jost_j_t_kay_a_c_thorisdottir_n_social_and_psychological_bas

that “a case can be made for saying that variations in experiential Openness are the major psychological determinant of political polarities” (p. 325, emphasis in original).

The conscientiousness personality dimension is associated with being organized, self-disciplined, orderly, efficient, dependable, responsible, hardworking, persistent, and likely to engage in rule-following behavior. These traits are reminiscent of Silvan Tomkins’ (1963) depiction of the “normative orientation,” which he associated with conservative and right-wing personality styles in general. Although fewer studies have examined conscientiousness than openness, some evidence suggests that it, too, is related to political orientation (e.g., Carney et al., 2008). Using both short and long versions of Big Five scales, Gosling and colleagues (2003) obtained weak but significant positive correlations between conscientiousness and self-reported conservatism (with r values ranging from .06 to .11) and weak but significant negative correlations between conscientiousness and self-reported liberalism (ranging from –.08 to –.13). Similar results were obtained in a study by Caprara and colleagues (1999) with regard to voting behavior in the 1994 Italian national election; people who scored higher on conscientiousness were slightly more likely to support the center-right party than the center-left party. Caprara, Barbaranelli, Consiglio, Picconi, and Zimbardo (2003) administered personality questionnaires to Italian members of regional, national, and European parliaments and found that center-right politicians scored higher on conscientiousness than did center-left politicians.

Caprara and colleagues (1999, 2003) also found that center-right politicians and voters scored higher on energy or (extraversion) than did center-left politicians and voters. It is conceivable that such differences reflect varying degrees of self-confidence, especially in light of the electoral success of the rightist National Alliance party headed by Silvio Berlusconi and Gianfranco Fini in Italy over the last several years. In the American context, Gosling and colleagues (2003) obtained no association between extraversion and political orientation. Insofar as this personality dimension captures how active, energetic, dynamic, enthusiastic, enterprising, and happy a person seems to be, it may be relevant to political success, although it may not be consistently related to leftversus right-wing differences in personality across time and place. In any case, this is an empirical question.

No consistent evidence indicates that the other two Big Five dimensions of personality, agreeableness and neuroticism (or emotional instability), are associated with political orientation per se. However, Gosling and colleagues (2003) did find that agreeableness was negatively correlated with scores on Pratto, Sidanius, Stallworth, and Malle’s (1994) social dominance orientation (SDO), which measures a preference for group-based inequality that tends

to be associated with other right-wing attitudes (see Chapter 12). This suggests that agreeableness, which involves warmth, kindness, trust, generosity, sympathy, and unselfishness, might be more prevalent among liberals than conservatives. Evidence also indicates that voters’ perceptions of candidates are strongly affected by their perceptions of candidates’ levels of agreeableness (as well as their levels of extraversion), regardless of whether those candidates are leftor right-wing (Caprara et al., 2002).

Neuroticism refers to emotional instability and is associated with anxiety, moodiness, and an inability to cope with stressful situations. Jost and colleagues (2003a) found that conservatives were more likely than liberals to worry about potential losses and to be sensitive to threat-related stimuli, but the effect size for this relation was considerably weaker when fear of threat and loss was measured in terms of neuroticism. Studies by Caprara and colleagues (1999, 2003) in Italy yielded no evidence that neuroticism was associated with preferences for center-left versus center-right political parties and candidates, but it may be relevant that third-party candidates were excluded from consideration in these studies. Effects of neuroticism may be more likely to emerge in the context of a wider range of political views, including those of third-party candidates. To assess this possibility, the present study included data on the percentage of votes cast for thirdparty candidates (i.e., Reform and Green party) as well as for Democratic and Republican candidates.

By integrating previous research on personality and political orientation with recent observations concerning the geographic clustering of political attitudes, a number of novel research questions can be generated and addressed. For example, are residents of politically liberal states higher, on average, in openness than residents of conservative states? Are residents of conservative states higher in conscientiousness than residents of liberal states? And, most critically, do statewide differences in personality explain voting patterns above and beyond standard sociological and political variables?

