirmed by a strong positive correlation between the two variables (r =.72).

The results in Figure 4.3 are a bit of a mixed bag. On one hand, changes in poverty and unionization are just barely related to changes in party (p.98)

Compositional Change and Political Change

Figure 4.3 Changes in Socioeconomic and Occupational Status in the States and Changes in Party Support in Presidential Elections, 1972–80 to 2004–12

Note: The dependent variable in all scatter plots is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4). The independent variables are measured as the change in their average value from 1972–80 to 2004–12.

(p.99) performance. Although this is a bit at odds with expectations, it bears remembering that the individual-level evidence presented in Figure 4.1 shows that related measures (family income and union household) were not strongly related to vote choice in 2012, and there is increasing evidence of income-related measures taking on less meaning for vote choice as both low- and high-income voters feel cross-pressured by party positions on cultural issues (Ortiz and Stonecash 2009). On the other hand, there are clear patterns in the data supporting the connection between professional occupational status, educational attainment, and presidential votes. As expected, Democrats gained the most ground in states with the greatest increases in professionals and the highly educated and lost the most ground in states with the lowest growth in these groups. This finding is perfectly in keeping with the idea that professionals and the highly educated are increasingly driven to the Democratic Party because of their generally liberal views on social and postmaterial issues (Coan and Holman 2008; Feldman and Johnston 2014; Judis and Teixeira 2004; Ortiz and Stonecash 2009).

Cultural Indicators

Cultural variables increasingly matter to voting behavior and election outcomes (Abramowitz and Saunders 2008; Judis and Teixeira 2004; Leege et al. 2002; Pieper 2011). There is no single state-level indicator that measures cultural liberalism or conservatism of the states, but it is possible to observe changes in demographic characteristics that are typically connected to different sides of the cultural divide. For instance, part of the explanation given for the gender gap and the marriage gap is that men and women differ in important ways on cultural issues (Kaufmann 2002, 2009), as do married and unmarried people (Gerskoff 2009), the expectation being that women and single people are more culturally liberal than men and married people. Perhaps the most obvious cultural fault lines are found in religious differences, primarily differences in religiosity (Olson and Green 2006). The data in Figure 4.4 provide a simple, bivariate look at how changes in variables connected to cultural issues (p.100) are related to changes in party support in presidential elections. The top two scatter plots focus on changes in the percentage of CVAP who are female (top left) and who are single and never married (top right). The two scatter plots on the bottom use estimates of the percentage of the state population who attend weekly religious services and the number of religious congregations per 10,000 state residents.12 Despite relatively strong individual-level evidence of substantively important relationships for these variables, there is very little suggestion of significant relationships in Figure 4.4. The slope for change in female percentage of CVAP is nearly flat and the correlation is barely greater than zero. In this particular case part of the explanation for the weak relationship could be that there is not much real change in percentage of female CVAP among the states. With the exception of Alaska and Hawaii, almost all states changed by less than 2 percentage points. Recall from earlier scatter plots that there tends not to be much differentiation on the dependent variable when substantial numbers of states are clustered as low levels of change. This is clearly the case with sex breakdowns but probably does not explain the lack of relationship between change in percentage of single CVAP and change in election outcomes (upper right quadrant), since there is much greater variation in the independent variable. Although there is a slight positive trend to the data on marital status, the correlation is quite small and not statistically significant (r =.16, p =.14). Interestingly there also is no real evidence of important effects from the measures of religious saturation. In the plot for change in congregations (lower left quadrant) there is a negative trend to the data (seemingly driven by Utah’s movement), but again the correlation is small and not significant (r = –.20, p =.14). The pattern is even weaker for the relationship between change in weekly attenders and change in election outcomes (r = –.06, p =.33). Simply put, there is no evidence here that changes in any of these variables are connected to changes in party support. Note, though, that this is different from saying that states with relatively high levels of religious commitment are not different from states with low levels of commitment, or that states with a high percentage of single people are the same politically as states with a low percentage of single people. It is just saying (p.101) (p.102) that changes

