He Said,She Said

Abstract

This paper examines how interpersonal social networks relate to the voting behavior of men and women. We argue that underlying the gender gap in voting is related to social processes that depend on the partisan and sex composition of networks. Analysis of the 2000 American National Election Study identified two ways that sex differences are relevant to network explanations of voting behavior. First, men have more sex homophily in their networks than women. As men are more likely to be Republican than women, this leads to different discussion environments for men and women. Second, men—and not women—are more likely to share the political opinion of women discussants, but only when they are pro-Bush and the remainder of the network is also supportive. The results support a social model of voting behavior that highlights the importance of social factors (in this case sex) other than just partisan differences.

Keywords

social networks voting behavior gender

There appear to be three ways in which social relations contribute to the maintenance of political differences.

First . . . it is necessary to have a social basis for a political interest. It would be difficult in contemporary America, for example, to maintain voting differences by sex, because there are few policy issues persisting over a period of time that affect men and women differently . . .

Second, a necessary condition for the persistence of political differences is their transmission to succeeding generations . . .

Finally, given the origin of a voting difference in one generation and the transmittibility of it to the next, another condition is necessary . . . [m]embers of the social groups involved must be substantially more in contact with one another, socially and physically, than they are with opposing groups. (Berelson, Lazarsfield, & McPhee, 1954, p. 74)

This epigraph from Voting is the classic statement on how social cleavages become politically significant. Although very few scholars directly challenge the Columbia researchers’ theoretical claims, their example of “voting differences by sex” is striking in light of the gender gap in voting that emerged in the last three decades of American politics (Ford, 2006; Howell & Day, 2000; Kaufmann & Petrocik, 1999; Wirls, 1986). Evidence shows that gender differences exist on policy matters (e.g., Box-Steffensmeier, DeBoef, & Lin, 2004) and that there is intergenerational socialization (e.g., Jenkins, 2013), satisfying of the theoretical conditions for a political cleavage. However, it is unclear whether men and women have socially segregated interactions necessary to produce “maintenance of political differences.”

While significantly advancing our understanding of gendered voting, previous research does not often consider the intersection of social networks and sex in explaining behavior (but see Djupe, McClurg, & Sokhey, 2013; Mendez & Osborn, 2010; Osborn & Mendez, 2011). Yet as Berelson and his colleagues recognized, not all social cleavages are politically relevant and among those that are, there is a social mechanism that sustains them. Their perspective—which we share—is that political fault lines are related in large part to the levels of contact within and between groups. As levels of within-group contact increase and between-group contact decrease, the expected political significance of a social cleavage should become more acute.

The problem of applying this argument to the gender gap in voting is that there are ample opportunities for men and women to interact. To be sure, men and women are not on equal ground with respect to the traits that predict political discussion. Men are less likely to identify women as political discussants than women are to identify men (Huckfeldt & Sprague, 1995; Klofstadt, McClurg, & Rolfe, 2009). Women are less inclined to offer political opinions to others—that is, political proselytizing—than are men, except when women candidates run (Hansen, 1997). And men tend to see women as less knowledgeable about politics, potentially downgrading their level of social influence (Huckfeldt, 2001; Mendez & Osborn, 2010; Mondak & Anderson, 2004). Nevertheless, sex does not on the surface appear to be as important in producing social segregation than other factors such as partisanship (Finifter, 1974; MacKuen, 1990), religion (Huckfeldt, Plutzer, & Sprague, 1993), and class (Marsden, 1987). This is because these other differences overlap with social institutions (such as churches and unions) that buttress within-group relations. The one social institution most apt to affect sex homophily in networks—marriage—may work to bring their preferences and behavior together (but see Mendez & Osborn, 2010; Osborn & Mendez, 2011; Stoker & Jennings, 1995).

This leads us to consider in more depth how social interaction is related to vote differences between men and women. We consider three different possibilities. The first is that there are no differences in the networks of men and women, implying that information from networks is unrelated to vote choices. The other two come from looking at the partisan and gender composition of networks and discussion dyads. Our evidence suggests that the “social process” underlying voting depends on two factors—the respondent’s partisanship and his or her network’s sex composition. Based on these results, we argue that the networks surrounding men and women provide them with different cues in the voting process, thus sustaining this important political cleavage in American politics.

How Does A Gender Gap Emerge in a World Not Strongly Segregated by Sex?

How does social interaction potentially affect sex differences in voting, especially in a world in which personal relationships are not strongly segregated by sex? Is it possible that social interaction has nothing to do with sex and vote choice? Are there patterns in discussion networks that get overlooked? Or, are some discussants more or less valued in networks because of their political and sex characteristics? The opening to this paper introduces the possibility that the gender voting gap may demonstrate the irrelevance of social segregation and interaction to the formation of political cleavages. However, that seems somewhat unlikely given the rich literature in political science showing that individual political behavior is strongly rooted in discussion networks (e.g., Rolfe, 2012; Sinclair, 2012). Consequently, exploring the relationship between social interaction and sex differences in voting presents itself as a good place to test hypotheses related to the network model of voting.