OVERVIEW OF THE PRESENT RESEARCH

The purpose of the present research is to determine whether regional differences in political orientation reflect differences in the personalities of the individuals living in a particular region. To investigate this issue, we examined connections between state-level personality scores on each of the Big Five personality dimensions and the percentage of votes cast for Democratic, Republican, and third-party candidates in the 1996, 2000, and 2004 Presidential elections.

Our predictions concerning the correspondence between state-level personality scores and voting patterns were based largely on extrapolations from research undertaken at the individual level of analysis (e.g., Caprara et al., 1999; Gosling et al., 2003; Jost et al., 2003a, 2003b; McCrae, 1996). As summarized earlier, previous studies have identified clear links between openness and conscientiousness and political orientation. The evidence connecting extraversion, agreeableness, and neuroticism to political orientation is less clear. Therefore, we expected that state-level openness and conscientiousness scores would significantly predict state-level voting patterns, but we made no explicit predictions about the other three personality dimensions. More specifically, we hypothesized that states with populations that are relatively high in openness and low in conscientiousness would be more likely to vote for Democratic candidates (Bill Clinton in 1996, Al Gore in 2000, and John Kerry in 2004) and less likely to vote for Republican candidates (Bob Dole in 1996 and George W. Bush in 2000 and 2004). Given the dearth of available research on the psychological correlates of preferences for independent and third-party candidates, we did not make explicit predictions about how state-level personality would be related to preferences for Ross Perot in 1996 and Ralph Nader in 2000 and 2004. However, given that Perot seemed to appeal most to right-of-center voters, whereas Nader mainly appealed to left- of-center voters, we expected the patterns of personality correlates to match those of the Republican and Democratic candidates, respectively.

Although we expect statewide personality differences to account for significant proportions of variance in voting patterns, it is reasonable to assume that the effects will be weaker after adjusting for standard sociodemographic characteristics and past election results. Consistent with the research literatures on electoral coalitions, migration patterns, political regionalism, and social influence (e.g., Axelrod, 1972, 1986; Bourgeois & Bowen, 2001; Brooks & Manza, 1997; Cohen, 2001; Erikson, Wright, & McIver, 1993; Heppen, 2003; Hero, 1998; Huckfeldt & Sprague, 1995; Latané, 1981; Vandello & Cohen, 1999), it is to be expected that variables like education, income, social class, ethnic diversity, proportion of state population residing in large cities, and past voting behavior could be related to both the personalities and political preferences of individuals living in a given region. To assess whether the relations between state-level personality and political preferences persist even after adjusting for such factors, we examined three models on regional voting patterns. First, the state-level personality model included the Big Five personality estimates for each state. Second, the sociodemographic model included for each state the median family income, percentage of residents with at least a college degree, percentages of whiteand blue-collar workers, percentage of Black residents, and proportion of

a state’s population living in cities with 1 million or more residents (U.S. Census, 2000). Third, the political model included the change in state-wide voting patterns from the two elections preceding the one under investigation (e.g., the difference in the Democratic vote from 1988 to 1992 to predict voting outcomes in 1996; Leip, 2005).1 Together, these data permitted analyses of the independent and shared contributions of each model on statelevel voting patterns, thereby enabling us to determine whether state-level personality accounted for unique variance in political preferences when variables from the other models were held constant.

Testing the predictions required that we obtain reliable and representative personality estimates for each state. We therefore needed a methodology that would (a) provide a sufficiently large sample to investigate our hypotheses,

(b) enable us to collect equivalent data from individual respondents around the country, and (c) provide a diverse sample of respondents that would be reasonably representative of the population of the United States. To achieve these aims, we decided to use the Internet as the vehicle for collecting personality data. Research on web-based studies indicates that Internet users may not be perfectly representative of the general population (Lebo, 2000; Lenhart, 2000), but Internet-based samples are much more diverse than convenience samples commonly used in social science research (Birnbaum, 2004; Gosling, Vazire, Srivastava, & John, 2004). Furthermore, researchers typically obtain very similar results across Internet and non-Internet samples, especially with regard to personality variables (e.g., Srivastava, John, Gosling, & Potter, 2003).

