Online aggression from a sociological perspective: An integrative view on determinants and possible countermeasures

The present paper introduces a theoretical model for explaining aggressive online comments from a sociological perspective. It is innovative as it combines individual, situational, and social-structural determinants of online aggression and tries to theoretically derive their interplay. Moreover, the paper suggests an empirical strategy for testing the model. The main contribution will be to match online commenting data with survey data containing rich background data of non- /aggressive online commentators.


Introduction
In the past years, online aggression in social media has attracted a lot of attention not only in the broader public but also in academia (e.g. Cicchirillo et al. 2015;Sydnor 2018). Studies show that offending, defaming, or threatening online comments posted by Internet users fundamentally negatively affect the targeted persons' well-being, social harmony, and democratic outcomes (e.g. Anderson et al., 2014;Bauman, 2013;Kwon and Gruzd, 2017). Accordingly, knowing why people aggress online is the first step to counter it. Although previous research on online aggression has been successful in suggesting and explaining single determinants driving aggressive online commenting (see studies in the State of Research below), (1) their interplay has hardly been studied due to the lack of an overarching theoretical framework and (2) socio-structural determinants have been largely ignored so far. Moreover, from a methods point of view, (3) there are no studies that systematically link digital commenting data to offline information on adult aggressors from the wider population. Hence, the present paper introduces a theoretical model that relates several determinants of online aggression to each other in a more general framework of sociological explanation. Based on the model, we aim to answer the following research questions: (1) Which individual determinants, situational determinants, and social-structural determinants drive online aggression? (2) How do various determinants relate to each other when producing aggressive online behavior? (3) Are there differences in online aggression between social-structural groups?
Answering such questions requires a specific empirical strategy. We intend to conduct a largescale quantitative survey in German-speaking Switzerland, including aggressive and nonaggressive online commentators. They are drawn from a large population of commentators having submitted to online commentary sections of a large Swiss media organization. We match their survey information with their commenting behavior, ranging from non-aggressive to frequently aggressive (this classification emerges from human/automated content analysis).
We will elaborate on the theoretical model and the planned empirical strategy in the following sections. First, however, we will describe in more detail the current state of online aggression (OA in the remainder of the paper) research.

State of research
In the literature so far, determinants of OA are explored primarily from three different perspectives: the individual-psychological, the situational, and the social-structural. All three perspectives are shortly reviewed here, from the fields of psychology, political science, and communication.

Psychological-individual determinants
From a psychological-individual perspective, OA can on the one hand be motivated by relatively stable psychological traits ("aggressors as antisocial individuals"). The underlying theory proposes that each individual has a unique personality and that associated traits motivate behavior and thus (online) aggression. For example, online aggressors score relatively higher in narcissism, psychopathy, and Machiavellianism (e.g. Abell and Brewer, 2014), might lack empathy (Steffgen et al., 2011), may be less open, low in self-control, and impulsive (Peterson and Densley, 2017), but also more depressive and shy (Bauman, 2013).
On the other hand, OA can be motivated by less stable individual emotions, beliefs, and goals ("aggressors as venting, convinced Internet activists"). For example, people in negative mood may troll (Cheng et al., 2017), being angry at unfair negotiators motivates to digitally aggress (Johnson et al., 2009), and car drivers vent their rage (Stephens et al., 2016). Also, online aggressors belief that they do not get caught and that their online con-ent is not permanently stored (Wright, 2013). Further, people participating in collective online outrage are motivated by moral heuristics and moral beliefs (e.g. based on moral disengagement theory by Faulkner and Bliuc, 2016) and punishing violators of social norms (based on social norm theory; Rost et al., 2016). Finally, online aggressors have goals. They spread political ideologies, seek thrill and fun, draw attention to social in-justice (Erjavec and Kovačič, 2012), or seek social standing, status, and recognition (e.g. Ballard and Welch, 2017).

Situational determinants
Research on situational determinants suggests that online aggressive individuals are influenced by properties of the digital media environment and the surrounding social and situational context ("aggressors as ordinary people, but situationallydriven"). The psychological-communicative Reduced cues approach (Sproull and Kiesler, 1986) argues that properties of online environments may cause toxic online disinhibition (Suler, 2004): people feel less restraint because of the absence of social-context cues, anonymity, invisibility, asynchronicity, or minimization of authority. This is explained either by deindividuation theories (Diener, 1980) or by the social identity model of deindividuation effects (SIDE) which argues that deindividuation triggered by reduced social cues and anonymity in online settings boosts the salience of individuals' social identity relative to their personal identity. Thus, if a group norm is salient (e.g. in an online forum), commentators will conform to it rather than engage in uncontrolled aggressive behavior (Reicher et al., 1995). SIDE is empirically supported in several settings (e.g. Hmielowski et al., 2014).
OA is also explained by social learning theories and situational social control. For example, perceiving flaming norms socializes people into flaming (Cheng et al., 2017). Also, people more likely aggress online if informal social controls from an effective community policy and peer pressure are lacking, predicted by routine activity theories of crime (Navarro and Jasinski, 2012), deterrence theory (Xu et al., 2016), or social norms (Álvarez-Benjumea and Winter, 2018). Similarly, people more likely aggress if they have become cyber-victims themselves (Quintana-Orts and Rey, 2018), receive comments challenging their beliefs  or threatening their face (Masullo Chen and Lu, 2017), or if public actors misbehave (Johnen et al., 2017;Rost et al., 2016). Finally, legal frameworks, ethical guidelines, and moderation strategies set up by online (news) platforms may be situationally influential (Ksiazek, 2015).