The “social calculus of voting” argument is that electoral choices are conditioned on environments created by interpersonal social networks (Huckfeldt, Johnson, & Sprague, 2004). The basic assumption of this model is that social interaction between family, friends, and acquaintances provides opportunities for people to learn about politics from trusted sources. Consequently, these discussions are presumed to provide information that reinforces or challenges a voter’s political predispositions through processes of learning or social conformity. Whatever the mechanism, the argument is that an individual’s votes should over time be more or less congruent with votes of the other networks members. This correspondence should be reflected at two levels of observation. The first is the composition of a person’s discussion network (e.g., the ratio of Republicans to Democrats in the network). Simply stated, we should see that people with majority Republican (or Democratic) networks should be more likely to vote in the same direction as the network. The second level of observation occurs at the level of the discussion dyad (the relationship between one respondent and one discussant), where the level of correspondence reflects characteristics of the discussant and respondent and accounts for the constraint imposed from the remainder of the network (Huckfeldt et al., 2004). Again, the expectation is that members of the dyad should have similar votes, contingent on all other dyads.

There are three different potential relationships between networks and voting behavior. The first possibility is that network processes are unrelated to the gender gap in voting. Support for this idea stems from the apparent weakness of social segregation between the sexes. As men and women have different political interests, it may not be surprising that a gender gap would arise as men and women formed political identities around their interests (Gilligan, 1982), even if they interact with each other. Therefore, even if men and women interact and exchange political information, they may be talking past each other and their interaction is unrelated to voting behavior. If this is the case, we would expect that men and women have similarly composed networks, at least when it comes to sex and partisanship. With respect to dyads, there should be no correspondence between a respondent’s sex, a discussant’s sex, and how the discussant’s partisan leanings impact the respondent’s partisan evaluations.

A second relationship is that the partisan dimensions of the network—but not the sex differences—are related to the gender gap. Most research focusing on social interaction in American politics often focuses on levels of political difference as the link between networks and choice (Klofstadt, Sokhey, & McClurg, 2012). A model focusing on political differences between men and women would accept that they interact with each other, but in a way that does not depend on the sex composition of the network. The expectation is that men and women on average have individual characteristics that produce different partisan preferences that their network partners share. In this scenario, we would expect few sex differences—that is, women associating more with other women, men with other men—but plenty of political segregation—Democrats associating more with Democrats, Republicans with Republicans. This means men are more likely to have Republican networks, women are more likely to have Democratic networks. However, the sex of the participants in the network is the same for men and women and the potential correspondence between the views of a person is unrelated to sex differences between respondents and their discussants.

We argue for a third possibility that a refinement is the original Bereleson et al. argument. This is basically a call to specify more precisely what is meant by the phrase, having a “social basis for a political interest.” We suggest that political differences in social interaction do not alone create cleavages. Instead, the issue is not whether there is social segregation (which is never absolute), but rather how much social segregation is necessary to sustain a politically relevant cleavage and under what conditions. Recognizing this clarification in the original argument creates flexibility in trying to understand how political cleavages work (i.e., why some are stronger than others). In this case, it suggests that for a cleavage to emerge and be sustained, there is an interaction between both social and political characteristics of people in a network. This model raises the possibility that the partisan balance of and influence in a network might also be a function of the sex of people of the network. For the reasons outlined by Berelson and his colleagues, we might expect stark political and social segregation to be necessary for networks to produce a gender voting gap. By examining differences between men and women, we are asking if degrees of social segregation are sufficient for producing sex differences in voting.

To the degree that differences exist among the sexes in terms of their access to professional advancement (Applebaum, Audet, & Miller, 2003), home responsibilities (Mezey, 1978; Sapiro, 1982), and types of organizational involvement (Burns, Schlozman, & Verba, 2001), we expect that this creates differential access to political information that is not just a function of network partisanship but of network sex composition as well. As a consequence, we hypothesize that women who have more female discussion partners are more likely to be in even more pro-Democratic networks; likewise, men with more male partners will be expected to be in more pro-Republican networks (Hypothesis 1).¹ The same logic suggests cross-sex dyads should exhibit different levels of opinion correspondence than same-sex dyads in the context of the entire network. Previous research shows that women discussants are perceived as less politically informed then men discussants. At the same time, Berelson et al. suggest that voting gaps are not driven solely by politics, but also by social group itself. This leads us to our second hypothesis: (a) men’s opinions of candidates will be less related to female’s partisanship relative to the rest of the network; and (b) women’s opinion will be more related to the views of their female partners relative to the rest of the network.