Compositional Change and Political Change

Figure 4.4 Changes in Cultural Indicators in the States and Changes in Party Support in Presidential Elections, 1972–80 to 2004–12

Note: The dependent variable in all scatter plots is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4). The independent variables are measured as the change in their average value from 1972–80 to 2004–12.

in these indicators over this period of time are not related to changes in party support. As will become clear in the next chapter, differences in the levels of some of these indicators at a given point in time are related to state differences in presidential election outcomes.13

Combined Effects of Demographic Characteristics

As interesting as these snapshots of the bivariate relationships are, they don’t put us in a position to be as definitive as we could be about the unique effects of each variable. To do this we need to evaluate these variables—or at least a subset of them—in the context of a multivariate model that takes into account the extent to which the characteristics have overlapping relationships to the dependent variable. The goal of this model is to develop a parsimonious set of variables that represent each of the three sets of influences but that also avoids excessive interitem correlation that could lead to high levels of collinearity.14 For instance, change in the percentage of nonwhite CVAP is selected from the set of race and ethnicity variables rather than specifying change in each of the component groups. In the socioeconomic and occupational status variable, change in poverty, change in unions, and change in percentage with advanced degree are utilized. Change in professionals is not used because it is strongly related to change in advanced degrees (r =.72), is never significant when entered in the same model as advanced degrees (whereas advanced degrees usually is significant), and in some models the correlation with advanced degrees leads to the conclusion that neither change in professional nor change in advanced degrees matters. From the cultural variables, change in congregations per 10,000 population was selected for the model, primarily because it is the most direct measure of one of the most important influences on cultural politics (religious influence),15 and it is less collinear with noncultural variables than change in percent female, change in percent single, or change in regular attenders.

The first step in the multivariate analysis is to look at the pared down set of five demographic characteristics together and then add dynamic (p.103)

Table 4.1 Compositional Change and Changes in Centered Democratic Vote in Presidential Elections, 1972–2012

Change in Centered Democratic Vote

Model 1

Model 2

Model 3

b/s.e.

Δ‎Y,Sx

b/s.e.

Δ‎Y,Sx

b/s.e.

Δ‎Y,Sx

Δ‎ % Foreign-Born

.853

2.33

.718

1.97

.633

.258

Δ‎ Internal Migration Index

4.809

1.39

4.88

1.41

2.612

2.625

Δ‎ % Non-White

.315

1.5

−.097

−.47

.165

.373

Δ‎ % Advanced Degree

2.407

3.6

2.16

3.22

2.14

3.18

.554

.446

.436

Δ‎ % Union

.341

1.4

.241

.99

.229

.94

.222

.207

.199

Δ‎ % Poverty

.603

1.4

.500

1.18

.450

1.06

.435

.518

.427

Δ‎ Congregations per 10,000

−.071

−.1

−.012

−.02

−.006

−.01

.387

.341

.345

Constant

7.64

−7.68

8.27

4.076

4.327

3.686

N

50

50

50

R2

0.559

.619

.618

Adj. R2

0.509

.555

.564

Root MSE

4.66

4.43

4.39

Note: All models estimated with robust standard errors. Bold = p < .05; bold italics = p < .01 (one-tailed); b/s.e. = slope/standard error; and Δ‎Y,Sx = change in the dependent variable for a standard deviation change in the independent variable. The dependent variable is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4).

versions of the two measures of migration—change in percent foreign-born and change in the weighted measure of internal migration—used in chapter 3. Table 4.1 presents the set of five demographic indicators in Model 1. Here we see results similar to those found in the separate bivariate scatter plots: both change in nonwhite population and the percentage with advanced degrees are significantly related to change in state-level presidential outcomes, but none of the other variables has a statistically significant effect on changes in votes, either individually or when tested together. The effect of race and ethnicity and education are positive, as anticipated, and the impact of changes in education is particularly pronounced. The second column in Model 1 presents an intuitively clear means of comparing the effects of the independent (p.104) variables: each coefficient in this column is the amount of change in the centered Democratic vote for a 1 standard deviation change in each of the independent variables. This measure puts the independent variables on equal footing in terms of central tendency and variance and permits a more intuitively clear estimation of relative effect than do the raw coefficients. By the standardized measure, change in advanced degrees has more than twice the impact as change in the nonwhite population: a standard deviation change in nonwhite population produces a change in centered Democratic vote of 1.5 points, while a standard deviation change in percentage with advanced degree leads to a change in vote of 3.6 points.16 These findings fit fairly well with expectations related to the changing racial and ethnic composition of state electorates (Judis and Teixeira 2004), as well as those related to the changing size of the well-educated professional class (Hood and McKee 2010; Judis and Teixeira 2004; Manza and Brooks 1999), the primary difference being that education is singled out here as the indicator of the sort of values (p.105) that may produce more socially liberal preferences of the professional class (Feldman and Johnston 2014).

