Party Polarization, Ideological Sorting and the Emergence

Party Polarization, Ideological Sorting and the Emergence of the Partisan Gender Gap Abstract We argue that the modern American partisan gender gap—the tendency of men to identify more as Republicans and less as Democrats than women—emerged largely because of masslevel ideological party sorting. Gender differences in preferences have persisted in several major policy areas since the dawn of modern polling. As the parties ideologically polarized at the elite level, the public gradually perceived this polarization and better sorted themselves into parties that matched their preferences. Existing preference differences between men and women caused this sorting to generate the modern partisan gender gap. Those with more general political awareness are more attuned to elite party polarization, and consequently the partisan gender gap developed earlier and is consistently larger among the most politically aware. We find support for this argument from decades of pooled American National Election Studies data and a new large dataset of decades of pooled individual-level Gallup surveys. 1 Introduction It is well known that currently in the United States men identify more as Republicans and less as Democrats than women do. While the press and the academic literature largely ignored the relationship between gender and party identification in the 1960s and 1970s, journalistic and scholarly attention to this topic has persisted since males disproportionately supported Ronald Reagan in 1980 (e.g., Mansbridge, 1986; Mueller, 1988; Wolbrecht, 2000; Kaufmann, 2006). Some contend that the modern partisan gender gap emerged because social changes, such as greater female labor force participation, liberalized divorce laws or feminist political socialization of younger generations, caused men’s and women’s policy preferences to increasingly diverge. Yet, public opinion scholarship finds that men have been more conservative than women on criminal justice, social welfare, and foreign policy more or less since the start of modern polling. (e.g., Shapiro and Mahajan, 1986; Kaufmann and Petrocik, 1999).1 Specifically, Shapiro and Mahajan’s meta-analysis finds that these gender differences in opinions were roughly stable from 1964 to 1983, the entire period they examined and the period when we find that the partisan gender gap emerges and shows its largest growth.2 What changed during this time were perceptions of party policy divergence. Since the 1960s, Democratic and Republican elites more consistently took liberal and conservative positions, respectively, on a wide range of prominent national political issues, including the size of the welfare state, crime, civil rights, and military aggressiveness (Carmines and Stimson, 1989; McCarty, Poole and Rosenthal, 2006; Noel, 2013). In response to this polarization, citizens gradually perceived greater ideological divergence between the parties’ issue positions (Layman and Carsey, 2002).3 We contend that the partisan gender gap emerged because perceiving greater ideological divergence between the parties made it easier for people to choose a party that matched their 1 The origins of these consistent differences in the policy preferences of men and women is beyond the scope of this paper. On this, see for example Huddy, Cassese and Lizotte (2008, 33-35). 2 We follow convention and label these sex differences as a “gender gap,” although it is more precisely called a sex gap (Beckwith, 2005). 3 The pattern of party ideological polarization at the elite and mass level is at least partially driven by liberal and conservative social movements and interest groups increasingly affiliating with the Democrats and Republicans, respectively. This is true on women’s rights issues (Mansbridge, 1986; Wolbrecht, 2000) as well as in many other areas (e.g., Karol, 2009). 1 preferences. A number of prominent studies show that elite polarization caused slow changes in mass-level party identification as people gradually sorted into parties based on those preferences (e.g., Levendusky, 2009; Abramowitz, 2010; Fiorina, Abrams and Pope, 2011). We assert that, because men hold more conservative positions than women on several prominent issues, such ideological sorting is likely to produce a partisan gender gap in which men and women disproportionately identify as Republicans and Democrats, respectively. While there are many reasons why the modern partisan gender gap developed, we argue that ideological sorting was the main mechanism causing its steady emergence between the 1960s and the 1990s. We test this explanation with two datasets, each with important, yet different, strengths. The biennial American National Election Studies (ANES) contain detailed questions about respondents’ partisan identification, policy preferences, and perceptions of the parties’ ideological stances in different issue areas. However, ANES surveys are too small and infrequent to precisely measure variation in the partisan gender gap over time or differences in its size between those with more and less political awareness. To remedy this, we assembled the largest dataset ever used to study the partisan gender gap by pooling individual-level responses to 1,822 Gallup polls that included questions about gender and party identification from 1953 through 2012. While lacking the ANES’s variety of political questions, the Gallup dataset gives us the statistical power to precisely track party identification separately by gender and education over time. By leveraging the different strengths of these datasets, we find support for our argument that the modern partisan gender gap emerged largely because party polarization made longstanding gender opinion differences more relevant to partisanship. 2 Existing Literature Males’ disproportionate support of Ronald Reagan in 1980 sparked scholarly interest in the partisan gender gap. Since that time, scholars have put forth several plausible explanations for its emergence. Many of these explanations posit changes over time in the difference between men’s and women’s policy preferences (e.g., Deitch, 1988; Inglehart and Norris, 2000; Edlund and Pande, 2 2002; Box-Steffensmeier, De Boef and Lin, 2004; Iversen and Rosenbluth, 2006). Some scholars propose that women’s policy preferences changed as they became more psychologically and financially independent from men. The Developmental Theory of Gender Realignment claims that women’s attitudes on cultural issues, such as freedom of self-expression and gender equality, moved substantially to the left in affluent countries because of their increased labormarket experience and greater economic independence from men (Inglehart and Norris, 2000). In support, Inglehart and Norris show that the partisan gender gap grew more in wealthy countries, particularly in younger cohorts that were socialized into politics later. Inglehart and Norris (2003) also show that a substantial portion of the partisan gender gap in wealthy countries is explained by differences in cultural attitudes toward postmateralism, support for gender equality and the role of government.4 Consistent with this argument, Kaufmann (2002) finds that American women’s attitudes on social issues have become increasingly correlated with their party identification over time. Also consistent are Conover’s (1988) and Manza and Brooks’ (1998) findings that women with a feminist consciousness have more liberal policy attitudes and are more likely to identify as Democrats (although see Cook and Wilcox (1991)). A related argument made by Iversen and Rosenbluth (2006, 2011) is that over time women increasingly preferred a larger welfare state because they needed it to sustain their economic independence from men. They cite public sector employment and subsidized daycare as examples of policies that make it easier for females to maintain their careers while raising children. Consistent with their hypothesis, they show that, in European countries, the gender gap in support for public employment and left-leaning political parties is larger among labor force participants. Likewise, Carrol (1988) and Manza and Brooks (1998) show, in the United States, that women who work outside the home and in more economically independent professions vote more Democratically. Another line of argument is that women began to prefer a larger welfare state over time because they became more economically vulnerable. Declining marriage rates and increasing divorce rates are often cited as sources of this increased economic vulnerability. Edlund and Pande (2002) show 4 Although Inglehart and Norris (2000) find these factors have little explanatory power over the partisan gender gap in the United States. 