Data

We examine these propositions with data from the 2000 American National Election Study. This version of the American National Election Study includes traditional questions used in the study of voting behavior and information on each respondent’s social network. In order to learn about discussion networks, respondents were asked for up to four names of the people with whom they talk politics.² Respondents then reported how they knew each discussant, how frequently they discussed politics, their perception of the discussant’s political knowledge, and their perception of the discussant’s vote choice.³ We organize these data in two ways. First, we examine the entire reported network of our respondents in order to address arguments about the entire network environment. Individual respondents are the unit of analysis, and the information on discussants is aggregated to represent the characteristics of the respondent’s entire set of named discussants. This enables us to examine the extent to which partisan characteristics of the network (the dependent variable) overlap based on respondent sex and partisanship and network gender composition (the independent variables). Second, we examine network dyads. Here the object is to consider the role of discussant characteristics—namely sex, our main independent variable—on candidate opinions more closely, all while still controlling for the rest of the network. Because this over-represents respondents with larger social networks (i.e., they have more dyads), we employ the procedure used by Huckfeldt et al. (2004) to correct for what they label social autoregression.⁴

Among the drawbacks of our study, two are especially worth noting. First, the absence of interviews for a random sample of named discussants limits our analysis to respondent perceptions of network interaction. Because respondent perceptions of a discussant’s vote preference depend upon both respondent partisanship and the distribution of voter preferences in the rest of the network (Huckfeldt et al., 2004), these data may understate levels of disagreement in networks and reflect reciprocal influence of the respondents on their network. Second, we are unable to account for the influence of factors specific to the 2000 election that might influence the distribution of political preferences across networks or between men and women. In both cases, these limits restrict our ability to determine causality, especially beyond this particular election. This means that we are best able to show, at this point, correspondence between network characteristics, respondent characteristics, and their intersection. Correlation does not, of course, lead to causation. However, it is a necessary condition in any social influence process. We discuss this more in the conclusion.

Sex and Political Homophily in Social Networks

We begin by comparing the political composition of social networks as reported by the men and women in our survey. Table 1 reveals significant differences between men and women in regards to the size of their political networks, the average level of political discussion, the average level of support for Bush, the number of women in a network, and the intimacy in the network. Compared to men, women have fewer network members and, on average, engage in less political discussion with their network members. Support for Bush also varies, with men’s networks showing more support for Bush than women’s networks. There are no differences in the proportion of Gore discussants in male and female networks. And, female respondents are also more likely to report having higher numbers of women in their networks and more intimate networks—meaning more family members—than male respondents. From these data we see that respondent sex is related to the political information contained in each respondent’s network. The distribution of partisan preferences has some similarities for men and women in that the average percentage of Gore voters is the same for men and women. Yet there are still sex and political differences in the networks of men and women.

Table 1.

Characteristics of Full Social Networks by Gender.

Network property	Respondent gender				Difference
	Male		Female
	Mean	SE	Mean	SE
Size	1.72	0.06	1.51	0.05	0.21*
Average political talk (scale 0-3)	1.34	0.04	1.24	0.03	0.11*
Average political knowledge (scale 0-2)	0.81	0.03	0.80	0.02	0.01
Average Gore support (scale 0-1)	0.25	0.01	0.24	0.01	0.01
Average Bush support (scale 0-1)	0.29	0.01	0.24	0.01	0.05*
Average female	0.22	0.01	0.31	0.01	−0.09*
Average network intimacy (scale 0-1)	0.25	0.01	0.32	0.01	−0.07*

p < .05, two-tailed t-test.

The first question we pursue is whether these differences are related to each other and, eventually, views on the candidates. Since women are more likely on average to be Democratic partisans than are males, it is possible that sex sorting—rather than only political sorting—is relevant to the network environments in which our respondents make political choices. To examine this possibility, we need to add respondent and discussant sex to the analysis to see if sex homophily is related to partisan cues embedded in networks. As seen in Table 2, levels of candidate agreement in dyads is related to the sex of both the respondent and the discussant. Fifty-three percent of male respondents with male discussants voted for Bush, indicating the highest level of political agreement. This compares to only 38.7% of male respondents’ with female discussants who voted for Bush. Among female respondents, we see that 47.6% of female discussants voted for Gore in comparison to 36.1% of male discussants. Importantly, women’s male discussants are relatively just as likely to be Republican as are men’s male discussants. The most important differences are attributable to female respondents reporting that male discussants have either “no” preference or support some “other” candidate. Although possible that this reflects respondents’ perception of discussant preferences more so than discussants’ actual preferences, what is perhaps most important is that these are the cues respondents “see” in the network.⁵

Table 2.

Discussant Political Preferences by Respondent and Discussant Gender.

Discussant vote	Female respondents		Male respondents s
Discussant vote	Female discussant	Male discussant	Female discussant	Male discussant
Percent Gore	47.6% (344)	36.1% (250)	48.6% (206)	34.5% (284)
Percent Bush	39.6% (286)	45.2% (313)	38.7% (164)	52.7% (434)
Percent other/none	12.8% (93)	18.6% (129)	12.7% (54)	12.8% (106)
Total	100% (723)	100% (692)	100% (424)	100% (824)

These analyses show that the sex composition of the network overlaps with the partisan cues a respondent is likely to see in his or her network. With more female discussants and male discussants (who are less likely to support Bush) female respondents are typically exposed to more pro-Gore information than pro-Bush information. The political cues men receive are different. In part, this is because their male discussants are more supportive of Bush, but it also reflects the fact that their networks are four times as likely to include other men rather than women. This supports the underlying logic of our hypotheses as the analysis at least makes clear that interpersonal political communication experienced by men and women is different.

