Political Network Size and Its Antecedents and Consequences

Abstract

Recent evidence supports the important political role that political network size and distribution plays at both the individual and system levels. However, we argue that the evidence is likely stronger than the current literature suggests due to network size measurement limitations in the extant literature. The most common approach to measuring political network size in sample surveys—the “name generator” approach—normally constrains network size measurement to three to six individuals. Because of this constraint, research often undercounts individual network size and also leads to a misrepresentation of the distribution of the underlying variable. Using multiple data sets and alternative measurement approaches, we reveal that political network hubs—individuals with inordinately large network sizes not captured by name generators—exist and can be identified with a simple summary network measure. We also demonstrate that the summary network size measure reveals the expected differences in communicative, personality, and political variables across network size better than name generator measures. This suggests that not only has prior research failed to identify network hubs, but it has likely underestimated the influence of political network size at the individual level.

Keywords:

• network hubs
• network size
• political discussion
• political knowledge
• political participation
• name generator

Acknowledgments

The authors would like to thank Paul A. Beck and Janet Box-Steffensmeier for their useful comments on a previous version of this article.

Notes

1. In order to conserve space given the large number of data sets we will employ, when possible we will reference published studies for methodological details rather than reiterate in this article. Other study details will be reported in notes at the point the data are addressed.

2. Hubs are empirically identified explicitly apart from the nature of the distribution only if a researcher dichotomizes the distribution into “hubs” and “not hubs.” Unfortunately, there has been little work to theoretically define the appropriate cut point that distinguishes these two categories (Vallabhajosyula, Chakravarti, Lutfeali, Ray, & Raval, 2009). Some studies have developed ad hoc cut points such as one eighth of the population (Himelboim, Gleave, & Smith, 2009), whereas others have operationalized hubs by their effect on system connectivity (Vallabhajosyula et al., 2009). Because of the ambiguity of the precise theoretical and empirical definition of a hub, and because often the definition of hub could vary depending on the specifics of the population or distribution, in this article we empirically emphasize the forms of distributions that contain hubs rather than attempting to define a given threshold for being defined a hub in a given data set.

3. There is some debate regarding the exact distribution that best fits social network size. For instance, McCormick, Salganik, and Zheng (2010) find a log-normal distribution a better fit than a power law for their data. Jackson (2008, p. 65) argues that while “it is hard to find networks that actually follow a strict power law,” it is true that “many social networks exhibit fat tails.” Our point here is not to argue for a specific skewed distribution, but to argue that a normal distribution should not be expected, and that the long tail of the distribution should be expected to be to the right.

4. Lin and associates (2001) critique name generators for studies of social capital and argue for the use of a “position generator.” “The general technique is to pose one or more questions about the ego's contacts (“names”) in certain social contexts or situations which may range from role or content (neighbors, important family or work matters) to closeness (confidences, intimacy, etc.), geographic limits, or for specific periods of time” (p. 62). The related “how many people do you know” approach asks questions about contact with groups with known distributions in society to estimate network size (e.g., McCormick et al., 2010). Although these approaches have advantages over single name generators in terms of breadth of network tapped, they share the limitation of requiring extensive survey questions and thus producing interviewer and cooperation effects much like those we describe later caused by the time demands of name generators. This is not a limitation of the network summary measure due to its brevity. Moreover, position generators and the “how many people do you know” approach are rare in the political communication literature. So, to save space, we do not discuss them in as much detail here as we have the more ubiquitous and established name generator approach.

5. Actually, the name generator in these data is not technically capped, but the name interpreter—the questions asked about named discussants—is capped at five. So, the only information that is available about network sizes greater than five is that the network size was some unspecified value greater than five. See Marsden (2003, Footnote 3) for a discussion.

6. The network size measure is based on dichotomizing each report to did not (0) or did (1) talk about the 2008 presidential election campaign with a given other group member, or “outdegree” in social network terminology (Scott, 1991, p. 72). Results are quite similar when a measure of “indegree” (the number of times a given individual is named by others as a political discussion partner) is used to construct the distribution.