Adding Migration to the Mix

The story so far fits with the factors that researchers have suspected were moving states toward and away from the Democratic Party: states that grew relatively more nonwhite than other states shifted somewhat toward the Democratic Party, and states with slower growth in the nonwhite population shifted toward the Republican Party;17 states that saw larger than average increases in the percentage of population with advanced degrees grew substantially more Democratic, while states with smaller than average increases in advanced degrees generally grew more Republican. These effects are impressive, but they may be masking the effects of other variables, in particular variables related to population migration. Importantly both foreign-born and internal migration are tied to changes in party support (see chapter 3) and are also related to changes in race, ethnicity, and levels of education. For instance, the correlation between change in the nonwhite population and change in the foreign-born population in the states is .86, and the correlation between change in percentage with advanced degrees and change in the weighted measure of internal migration is .49. So some of the effects picked up in Model 1 may be attributed to the process that helped produce changes in race, ethnicity, and education, and at the same time, the migration processes may have some independent effects of their own.

The exploration of the relationship between migration and changes in political support presented in chapter 3 focused on how much political change there was in states with relatively low and high levels of foreign-born migrants and liberal internal migrants. The idea was that states with higher levels of population born outside the United States or in other states would have experienced more political change as a result of migration. And indeed the analysis in chapter 3 shows that there are substantively important relationships between the measures of migration (p.106) and electoral change. However, that analysis examined the relationship between levels of migration and political change, not changes in levels of migration and political changes. The point of this chapter is to examine changes in population characteristics, including changes in levels of migration. There was some existing level of foreign-born and internal migrant population in the states in the 1970s and 1980s that is highly correlated with levels in the 2000s, and to be confident that changes in migration influence changes in election outcomes we need to measure the level of migration in the 2000s relative to the levels forty years ago. The “change” operationalization is consistent with the measurement of other variables examined so far in this chapter and should increase the level of confidence we can have that the results for migration represent a dynamic process. This is easy enough for the foreign-born population, by simply subtracting the average percentage of foreign-born in 1972–80 from the average in 2004–12. It is a bit more difficult to measure change in the index of internal migrants weighted by the ideology of migrants’ birth states, since that index uses a lagged measure of state ideology that is not available prior to the mid-1970s. Still it is possible to calculate what the weighted internal migrant measure would have been in the late 2000s if migration patterns had not changed since the earlier time period. To do this Equation 3.1 is recalculated using the internal migration patterns that existed in 1980 rather than in 2008, and the difference between the two can be used as a measure of the change in the political direction of internal migration. (Positive values indicate a more liberal internal migration pattern in the later period.)

Model 2 combines the demographic and migration variables, producing some interesting changes. The most apparent among these is that when entered in the same model, change in percent nonwhite and change in the foreign-born population lose statistical significance. As pointed out earlier, these two variables overlap a lot, and including them in the same model introduces substantial collinearity.18 Given that the substantive impact of the foreign-born coefficient is barely changed from a model that includes just the migration variables,19 while the slope for nonwhite population is reduced to virtually zero (in fact is slightly negative), it appears that the effect of nonwhite population is swamped by adding change in (p.107) foreign-born population to the model. This does not mean that change in the racial and ethnic makeup of the states has no bearing on change in political outcomes, as part of the impact of changes in the foreign-born population stems from the fact that foreign-born migration is a major source of change in racial and ethnic composition. Despite its sustained substantive effect in Model 2, the coefficient for change in foreign-born population loses significance if change in nonwhite population is left in the model. When change in percent nonwhite is dropped from the model (Model 3), the slope for foreign-born population regains significance, and the model fit is relatively unaffected.20 It is important to understand that this diminution of the effects of change in nonwhite population is not in response to some coincidental overlap between two variables; it is in response to specifying one of the important processes that leads to changes in the racial composition of the electorate: changes in foreign-born migration patterns.