This descriptive support for the argument raises two questions for deeper empirical analysis. First, what is the role of respondent characteristics in explaining these differences? Are male and female respondents more important than individual partisanship in structuring these network environments? Second, how important is discussant sex in understanding how the respondent relates his or her preferences? Are men more likely to take into account preferences of other men, or do they treat women’s preferences the same way? These are the questions we take up in the next two sections.

Respondent Characteristics and Their Relationship With Network Homophily

We next ask which individual respondent traits are most relevant to understanding this relationship between sex and partisan composition of the network. This is important to understand because respondents are not simply “subjected” to their network environment, but also play a role in constructing it. Our approach is to regress the percent of a respondent’s network support for Gore and Bush on respondent partisanship, the proportion of women in the network, and an interaction between the two. Individual level controls are added for education, age, union membership, marital status, and respondent sex, the last of which is also interacted with the network gender. Since people with larger political networks have more opportunities for exposure to disagreement, we also control for the size of the network. The results reported in Table 3 generally demonstrate that the political composition of the network is associated with respondent partisanship.

Table 3.

The Effect of Respondent Partisanship and Network Gender Composition on Network Political Composition, Ordinary Least Squares Regression.

	Percent of network supporting Gore		Percent of network supporting Bush
Variables	Coefficient	SE	Coefficient	SE
Female	−0.01	0.02	−0.01	0.02
Education	0.01	0.00*	0.00	0.00
Age	0.00	0.00	0.00	0.00
Union member	0.07	0.02**	0.02	0.02
Married	−0.02	0.02	0.04	0.02**
Partisanship	0.03	0.00**	−0.05	0.00**
Network size	0.07	0.01**	0.11	0.01**
Network Sex	0.34	0.04**	0.03	0.04
Female × network Sex	−0.08	0.04	0.05	0.05
Partisanship × network Sex	0.08	0.01**	−0.06	0.01**
Constant	−0.06	0.05	0.01	0.05
N	1,595		1,595
F	97.32**		101.30**
R²	0.38		0.39
Root mean squared error	0.29		0.30

Note: If a coefficient did not attain statistical significance, we rounded to two decimal places. If a coefficient attained statistical significance, we rounded to the highest possible decimal place up to two digits.

Dependent variable: Average support for candidate in the network, all respondents.

p < .05, two-tailed. **p < .01, two-tailed.

Because we are most interested here in the role of respondent sex and partisanship, we focus our discussion on their interaction with network gender in shaping the partisan information in the respondent’s network. As this requires us to focus on interactions and because Friedrich (1982) shows that the statistical significance of a variable depends on the value of the other variable in an interaction term, we evaluate the significance of all possible combinations of these variables to more accurately understand these relationships. We do this by evaluating the impact of having more women in the network for men and women separately and then taking the first derivative on the ordinary least squares equation with respect to respondent sex and partisanship, separately. This gives us an estimate on how network support for Gore (and Bush) is related to the number of women in the network for respondents of different sex and partisanship. The coefficients and confidence intervals that result from this process are displayed in Figure 1.

Figure 1.

The relationship of sex composition in the network with the political composition of social networks for men and women.

Consider first the case of how many Gore voters are in a respondent’s network. For both males and females, each additional female discussant is positively associated with having more support for Gore in the network. If a man with 0% females in his network is compared to a man with 20% (e.g., 1 of 5 discussants is a female), this estimate suggests that the expected difference between their networks’ support of Gore would be around 7%. The estimated coefficient is slightly smaller for female than for male respondents, though the difference between them is not significantly different. Assuming the same scenario, adding a female is associated with a difference that is approximately 5%. Respondent partisanship is even more strongly related to the level of Gore voters in the network. For all types of partisans—even self-identified Republicans—networks with more women are more likely to have more support (on average) for Gore. There are differences between each partisan category that are statistically significant and consistent with what we might expect.

A different pattern appears in the results for the model of Bush voters in the network—the estimated coefficient for females is larger than that for males; Republicans with more women in their network have more Bush voters than do independents and Democrats, respectively. More interesting is that the inclusion of more women in the network does not always mean there are Bush voters in the respondent’s network (as it does in the Gore model). The size of the coefficient for having more females in a network is unrelated to respondent sex in the Bush model. This is different than in the Gore model, where the number of females in the network increases Gore support for men and women. Together the results imply that respondent sex is not important for understanding the relationship between the political and gender composition of network; any variability in the correlation of number of women and Bush supporters in the network is a function of something else. That “something else” is respondent partisanship. For Republican self-identifiers, having more female discussants in the network is related to having more Bush supporters. Independent respondents exhibit the same directional relationship, but the effect is small and close to statistical insignificance. Democratic respondents, however, actually have fewer Bush supporters when there are more females in the network.