7. This study was based on a nationally representative sample of 500 U.S. adult respondents aged 18 or older. We conducted five independent surveys—one per week for 5 weeks beginning on October 4, 2009—each with a representative national sample of 100 respondents. All surveys began at the end of the given week, and data collection for each took only a matter of a few days. The last survey in the sequence went into the field after November 3, 2009, which was election day. In effect, this is a repeated cross-sectional design with five identical surveys conducted in the course of 1 month. Surveys were conducted online by the company YouGovPolimetrix. YouGovPolimetrix employs a unique “sample matching” methodology (see http://www.polimetrix.com/documents/YGPolimetrixSampleMatching.pdf). Slightly over half (53.4%) of our respondents were female. The majority were White (73.2%), followed by Hispanic (11.2%), Black (10%), and 2% or fewer of other ethnic groups. Nearly 60% of our respondents were married. The mean respondent was 47 years old (SD = 16) and had a family income of $50,000–$59,999. Our sample consisted of 34% self-identified Democrats, 28% Republicans, and 31% Independents. These descriptive data match very closely with data from the U.S. Census Bureau's American Community Survey (2006–2008) and the 2008 American National Election Study (a highly respected national face-to-face survey).

8. The exact wording of the questions across these studies varied, as did the time frame to which the question referred. These factors likely contributed to the differences in mean political network size.

9. We should note that across these studies we excluded a total of three respondents for reporting values of network size that we felt must be in error or greatly exaggerated. In the Qualtrics study, one individual reported a political network size of 1,024 (whereas the next highest reports were two individuals reporting network sizes of 100). In the Polimetrix study, one individual reported a network size of 400 and another reported a network size of 1,200 (whereas the next highest report was one individual reporting a network size of 100). These three individuals were excluded from further analyses, although we acknowledge that it is a judgment call to retain or exclude any given case of extremely high network size. For these instances, since two of the three reports were an order of magnitude larger than the highest network sizes otherwise recorded (which ranged from 50–100 across the studies), and represent about one tenth of 1% of these two studies (and an even smaller proportion of all of the survey respondents), we feel comfortable with these decisions.

10. The exception is the student activity group study, for which network size is measured only with reference to other members of a single group of limited size, and which therefore excludes the vast majority of possible connections, including strong ties such as family members and essentially anyone who is not in college. Moreover, since college students tend to be less politically involved than older adults in general, it is not surprising that the network sizes in this study do not match those of the other six, which are general population studies.

11. Since it is possible that A reports talking politics with B, but B does not report talking politics with A, the size of the circles in this figure does not perfectly match the number of lines attached to that circle. The size of the circles, instead, reflects most closely the concept of self-reported network size as discussed throughout this article.

12. Ideally we would have capped our name generator at six, but as we have noted doing so takes up a large amount of survey time. Readers should keep in mind that most name generator studies are capped at fewer than six names, including but not limited to Mutz's Spencer Foundation survey (capped at three; Mutz, 2002b), the 2000 ANES (capped at four; Nir, 2005), and the Electronic Dialogue project (capped at four; Cappella, Price, & Nir, 2002).

13. This should not be surprising given what we have already described regarding the implications of survey-related variables such as respondent cooperativeness and interviewer effects for name generator estimates of network size. Other studies have looked very closely at the issue of overreporting of zeros in name generator data gathered by the General Social Survey and concluded that in the 2004 GSS there are approximately 208 “excess zeros” out of 356 zeros in the data. This is an astounding 58% of respondents who offered no names for the name generator being estimated as actually being a non-zero value (McPherson, Smith-Lovin, & Brashears, 2009, pp. 674–675).

14. We used a log transformation to account for the non-normal distribution of this variable (see also Gil de Zúñiga & Valenzuela, 2011). We discuss the value of this log transformation in more detail later in the article. The alternative of using a non-parametric correlation (i.e., Spearman's rho) produces almost identical results: ρ _T1T2 = .42, ρ _T2T3 = .58, and ρ _T1T3 = .44.

15. We conduct bivariate analyses because the control variables available to us differ from study to study, and because our emphasis is not on assessing the unique relationship of any given variable with network size, but confirming that network size is correlated with variables in the way we would expect it to be. In effect, a central purpose of these analyses is to establish construct validity for the summary network size measure.

16. We acknowledge that our classification of network size predictors versus outcomes, at least in some cases such as the link between frequency of discussion and network size, is subject to debate. In other cases—for instance, personality traits predicting network size as an outcome—the causal logic is more defensible. In any case, our goal here is not to establish causality but instead merely association.