In addition to change in foreign-born population, change in the internal migrant population and change in percentage with advanced degrees also have significant influences on changes in Democratic support in Model 3. Among these variables, change in advanced degrees has the greatest impact, followed by change in foreign-born, and then by change in internal migrants. For changes in advanced degrees and changes in percentage of foreign-born these are fairly stable effects, but the standardized effect of change in weighted internal migration drops off a lot when the other demographic variables are added (see note 19), probably due in part to its relationship to change in advanced degrees (r = .49).

Party and Ideology

To this point my analysis has focused primarily on changes in population demographic characteristics, mostly those that are easily defined and measured by objective methods. However, as discussed in chapter 2, it is important to take into account changes in political tastes at the state level, changes that can be linked to changes in voting behavior. In particular it is important to measure changes in state-level ideology and party (p.108) identification. Certainly it must be the case that one of the reasons states drift toward supporting Democratic or Republican presidential candidates is that they have experienced changes in the underlying distribution of ideological or partisan preferences, especially given the importance of party and ideology as structuring factors in existing research on presidential outcomes in the states (Holbrook 1991; Rabinowitz et al. 1984; Rosenstone 1983). One of the difficulties in analyzing changes in party and ideology, especially as those changes relate to changes in election outcomes, lies in finding appropriate measures that are empirically distinct from the thing we want to explain: election outcomes. Most contemporary measures of state ideology, for instance, incorporate election outcomes into the measurement itself, either by measuring the ideological tendencies of elected officials from the states (Berry et al. 1998, 2010; Holbrook and Poe 1987) or, for both party and ideology, by using the election results or state demographic characteristics, along with public opinion data, to generate state-level estimates (Enns and Koch 2013; Pacheco 2011). For instance, the measure of state ideology used in chapter 3 to estimate the ideological tendencies of home states for internal migrants is based on hundreds of thousands of survey responses gathered across the fifty states but utilizes presidential election results and state demographic measures as part of a poststratification strategy to help provide more accurate estimates of state ideology (Enns and Koch 2013). This makes it an ideal measure of ideology for the analysis in chapter 3, where I was primarily interested in capturing the political context of the birth states for internal migrants at a given point in time. However, when explaining change in elections over time it is unwise to use a measure that incorporates election outcomes, since at least part of the dependent variable will be embedded in the independent variable. Likewise, because the Enns and Koch (2010) and Pacheco (2011) measures of party and ideology rely on state demographic characteristics in the estimation process, they are not ideal for models that include demographic characteristics, such as the one tested in this chapter.

The best alternative for this analysis is to utilize responses to party identification and ideology questions from national public opinion (p.109) surveys, aggregated at the state level. This method has been used by others (Carsey and Harden 2010; Wright et al. 1985), but none of the existing measures covers the time period studied here. The data I use are based on the raw—unadjusted—survey data from Enns and Koch’s (2013) project on state-level public opinion, which used hundreds of surveys and hundreds of thousands of responses to create poststratified measures of party and ideology.21 For each state in each election cycle I created a net Democratic Party identification (% Democratic identifiers minus % Republican identifiers) and a net liberal identification (% liberal identifiers minus % conservative identifiers) measure based on the raw state-by-state survey marginal percentages. There are some discontinuities in the data, and many steps were used to create these measures, all of which are summarized in the appendix.