Comparing these results to those from the Gore model paints an interesting picture. For Republicans and Independents, networks that have more females are associated with a network with more discussants expressing a vote choice. This is what we call a politicization relationship. For Democrats, their networks have more Gore voters and (slightly) fewer Bush voters. This is consistent with more agreement in their network.⁶

When it comes to the political information in a respondent’s network, sex differences are part of the story. First, there is no evidence that networks are “randomly” constructed. Second, these differences are more strongly related to respondent partisanship rather than respondent gender. This is to say that the network differences that exist between men and women is more strongly related to the fact that more men are Republicans and women are Democrats than it is that men and women only seek same-sex discussion partners. Finally, even though men and women do not seem to be seeking same-sex discussion partners, the number of females in the network is related to the partisan character of the network.

These results suggest that the gender gap in voting is partly rooted in network composition, though by the simple path indicated by Berelson et al. When it comes to sex differences, men are less likely to talk politics with women. However, it seems that party interacts with average number of women in a network to further exaggerate political differences, especially for men. Importantly, there is no evidence to suggest that if women only spoke about politics with each other that the cleavage would be stronger. Our evidence suggests that strict segregation is not a necessary condition for a cleavage to emerge.⁷

Discussant Sex and Discussant Partisanship in Candidate Evaluation

The political composition of networks is not differentiated solely, or even primarily, by respondent sex. However, the evidence shows that discussants’ sex is related to the overall partisan environment in the network. This implies that discussant sex may serve as a way of alerting respondents—either by accident or on purpose—to how a person fits with their own politics. For men, having more women in their network polarizes their network. This may explain why men’s networks exhibit more sex homophily in Table 2—bringing women into the network cuts at the natural tendency of people to avoid disagreeable political discussants (Festinger, 1957). More subtly, this process may be gendered as men and women react differently to sex composition in the network. In particular, men are more sensitive to the combined role of partisanship and gender in screening the Republican parts of their networks while women are more responsive in the Democratic end.

Overall, this implies that the relative weight given a particular discussant may be a function of discussant sex. One way to think about this is that men and women may not adjust their candidate evaluations the same depending upon the discussant’s sex; having female Democrats and male Republicans in the network is one thing, but listening to them is another completely. We must, therefore, consider the possibility that the exchange of political information has few or no consequence when it either crosses sex lines or, alternatively, comes from women.

With this in mind, we specify a model in which changes in candidate evaluations are a function of respondent preferences, discussant preferences, and the preferences of everyone else in the network.⁸ The basic unit of analysis here is the discussant-respondent dyad. Following the lead of Huckfeldt et al., the baseline autoregressive model examines change in the respondent’s evaluation of the candidate as a function of the impact of the discussant’s partisan preference, the respondent’s partisanship, the partisanship of the other network members, and the respondent’s original candidate evaluation. There are two measures of social context in our analysis: (a) the discussant’s political preference and (b) the amount of agreement with the discussant in the remainder of the network. It is the interaction of these terms for same and different sex dyads in which we are most interested. The logic of this analysis is that we are trying to understand the correspondence of particular discussants with the opinion of respondents in the context of their overall network. Our twist on this model is that we then examine whether sex homophily in the dyad is related to the measures of correspondence between change in evaluation of the candidates.

We cannot claim that the estimates of social influence are free from endogeneity (Kenny, 1994) or selection bias in the context of available data. However, the model includes several features that minimize these concerns. First, the dependent variable is the change in the respondent’s evaluation of Al Gore (and George Bush, separately) from the preelection survey to the postelection survey, coded so that positive numbers indicate increasing affection for the two candidates. We argue that this helps mitigate selection bias as candidate evaluations are less likely to drive partisanship or friendship selection, than vice versa. Second, including the residual partisan composition in the network reduces the effect of omitted variable bias and provides a more realistic view of the network—that is, the relationship between one discussant and a respondent occurs in the context of the other relationships in the network. Finally, by including a preelection measure of the candidate evaluation, we account for potential regression-to-the-mean effects. Despite these precautions, we must acknowledge the limitations of the data and look to future work that has different data to attack this issue more directly.⁹

We estimate the model separately for (a) change in feeling toward Gore and Bush and (b) all permutations of respondent and discussant sex for a total of eight different models. We specifically look at whether our measures of social context depend upon sex homophily in the respondent–discussant dyad. Interpreting social influence in this model depends on evaluating an interaction between each discussant’s vote preference and the preferences of the rest of the network. With two candidates, four types of gender-discussant dyads, and four values of the interacting term (the residual network can have 0, 1, 2, and 3 as possible values), we have 32 different statistical tests to evaluate. We report the results from the eight statistical models in Tables 4 and 5 and focus our discussion on the summary of the interaction tests in Table 6 and the graph of coefficient effects in Figure 2.

Table 4.

Effect of Discussant Vote Choice on Respondent Evaluation of Gore by Respondent and Discussant Gender, Ordinary Least Squares Regression with Clustered Standard Errors.