3 that the partisan gender gap is larger in states where divorce is more prevalent and that, in panel data, marriage and divorce make men and women more Republican and Democratic, respectively. Similar to this, Iversen and Rosenbluth (2006) find that the partisan gender gap is larger in the unmarried population in European countries. Several studies also show that women tend to perceive the economy more negatively than men do (Miller, 1988; Ladd, 1997; Chaney, Alvarez and Nagler, 1998). Using aggregated time-series data, Box-Steffensmeier, De Boef and Lin (2004) show that the U.S. partisan gender gap increases when economic performance wanes and the number of economically vulnerable single women increases. Yet, there are reasons to question whether women’s changing policy preferences fueled the partisan gender gap’s emergence. If changes in women’s social or economic policy attitudes caused the partisan gender gap, we should see growing gender policy preferences differences at the same time when the partisan gender gap emerged. However, gender differences in policy preferences predated the origins of the partisan gender gap and remained relatively constant in the 1970s and 1980s when the partisan gender gap grew the most (Shapiro and Mahajan, 1986; DiMaggio, Evans and Bryson, 1996; Fiorina, Abrams and Pope, 2011). Also inconsistent with this explanation, a number of studies also show that the partisan gender gap developed at least as much because of changes in the men’s partisanship as changes in the women’s partisanship (Wirls, 1986; Kaufmann and Petrocik, 1999; Norrander, 1999). On what issues do men and women differ? Shapiro and Mahajan’s (1986) meta-analysis of gender differences on 962 issue questions asked from 1964 through 1983 shows that men were significantly more conservative than women on the size of the welfare state and issues related to the use of force, such as national defense and criminal justice policy. Kaufmann and Petrocik (1999) also show that men have held more conservative social welfare positions than women since at least the 1950s. Other studies also find gender differences on issues related to the use of force (e.g., Smith, 1984; Mansbridge, 1985; Kaufmann, 2006; Eichenberg and Stoll, 2012), which date 4 back at least to the 1940s (Ladd, 1997, 116).5 If men’s and women’s issue preferences have differed in several areas since at least the 1960s, why did the partisan gender gap not form earlier? In the next section, we argue that this is because people did not perceive that the parties were sufficiently differentiated on the issues over which men and women disagreed. We are not the first to relate party polarization to development of the partisan gender gap. Activists in the 1980s publicized the growing differences between the Democratic and Republican platforms on women’s rights issues like abortion and the Equal Rights Amendment (ERA) as an explanation for the partisan gender gap (Bonk, 1988; Costain, 1988; Wolbrecht, 2000). However, a problem with this explanation is that men and women usually report similar preferences on these issues (Mansbridge, 1986; Seltzer, Newman and Leighton, 1997; Fiorina, Abrams and Pope, 2011). In addition, the partisan gender gap predates changes in the national party’s positions on these social issues (Norris, 2003; Kaufmann, 2006; Wolbrecht, 2000) and campaign appeals about traditional women’s issues do not increase the partisan gender gap (Hutchings et al., 2004). 3 Theory and Hypotheses Since the 1960s, elite level ideological polarization of the parties led the public to gradually perceive that the parties were getting more polarized. This caused some members of the public to become aware that their issue preferences were inconsistent with their party preferences. Some of these people changed their party identification in order to sort into a party that better matched their issue preferences.6 Because men consistently hold more conservative policy views in many major issues areas, such ideological party sorting caused a partisan gender gap to emerge and grow 5 Findings that, because they are framed more around gender, other types of policy attitudes have become more related to attitudes about gender equality (e.g., Winter, 2008) are consistent with our argument that people’s attitudes on all major issues became more correlated with partisanship and with each other as people ideologically sorted. Partisanship and attitudes in all areas became more highly correlated with each other, while gender preference differences on women’s rights issues remained substantially smaller than on other topics. But this does mean that some of the issues where there was a gender preference gap, while not traditionally associated with women’s equality, were now thought of as gendered to many people. 6 Our argument does not preclude that some people adjusted their preferences to match their partisanship. We agree with Layman and Carsey (2006) that elite ideological party polarization causes some of both processes to occur. 5 over time. Previous work establishes that party identification is a sturdy attitude that only changes very gradually in response to major political change (e.g., Converse, 1964; Zaller, 1992). Thus, we expect changes in party identification in response to elite party polarization to be gradual as well. A precondition to sorting on the basis of elite party polarization is knowledge of elite party polarization. Previous work shows that the more politically aware are more knowledgeable of elite political positions, and particularly changes in the positions of political elites over time (e.g., Converse, 1964; Zaller, 1992). So, if the partisan gender gap is ultimately driven by noticing the increasing ideological polarization of the two parties, it should emerge earlier and be larger among the most politically aware. This argument leads to five hypotheses that we test in this paper. First, we expect to observe a larger partisan gender gap among those who perceive greater ideological divergence between the parties. Thus, measures of overall political awareness, such as education and political knowledge levels, should positively associate with perceptions of partisan ideological divergence and with the partisan gender gap. Second, we expect the association between issue preferences and partisanship to increase over time and at all times to be larger among the more politically aware. Third, we don’t expect gender differences in policy preferences to grow before or during the period when the partisan gender gap emerges. Fourth, because elite party polarization happened slowly over several decades and party identification is a sturdy attribute, the partisan gender gap should emerge and grow gradually over decades, not suddenly in response to any specific event. Fifth, as a consequence of this, the partisan gender gap should emerge earlier, and be consistently larger, among the most educated subset of the population, who are better able to perceive elite ideological position-taking. This argument is most similar to, and builds on, Kaufmann and Petrocik’s (1999) claim that the partisan gender gap was caused by an increase over time in the influence of social welfare preferences on partisan identification. However, we move beyond Kaufmann and Petrocik both theoretically and empirically. Theoretically, we draw on the ideological sorting literature and argue that people increasingly sorted into parties that matched their preferences on all prominent political domains, not just their social welfare preferences. We also explain and demonstrate that 6 political awareness and perception of party polarization are key steps in the causal process.7 Empirically, while Kaufmann and Petrocik (1999) examine only two years of ANES data (1992 and 1996), both after the gender gap had emerged, we employ four decades of pooled ANES data and five decades of polled Gallup data. 4 Data We test our predictions using two datasets, each with its own different strengths: the ANES and a new individual-level dataset of pooled responses to Gallup polls from 1953 through 2012. As the ANES is used widely in political science, we will not describe it in detail here, but simply highlight its advantages for our project. Since 1970, the ANES asked respondents about their perceptions of the positions of the Democratic and Republican parties in a variety of issue domains. This allows us to directly relate perceptions of polarization to the partisan gender gap. The ANES also contains detailed preference questions on a variety of major policy areas during the decades when this gender gap arose. This allows us to study how the gender gap in issue preferences changed before and after the emergence of the modern partisan gender gap. While most existing literature on the partisan gender gap uses only the ANES as its public opinion dataset (e.g., Wirls, 1986; Chaney, Alvarez and Nagler, 1998; Kaufmann and Petrocik, 1999; Norrander, 1999; Edlund and Pande, 2002; Kaufmann, 2002; Norris, 2003; Kaufmann, 2006), the ANES only contains between 1,000 and 3,000 respondents every two years (at best) since 1952. A few studies use the GSS instead (e.g., Wirls, 1986; Deitch, 1988), but it has a roughly similar sample size and has been conducted at about the same frequency (usually every two years) since 1972. This makes it unlikely that we would detect a partisan gender gap in these datasets that is statistically significant at conventional levels when only a small partisan gender gap is appears in the population. This also limits our ability to detect whether the size of partisan gender gap depends on attributes like education. To measure exactly when the partisan gender gap emerged, 7 In the existing literature, Ladd (1997) observes that the partisan gender gap is larger among the more educated in the 1980s and 1990s, but does not look at earlier time periods or changes over time, nor does he connect this phenomenon to partisan sorting and perceptions of polarization. 