Figure 4.5 presents the bivariate relationships between changes in state-level party identification and ideology and changes in presidential voting. The top panel shows the influence of changes in party identification on changes in votes. There are a couple things to note here. First, focusing just on the horizontal axis, Democrats lost ground in party identification in the overwhelming majority of states, gaining strength in only eight states. This reflects a longer-term secular decline that began in the mid-1960s and gained momentum in the 1980s (Meffert et al. 2001; Petrocik 1987). Second, Democrats generally made their greatest electoral gains in states in which net Democratic identification increased or decreased relatively little, and they suffered their greatest losses in states in which net identification declined substantially. With the exception of Iowa, the other seven states where Democrats increased their identification advantage (Connecticut, Delaware, Hawaii, Illinois, New Hampshire, New York, and Vermont) are all states where Democrats made significant gains. Further, on average, Democrats gained 2.6 points in states with above-average change in Democratic identification and lost 1.8 points in states with below-average change in Democratic identification. This pattern is further reflected in the correlation coefficient (.44), which indicates a moderate positive relationship. The impact of changes in ideology on changes in votes is presented in the bottom (p.110)

Compositional Change and Political Change

Figure 4.5 Changes in Partisan and Ideological Predispositions and Changes in Party Support in Presidential Elections, 1972–76 to 2004–12

Note: The dependent variable is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4). The independent variables are measured as the change in their average value from 1972–80 to 2004–12.

(p.111)

Table 4.2 Changes in State Political Predispositions and Changes in Centered Democratic Vote

Change in Centered Democratic Vote

Change in Ideology

Change in Party Identification

Model 1

Model 2

Model 3

Model 4

b/s.e.

b/s.e.

b/s.e.

b/s.e.

Δ‎ Liberal Advantage

0.318

0.101

0.122

0.108

Δ‎ Democratic Advantage

0.068

0.039

0.069

0.074

Δ‎ % Foreign-Born

0.793

−0.574

−0.456

0.245

0.402

0.711

Δ‎ Weighted Internal Migration

3.947

6.152

8.009

2.529

3.831

7.189

Δ‎ % Advanced Degree

1.798

1.953

3.694

0.495

0.624

1.085

Δ‎ % Union

0.168

0.889

−0.725

0.211

0.295

0.479

Δ‎ % Poverty

0.21

1.379

2.571

0.431

0.552

0.899

Δ‎ Congregations per 10,000

−0.012

0.308

−0.631

0.409

0.858

1.209

Constant

4.182

−6.184

−6.136

37.088

1.146

4.416

5.168

9.605

N

50

50

50

50

R2

0.295

0.641

0.443

0.493

Adjusted R2

0.265

0.570

0.366

0.422

RMSE

5.695

4.354

7.124

11.656

Note: All models estimated with robust standard errors. Bold = p < .05; bold italics = p < .01 (one-tailed); and b/s.e. = slope/ standard error. The dependent variable in Models 1 and 2 is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4); the dependent variable in Model 3 is change in the average net liberal identification from 1972–80 to 2004–12; the dependent variable in Model 4 is change in the average net Democratic identification from 1972–80 to 2004–12.

scatter plot in Figure 4.5. Akin to party identification, net liberal identification declined over time in most states, increasing in just nine states,22 and the scatter plot and correlation (r = .53) show a moderate positive relationship between change in ideology and change in votes. On average, Democrats gained 3.5 points in states with above-average change in liberal identification and lost two points in states with below-average change in liberal identification.

The final part of the analysis focuses on the combined effects of party, ideology, and demographic characteristics on presidential outcomes, as well as the extent to which changes in demographic characteristics influence changes in party and ideology. The first two models in Table 4.2 provide a test of the impact of party and ideology together, as well as in a model alongside demographic change variables. When considered together, impact of changes in party identification is diminished and not statistically significant, while ideology maintains both substantive and statistical significance.23 However, when demographic (p.112) variables are added to the model, both party and ideology are reduced to trivial influences. This is because their explanation of vote change overlaps with the statistical explanation provided by the set of demographic characteristics. Models 3 and 4 in Table 4.2 provide a sense of how much of the observed change in party and ideology is attributable to changes in demographic characteristics. In the case of political ideology, as a group, changes in demographic variables account for about (p.113) 37% of variance in changes from the 1970s to the 2010s, and changes in advanced degrees, poverty rate, and union membership stand out as particularly important, with change in poverty rate also making a significant contribution. In the case of changes in party identification (Model 4), demographic characteristics explain slightly more variation than for ideology, and changes in advanced degrees and poverty are also important influences.24 These results are important in illustrating that part of the mechanism by which changes in demographic characteristics lead to changes in election outcomes is by altering the underlying political predilections of the states.