	Male respondent				Female respondent
	Male discussant		Female discussant		Male discussant		Female discussant
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
Partisanship	7.01	0.60***	6.79	0.90***	6.16	0.66***	8.89	0.61***
Discussant Gore vote	7.61	2.46***	9.53	3.35***	9.00	2.56***	7.67	2.44***
Partisanship × discussant	−0.69	0.83	−0.21	1.09	0.14	0.91	−1.64	0.78**
Residual Gore support	4.06	1.59**	7.44	2.77***	5.83	1.39***	2.38	1.55
Discussant vote × residual	−3.22	1.68*	−5.84	2.86**	−4.47	1.83**	−1.21	2.00
Preelection thermometer	−0.72	0.03***	−0.76	0.05***	−0.76	0.04***	−0.84	0.03***
Constant	31.48	31.48	31.75	3.38	34.72	2.67***	41.59	2.55***
N	737		379		626		639
F-test	109.22***		54.95***		86.01***		192.06***
R²	0.55		0.59		0.57		0.66
Root mean squared error	19.21		19.21		18.76		18.51

p < .10, two-tailed test. **p < .05, two-tailed test. ***p < .01, two-tailed test.

Table 5.

Effect of Discussant Vote Choice on Respondent Evaluation of Bush by Respondent and Discussant Gender, Ordinary Least Squares Regression with Clustered Standard Errors.

	Male respondent				Female respondent
	Male discussant		Female discussant		Male discussant		Female discussant
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
Partisanship	−5.81	0.71***	−5.22	0.89***	−6.49	0.72***	−7.06	0.83***
Discussant Bush vote	5.20	2.59**	8.26	3.80**	5.69	2.64**	8.78	3.02***
Partisanship × discussant	−1.15	0.87	−1.49	1.35	−0.08	0.88	−1.02	1.04
Residual Bush support	4.90	1.51***	4.34	1.64***	4.42	1.57***	3.93	1.68**
Discussant vote × residual	−2.30	1.90	−0.26	2.30	1.33	1.90	1.15	2.01
Preelection thermometer	−0.78	0.03***	−0.82	0.05***	−0.77	0.03***	−0.87	0.03***
Constant	39.57	2.67***	39.84	3.37***	37.34	2.62***	41.40	2.76
N	741		379		626		636
F-test	134.82***		66.84***		101.56***		210.40***
R ²	0.64		0.64		0.58		0.68
Root mean squared error	17.53		18.36		20.55		20.18

p < .10, two-tailed test. **p < .05, two-tailed test. ***p < .01, two-tailed test.

Table 6.

Coefficients, T-Values, and Statistical Significances for Discussant Vote Based on Residual Network Support and Dyad Gender Homophily.

Dyad	Residual support	Coefficient	T-value	Statistical significance
Dyad	Residual support	Coefficient	T-value	p < .01	p < .05	p < .10
Change in Gore feeling thermometer
Male respondent, male discussant	0	9.61	3.90	x
	1	6.39	4.23	x
	2	3.17	1.36
	3	−0.05	0.01
Male respondent, female discussant	0	9.53	2.84	x
	1	4.69	2.06		x
	2	−0.15	0.04
	3	−4.99	0.78
Female respondent, male discussant	0	9.00	3.51	x
	1	4.53	2.49		x
	2	0.06	0.02
	3	−4.41	1.07
Female respondent, female discussant	0	7.67	3.14	x
	1	6.46	3.82	x
	2	5.25	1.89			x
	3	4.04	0.80
Change in Bush feeling thermometer
Male respondent, male discussant	0	5.20	2.01		x
	1	2.90	1.88		x
	2	0.60	0.26
	3	−1.70	0.43
Male respondent, female discussant	0	8.27	2.17		x
	1	8.01	3.05	x
	2	7.75	2.46		x
	3	7.49	1.54
Female respondent, male discussant	0	5.69	2.15		x
	1	4.36	2.45		x
	2	3.03	1.18
	3	1.70	0.41
Female respondent, female discussant	0	8.78	2.91	x
	1	9.93	5.06	x
	2	11.08	4.28	x
	3	12.23	2.91	x

Figure 2.

Impact of discussant on change in Gore and Bush feeling thermometors based on residual support for the discussant’s candidate choice in the network.

Each interaction coefficient listed in Table 6 shows our estimate of discussant influence on the change in feeling toward one of the candidates under different network conditions. There is strong evidence that change in attitudes toward the candidates are embedded in social processes, with 20 of 32 of the coefficients obtaining some level of statistical significance of at least the .10 level. By itself, evidence of a relationship between respondent and discussant evaluations is not that surprising given previous research (Huckfeldt et al., 2004). From our perspective, the comparison across different types of discussion dyads is more interesting.