7 and among whom, we need bigger samples. One way to achieve this is with more frequent surveys that, when pooled, produce large enough sample sizes to measure the gap with more precision. To this end, we assembled the largest dataset ever used to study the partisan gender gap. While commercial survey data previously has been used to model the dynamics of men’s and women’s presidential approval (Clarke et al., 2005) and partisanship (Box-Steffensmeier, De Boef and Lin, 2004), we are aware of no previous paper that has used individual-level data to do so. In Section 7.1 in the Appendix, we describe our collection of individual-level responses from every poll conducted by the Gallup Organization from 1953 through 2012 that (1) asked about presidential approval, party identification, and/or ideology and (2) is contained in the Roper Center iPOLL database.8 There are 2,246,369 observations from 1,822 surveys that ask a nationally representative sample about their gender and partisan identification. Because at least 13, and often substantially more, Gallup polls are available in every year our dataset covers, we observe the party identification of tens of thousands of respondents per year. This gives us greater statistical power to precisely measure how the partisan gender gap changes over time and to examine subgroups (e.g. college graduates) separately, while maintaining adequate sample sizes.9 There are also drawbacks to our Gallup dataset. Most importantly, many demographic characteristics and political attitudes that we would like to observe are asked sporadically, if at all. Given the substantial cost of collecting each additional variable, we only collected variables that were asked relatively consistently over time.10 The only attitudes we collected are presidential approval, party identification, and, since 1992, ideological self-placement. Thus, we are unable to examine issue preferences over time using Gallup data. Gallup data also do not contain the ideal measure of political awareness. While political awareness is probably best measured by a short quiz of basic political facts (Price and Zaller, 1993), such quizzes are almost never asked in Gallup surveys. Instead, we use education as a proxy for respondents’ awareness of changing elite party positions and ability to understand the consequences of 8 We start in 1953, because Dwight Eisenhower’s was the first presidency during which Gallup used probability sampling exclusively (Berinsky, 2006). 9 Our Gallup series begins well before the emergence of the modern gender gap, unlike the aggregate time series of CBS/New York Times polls used by Box-Steffensmeier, De Boef and Lin (2004), which starts in 1979. 10 See Section 7.1 in the Appendix for a description of the variables we collected. 8 those positions for their own party affiliation. Because education is strongly correlated with media exposure and overall political sophistication (Fiske, Lau and Smith, 1990; Price and Zaller, 1993; Delli Carpini and Keeter, 1996), it is often used as a proxy for these types of attributes in public opinion scholarship (e.g., Carmines and Stimson, 1989; Zaller, 1994; Berinsky, 2007; Hayes and Guardino, 2011). For instance, Carmines and Stimson (1989, ch. 5) contend that the politically sophisticated are more likely to notice long term changes in party positions, and use education, in addition to political information questions, to measure political sophistication. Gallup asks about party identification in a slightly different manner than the ANES, GSS, and CBS/New York Times polls. Gallup asks “In politics, as of today, do you consider yourself a Republican, Democrat or Independent?”11 Abramson and Ostrom (1991) argue that the Gallup question, which is also used in Erikson, MacKuen and Stimson’s (2002) analysis of “macropartisanship,” introduces more short-term political and economic considerations than the partisanship wording in the ANES. Thus, our study lies somewhere between a study of voting preferences and of ANES-style party identification. While important, these differences should not be overstated. All three of these variables—presidential voting, ANES party identification, and Gallup party identification—are substantially correlated both within individuals and in their aggregate movements over time (MacKuen, Erikson and Stimson, 1992; Kaufmann and Petrocik, 1999, 868-870). A related complication is that Gallup did not always follow-up with Independents and ask whether they lean towards a party. Gallup asked Independents about their leanings occasionally in the 1950s, almost never in the 1960s and 1970s, sometimes in the 1980s, and then regularly from the 1990s on. Because we cannot consistently observe Independent leaners in the Gallup data, we code them as Independents in our main analysis. However, because there is some evidence that men are more likely than women to label themselves as Independent leaners rather than partisans (Norrander, 1999), we also examine the robustness of results to including leaners as partisans on the subset of surveys that ask about leanings. A final concern with the Gallup data is that Gallup changed from using in-person surveys to phone surveys over time. When both modes were used during the late 1980s and early 1990s, more 11 Gallup also occasionally omits “as of” from the question. 9 respondents identified as Republicans when the survey was administered by phone than in person (Green, Palmquist and Schickler, 2002). Thus, care needs to be taken when looking at long run changes in the partisan gender gap to account for any differences that were caused by this change in survey mode. 5 Results 5.1 Perceptions of Party Polarization and the Partisan Gender Gap We first examine how perceptions of party polarization evolved over the time period when the partisan gender gap emerged. Figure 1 graphs the mean difference in respondents’ assessments of the policy stances of the two parties in ANES surveys from 1970 to 2004. The dependent variable is the difference in respondents’ assessments of the average conservatism of the Republican Party’s and Democratic Party’s positions on all available issues, including domestic welfare spending, law and order, racial policy, and gender-related issues. A value of zero corresponds to an assessment that the issue positions of both parties were equally conservative, while a value of one corresponds to an assessment that the Republican Party’s issue positions were maximally conservative and the Democratic Party’s issue positions were maximally liberal. We separate college graduates (white dots) from those with less education (black dots) to examine how political awareness affects perceptions of ideological differences. While there is year-to-year variation, Figure 1 shows that assessments of party polarization were significantly larger in the 1980s and 1990s than in the 1970s among both the high and low education groups. In all years, the college educated assess more polarization than the less educated.12 Column 1 of Table 1 shows that the partisan gender gap increased over the same time interval that assessments of issue polarization increased. The dependent variable for all of the regressions in Table 1 is a respondent’s seven-point partisan identification recoded so that it ranges from zero 12 Section 7.4 in the Appendix shows similar patterns when we account for changes in the issues that are surveyed in different years. 10 Differences in Ideological Assessments of Parties’ Issue Positions 0 .1 .2 .3 .4 .5 Figure 1: Assessments of Polarization in the Parties’ Issue Positions 1968 1972 1976 1980 1984 1988 Survey Year 1992 1996 2000 2004 College Graduates Non−College Graduates Black Lines Indicate 95% Confidence Intervals Grey Lines Indicate Linear Trend (“Strong Democrat”) to 1 (“Strong Republican”). The partisan gender gap almost tripled from -.013 (std. err. = .007) to -.034 (std. err. = .007) between the 1970s and 1980s. By the 1990s, it had nearly doubled again to -.064 (std. err = .007).13 Our first hypothesis predicts that the partisan gender gap will be larger when people perceive greater differences in the parties’ policy stances. Consistent with this prediction, column 2 of Table 1 reports a significant negative interaction between the female indicator and the respondent’s polarization assessment. Given that the interquartile range of polarization assessments is .000 to .367, moving from the 25th to the 75th percentile in polarization assessments is associated with the partisan gender gap increasing from -.010 (std. err. = .005) to -.062 (std. err. = .005). Column 3 shows that moving from the 25th to the 75th percentile in the polarization assessment is associated 13 Following Kam and Franzese’s (2007) advice, when interpreting results from this table and all subsequent tables with regression models containing interaction terms, we transfer the coefficients when necessary into quantities of substantive interest, e.g., the marginal effects at different levels of the conditioning variable. Here, our quantity of interest is the partisan gender gap among different segments of the data, such as in specific decades and/or at certain education levels. In addition, when interpreting interaction terms in this and subsequent tables, we will again follow Kam and Franzese’s (2007, 49) advice by interpreting the statistical significance of the interaction term’s coefficient as indicating whether the slope of the main variable is significantly different when the conditioning variable moves from 0 to 1, which is the range of all our conditioning variables. 