A Network Model of Negative Campaigning: The Structure and Determinants of Negative Campaigning in Multiparty Systems

Reciprocity: Retaliation or “Tit for Tat”

It is often regarded that social relation tends to evolve from an asymmetric to a symmetric relationship (Snijders, 2011; Wasserman & Faust, 1994). This intuitive notion of reciprocity fits particularly well for some of the existing perspectives on negative campaigning, as political actors are more likely to go negative toward opponents that have previously attacked them (Damore, 2002; Lau & Rovner, 2009). Reciprocity is expected to govern the decision calculus of political actors if a failure to counterattack invites the perception that the target of negative campaigning is less competent and less credible (Lau & Pomper, 2004). Also, it is suggested that retaliation against competitors enables those being attacked to offset the relative losses by damaging the competitors’ standing (Dolezal et al., 2016). Yet, the notion of reciprocity has rarely been considered in the previous literature (Lau & Pomper, 2004; but see De Nooy & Kleinnijenhuis, 2013). This leads to the following hypothesis:

Structural Balance: Attacking “Friend’s Enemies” or “Enemy’s Friends”

A second fundamental principle that governs human behavior is the concept of structural balance (Cartwright & Harary, 1956; Heider, 1946). The notion of structural balance suggests that an even number of attacking dyads create balanced states in a triad, where a triad is defined as a set of all possible relationships among three “nodes” in a network (i.e., i-h, i-j, and j-h relations). In the present context, a group of three actors is considered in a balanced state if two of these actors are in the following scenarios: (a) they have a positive relation with each other and jointly attack a third actor (“attacking my friend’s enemy”), or (b) they have a negative relation with each other and only one of the actors attacks a third actor (“attacking my enemy’s friend”). Hence, actor i’s relation with actor j, and their relation with actor h are mutually interdependent. In other words, friends can make common enemies, and common enemies can make friends. The principle of structural balance captures how endogenously driven dynamics could lead to a different cost-benefit calculus compared with a two-party system, as demonstrated in De Nooy and Kleinnijenhuis’s (2013) seminal investigation of the 2006 Dutch national election campaign. Their analysis suggests that the decision to go negative is at least partly governed by the principle of “supporting your friends’ friends and attacking your friends’ enemies” (De Nooy & Kleinnijenhuis, 2013, p. 123). Yet, to the best of our knowledge, this empirical regularity has not been replicated in a different multiparty electoral context. We thus conjecture the following:

Transitivity: Hostility “Roll-Over”

Network analyses often entail the concept of transitivity (“closing the triangle”), such that a tie between an actor i and an alter j is more likely given ties between actor i and actor h, and h and j (Snijders, 2011; Wasserman & Faust, 1994). With a specific view on negative campaigning, this implies that a group of three or more actors is more likely to attack one another, creating triads that consist of all negative signs. From the perspective of the initial attacker (e.g., actor “i” in an i → h → j triad), attacking one’s opponents (e.g., i → h attack relation) might unintentionally help the other competitor j, as disaffected voters (who initially supported h) may turn toward j rather than the initial attacker (Elmelund-Præstekær, 2008; Walter, 2014). In other words, in a multiparty system, an enemy’s enemy is not necessarily a friend. Such a situation creates incentives for actors to go negative toward multiple opponents—including the enemy’s enemy (e.g., i → j as well as i → h attack). Also, from the point of view of the attackee (e.g., actor “h” in an i → h → j triad), to the extent that attacking other candidates (e.g., h → j attack relation) is more effective in compensating the relative losses than retaliating the initial attacker, such “roll-over” or “transitivity” of negativity may be preferable (e.g., Dolezal et al., 2016).