Examining the results of the model of changing attitudes toward Gore, we see statistically significant coefficients from discussants of each sex to respondents of each sex. However, those relationships are restricted to networks in which there are one or fewer additional Gore discussants. As can been seen in Figure 2, it is in that area where there are the least differences between the different types of dyads. In other words, the relationship between discussant partisanship and respondent attitudes toward Gore does not seem conditioned by the sex of the respondent or discussant. There are two additional observations to be made here. First, as predicted by the social autocorrelation model, the contribution of any single discussant declines as the network becomes politically homophilus. Second, the declining effect of the discussant influence is less steep for same-gender pairs. The effect is not statistically significant for male-male pairs, but is for female–female pairs (p < .10) with up to two additional supportive discussants. That is, women’s opinions are more related to the partisanship of their women discussants, and vice versa for men. All in all, the implication is that women and men seem to place at least slightly more emphasis on their same-sex discussants in the context of more political cohesive networks, at least when it comes to their evaluation of Gore.

The results of the model of evaluations of Bush are different. We still see a pattern of diminishing returns from the contribution of each discussant as the Bush support in the network grows. Thus, with more Bush support in the network, the influence of a single Bush supporter declines but only when men are the discussant. As a man’s network becomes more pro-Bush, he gives less weight to the opinion of his discussants; that is especially true for male–male dyads where the individual discussant is not influential after even one other Bush voter is included in the network. A different story emerges when we consider female discussants. For men, there are no meaningful “declining effects.” If the network has no other discussants backing Bush or only one other, that single female Bush supporter has a strong positive effect on evaluations of both. If the respondent is a woman, having more Bush-supporting networks actually increases the impact of a pro-Bush female discussant. Whether this reflects men being more selective of the women in their network or if they are more on “the same page” is unclear.

Our interpretation of these complicated results is relatively straightforward. When women are Democrats, their impact on respondent’s attitudes toward the candidates is in some sense unremarkable—it is the same as any other discussant. However, when they act “out of character”—in this context meaning that they are Republican, the minority self-identification—they have an opportunity to exert significant influence. We argue that this occurs because in those cases, the socially supplied information women provide is more novel and therefore more important. All in all, this story does not seem to be about gender and politics alone, but about their combined impact. We are hesitant to draw conclusions about causality on the basis of this evidence, as endogenous influence and network selection serve as potential sources of statistical bias. However, the evidence is consistent with the argument that sex differences are related to the correspondence between the discussant’s partisanship and changes in the respondent’s view of the candidates. The most consistent pattern we see is that the partisanship of female discussants is not differently weighted than that of male discussants, under most conditions.

Summary and Discussion

Our results suggest that the gender gap in voting is related to social networks, but in ways that are not solely a function of partisan biases in people’s conversations. This is a potentially important amendment to the study of the social foundations of political cleavages. Cleavages are supported by social interaction, but not exactly as characterized by Berelson et al. This is because our results, in part, show that political identities do not need clear social segregation to emerge or be sustained. At the same time we find that even a little social segregation—perhaps driven by a mix of partisan selection and the way that discussant sex can shape the way their preferences are perceived—can seemly be enough to produce meaningful sex cleavages in the electorate. At first blush, this implies that social networks are either not important or play no unusual role in explaining the foundations of voting differences between men and women. Upon closer examination, however, there are differences in male and female networks that are related to differences in their behavior.

These results shed light on the questions that motivated this research initially, starting with our understanding of how political cleavages are related to social organization. Even though men and women are not isolated from each other, the combination of sex serving as a cue for politics and the reluctance of men to include women in the networks is seemingly enough to support a political cleavage. Two potential conclusions flow from this observation. First, social segregation is a sufficient, but not necessary, condition for political cleavages. Or to state it differently, a little bit of segregation—even by one of the groups—may be enough for significant political cleavages to emerge. This is a significant amendment to the conditions outlined by Bereslon et al. The fact that women are usually less valued in the influence process only strengthens this conclusion. Second, the absence of social segregation does not mean that political cleavages emerge solely as a function of individual traits. Though earlier research on the gender gap in voting has considered the position of society (usually with respect to marriage and employment), the focus has been on how this manifests itself in different preferences. These results imply that there is also a “social logic” to these decisions that are important.

Beyond these implications for the study of the gender gap and the emergence of meaningful cleavages, there is an important lesson for research on social network models of politics. The vast majority of that literature is interested in the partisan content of social interaction, potentially overlooking the role of other social traits in shaping behavior. In this case, the data demonstrate that it is not just the partisan preferences of discussants that matter. Instead, the sex of both the respondents and the discussants is an integral part of the theoretical model. We think that gender represents an interesting, if complicated, case for exploring heterogeneity in a baseline model of social network effects. However, this only begins to scratch the surface of potentially interesting social processes that can intervene in the communication of political information in our networks. We believe that the results in this paper point to other opportunities for unraveling how different social organizations, contexts, and identities can help us better understand political behaviors like voting.

Footnotes

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Notes

Author Biographies

Scott D. McClurg is Professor of Political Science and Journalism at Southern Illinois University, Carbondale.

Michelle L. Wade is an assistant professor at West Chester University. Her research focuses on identity and political participation in community organizations, particularly neighborhood associations in urban areas.

Maja V. Wright-Phillips is a doctoral candidate at Southern Illinois University. Her research focuses on identity, political activism and political participation, particularly in the U.S. military veteran population.