11 Table 1: Assessments of Party Polarization and Partisan Identification (ANES, 1970 - 2000, 2004) N Female Female X Year ≥ 1980 Female X Year ≥ 1990 Polarization Assessment Polarization Assessment Year ≥ 1980 Polarization Assessment Year ≥ 1990 Female X Polarization Assessment Female X Polarization Assessment Female X Polarization Assessment College Graduate Female X College Graduate Political Information Female X Political Information (1) 31833 (2) 28115 (3) 28115 (4) 31663 (5) 31833 (6) 27974 -0.013 -0.011 -0.004 -0.024 0.010 -0.001 (0.007) (0.005) (0.008) (0.004) (0.008) (0.009) -0.022 0.007 (0.010) (0.013) -0.029 -0.035 (0.009) (0.013) 0.124 0.088 0.062 (0.013) (0.024) (0.014) X 0.035 (0.034) X 0.013 (0.030) -0.142 -0.057 -0.120 (0.017) (0.032) (0.018) -0.134 X Year ≥ 1980 (0.044) 0.046 X Year ≥ 1990 (0.040) 0.100 0.072 (0.007) (0.008) -0.044 -0.024 (0.010) (0.012) 0.145 0.084 (0.010) (0.013) -0.069 0.001 (0.014) (0.017) Notes: The dependent variable is the 7-point party identification scale recoded so that it ranges from 0 (“Strong Democrat”) to 1 (“Strong Republican”). 7-point assessments of a party’s issue positions are recoded so that they range from 0 (“Most Liberal”) to 1 (“Most Conservative”). “Polarization Assessment” is constructed by subtracting the average issue position of the Democratic party from the average issue position of the Republican party. “Political Information” is constructed by averaging the interviewer’s assessment of the respondent’s “politics and public affairs” information and ranges from 0 (“very low”) to 1 (“very high”). All regressions also include year fixed effects and observations are weighted by their sample weight. Robust standard errors are reported in parentheses. 12 with the partisan gender gap increasing from -.004 (std. err. = .008) to -.020 (std. err. = .009) in the 1970s, .003 (std. err. = .010) to -.073 (std. err. = .009) in the 1980s, and -.033 (std. err. = .010) to -.100 (std. err. = .010) in the 1990s. Thus, in all three of these decades, assessments of party issue polarization are positively associated with the size of the partisan gender gap, and that association is larger in the 1980s and 1990s, when this gender gap is larger. Our contention is that political awareness leads to greater assessments of party polarization, which in turn causes a partisan gender gap. This implies that we should observe negative interactions between the female indicator and measures of overall political awareness, which attenuate once we control for the interaction between the female indicator and assessments of polarization, because the affect of awareness flows through perceptions of polarization. Columns 4-6 of Table 1 show that this is in fact the case. We test two different measures of political awareness: an indicator for whether the respondent graduated from college (which we will also use in the Gallup data analysis) and the ANES interviewers’ assessments of each respondent’s level of political information.14 If education and interviewers’ assessments of respondents’ political information are both good proxies for general political awareness, both should have similar effects on the gender gap. That is what we find. Results in column 4 indicate that the partisan gender gap is about three times larger among college graduates (-.065, std. err = .010) than among non-college graduates (-.021, std. err = .004). Column 5 shows that the partisan gender gap is also positively associated with interviewers’ political information ratings (-.006, std. err = .005 at the 25th percentile to -.040, std. err = .008 at the 75th percentile). However, column 6 demonstrates that when polarization assessments are included as an explanatory variable, the magnitude of the interaction between the female indicator and college indicator is reduced by about 50 percent and the interaction between the female indicator and the interviewers’ political information assessments switches signs and is statistically insignificant. In contrast, the interaction between the female indicator and polarization assessments is similar to what it was in column 2. Political awareness appears to affect the size of the partisan gender gap through polarization assessments. 14 While we would prefer to use responses to a battery of political fact questions to measure political information, the ANES did not consistently ask such questions in the 1970s. In contrast, the interviewer rating variable is available in every ANES and has been used previously as a political information proxy (e.g., Bartels, 1996) 13 5.2 Issues Preferences and Partisanship Over Time Our second hypothesis is that the association between issue preferences and partisanship is always higher among the more politically aware, but increases over time among those of all political awareness levels. To investigate this, we regress party identification on three separate indices of ANES issue questions: social welfare preferences, gender role preferences and all other preferences separately by decade. Our expectation is that sorting will cause the association between issue preferences and partisan identification to increase over time. Because college graduates were aware of polarization earlier and continued to assess more polarization as time passed, we also expect to observe a stronger association between the issues positions and party identification of college graduates than non-college graduates. When running these regression we include the interaction with an indicator for whether the respondent is female and our three issue position measures to account for the possibility that men and women may weight their issue positions differently when forming their partisan identification (Kaufmann and Petrocik, 1999). Consistent with Kaufmann and Petrocik (1999), Table 2 shows that the association between social welfare preferences and partisan identification more than doubles over time. However, the fact that we see a similar pattern when looking at all preferences other than social welfare and gender roles suggests that this wasn’t unique to social welfare. Preferences became increasingly correlated with partisanship in most issue areas. In both social welfare and the index of all other policy preferences, we observe stronger associations among college graduates. In contrast, on gender role preferences, we always observe weaker relationships between preferences and partisanship and smaller differences between college graduates and non-college graduates. Finally, also contrary to Kaufmann and Petrocik (1999), we see little difference between men and women in how issue preferences relate to partisan identification. While we need to be cautious drawing conclusions about causal direction from this type of analysis because of possible endogeneneity, the similar associations between preferences and party identification among men and women is at least inconsistent with claims that either men place greater weight on social welfare preferences or that women place greater weight on gender roles in forming their party 14 Table 2: Associations between Issue Positions and Partisan Identification over Time and by Educational Attainment (ANES, 1972 - 2008) (1) Included Years College Graduates N Female (2) 72-78 Yes 1143 -0.010 (0.046) Social Welfare Preference 0.288 (0.050) Gender Roles Preferences 0.029 (0.047) All Other Preferences 0.420 (0.077) Female X 0.011 Social Welfare Preference (0.075) Female X 0.090 Gender Roles Preferences (0.071) Female X -0.017 All Other Preferences (0.114) No 5823 0.003 (0.024) 0.171 (0.021) -0.023 (0.018) 0.248 (0.034) 0.007 (0.028) 0.029 (0.024) -0.017 (0.045) (3) 80-84, Yes 1230 (4) 88 No 4984 (5) (6) (7) (8) 00, 04, 08 Yes No 872 2331 90-98 Yes 1957 No 5744 -0.144 -0.065 0.003 0.009 (0.064) (0.032) (0.044) (0.033) 0.367 0.171 0.653 0.483 (0.073) (0.033) (0.061) (0.041) 0.089 0.003 0.203 0.013 (0.051) (0.024) (0.039) (0.024) 0.647 0.349 0.754 0.363 (0.080) (0.040) (0.060) (0.042) 0.001 0.012 -0.116 -0.038 (0.109) (0.042) (0.094) (0.053) 0.111 0.071 0.016 0.044 (0.076) (0.031) (0.064) (0.030) 0.162 0.049 0.007 -0.045 (0.129) (0.053) (0.094) (0.054) -0.083 0.075 (0.059) (0.042) 0.592 0.643 (0.105) (0.061) -0.132 -0.005 (0.066) (0.036) 0.865 0.437 (0.099) (0.056) 0.179 -0.209 (0.154) (0.090) 0.280 0.094 (0.097) (0.051) -0.089 -0.042 (0.147) (0.078) Notes: 7-point partisan identification is recoded so that it ranges from 0 (“Strong Democrat”) to 1 (“Strong Republican”). Issue positions are recoded so that they range from 0 (“Most Liberal”) to 1 (“Most Conservative”). All regressions also include year fixed effects and observations are weighted by their sample weight. Robust standard errors are reported in parentheses. identification. 