The Internet is still a rather novel methodological tool, and no other studies have used this methodology to examine regional differences in personality. Therefore, we were concerned that the findings might not be reliable. Our concern was particularly acute for the early analyses because, in

1 The political variable chosen to represent the impact of state-wide political behavior was a measure of change in voting patterns, or the difference in the percentages of votes from the two previous elections. This variable sheds light on whether change in state-wide Democratic votes between 1988 and 1992, for example, influenced votes for the Democratic candidate in 1996. A political variable like this is more illuminating than the percentage of votes in the previous election because (a) state-wide presidential voting patterns are very highly correlated across elections, and (b) a political change variable measures variations in the political behaviors of state residents. Thus, if changes in previous voting patterns predict the percentage of state-wide votes, this would suggest that there is a shift in the overall political orientation of state residents.

1999, the Internet was not nearly as widely used as it is today (Lebo, 2000; Lenhart, 2000). To determine whether the findings were generalizable, we gathered two independent personality samples at two different time periods and sought to replicate the findings across samples. Data for Sample 1 were collected from January 1999 to December 2000, and data for Sample 2 were collected from January 2003 to December 2004. Big Five state-personality estimates from Sample 1 were used to predict voting patterns in the 1996 and 2000 elections, and estimates from Sample 2 were used to predict voting patterns in the 2004 election. This approach enabled us to determine whether the effects of state-level personality on voting patterns would replicate across the independent samples and, simultaneously, to evaluate the viability of our methodology.

Cross-level Analysis Considerations

When working with variables that can be measured at multiple levels of analysis (e.g., at the levels of individuals and states), researchers may be tempted to generalize findings from one level to another. However, although findings at one level can match findings at another level, the different levels are logically independent, so that generalizations across levels are not always warranted. The well-known ecological fallacy refers to one class of such mistakes, in which aggregate-level findings are incorrectly generalized to individual-level phenomena (Duncan, Cuzzort, & Duncan, 1961; Robinson, 1950; Shively, 1969). For example, Robinson (1950) showed that the ecological correlation between the percentage of foreign-born state residents and the percentage of illiterate state residents was –.53, but that the individual correlation between foreign-born status and illiteracy was .12. In this example, the ecological correlation suggests that illiteracy rates are higher in states where there are fewer foreign-born residents than native-born residents; however, it does not follow that illiteracy is higher among native-born individuals than foreign-born individuals. Indeed, the individual correlation reveals just the opposite. A similar, albeit less common error is the individualistic fallacy, in which findings from individual-level analyses are assumed to generalize to aggregate-level analyses; for these reasons, it cannot be assumed that the relationships between personality and political ideology identified at the individual level will generalize to state-level analyses. Of course, findings at one level may in fact generalize to another level, and they often do; the ecological and individualistic fallacies merely highlight the fact that the two levels are logically independent.

The decision to rely on ecological or individual levels of analysis rests primarily on how researchers intend to use the variables and the level of analysis they are most concerned with describing (Shively, 1969). The pres-

ent research is concerned with the connection between regional personality and political orientation at the state level. Therefore, we used aggregate-level data to examine the relationship between regional personality and political orientation at the aggregate level. We merely used previous individual-level findings to identify promising candidate traits for our analyses.

Method

Procedure. The personality data were part of an Internet personality project, a personality study of volunteers assessed over the World Wide Web (for details see Gosling et al., 2004; Srivastava et al., 2003). The website used is a noncommercial, advertisement-free website containing a variety of personality measures. Potential respondents could find out about the site through several channels, including search engines or unsolicited links on other websites. For both samples, respondents volunteered to participate in the study by “clicking” on the personality test icon and were then presented with a series of questions about their personalities, demographics, and state of residence. After responding to each item and submitting their responses, participants were presented with a customized personality evaluation based on their responses to all the items.

Participants. Only participants indicating that they lived in the United States were included in the analyses. Data for Sample 1 were available for 238,709 participants (58% female); the average age was 24.1 years (SD = 9.6 years). Of those who disclosed information about their ethnicity, 11,580 (4.9%) respondents were Asian; 7,651 (3.2%) were African American; 5,073 (2.1%) were Latino; 3,974 (1.7%) were Middle Eastern; 201,148 (84.8%) were White; and 7,700 (3.2%) indicated “Other.” Data for Sample 2 were available for 273,685 participants (59% female); the average age was 24.9 years (SD = 10.7 years). Of those who disclosed their ethnicity, 13,018 (4.8%) respondents were Asian; 15,667 (5.8%) were African American; 16,025 (5.9%) were Latino; 2,711 (1.0%) were Middle Eastern; 207,849 (76.8%) were White; and 15,315 (5.7%) indicated “Other.”