References

Applebaum

S. H.

Audet

Miller

J. C.

(2003). Gender and leadership? Leadership and gender? A journey through the landscape of theories. Leadership and Organization Development Journal, 24, 43-51.

Berelson

B. R.

Lazarsfield

P. F.

McPhee

W. N.

(1954). Voting: A study of public opinion formation in a presidential campaign. Chicago, IL: University of Chicago Press.

Box-Steffensmeier

J. M.

DeBoef

Lin

(2004). The dynamics of the partisan gender gap. American Political Science Review, 98, 515-528.

Burns

Schlozman

K. L.

Verba

(2001). The private roots of public action: Gender, equality, and political participation. Cambridge, MA: Harvard University Press.

Burt

R. S.

(1985). General social survey network items. Connections, 8, 119-123.

Burt

R. S.

(1992). Structural holes: The social structure of competition. Cambridge, MA: Harvard University Press.

Djupe

P. A. D.

McClurg

S. D.

Sokhey

A. E.

(2013). The political consequences of gender in social networks. Unpublished manuscript.

Festinger

(1957). A theory of cognitive dissonance. Stanford, CA: Stanford University Press.

Finifter

(1974). The friendship group as a protective environment for political deviants. American Political Science Review, 68, 607-625.

10.

Ford

(2006). Women and politics. Boston, MA: Houghton Mifflin Company.

11.

Friedrich

R. J.

(1982). In defense of multiplicative terms in multiple regression equations. American Journal of Political Science, 26, 797-833.

12.

Gilligan

(1982). Concepts of self and morality from in a different voice. In Kolmar

W. K.

Bartkowski

(Eds.), Feminist theory: A reader (pp. 368-372) (2nd ed.). New York, NY: McGraw-Hill.

13.

Hansen

S. B.

(1997). Talking about politics: Gender and contextual effects on political proselytizing. Journal of Politics, 59, 73-103.

14.

Howell

S. E.

Day

C. L.

(2000). Complexities of the gender gap. Journal of Politics, 62, 858-874.

15.

Huckfeldt

(2001). The social communication of political expertise. American Journal of Political Science, 45, 425-438.

16.

Huckfeldt

Johnson

P. E.

Sprague

(2004). Political disagreement: The survival of diverse opinions within communication networks. Cambridge, UK: Cambridge University Press.

17.

Huckfeldt

Plutzer

Sprague

(1993). Alternative contexts of political behavior: Churches, neighborhoods, and individuals. Journal of Politics, 55, 365-381.

18.

Huckfeldt

Sprague

(1995). Citizens, politics, and social communication. Information and influence in an election campaign. New York, NY: Cambridge University Press.

19.

Jenkins

(2013). Mothers, daughters, and political socialization: Two generations at an American women’s college. Philadelphia, PA: Temple University Press.

20.

Kaufmann

K. M.

Petrocik

J. R.

(1999). The changing politics of American men: Understanding the sources of the gender gap. American Journal of Political Science, 43, 864-887.

21.

Kenny

C. B.

(1994). The microenvironment of attitude change. Journal of Politics, 56, 715-728.

22.

Klofstadt

McClurg

S. D.

Rolfe

(2009). Measurement of political discussion networks: A comparison of two “name generator” procedures. Political Opinion Quarterly, 73, 462.

23.

Klofstadt

Sokhey

A. E.

McClurg

S. D.

(2012). Disagreement about disagreement: How conflict in social networks affects political behavior. American Journal of Political Science, 57, 120-134.

24.

MacKuen

(1990). Speaking of politics: Individual conversational choice, public opinion, and the prospects for deliberative democracy. In Ferejohn

Kuklinski

(Eds.), Information and democratic processes (pp. 23-58). Champaign-Urbana: University of Illinois Press.

25.

Marsden

P. V.

(1987). Core discussion networks of Americans. American Sociological Review, 52, 122-131.

26.

Mendez

Osborn

(2010). Gender and the perception of knowledge in political discussion. Political Research Quarterly, 64, 269-279.

27.

Mezey

S. G.

(1978). Does sex make a difference? A case study of women in politics. The Western Political Quarterly, 31, 492-501.

28.

Mondak

Anderson

M. R.

(2004). The knowledge gap: A reexamination of gender-based differences in political knowledge. Journal of Politics, 66, 492-512.

29.

Osborn

Mendez

(2011). Two become one? Spouses and agreement in political opinions. American Politics Research, 39, 783-803.

30.

Rolfe

(2012). Voter turnout. New York, NY: Cambridge University Press.

31.

Sapiro

(1982). Private costs of public commitments or public costs of private commitments? Family roles versus political ambition. American Journal of Political Science, 26, 265-279.

32.

Sinclair

(2012). The social citizen. Chicago, IL: University of Chicago Press.

33.

Stoker

Jennings

M. K.

(1995). Life-cycle transitions and political participation: The case of marriages. American Political Science Review, 89, 421-433.

34.

Wirls

(1986). Reinterpreting the gender gap. The Public Opinion Quarterly, 50, 316-330.