5.3 Gender Differences in Issue Preferences Over Time Next we test our third hypothesis: that the preference differences between men and women do not notably grow during the period when the partisan gender gap is emerging. The dependent variable in this analysis is a respondent’s issue position preferences, rescaled to run from 0 (most liberal) to 1 (most conservative). The independent variables are an indicator for whether the respondent is female, the interaction between the female indicator and an indicator for whether the survey occurred in 1980 or later, the interaction between the female indicator and an indicator for whether the survey occurred in 1990 or later, the interaction between the female indicator and an indicator for whether the survey occurred in 2000 or later, and survey-year fixed effects. Standard errors are clustered by respondent to account for non-independence across issue positions. Table 3 shows that the issue preference gender gap and the partisan gender gap did not evolve 15 Table 3: Gender Differences in Issue Positions, Ideology, and Economic Evaluations (ANES, 1970 - 2008) (1) All Issue Positions (2) Social Welfare N 514313 Female -0.023 (0.003) -0.007 (0.004) -0.011 (0.004) -0.005 (0.004) Dependent Variable 169393 (3) Foreign Relations 74, 78, 02 43809 -0.045 (0.007) 0.016 (0.008) -0.016 (0.005) 0.006 (0.005) -0.051 (0.009) 0.019 (0.011) 0.011 (0.007) -0.030 (0.012) No Questions Asked Female X Year ≥ 1980 Female X Year ≥ 1990 Female X Year ≥ 2000 (4) (6) (7) 88493 (5) Gender Roles 70 86, 02 25459 Abortion 70 74, 02 30024 ERA 70-74, 82-08 4589 -0.013 (0.005) -0.013 (0.008) -0.023 (0.008) 0.011 (0.011) 0.035 (0.009) -0.026 (0.012) 0.001 (0.010) -0.039 (0.011) -0.022 (0.008) 0.006 (0.011) 0.003 (0.011) 0.020 (0.014) 0.011 (0.013) Race 02 (8) (9) (10) Personal Finances Ideology Last Year Next Year 78 02 27353 33114 28998 -0.010 (0.004) -0.008 (0.006) -0.007 (0.005) 0.005 (0.006) -0.046 (0.010) -0.012 (0.013) 0.003 (0.012) 0.014 (0.013) -0.024 (0.009) -0.014 (0.012) -0.002 (0.010) 0.000 (0.011) Notes: Issue positions and ideology are recoded so that they range from 0 (“Most Liberal”) to 1 (“Most Conservative”). Economic evaluations recoded so that they range from 0 (“Least Favorable”) to 1 (”Most Favorable”). “No Questions Asked” indicates years in which no questions were asked on that specific issue dimension. All regressions also include year fixed effects and observations are weighted by their sample weight. Robust standard errors that are clustered by respondent are reported in parentheses. concurrently. Column 1 reports results when pooling across all issue position questions, and shows that males are .022 (std. err. = .004) units more conservative than females in the 1970s. The issue preference gender gap does not significantly change between the 1970s and 1980s.15 Between the 1980s and the 1990s, the gap statistically significantly grows to .041 (std. err. = .003). This indicates that the gender gap in issue preferences grew mostly after the partisan gender gap experienced its largest growth. Breaking things down by specific issue areas shows that the initial 1970s preference gaps are mostly concentrated among social welfare and foreign relations preferences. If anything, men’s and women’s policy preference on these issue areas become more similar over time. In contrast, men’s and women’s racial policy preferences are only slightly different in the 1970s and then become increasingly distinct over time. In three areas more explicitly related to gender, women were more liberal than men only on abortion in the 1970s. There was no difference in opinions on the Equal Rights Amendment and women actually held more conservative views than men on gender roles 15 Because the specific issue questions change over time, one concern is that this regression confounds changes over time in the issue preference gender gap with variation in gender differences across the different issues that are surveyed in any particular year. One way to account for this is to also include question fixed effects and question fixed effect interacted with the gender indicator. When we run this regression, we find that the gender preference gap increased by 0.016 (std. err. = 0.006) units between the 1970s and 1980s. A downside of this approach is that it uses only the small number of issue position questions that are asked consistently over time. 16 in the 1970s, though that gap shrunk by the 1980s and reversed in the 2000s.16 Possibly as a result of the clearer ideological choices offered by the two parties as a result of elite polarization, women and men’s ideological self-labels became steadily more distinct over this period. Consistent with Norrander and Wilcox (2008), the gender gap in ideology more than doubles between the 1970s and the 1990s, from 0.010 (std. err. = 0.004) to 0.026 (std. err. = 0.004). Finally, because some scholars have argued that growing gender differences in economic vulnerability were a cause of the partisan gender gap’s emergence, columns 9 and 10 look at gender differences in perceptions of their personal finances in the current year and expectations for next year. There are large gender differences in all three decades, but only ambiguous evidence that the differences are growing over time. For both current and prospective personal financial perceptions, the gender gap grew between the 1970s and 1980s, but neither increase was statistically significant. There was essentially no change between the 1980s and 2000s. Overall, Table 3 indicates that there are substantial policy preference differences that predate the emergence of the partisan gender gap. These preference differences largely hold steady throughout most of the partisan gender gap’s emergence, then become somewhat larger in the 1990s. The pre-existing preference gender gaps are largely (though not entirely) among issues less directly associated with women’s rights: social welfare spending, foreign relations, and racial policy. The only issue area in which gender differences significantly increased between the 1970s and 1980s was racial policy, where the difference modestly increased by .013 (std. err. = .008). There is a significant negative interaction between being female and the post-1980 indicator when gender role attitudes are the dependent variable, but that reflects a reverse gender gap in the 1970s on this topic largely disappearing by the 1980s. 5.4 The Gradual Emergence of the Partisan Gender Gap Our fourth hypothesis is that, because elite party polarization happened slowly over several decades and party identification is a sturdy attribute, the partisan gender gap should grow grad16 Questions on the ERA were not asked in the ANES after 1980. 17 ually over these decades. Existing studies using ANES data give varying answers as to when, and how suddenly or gradually, the gender gap first emerged.17 However, they do agree several broad points. In the ANES, the genders start moving toward the modern pattern in the 1960s, with the party identification gap first reaching statistical significance in the 1970s, and becoming consistently statistically significant in every survey by the late 1980s (e.g., Miller, 1988; Kaufmann and Petrocik, 1999; Norrander, 1997, 1999). This is the best picture surveys like the ANES can give. It is a picture in which the long-term trend is clear with a lot of variation from survey-to-survey. Given the sample size, it is not possible to tell whether a number of small partisan gender gaps in the ANES in the 1960s, 1970s and 1980s is evidence that a partisan gender gap is present in the population and whether often large yearto-year fluctuations reflect real changes in the partisan gender gap or just sampling variation. Consequentially, the literature often faces a tension between defining the partisan gender gap in terms of substantive and statistical significance. A focus on statistical significance risks missing the partisan gender gap’s emergence, as a modest partisan gender gap will not be statistically distinguishable from zero in a survey with the ANES’ s sample size. But a focus on point estimates risks overfitting to explain sampling error. In contrast, our pooled Gallup dataset contains surveys that are frequent and numerous enough, and thus a sufficiently large overall sample size, that we can better identify when the partisan gender gap first emerged and its size over time. The smaller standard errors on our estimate of the gap allow us to focus more on its substantive magnitude, while still remaining cognizant of the potential for sampling error. We can (and do) move beyond previous work by reporting convincing evidence on both the statistical significance and the substantive magnitude of the gender gap over time. To analyze the partisan gender gap over time with the Gallup dataset, we construct the sampleweighted average partisanship of men and women, respectively, within each survey s in our sam17 For example, Kaufmann and Petrocik (1999) identify 1964 as the origin of the partisan gender gap because a higher percentage of women have identified as Democrats than men in presidential ANES surveys since then. However, Norrander (1999) notes that this is partially an artifact of men being more likely to initial identify as an Independent than women; she shows that throughout much of the 1970s women were more likely to identify both as Democrats and Republicans than men. When leaners are classified as partisans, there is a statistically significant partisan gender gap in the ANES of about four percentage points that first appears in the 1972 and 1974, largely disappears in 1976 and 1978, only to reemerge again in 1980 (Norrander, 1997). 