Changing Populations and Changing Election Outcomes

A frequently offered explanation for changes in state-level voting patterns points to growth in certain population groups—usually racial minorities and well-educated professionals—as well as changes in the geographic distribution of those groups. Put very simply, the proposition is that states have become more Democratic or Republican over time in response to changes in their population characteristics. This idea has been subjected to empirical scrutiny here and largely finds a lot of support. But all population changes are not of equal consequence. The evidence I presented points to three important sources of political change in the states: changes in the percentage of the state CVAP who are foreign-born, changes in the pattern of internal migration among the CVAP, and changes in the percentage of state CVAP who are educated beyond a four-year college degree. States with significant gains in foreign-born population increased their support for Democratic presidential candidate, while states with declines in or low levels of foreign-born growth tended to move toward Republican presidential candidates. States whose patterns of internal migration shifted toward more liberal source states grew more Democratic, while states that shifted to more conservative source states moved toward Republican candidates. And states with relatively large (p.114) increases in the share of the population with more than a four-year college degree became markedly more supportive of Democratic presidential candidates, while those with slower growth in the highly educated population became more supportive of Republican candidates. The impact of changes in advanced degrees cannot be overstated: in every instance—whether explaining changes in votes or changes in party and ideology—changes in the percent with advanced education have had an important effect on all outcomes.

I have not discussed two hypothetically important sources of political change—changes in race and ethnicity and changes in the professional occupation class—as having substantial empirical effects. This is not because these variables are unimportant but because their effects are largely subsumed by two other variables: change in foreign-born population (strongly correlated with change in nonwhite population) and change in advanced degrees (strongly correlated with change in professional population). Changes in racial composition and professional occupations are related to changes in presidential votes. In fact when Model 3 of Table 4.1 is reestimated with change in nonwhite population substituted for change in foreign-born and change in percentage of professionals is substituted for change in percentage with advanced degrees, both variables are significant and show that Democratic fortunes have grown in states with the largest increases in percent nonwhite and percent professional and decline in states with the slowest growth in nonwhite and professional population (analysis not shown). However, substituting these variables comes at substantial cost in explanatory power, with the adjusted R2 dropping from .56 to .44. So it is not that changes in race, ethnicity, and professional occupations don’t matter to changes in state-level presidential election outcomes—they do—but that their effects are captured as part of the effects of changes in foreign-born and highly educated shares of the population.

Figure 4.6 provides a sense of how well the model tested here explains actual changes in presidential election outcomes in the states. The vertical axis is the actual change in centered Democratic share of the two-party vote in the states from 1972–80 to 2004–12, and the horizontal axis is the predicted change, based on the regression results presented in Model 2 of (p.115)

Compositional Change and Political Change

Figure 4.6 Predicted and Actual Changes in Centered Democratic Vote Share, 1972–80 to 2004–12

Note: The dependent variable is change in estimated Democratic support from 1972–80 to 2004–12, based on the trend in the Democratic share of the two-party vote, centered around the national two-party division (see Figures 1.3 and 1.4). The independent variable is the predicted change in estimate Democratic vote based on the regression slopes from Model 2 of Table 4.2, multiplied times the values of all of the corresponding independent variables.