Social-structural determinants
Research on social-structural determinants is very scarce. It includes socio-demographics, social group memberships, and structural positions and relations. Accordingly, OA may differ by cultural and national backgrounds (Shapka et al., 2018), gender (Ballard and Welch, 2017;Bauman, 2013;Shapka et al., 2018), and age (Bauman, 2013;Shapka et al., 2018). Also, incivility on Twitter is higher in areas of low socioeconomic status (SES), low social capital potential (i.e. potential for interconnected citizen networks), and low indistrict partisan polarization (Vargo and Hopp, 2017).
Finally, (few) structural and sociodemographic factors are considered in the social media cyberbullying model (SMCBM) model by (Lowry et al., 2016).

Gaps
Reviewing the literature on OA, several gaps emerge. Theoretically, there is, first, no overarching theoretical framework integrating the determinants suggested. Accordingly, theoretical approaches to cyberbullying are "sparse and piecemeal" (Espelage et al., 2012: 49) and "have received scant conceptual development" (Runions, 2013: 751). Hence, a major task of future research is to develop "a comprehensive theoretical model that might ground the conversation about cyber aggression and violence" (Peterson and Densley, 2017: 197). At best, such a model addresses the "interaction between micro, meso, and macro levels of explanation" in order to overcome research's current lack of "continuity and coherence" (Peterson and Densley, 2017: 197).
Second, there is a need to relate OA more systematically to social-structural factors. Up to now, information on aggressors and their aggression-benefiting circumstances is limited (Coe et al., 2014: 675;Peterson and Densley, 2017:195). Especially with regard to potential "civility divides" (Vargo and Hopp, 2017: 26), exploring socio-demographic and socio-economic determinants (such as gender, age, education, or prestige) enables to empirically test whether "those equipped with economic and social privilege in the off-line realm may disproportionally gain value from online deliberation, while those with diminished economic and social resources may interact in a hostile, uncivil, (…) strata of the Internet" (Vargo and Hopp, 2017: 24; also see Cicchirillo et al., 2015).
Third, there are no studies that systematically link digital commenting data to offline information in a large sample of adult aggressors. Most studies only use natively online data. If offline information is collected at all, then it is linked to OA intentions, self-reports, or experimental triggers, at best.