18 ple.18 We take a non-parametric approach in which each gender’s partisanship at time t is estimated by taking a weighted average of surveys that occur in close proximity to time t. A key consideration when constructing these weighted averages is determining how much weight is given to a specific survey s based on the proximity of the date of the survey, labeled ts , to time t. We define ts as the midpoint of when survey s was in the field. Based on the results of a leave-one-out cross-validation procedure, we use a Epanechnikov kernel function with a bandwidth of 100 days to construct these weighted averages throughout our analysis.19 The top panel of Figure 2 presents the evolution of partisanship by gender from 1953 through 2012. Several trends stand out from this broad overview. As in other wealthy countries in the 1950s, women were slightly more likely to identify with the more conservative party (Ladd, 1997). However, the partisanship of men and women was quite similar from the early 1960s through the mid 1970s. Although both men and women became more Republican from the late 1970s to the early 1990s, men did so at a faster rate. As a result, a substantial partisan gender gap emerged over this time period that remains in place through 2012. While both men and women are more Republican on phone surveys than on in-person surveys, the partisan gender gap is slightly larger on phone surveys than on in-person surveys conducted over the same time period. The bottom panel of Figure 2 shows that there wasn’t much of a partisan gender gap before the late 1970s. The most rapid change in the partisan gender gap occurred between 1977 and 1980, when the partisan gender gap went from being nearly zero to about 4 p.p. This contrasts with Box-Steffensmeier, De Boef and Lin (2004), who show a reverse partisan gender gap in CBS/New York Times polls in early 1979, followed by a rapid increase in 1979 and 1980 in the difference in the percentage of women relative to men who identify as Democrats. From 1980 until 1997, the partisan gender gap oscillated between about 3 p.p. to 6 p.p. on in-person Gallup surveys. It also consistently remained significantly different than zero, with the exception of a couple of instances where relatively less data caused the standard errors to spike. During the 1990s, the partisan gender gap was about 2 p.p. higher on phone surveys than on in-person surveys. The 18 19 Figure A.7 in the Appendix plots both of these quantities for each s in our sample. See Section 7.2 in the Appendix for more details. 19 partisan gender gap remained about 7 p.p. on phone surveys in the 2000s.20 Figure A.5 in the Appendix compares the estimated partisan gender gap in the Gallup data and every ANES and GSS survey during this period. When viewed together, all three series appear to follow the same trend. The partisan gender gap estimate from in-person Gallup surveys is within the ANES 95% confidence interval 20 out of 22 times and the GSS 95% confidence interval 19 out of 21 times. Likewise, the partisan gender gap estimate from phone Gallup surveys is within the ANES 95% confidence interval 8 out of 9 times and the GSS confidence interval 13 of 14 times. This is roughly the proportion of the Gallup estimates that would be expected to fall outside the 95% confidence intervals due to sampling error. This reassuringly suggests that, despite the differences in question wording, all three surveys capture a similar construct. The main difference is that the smaller sample sizes in the ANES and GSS produce much larger confidence intervals, which often made them unable to detect a significant partisan gender gap in years where it was present. For example, ANES and Gallup generate nearly identical point estimates of the partisan gender gap in late 1980s, but we can only reject the null of no partisan gender gap using Gallup data. An implication of Figure A.5 is that the development of the partisan gender gap is a smoother process than one might conclude from the ANES or GSS. Several surges and swoons in the gap’s size in the ANES or GSS, which one might be tempted to imbue with political importance if one considered these surveys alone, now appear to be mere sampling variation around the gradual trend. Whether caused by partisan sorting or some other mechanism, it is possible that the partisan gender gap’s growth was driven by specific political events. Looking at racial preferences and using ANES data, Carmines and Stimson (1989, ch. 6) argue that 1964 was a “critical moment” that triggered a lot of immediate mass-level party sorting, followed by slower party sorting in the years after. Similarly, Kaufmann and Petrocik (1999) identify the 1964 and 1980 presidential campaigns as instances where the Republican Party’s positions on social welfare policy moved substantially 20 As we noted earlier, leaners are treated as Independents because the question about which party Independents lean toward was not asked in all surveys. Section 7.5 in the Appendix examines how results change if we classify Independent leaners as partisans in the subset of Gallup surveys that followed up by asking Independents about their leanings. Results show that our key findings on when the partisan gender gap emerged and the presence of educational heterogeneity in the partisan gender gap hold when we classify Independent leaners as partisans. 20 .35 Local Weighted Average Partisanship Level Higher Values = More Republican .4 .45 .5 .55 Figure 2: Smoothed Partisanship Level by Gender in Gallup Surveys 01jan1953 01jan1961 01jan1969 01jan1977 01jan1985 01jan1993 01jan2001 01jan2009 01jan1957 01jan1965 01jan1973 01jan1981 01jan1989 01jan1997 01jan2005 01jan2013 Survey Date Males (Phone) Females (In Person) Females (Phone) 0 −.05 −.1 Women’s Minus Men’s Partisanship Level .05 Males (In Person) 01jan1953 01jan1961 01jan1969 01jan1977 01jan1985 01jan1993 01jan2001 01jan2009 01jan1957 01jan1965 01jan1973 01jan1981 01jan1989 01jan1997 01jan2005 01jan2013 Survey Date Gallup In Person Gallup In Person 95% CI Gallup Phone Gallup Phone 95% CI 21 to the right, raising the salience of those positions and thus causing the gap’s growth. However, Figure A.2 in the Appendix illustrates that both the size and slope of partisan gender gap appear to be very similar before and after these campaigns. In Section 7.3 in the Appendix, we also apply a formal statistical test and find no evidence that either the level or the slope of the partisan gender gap changed before or after the 1964 and 1980 presidential campaigns. The evolution of the partisan gender gap appears to be similar before and after 11 of 13 presidential campaigns between 1960 and 2008, with the exceptions being 1976 and 1996. Thus, our large Gallup dataset reveals that the sorting leading to the partisan gender gap was a slow process of secular realignment (Key, 1959). In Carmines and Stimson’s (1989, 139) typology, we find that the sorting process follows more of a “secular realignment” than a “dynamic growth” pattern.21 In our view, the partisan gender gap’s gradual emergence is a natural consequence of voters responding to the policy polarization of the parties. Most of this elite polarization happened gradually. The 1964 and 1980 presidential elections and other prominent events did sometimes send unusually strong signals of growing elite polarization. Yet because party identification is a very sturdy disposition, new information (even if delivered suddenly) only slowly led to changes in basic party allegiances (see Green, Palmquist and Schickler, 2002). 5.5 Politically Awareness and the Partisan Gender Gap We have found that the more politically aware notice party polarization and sort themselves ideologically into parties earlier and more thoroughly than the less politically aware. Thus, our fifth and final hypothesis is that this leads the partisan gender gap to emerge earlier and remained larger among the more politically aware. We can measure general political awareness in the Gallup data with education level. We restrict our analysis to in-person surveys in this subsection to ensure the results aren’t driven by changes in survey mode. The top row of Figure 3 compares gender differences in the partisanship of college graduates (left) to non-college graduates (right). The slow and steady growth in the aggregate partisan 21 Of course, in addition to using a larger dataset, we are examining a different question than Carmines and Stimson (1989) did. They looked only at sorting on racial issues, while our focus is sorting on a range of issues that produced the partisan gender gap. 22 gender gap displayed in Figure 2 masks large differences across education levels. Among college graduates, in the 1950s, male and female partisanship was similar. Yet men were more resistant than women to the pro-Democratic macropartisanship trends of the 1960s and 1970s, and women were almost entirely unmoved by the pro-Republican macropartisanship trend in the 1980s. For those without a college degree, there is no sign of a partisan gender gap before 1980. During the 1980s, both sexes were influenced by the overall pro-Republican macropartisanship trend. But men embraced it more, creating a significant