Empirically, a considerable number of transitive triads are not uncommon in networks with negative relations, especially when actors face uncertainties and competing interests among multiple competitors (e.g., Maoz, Terris, Kuperman, & Talmud, 2007). Much like negative campaigning in a multiactor environment, transitive triads within such a setting can potentially maximize the short-term gains of individual actors through their strategic actions. We therefore suggest that, controlling for other factors, transitivity is an integral aspect of negative campaigning in multiparty contexts:

Preferential Attachment: “Cumulative Hostility”

Empirical studies of social networks have frequently shown that a small subset of influential actors are highly connected with others, whereas a large number of actors have few connections, resulting in a power-law distribution of degrees (“preferential attachment”; Barabási & Albert, 1999; Newman, 2001; Snijders, 2011). We expect a similar preferential attachment structure in negative campaign networks but only for incoming attacks. That is, few actors receive multiple attacks, whereas there is no tendency to disproportionately send out attacks. Extensive attacks over a large number of actors are susceptible to backlash and retaliation (Lau & Rovner, 2009), and potentially detrimental to postelection bargaining processes common in multiparty contexts (Elmelund-Præstekær, 2010). Therefore, initiating a large number of attacks is comparatively unappealing, rendering it unlikely that a particular actor dominates the attacks. Conversely, and in line with the analysis of Skaperdas and Grofman (1995), attacks are often concentrated on few candidates (see also Djupe & Peterson, 2002; Haynes & Rhine, 1998). From an attacker’s perspective, any attack directed at important and, hence, more visible candidates is more likely to be covered by news media (e.g., Haynes & Rhine, 1998). Consequently, we expect preferential attachment to be a significant predictor of tie formation in a mediated attack network, such as follows:

Temporal Dynamics of Negative Campaigning

A detailed view on how negative campaigning evolves over time further illuminates the underlying logic of negative campaigning. Our general expectation here is that the “network of previous statements . . . provide[s] the context for new statements” (De Nooy & Kleinnijenhuis, 2013, p. 112). This is based on the assumption that politicians would react to other politicians’ attacks in a “timely” fashion. Moreover, such behavior is likely not to take into account the complete history of their competitors’ prior attacks but to rely on the most recent interactions (e.g., Dolezal et al., 2016; Kleinnijenhuis & de Nooy, 2013). Likewise, we expect a tendency toward structural balance and transitivity only when a sequence of attacks followed by support (for structural balance) or attack (for transitivity) among two actors is sufficiently recent to warrant a reaction (De Nooy & Kleinnijenhuis, 2013). Therefore, with regard to the temporal sequence of endogenous dynamics in an attack network, we predict the following:

We present substantive interpretations and visualizations for each of the hypothesized network structures predicting the attack instances in mediated negative campaign networks, in terms of both their static forms (H1-H4) and their dynamic versions (H5-H7) in Table 1.

OPEN IN VIEWER

Table 1.

ERGM/TERGM Parameters, Associated Configurations, and Interpretations.

Parameters	Network configurations and interpretations
Reciprocity (H1)		Propensity to retaliate given alters’ attacks
Delayed reciprocity (H5)		Propensity to retaliate at time t given alters’ attacks at time t − 1
Structural balance (H2)		Propensity for attacking alters based on sequences of “support-attack” (i.e., attacking friend’s enemy) or “attack-support” (i.e., attacking enemy’s friend)
Delayed structural balance (H6)		Propensity for attacking alters based on sequence of “support-attack” or “attack-support” that are formed at time t − 1
Transitivity (H3)		Propensity for transitive closure with multiple attack relations
Delayed transitivity (H7)		Propensity for transitive closure with multiple attack relations formed at time t − 1
Preferential attachment, in-ties (H4)		Propensity to be attacked from multiple alters (i.e., popularity)

Note. The thick lines represent the focal attack relations being modeled. All black lines represent “attack” relations, whereas blue lines represent “support” relations. Solid lines represent relations at a given time point (t), and dashed lines represent relations formed at t − 1. ERGM = exponential random graph model; TERGM = temporal ERGM.