Measures

Respondent location. Given our interest in the relations between state-level personality and voting patterns, respondents in both samples were asked to indicate the state in which they currently reside. Respondents could choose from a list of all 50 states as well as Washington, D.C.

Personality. We used the Big Five Inventory to assess personality (BFI; John & Srivastava, 1999) in both samples. The BFI consists of 44 short statements designed to assess the prototypical traits defining each of the Big

Five dimensions. Using a 5-point Likert-type rating scale with end-points at 1 (Disagree Strongly) and 5 (Agree Strongly), respondents indicated the extent to which they agreed with each statement. The BFI scales have shown a robust factor structure, substantial internal and temporal reliability, and considerable convergent and discriminant validity with other Big Five measures (Benet-Martinez & John, 1998; Gosling et al., 2003; John & Srivastava, 1999).

Voting patterns. To assess regional political preferences, we gathered voting data for each of the 50 states and Washington, D.C., from Dave Leip’s Atlas of U.S. Presidential Elections, an online database consisting of Presidential election results obtained from publications by official election agencies within each state (i.e., Secretary of State offices, State Board of Election offices, Congressional Quarterly, and the U.S. National Archives and Records Administration). For the 1996, 2000, and 2004 Presidential elections, we obtained data on the percentage of votes cast in each state for major party candidates (Clinton, Dole, Gore, and Bush) and third-party candidates (Perot and Nader).

Sociodemographic characteristics. To determine the unique contribution that state-level personality has on voting over and above sociodemographic variables, we gathered relevant data for each state from the U.S. Census Bureau (2000). On the basis of previous research on political regionalism and political coalitions (Agnew, 1987a; Axelrod, 1972, 1986; Brooks & Manza, 1997; Heppen, 2003; Hero, 1998; Shelly, Archer, Davidson, & Brunn, 1996), which has identified several robust sociodemographic predictors of state-wide voting patterns, we obtained data on median family income, percentage of the population 25 years and older with at least a college degree, percentage of working-aged individuals with white-collar jobs, percentage of workingaged individuals with blue-collar jobs, percentage of population that is African American, and proportion of the population living in a city with one million or more residents.

Sample Characteristics, Reliability, and Data Aggregation

Sample characteristics. To ensure that each state was fairly represented in terms of geographic region, we correlated the number of respondents from each state in Sample 1 with the population for each state using data from the U.S. Census Bureau (2000). The number of respondents from each state in our sample was directly proportional to the population of each state, r = .98.

Past research on Internet-based surveys suggests that many ethnic minority groups are under-represented on the Internet (e.g., Lebo, 2000; Lenhart, 2000). To determine whether our sample under-represented particular ethnic groups, we compared the ethnic population of respondents from our

Internet sample with the ethnic population of residents in each of the 50 states. Specifically, we correlated the percentage of respondents from each ethnic group in the Internet samples with the percentage of the population of that group within each state. For example, we correlated the percentage of Asian respondents from each state with the actual percentage of Asians in each state. These analyses indicated that our Internet-based samples were remarkably representative of the population at large. Specifically, the correlations for Samples 1 and 2, respectively, were: .99, .97 for Asians; .57, .93 for African Americans; .96, .96 for Latinos; and .98, .87 for Whites. Thus, with the exception of African Americans in Sample 1, the ethnic composition of our samples matched almost perfectly the ethnic populations in each state.