partisan gender gap, albeit one that was still smaller than among college graduates. The left graph on the bottom row of Figure 3 compares the size of the partisan gender gap for college graduates and non-college graduates. The right graph on the bottom row of Figure 3 plots the difference in the size of the partisan gender group between these two groups. The solid black line shows the difference in the size of the estimated gap, with the dotted lines bounding the 95% confidence interval on that estimate. We observe that the partisan gender gap was significantly larger among college graduates than among non-college graduates from the early 1970s until the dataset ends in mid 1990s.22 While we contend that the partisan gender gap is larger among college graduates because of their greater awareness of elite polarization, there are many of variables, such as race, age, region, and economic circumstance, that also associate with being a college graduate. It is possible that one of these other variables drives the relationship between education and the size of the partisan gender gap.23 To control for the effect of a number of these other variables that associate with education, we estimate Equation 1. Our key parameter of interest is θ, which measures the difference in size of the partisan gender gap among those who have and have not graduated from college. We control for a vector of other characteristics that may influence the partisanship of respondent i on survey s, labeled Xs,i , as well as the interaction between these covariates and a female indicator, labeled Females,i and a college-graduate indicator, Colleges,i . Finally, we include 22 Figure A.6 in the Appendix shows that we generally cannot reject the null of no difference in the size of the partisan gender group among college graduates and non-college graduates in the ANES or GSS. However, the Gallup point estimate is within the 95% confidence interval of every ANES and GSS survey except the 1966 ANES. 23 Tables A.13-A.16 in the Appendix show the bivariate relationship between the size of the partisan gender gap and all of our demographic variables. 23 01jan1961 01jan1965 01jan1969 01jan1977 01jan1985 01jan1993 01jan1973 01jan1981 01jan1989 01jan1997 Survey Date Male College Graduates Non−College Graduate 95% CI Non−College Graduate Locally Weighted Average (c) Partisan Gender Gap by Education Level College Graduate 95% CI College Graduate Locally Weighted Average 01jan1953 01jan1961 01jan1969 01jan1977 01jan1985 01jan1993 01jan1957 01jan1965 01jan1973 01jan1981 01jan1989 01jan1997 Survey Date (a) College Graduates Female College Graduates 01jan1953 01jan1957 01jan1961 01jan1969 (b) Non-College Graduates Male Non−College Graduates 01jan1977 01jan1985 01jan1993 01jan1973 01jan1981 01jan1989 01jan1997 Survey Date Female Non−College Graduates 01jan1965 95% CI (d) Difference in Partisan Gender Gap by Education Level Locally Weighted Average 01jan1953 01jan1961 01jan1969 01jan1977 01jan1985 01jan1993 01jan1957 01jan1965 01jan1973 01jan1981 01jan1989 01jan1997 Survey Date 01jan1957 Figure 3: Partisanship by Gender and Education (Gallup) 01jan1953 .6 −.24−.22 −.2 −.18−.16−.14−.12 −.1 −.08−.06−.04−.02 0 .02 .04 .06 .08 .1 .12 Average Partisanship Level Higher Values = More Republican .45 .5 .55 Difference in Women’s Minus Men’s Partisanship Level .4 .5 Average Partisanship Level Higher Values = More Republican .4 .45 Difference in Women’s Minus Men’s Partisanship Level College Graduates and Non−College Graduates .35 .1 .05 0 −.05 −.1 −.15 −.2 24 survey fixed effects, labeled αs , to account for features of the political environment that affect the overall partisanship of the population at time ts . If θ is capturing the effect of political awareness, we expect our estimate of θ to be relatively unaffected by the inclusion of these demographic variables in our regression. P rtnshps,i = αs + βXs,i + (γ + δXs,i )Females,i + (λ + κXs,i )Colleges,i + θFemales,i Colleges,i + s,i (1) Table 4 shows how estimates of educational differences in the partisan gender gap vary over time. Each cell entry presents an estimate of θ from a separate regression over the specified time period and including the specified controls. To give each regression more statistical power to accommodate the controls, we group the data into 4-year periods. We also face a trade-off between adding more controls and not discarding data. Because Gallup did not always ask all of the demographics that we would like to use, we are forced to drop more surveys as we increase the number of controls. The regressions reported in column 1 include all surveys. In columns 2 and 3, we use all surveys that contain a set of “baseline controls” that are included in nearly every Gallup poll: race, decade of birth, and region. Columns 4-6 also control for income, and drop surveys where income was not asked, such as all surveys prior to the late 1950s. Finally, columns 7-9 also include marital and employment status, which requires dropping surveys prior to the late 1970s, when Gallup started asking these questions regularly. We generally find that educational differences in the size of the partisan gender gap attenuate only slightly when controls are included. Comparing columns 2 and 3 shows that educational differences in the partisan gender gap decline only slightly when we control for the respondent’s race, region of residence, and decade of birth. Comparing columns 5 and 6 shows that adding income as an additional control has little effect on the point estimates. Likewise, comparing columns 8 and 9 shows that our estimates of θ are roughly the same when we control for employment and marital status. The robustness of our θ estimates to controls is consistent with our claim that college graduation status proxies for political awareness, rather than other attributes like income 25 Table 4: Effects of Controls on Educational Heterogeneity in the Partisan Gender Gap (Gallup) Controls: Baseline Controls + Income + Employment & Marital Status Year of Survey: 1953 - 1956 1957 - 1960 1961 - 1964 1965 - 1968 1969 - 1972 1973 - 1976 1977 - 1980 1981 - 1984 1985 - 1988 1989 - 1992 1993 - 1996 (1) (2) (3) (4) (5) (6) (7) (8) (9) No No No No Yes No No No Yes No Yes Yes No No Yes No Yes Yes No No No No No No No No Yes 0.104 (0.041) 0.017 (0.051) N = 5045 -0.016 (0.009) N = 134181 -0.041 (0.009) N = 132281 -0.015 (0.008) N = 110459 -0.033 (0.007) N = 126698 -0.038 (0.007) N = 146539 -0.047 (0.007) N = 117532 -0.044 (0.013) N = 30857 -0.047 (0.012) N = 35204 -0.046 (0.010) N = 44083 0.025 (0.053) 0.000 (0.010) N = 106871 -0.014 (0.010) N = 107234 -0.011 (0.009) N = 147305 -0.040 (0.008) N = 141313 -0.021 (0.008) N = 113332 -0.044 (0.007) N = 135406 -0.045 (0.007) N = 150859 -0.053 (0.007) N = 122632 -0.058 (0.013) N = 31354 -0.048 (0.012) N = 36024 -0.048 (0.010) N = 45246 0.001 0.006 (0.010) (0.009) N = 104018 -0.015 -0.014 (0.010) (0.010) N = 98816 -0.022 -0.015 (0.008) (0.009) N = 138148 -0.035 -0.038 (0.007) (0.008) N = 139290 -0.019 -0.015 (0.007) (0.008) N = 112205 -0.040 -0.034 (0.006) (0.007) N = 132560 -0.047 -0.038 (0.006) (0.007) N = 149332 -0.056 -0.046 (0.006) (0.007) N = 120517 -0.067 -0.045 (0.012) (0.013) N = 31267 -0.051 -0.047 (0.011) (0.012) N = 35878 -0.050 -0.046 (0.009) (0.010) N = 44966 -0.023 (0.008) -0.035 (0.008) -0.018 (0.007) -0.038 (0.006) -0.047 (0.006) -0.057 (0.006) -0.065 (0.012) -0.052 (0.011) -0.050 (0.009) -0.013 (0.010) -0.036 (0.009) -0.014 (0.009) -0.031 (0.007) -0.037 (0.007) -0.046 (0.008) -0.049 (0.007) -0.058 (0.006) -0.045 (0.014) -0.065 (0.012) -0.049 (0.013) -0.057 (0.011) -0.043 (0.011) -0.048 (0.009) -0.039 (0.009) N = 70881 -0.049 (0.007) N = 107912 -0.044 (0.013) N = 30245 -0.051 (0.013) N = 32165 -0.044 (0.010) N = 43506 -0.043 (0.010) -0.050 (0.008) -0.043 (0.015) -0.054 (0.014) -0.038 (0.011) Notes: Each cell reports the estimate and standard error of θ when estimating equation 1 over the specified time period with the specified set of controls. Baseline controls are an African-American indicator, decade of birth indicators, and region of residence indicators. Income controls include indicators for being in the top 20th and 50th percentiles of household income in a survey’s sample. Employment and marital status controls include indicators for being married, being employed full time, and being employed part time. Observations are weighted by their sample weight. Robust standard errors are reported in parentheses. 