Coding of Campaign Coverage and Construction of the Mediated Attack Network

All of the campaign coverage appearing in eight newspapers—three quality papers (Der Standard, Die Presse, Salzburger Nachrichten), two midrange papers (Kurier, Kleine Zeitung), and three tabloids (Kronen Zeitung, Österreich, Heute)—was collected daily starting from August 19th until Election Day (on September 29, 2013) from a database provided by the Austrian Press Agency (APA). Campaign communication during this 6-week period constitutes the bulk of campaign-related interactions. The content of the newspapers was coded at the sentence level, including political actors (parties and individuals), as the subject and object of each statement, and the “topic” of the statement (Van Atteveldt, 2008), alongside the valence of statements that reflects the relationship between the actors (i.e., positive, neutral, or negative). For instance, the statement, “(Frank) Stronach accused (Werner) Fayman of having ‘sold out the workers to the banks’ and to be responsible for higher unemployment” was coded as follows: Frank Stronach is the object actor, Werner Fayman is the subject actor, and their relationship (from Stronach to Fayman) is −1 (“attack”). We provide additional examples of how actual sentences are coded in Online Appendix A. The manual coding of the data was performed by seven native German-speaking coders who were extensively trained on sample material. Intercoder reliability was satisfactory with Krippendorf’s alpha scores of .87 for the subject actor, .87 for the object actor, and .76 for the valence of statements (see Kleinen-von Königslöw et al., 2015).

As we control for party affiliation and list placement to establish a plausible baseline, we limit the analysis to actors with known party affiliations, resulting in 5,968 negative statements out of all 10,720 eligible statements (corresponding to 1,919 unique news articles out of 2,963) and 2,319 positive statements made either by party organizations or by individual politicians. Among 5,968 negative statements, 3,899 statements were dropped as they did not reference individual politicians. The remaining 2,069 negative statements were used to create the final attack network. This accounts for 796 unique articles (M = 18.95, SD = 9.05), with an average of 49.26 attack instances (SD = 34.82) per day. (The daily trends are visualized in Figure A1 in Online Appendix A.) A supplementary analysis of attack relations involving party entities (presented in Online Appendix B) did not substantially change the main conclusions.

Based on the 2,069 negative statements, we derived the mediated “attack” relations as an actor-actor attack matrix, featuring individual politicians who appear in the media coverage (N = 164). As there were only few cases of dyads (N = 164) with more than one attack instance (out of 164 × 163 = 26,732 possible pairs), we define the mediated attack network as binary, such that the cell entry of an attack dyad Xij is 1 (=presence of an attack) and 0 for otherwise. ¹ While the density of a count-based attack network is 0.076, the density of the dichotomized network is 0.011, thus dichotomizing the data does not greatly alter the underlying matrix. Nevertheless, readers should bear in mind that our binary conception of mediated attacks may have some nontrivial implications for understanding the phenomenon in question (see the Discussion section). Figure 1 below presents the cross-sectional network of attack patterns, and attack patterns based on reciprocity, structural balance, and transivity are presented in Online Appendix A.

Analysis Strategy

We employ an exponential random graph model (ERGM) to investigate several and potentially competing determinants of the mediated attack networks. ERGMs allow modeling the interdependence structure of the data, where each network tie is treated as a “random” variable x_ij = 1 if observed or x_ij = 0 otherwise (Lusher, Koskinen, & Robins, 2013; Robins, Pattison, Kalish, & Lusher, 2007). The central idea of this modeling strategy is to assume the observed network to be a single realization out of all possible networks with the same nodes. The probability of a network tie is expressed as a function of substantive dependence structures (Lusher et al., 2013; Robins et al., 2007), or network statistics (as represented in Table 1 above), that collectively differentiate between the possible networks of the same size that could have been observed and the empirically observed network. For instance, if the theory implies that reciprocity or transitive triads exist in a network, then we can compare the respective network statistics from the observed network to the distribution of the same network statistics from simulated networks (Desmarais & Cranmer, 2012). If the hypothesized dependency structures re-produce the global structural properties of the observed network in such simulated distributions, then the hypothesized mechanisms are regarded as the accurate and plausible representations of the underlying tie-generating processes.

We first investigate whether our static representation of the endogenous network structures (H1-H4) predict the negative campaigning patterns over the entire time frame using a cross-sectional ERGM. Next, in order to explore the temporal dynamics (H5-H7), we move to a temporal ERGM (TERGM), a time-series extension of the ERGM framework (Desmarais & Cranmer, 2012; Hanneke, Fu, & Xing, 2010). For TERGM, we create a longitudinal panel series of mediated attack networks by subsetting all of the mediated attack instances to weekly (N = 6 weeks) and daily (N = 42 days) observations. In applying TERGM, we assume that each observation at time t is conditionally dependent upon the network at time t − 1 but not on any other previous networks. Analyses were implemented using the ergm package (Handcock et al., 2013; Hunter, Handcock, Butts, Goodreau, & Morris, 2008) and the xergm package (Leifeld, Cranmer, & Desmarais, 2016).