Theoretical model
Here, we introduce an integrative model that relates a multitude of determinants to each other in a general framework of sociological explanation, also explicitly theorizing social-structural determinants. This model builds on the ideas of structural individualism (Coleman, 1994) and the model of frame-selection (Esser, 2001;Kroneberg, 2011).
Basically, structural individualism aims at dissecting social phenomena into its constitutive parts, that is meaningful decisions of individual actors. These decisions, however, are embedded in a configuration of social structures and institutions. This social context, in turn, affects (if correctly perceived) actors' goals, beliefs, and opportunities, which then guide their behavior (Maurer and Schmid, 2010;Udehn 2001). From this perspective, OA comments are defined as individual decisions (actions) which are in a first step explained by both characteristics of the individual (e.g. beliefs) and situational parameters (e.g. others' behaviors). In a second step, individual determinants are related to social-structural background. The relationship between these two sets of determinants can be thought of in several ways: social context conditions may structure the set of behavioral alternatives available, the behavioral costs, and an individual's preferences, attitudes, and body of knowledge. Theoretically, this can be explained by learning theories (Bandura 1977) or social production function theory (Ormel et al. 1999).
It needs to be specified, then, how individual decisions come about. This is important because the theory of action chosen has an impact on which individual and situational determinants can be taken into account. Instead of relying on a rather simple rational-choice approach for explaining individual decisions, we opt for the more elaborate model of frame-selection (MFS) as introduced by Kroneberg (2011Kroneberg ( , 2014. In classical rationalchoice theory (Opp, 1999), it is assumed that actors choose those behavioral alternatives which they expect to best fulfill their preferences given certain behavioral constraints. Thus, behavior is a function of individual goals (evaluative beliefs, including egoistic just as prosocial goals), beliefs about the consequences of decisions, and behavioral constraints (the latter two are often summarized as descriptive beliefs). However, rational-choice theory is silent about which descriptive and evaluative beliefs are active in a specific decision situation. Therefore, MFS explicitly incorporates the process of the definition of the situation (Esser, 1996). In this process, actors subjectively define which kind of situation they are actually facing (which may -in contrast to rational-choice theory -deviate from "objective" situational requirements). They do so by synchronizing given situational cues with internalized knowledge about typical situations (frames). Hence, descriptive and evaluative beliefs guiding behavior are not taken for granted but depend on actors' subjective perceptions of the situation. This means that behavioral differences between (groups of) actors do not simply result from individual or situational differences, but from interactions between individual and situational characteristics.
Based on these theoretical considerations, we propose the following explanatory model of OA (Figure 1): In this model, OA behavior results from individuals' definitions of potential online commenting situations. Such definitions represent a situation's general meaning and thus determine which individual beliefs are activated and which situational constraints are perceived by the actor. How situations are defined depends on two sets of factors: (1) situational determinants comprise all relevant characteristics of the situational context and thus are in principle identical for all actors in the same situation (but still differently perceived).
(2) individual determinants comprise all descriptive (representations of current states of the world) and evaluative beliefs (representations of desired states of the world) of an individual and thus do not vary across situations for a specific actor. The interactive relationship between individual and situational determinants can be understood in two ways. Straightforwardly, it means that those individual beliefs (and opportunities) guide behavior which are activated by certain situational cues. This differs according to the overall set of beliefs internalized by the individual. However, if some descriptive or evaluative beliefs are strongly internalized and thus chronically active, they can prompt a certain definition of the situation (and thus action) irrespective of the situational conditions given (possible misperception). As mentioned above, we assume descriptive and evaluative beliefs to be tied to social-structural determinants. In accordance with structural individualism, sociological factors such as socio-economic or demographic attributes are reflected in individual determinants. Hence, social-structural groups are expected to be similar in terms of certain beliefs. Overall, the model emphasizes that OA does neither result from characteristics of the individual, nor from characteristics of the situation, but rather from the interplay of these two.

Empirical approach
The empirical study seeks to collect data on individual, situational, and social-structural determinants of OA behavior. Therefore, we intend to conduct an online survey in German-speaking Switzerland with four different groups: frequent OA commentators, occasional OA commentators, non-OA commentators, and non-commentators. Group-differences in determinants, then, allow to assess determinants' relative effect on OA behavior. However, sampling OA commentators is not easy because it is a relatively rare behavior. Thus, we apply an elaborate, two-step sampling strategy: First, in order to sample OA and non-OA commentators, we use the unique opportunity to collaborate with a large Swiss media corporation. We will use a large dataset of news comments submitted to its website (including meta-data such as time of submission). The dataset includes moderated comments: comments considered as being non-aggressive by moderators (and were published in the commentary section) and comments considered as aggressive (and were not published). By employing human/automated content analysis of all comments, we identify the following groups and assign all commentators to one of them: frequently aggressive commentators, occasionally aggressive commentators, and nonaggressive commentators. From each group, we invite around 1500 people to participate in the survey. Second, in order to sample persons who do not engage in online commenting at all (noncommentators), we use a random sample of the resident population of German-speaking Switzerland.

Social-Structural Determinants
Aggressive Online Behavior

Definition of the Situation
(activated beliefs, perceived constraints) Particular attention is given to data protection and the ethics of recruiting. First, all the comments and meta data received by the Swiss media corporation is principally public data, thus principally searchable and retrievable. This is because commentators submit their comments to news platforms in the knowledge that their comments get principally published (even in cases where comments are ultimately not published by moderators). Beyond, this data set is given to us in an anonymized form. Thus, privacy concerns can be excluded. Second, not the authors but the Swiss media corporation invites the commentators to participate in the survey (as the e-mail addresses of commentators are only available to the corporation but not to us). Third, by forming groups of commentators (see above) whereby individuals in each group receive group-specific online surveys, the survey data of individuals will only be connected to the affiliation to these groups but at no time to individual comments or commentators. This makes it impossible to identify single individuals in the resulting data set. Fourth, an ethics approval will be sought in the process of designing the survey. Our approach of matching online data with survey data allows to combine behavioral data with a broad range of -so far scarcely collectedindividual, social-structural, and situational determinants of OA. While individual and socialstructural determinants will mainly be measured in the survey, most situational determinants will be measured through aggregating user-generated comments and meta-data.

Conclusion
The preceding paper introduced a novel, sociologically informed theoretical framework integrating a broad set of determinants of aggressive online commenting behavior. Furthermore, it suggested an empirical strategy allowing to disentangle the effects single determinants by matching online data with survey data.