Table 4.2. The predicted changes are calculated based on state outcomes on each of the independent variables and the slopes for those independent variables. Generally speaking, the actual outcomes track pretty closely with the predicted outcomes: states with changes in characteristics that are associated with increased Democratic support generally moved in the Democratic direction, and states whose populations changed in ways that augur for increased Republican support generally moved in that direction. There are a few states that are substantially more off diagonal than other states: Hawaii, Vermont, and especially Mississippi all became more Democratic than expected, given their population changes during this time period. It is usually best not to become too preoccupied with (p.116) explaining specific data points. Still it is worth pointing out that part of Hawaii’s deviance is probably due to a “favorite son” effect for President Obama, who was born in Hawaii, since the home-state variable used when estimating the trend in Democratic support only controlled for the state from which the candidates were running. Mississippi is a bit of a mystery, since none of its outcomes on the independent variables would point to any reason it should move toward the Democratic Party. At the same time, Mississippi’s increase of 4 points in Democratic support should not be confused with Democratic candidates having any chance of winning there or with Mississippi being a Democratic state. From 2004 to 2012 Democratic candidates ran 8.6 points below their national vote share in Mississippi, and only sixteen states have been less supportive of Democratic candidates. In short, Democrats have improved their lot in Mississippi a bit but still stand very little chance of winning.

It is also interesting to note that certain variables did not show up to the dance. For instance, in both the bivariate and multivariate models there is no sign of anything resembling a relationship between change per capita in religious congregations and change in presidential voting patterns. Similarly, although changes in party identification and ideology are related to changes in votes in the bivariate analysis, their effect is minimal in the multivariate analysis. In the case of party and ideology, part of the explanation for minimal effects is found in the overlap both variables have with the other independent variables. For change in religious congregations, however, the explanation might be that there simply wasn’t very much meaningful change. To be sure all states changed somewhat, but states that were very religious by this metric in the 1970s were also very religious in the 2010s. While there has been absolute change in religious saturation in the states, there hasn’t been a lot of relative change. In fact the correlation between religious congregations per 10,000 residents in the 1970s and in the 2010s is .93. Given the limited change in relative religiosity, it is perhaps not surprising that change in religiosity is not related to change in votes.

Chapter 5 takes a different perspective on explaining change in party support, looking not at changes in the values of independent variables but (p.117) at changes in the relationships between independent and dependent variables. So, for instance, state religiosity (measured with congregations per capita) might not have an impact on changes in votes when measured as a change variable, but its effect on election outcomes in any given election cycle may have changed over time in such a way that election outcomes are different now than they would have been if the relationship had not changed. The same may be true for other variables in the analysis. It may well be that some variables that were not connected to change in votes in this chapter still play a role in shaping changes in party support as a consequence of how their relationship to votes has changed over time.

Notes:

(1.) Moreover group characteristics continue to play an important role not just in voting studies but also throughout the discipline of political science. Some evidence of the relevance of groups to political science can be gleaned from the number of sections of the American Political Science Association with an explicit focus on group politics. Six of the forty-six sections focus on some aspect of group politics: Class and Inequality; Migration and Citizenship; Race, Ethnicity, and Politics; Religion and Politics; Sexuality and Politics; and Women and Politics Research.

(3.) Chaney et al. (1998) find that the growth of the gender gap in the 1980s was partly in response to women holding more pessimistic views of the economy and attaching greater weight to economic issues.

(4.) The 2012 ANES had both a face-to-face and an Internet sample, but the face-to-face sample does not include the measure of occupational status used here. The Internet sample was drawn from the universe of eligible voters and included both pre- and postelection interviews, with an overall sample size of 3,860. The sampling weight designed for the Internet sample was used in this analysis.

(5.) One concern with these measures of socioeconomic status is that they overlap a lot with race and ethnicity (whites are more likely than nonwhites to have high income, high level of education, and professional occupation), so these simple estimates might overstate the relationship. However, when the analysis is restricted to white voters, the same general patterns of support appear for income and occupation (though all levels are less likely than nonwhites to vote Democratic), and for education those with advanced degrees stand out from all other groups as the most likely (.60) to vote for Obama.

(6.) The ANES demographic group breakdowns are taken from http://electionstudies.org/nesguide/2ndtable/t9a_1_1.htm. Exit poll results are taken from the Roper Center, http://www.ropercenter.uconn.edu/polls/us-elections/how-groups-voted/.

(7.) In previous ANES surveys it was possible to get finer affiliation distinctions. As of this writing, those finer distinctions are not yet available for the 2012 study.

(8.) Nonwhites constitute, on average throughout the time period studied here, 23% of the CVAP in southern states and only 16% in nonsouthern states.