Measures

We operationalize the reciprocity by whether a pair of actors has mutual “attack” ties, where a directed attack from actor i to actor j, as well as a counterattack from actor j to actor i, exists at the same time. For its temporal version, delayed reciprocity is defined as a situation where a competitor i initiates an attack on actor j at time t − 1, and actor j retaliates at time t (Leifeld et al., 2016). For structural balance, we derive a summary measure of two paths between a pair of nodes i and j through h (i.e., the existence of the two paths through i → h → j), where only one of its constituent paths is a positive tie (e.g., i → h or h → j) and the other is a negative tie (see Table 1 for the respective diagrams). For its temporal version, we use the identical configuration but measured at time t − 1, expecting that such relationships would increase the likelihood of an attack from i to j at time t. Transitivity was operationalized using the geometrically weighted edgewise shared partner (GWESP) statistic (Robins et al., 2007; Snijders, Pattison, Robins, & Handcock, 2006), which represents the tendency of direct attack ties to have multiple shared third attacking relations. Similar to the transitivity, the preferential attachment is operationalized using a geometrically weighted in-degree (GWD-in) and geometrically weighted out-degree (GWD-out) distribution statistic, which signify differential attack activities across the network (for details, see Hunter, 2007). We fixed the decay parameters of each geometrically weighted term to 1.5 (GWESP), 1.8 (GWD-in), and 0.1 (GWD-out), respectively, to reduce the computational burden. These values were obtained by iteratively optimizing the model fit based on Akaike information criterion (AIC)/Bayesian information criterion (BIC), and the goodness-of-fit test. Lastly, for the temporal measure (delayed transitivity), the number of multiple shared attack relations at time t − 1 is used to predict the probability of ties at time t.

Next, we define politicians’ electoral prospects (and therefore their chances of being featured in media) by calculating the difference between candidates’ list placements and their respective parties’ expected number of seats (M = −0.106, SD = 0.626, ranged from −1 to 1). ² The variable is equal to 1 if a candidate is placed first on his or her party’s list and becomes negative if a candidate is outside of the expected range of winning seats. Dummy variables for gender (1 being “female,” 21.89%), whether politicians were running for election or not (1 running for election, 81.66%), and incumbency status (1 being in the national parliament, 34.91%) were also created. In addition, we control for party leader (1 being “party leader,” 3.55%) as they are more likely to be the targets as well as the sponsors of negative appeals (Tresch, 2009).

We control for two dyadic-level homophily: one based on party-level ideology and the other based on politicians’ regions of origin. In order to tap ideological distance, we employ the parties’ left-right placements from a survey of parliamentary candidates (Müller, Eder, & Jenny, 2014). Politicians were assigned their respective parties’ mean scores (BZÖ: 7.64, FPÖ: 8.96, Grüne: 2.13, ÖVP: 6.40, SPÖ: 3.10, Team Stronach: 7.06), and ties were regarded as more homophilous if attacker-attackee dyads have smaller differences in their ideology. ³ It is expected that actors with a greater ideological distance are more likely to attack one another (De Nooy & Kleinnijenhuis, 2013; Lau & Rovner, 2009; Walter, 2014). Likewise, based on the candidates’ regions of origin, we define regional homophily as the attacker-attackee dyads in the same federal region. Our expectation is that politicians from the same region are more likely to attack one another, or at least that such attacks are more likely to receive media coverage. ⁴

Finally, we control for affiliation with current governing parties. One of the most consistent findings in the literature is the effect of coalition potential and proximity to governmental power (Elmelund-Præstekær, 2010; Walter & van der Brug, 2013), which suggests that parties with greater coalition potentials (previous government experience and larger parties) are less likely to attack others, whereas smaller, less established parties are more likely to rely on negative campaigning. In the Austrian context, the two largest, catch-all parties (ÖVP and SPÖ) have frequently been in a “grand coalition” in the postwar period. Therefore, we expect that politicians of ÖVP and SPÖ are more likely to be attacked by other politicians, whereas they are less likely to attack politicians of other parties.