26 or labor market experience.24 An alternative explanation for this pattern is that generational replacement was changing the composition of the college educated. In younger generations, education opportunities grew especially for women, as many colleges became coeducational or made the education women received more equal to that of men. The nature of the college experience was also different for younger generations. Women’s equality became a larger part of the curriculum and socialization students received at many colleges. The Development Theory of Gender Realignment argues that the partisan gender gap results from the socialization women received after the transformation of sex roles, much of which may have happened at college (Inglehart and Norris, 2000). It is reasonable to wonder whether the difference between the college educated and others is simply the result of younger college graduates have more gender party identification differences because of either different socialization or different types of people attending college, rather than the college educated of all generations being more aware of growing elite ideological divisions. We examine this by tracking birth cohorts over time in the Gallup dataset. Figure 4 presents the partisan gender gap over time separately for those born in each decade from the 1880s to the 1960s for college graduates (top panel) and those with less education (bottom panel).25 Except for those born in the 1890s, the partisan gender gap grew on a similar gradual trend among college graduates born in each different decade. Even though those who went to college in the early and the late decades of the twentieth century were selected in very different ways, and had very different college experiences, the partisan gender gap evolved in a similar manner. Among the non-college graduates, changes are a bit more noisy, but again a similar basic trend over time is present among all decade cohorts. Among almost all cohorts, the partisan gender gap is smaller and grows more slowly among the non-college educated than among the college educated.26 Another concern is that even if overall gender differences in issue preferences were relatively 24 Table A.12 in the Appendix shows that including leaners as partisans slightly increases the difference in partisanship between college and non-college graduates in the Gallup data. 25 For those who would like to examine the standard errors on these estimates, Table A.18 in the Appendix provides the data from Figure 4 in table form. 26 Table A.8 extends this analysis through 2012 for people born between the 1920s and 1950s using phone survey data. 27 Figure 4: Partisan Gender Gap by Birth Cohort Across Time (Gallup) −.12 Partisan Gender Gap −.08 −.04 0 .04 (a) College Graduates 1953−1962 1963−1972 1973−1982 Survey Year 1880s 1910s 1940s 1982−1992 1890s 1920s 1950s 1993−1997 1900s 1930s 1960s −.12 Partisan Gender Gap −.08 −.04 0 .04 (b) Non-College Graduates 1953−1962 1963−1972 1973−1982 Survey Year 1880s 1910s 1940s 1890s 1920s 1950s 28 1982−1992 1900s 1930s 1960s 1993−1997 stable in the general population, the issue preferences of men and women who graduated from college could have become increasingly different over time. However, Table A.17 in Appendix suggests that this wasn’t generally the case. On most issue areas there was a similar gender gap in issue preferences among college graduates and non-college graduates. The exception was on gender roles and the ERA. On these issues, college educated men were more conservative than college educated women, while non-college educated men were more liberal than non-college educated women. However, this difference shrunk in the 1980s while the partisan gender gap expanded, making this unlikely to be what caused the partisan gender gap to be largest among college graduates. In summary, the difference in the partisan gender gap between college graduates and those with less education does not appear to be driven by generational replacement. Instead, the evidence is more consistent with our explanation. Namely, that the gender gap emerged earlier and is consistently larger among the college educated of all cohorts because they were more likely to be aware of the growing elite ideological polarization of the parties, understand its significance, and sort into a party that matched their policy preferences. 6 Conclusion While the modern partisan gender gap has attracted a lot of attention, scholars do not agree why it emerged. Our explanation is the following. Since the dawn of modern polling, men have consistently held more conservative views than women on social welfare, foreign policy, and racial issues. Starting in the 1960s, party elites increasingly ideologically diverged. People gradually perceived this ideological divergence and, in response, adjusted their party identification to sort themselves into parties that better matched their preferences. Because of their pre-existing preference differences in these areas, this sorting caused men to move disproportionately toward the Republicans and women disproportionately toward the Democrats. This paper tests five hypotheses implied by this theory and generally finds the evidence to be consistent with all five. An important limitation of our explanation is that ideological sorting does not explain anywhere 29 close to all the variation in macropartisanship over time. There are powerful overall trends that influence all genders and education levels to some extent.27 However, those overall trends don’t preclude the secular gender divergence or our explanation of it. Also, while our theory is that ideological party sorting explains the gradual growth of the partisan gender gap from the 1960s to the 1990s, it does not explain all historical variation in it. For instance, our theory does not explain why there was a small reverse partisan gender gap in the 1950s. Ideological sorting caused much of the secular growth in the partisan gender gap over these three decades, but we don’t claim that it is the only thing influencing it. How does our evidence fit with the literature’s other explanations of the partisan gender gap’s emergence? Much of the evidence that we uncover is inconsistent with the economic vulnerability explanation. The partisan gender gap first emerged because women with the most human capital became more Democratic in the 1970s. These women are, on average, at less financial risk from macroeconomic downturns, divorce, or other hardships. Also, we find that gender differences in social welfare attitudes and personal economic assessments among college graduates are indistinguishable from those among non-college graduates, even though the gender gap is substantially larger among the former group. In addition, while gender differences in social welfare preferences were substantial, differences on foreign relations and racial issues were similarly-sized. Finally, gender differences in assessments of personal economic well-being remained constant between the 1970s and 1980s. All of these findings are inconsistent with at least some interpretations of the theory that increases in women’s relative economic vulnerability are the main driver of the partisan gender gap’s emergence. Our evidence is at best only partially consistent with the Development Theory of Gender Realignment and related explanations. While the partisan gender gap developed earlier—and is consistently larger—among those with more education, this pattern does not appear to be caused by more-educated women adopting more liberal views on post-materialist social issues. The previous literature and our analysis indicate that gender differences on social issues were 27 See work like Erikson, MacKuen and Stimson (2002) for analyses of the cause of overall trends in macropartisanship. 30 small to nonexistent and did not grow during the time when the partisan gender gap emerged. Also inconsistent with this explanation, the emergence of the partisan gender gap appears not to be the product of generational replacement and/or changes in how young women are socialized. We find similar patterns in all birth cohorts, even though the socialization provided by colleges changed dramatically over time. The contention by some activists in the early 1980s that the right turn in the policy positions of the national Republican Party created the partisan gender gap also has substantial truth to it. However, it neglects the fact that this right turn was a continuation of the gradual ideological polarization of the two parties that began in the 1960s and 1970s. As the polarization went on longer and grew in size, more people noticed it and the partisan gender gap became larger and somewhat more widespread. Also, it was elite polarization and mass-level sorting based largely on issues less directly connected with gender that were the main drivers of this phenomenon because those are the issues with large gender preference differences. Finally, our analysis helps to bridge the divide between the literature on the partisan gender gap and the most well established general theories of public opinion and party identification. One of the most prominent scholarly claims about American public opinion is that people with different levels of political awareness comprehend politics and respond to new political information differently (e.g., Converse, 1964; Zaller, 1992). Another major claim is that, because party identification is a social identity, it changes gradually, even when important new information is received (e.g., Campbell et al., 1980; Green, Palmquist and Schickler, 2002). The fact that party identification gradually responded to party polarization, and did so differently depending on political awareness, is consistent with these two seminal arguments. New theories of mass political behavior are not required to explain the development of the modern partisan gender gap. Rather, its formation is an understandable consequence of men’s and women’s divergent policy preferences, the polarization of the party system, and the differences in how those with different amounts of political awareness learned about and responded to this changing political landscape. 31