(9.) If change is measured by first centering all of the variables on their mean levels at different points in time—so that, at each point in time, the variable is measured as how far above or below the fifty-state average a given state is—the regression slopes and correlations presented in this chapter stay exactly the same. The only thing that changes is the constant term in the regression models.

(10.) Data for this variable are taken from unionstats.com (Hirsch and MacPherson 2014) and represent the share of the percentage of each state’s nonagricultural wage and salary employees who are union members. Note that this is not based on the CVAP.

(11.) One-tailed tests are appropriate here, given that positive trends are expected in both cases.

(12.) These data were collected by the U.S. Religion Census (www.usreligioncensus.org), which provides estimates for 1971, 1980, 1990, 2000, and 2010. Data for years between these dates are estimated using linear interpolation, and the 2010 data are used for the 2012 election. The link to individual data reports is at http://www.rcms2010.org/compare.php.

(13.) The set of variables explored above, and the narrower set explored below, surely do not exhaust all possible demographic influences. Data were gathered on several other indicators, including percent single female, percent white male, percent with less than twelve years of education, percent between the ages of eighteen and thirty, percent in creative, supercreative (Florida 2004), and managerial occupations. The list could go on and on, but it really shouldn’t. In the end these additional indicators add nothing of substance—either because their effect is already capably picked up by other variables or because they just bear no relationship to votes—and including them, and other variables, would only serve to offend my (and probably the readers’) parsimonious sensibilities and clutter up the analysis.

(14.) Collinearity at extreme levels can lead to substantially inflated standard errors for the slope estimates—increasing the likelihood of concluding that there is no relationship when in fact there is one—and, in some cases, sign switching of coefficients that makes no sense and results in theoretically perverse conclusions.

(15.) Both attendance and number of congregations are based on responses to surveys of religious organizations. My sense is that it is easier for those organizations to provide more accurate estimates of the number of congregations in their state than the number of regular attenders.

(16.) Returning to the issue of overlap between change in professionals and change in percentage with advanced degrees, there is more evidence that including change in professionals adds nothing to the model. When change in professionals is added to the model, it is not statistically significant, but change in advanced degrees remains significant and also remains the most important variable. When advanced degrees is dropped from the model and professionals replaces it, change in professionals is statistically significant, but the overall fit of the model drops off appreciably (adjusted R2 = .34).

(17.) As I pointed out earlier, all states saw an increase in the size of the nonwhite population.

(18.) The tolerances for change in nonwhite percentage and change in foreign-born percentage are .20 and .22, respectively.

(19.) In a model including just the migration variables, both change in internal and foreign migration were statistically significant (slopes 11.48 and .90, respectively) and the model-adjusted R2 was .39. The standardized coefficients indicate that internal migration (Δ‎Y,Sx = 3.32) is slightly more important than foreign-born migration (Δ‎Y,Sx = 2.47).

(20.) By contrast, the fit for the model is diminished if change in foreign-born is dropped in favor of change in percent nonwhite (R2 drops to .59 and adjusted R2 drops to .53).

(21.) Peter Enns very generously provided raw survey percentages, as well as various versions of the poststratified measure he created with Julianna Koch.

(22.) One of those nine states with an increase in net liberalism is North Dakota, which brings up an important point: an increase in net liberal identification doesn’t necessarily mean a state is more liberal than conservative—just that the balance has changed somewhat in the direction of liberal identification. For instance, net liberal identification in North Dakota moved from –23 (a 23-point disadvantage) in the 1970s to –21 (a 21-point disadvantage) in the 2000s. This registers as a gain in net liberal identification, but North Dakota is by no means a liberal state.

(23.) When considered in isolation in bivariate models, the regression slope for change in party identification was .19 and that for ideology was .39. The fact that party affiliation loses strength and significance means that there is significant overlap between these two variables. Indeed the correlation between changes in party affiliation and ideology at the state level is .65.

(24.) It should be noted that change in internal migration is significantly correlated with both change in party identification (r =.39) and ideology (r =.48) in bivariate models. However, these effects are washed away when change in advanced degrees (which is correlated with change in internal migration) is entered in the model.