Limitations and Future Research

Throughout the article, we highlighted the notion that there are important structural effects in multiparty campaign environments compared with their two-party counterparts. While the analysis has important implications for our understanding of the nature of negative campaigning in multiparty systems, it is worth revisiting several methodological details and limitations of the current approach. First and foremost, while we acknowledge that our study relies on a single case that is not necessarily representative of all other electoral contexts, it should be noted that the theoretical predictions and empirical approaches are not bound to the specific context of the study, not least as Austria is often considered a fairly ordinary case of a European multiparty system (Müller, 2006). To be sure, only a replication beyond this particular context would provide firm evidence on the generalizability of our findings, but we are confident that these empirical regularities should stretch beyond the political system of Austria (e.g., De Nooy & Kleinnijenhuis, 2013). Future studies would benefit from a comparative perspective that employs a network model across different electoral contexts and over time.

Second, it should be noted that in “mediated attack networks,” individual politicians, political parties, and media are jointly responsible for producing attack patterns (De Nooy & Kleinnijenhuis, 2013). While identifying how media systematically and disproportionally misrepresent the direction and extent of negative campaigning is an important topic of study (Ridout & Smith, 2008), doing so requires nontrivial assumptions regarding the availability of data and the soundness of inferential logic underlying such analyses. Ideally, one would wish to systematically compare the mediated attack behavior (media coverage) and their “unmediated” counterparts (the source of such media coverage) for a set of political actors, thus assessing levels of disproportionality.

Third, following our theoretical focus, our analysis considered negative campaigning instances only among individual politicians. Given a nonnegligible share of attack relations that involve party entities, this appears to have nontrivial implications for the current analyses. Yet, a supplementary analysis (presented in the Online Appendix) suggests that we find similar evidence of endogenous dynamics even for attacks involving different “modes” of actors (a mix of individuals and party entities), largely confirming the conclusions reported here.

We close by acknowledging that our binary conception of attack instances limits our understanding of the endogenous dynamics of negative campaigning to a certain degree. Our theoretical account and analysis assume that any number of attacks is essentially identical with each other compared to “no attacks at all.” Although this may seem counterintuitive—especially as media attention toward attacks or support relations among politicians varies quite dramatically—readers should bear in mind that we are interested in “why” politicians attack their competitors and “when” such attacks are reported. This is mirrored in our use of the ERGM/TERGM framework, which generally aims to explain the principle of network partner “selection” and their temporal dynamics. To be sure, attacks among politicians vary in their level of severity, yet we expect the underlying logics of “why” and “when” they attack to not differ by the amount of attacks.

Overall, our analysis lends support to the notion that negative campaigning in multiparty systems is complex and interdependent. Traditional accounts of attack patterns during election campaigns have emphasized individual politicians’ or parties’ strategic calculus to maximize their electoral gains and minimize negative consequences. While our theoretical perspective and empirical analysis do not imply that such factors are not important, our analysis highlights the need for developing a more nuanced understanding that takes into account complex interdependencies to advance our understanding of negative campaigning. Our study thus speaks to the question of how politicians’ decisions to attack are shaped by others’ attack behavior. Taking such interdependencies into account sharpens our understanding of candidate behavior and may inform strategic decisions of political actors.

On a more practical level, recent accounts suggest that while media coverage of party-based electoral competition is declining in Western Europe (e.g., Garzia, 2013), individual candidates and their characteristics have increasingly become the focus of electoral campaigns (Zittel, 2015; Zittel & Gschwend, 2008), not least because individual campaigns have become a viable option in the age of social media (Hermans & Vergeer, 2013). While incentives for engaging in negative campaigning that result from the nature of the electoral competition systematically structure interaction patterns among individual actors, this study has provided evidence that individual political actors are political entities in their own right—even in party-centered political systems. Therefore, understanding individual candidates’ campaign behavior vis-à-vis party-level explanations becomes increasingly important.