Scale reliabilities. Reliabilities of the Big Five personality dimensions at the individual and state levels were of particular interest in this study. If there were problems administering the BFI on the web (e.g., random or unreliable responses), the internal reliabilities of the five scales would be low. Moreover, because we were developing personality estimates for each state using the BFI, it was important to determine whether the personality scales were reliable at the state level. We found that the scales were reliable across both levels and samples, with alpha (α) reliabilities very similar to those reported in previous research (see John & Srivastava, 1999). Specifically, the coefficient

αat the individual level for Samples 1 and 2, respectively, were: .79 and .70 for openness; .82 and .82 for conscientiousness; .86 and .85 for extraversion;

.80 and .80 for agreeableness; and .83 and .83 for neuroticism. The coefficient

αat the state level for Samples 1 and 2, respectively, were: .95 and .94 for openness; .86 and .91 for conscientiousness; .88 and .88 for extraversion; .85 and .91 for agreeableness; and .94 and .89 for neuroticism.

Data aggregation. Given that our Internet samples accurately represented each of the 50 states in terms of both overall and ethnic population and that state-level responses to the BFI were reliable, it was reasonable to aggregate participants’ responses to develop personality estimates for each state. For Samples 1 and 2, we scored the five personality scales at the state level using the unit-weighted scale scores of every respondent from each state.

Results

Our primary objectives were to determine: (a) whether state-level personality was related to the percentage of votes cast for each Presidential candidate in each election, and (b) the extent to which state-level personality accounted for voting patterns after adjusting for the effects of sociodemographic and political variables. We addressed these issues in a series of multiple regression analyses that examined the independent and combined effects of three

models (state-level personality, sociodemographic, and political) on the percent of votes cast for each candidate in each election period.2 Because we were concerned with the degree of covariance between state-level personality and political orientation that arises through self-selection and social influence processes, r was selected as the measure of effect size.3 All numerical variables were standardized prior to conducting analyses.

Democratic Candidates

State-level personality model. Analyses of the percentages of Democratic votes cast in the three elections revealed that the state-level personality model shared significant proportions of variance with votes cast for Clinton (in 1996), Gore (in 2000), and Kerry (in 2004) (R = .634, .682, and .718, respectively). As shown in the top portion of Table 13.1, the partial regression coefficients for each of the Big Five personality dimensions were strikingly similar across all three elections. As hypothesized, openness was positively

2 Votes for Democratic candidates are strongly negatively related to votes for Republican candidates. Therefore, we also analyzed the data using two bipolar political-preference variables as the criteria: a Democrat versus Republican variable (excluding the percent of votes cast for third-party candidates) and a third-party versus major parties variable. The results from these analyses were virtually identical in every respect to the results reported here, in which the raw percent of votes cast was used as the primary criteria. We used the raw percentage of votes cast for each candidate as the criteria, instead of the bipolar political preference variables because these analyses are more straightforward to interpret.

3 Although r2 is commonly thought to reflect the percent of shared variance between measured variables, this interpretation is correct only in certain circumstances. Jensen (1980) provided clarification for the difference between r2 and r by pointing out that r2 reflects the percent of variance in one variable that can be predicted by another variable, whereas r reflects the percent of variance common in two variables. Ozer (1985) extended this work by describing two analytic models, the variance decomposition and variance composition models, to delineate the appropriate conditions for using r and r2 as effect size estimates. The variance decomposition model assumes that the predictor variable is contained within the criterion variable, and requires that the predictor variable is corrected for attenuation (which is seldom the case). In this model, r2 represents the amount of variance in the criterion variable that is accounted for by the predictor variable, and should therefore be used as the effect size estimate. In contrast, the variance composition model assumes that a latent variable is responsible for the covariance between measured variables. In this model, r is the appropriate effect size estimate because it represents the covariance of measured variables by a latent variable (or the percent of variance common in the variables measured). For detailed discussions of the uses and misuses of r and r2, see D’Andrade and Dart (1990), Jensen (1971, 1980), and Ozer (1985).

related to the percentage of votes cast for each of the Democratic candidates, and it accounted for the most variance. Conscientiousness was a significant negative predictor of votes cast for each candidate. In addition, extraversion was a positive predictor of the percentage of votes cast for the three Democratic candidates. Although we made no specific predictions about the effects of neuroticism, it significantly predicted the percentage of votes for Kerry (but not for Clinton or Gore).

Sociodemographic model. The sociodemographic model also shared substantial proportions of variance with the percentages of votes cast for

Table 13.1 Multiple regression analyses of the state-level personality, sociodemographic, and political models for the Democratic Presidential candidates in the 1996, 2000, and 2004 elections.