Regression, Poisson regression, or negative binomial regression for panel data, as our dependent variable is (if not transformed) a count variable, or can be transformed into a binary variable that indicates whether a person is an aggressor or not. The results are similar with the results that follow and will therefore not be presented here.ResultsIn accordance with Hypothesis 1, the data substantiate that online aggression in social media is a more frequent phenomenon than in the non-digital context. In the analyzed online petition platform we find 197,410 aggressions according to our definition. 20.62 of all comments entail a minimum of one order AZD-8055 aggressive expression (Fig 1). In 9 of all comments we find two, up to fifteen, aggressive expressions. On the petition level, only 11 of all petitions include no aggressions. 34 include a negligible amount of aggressions from 1, up to 10. 37 include 11 up to 100 aggressions. 16 include 101 up to 1,000 aggressions. 2 include 1,001, up to 25,360, aggressions. Even if the prevailing majority of commenters make no use of aggressive language in social media, the numbers demonstrate that online aggression occurs not only in a vanishing minority of comments or petitions (compared to the observed vanishing minority of max 4 of bystanders aggressively sanctioning in the non-digital context [49]). This supports the claim that in social media, aggressive sanctioning behavior is a relatively frequent phenomenon because it takes place in low-cost situations. We now move to the presence of selective incentives and intrinsically motivated actors in social media. The descriptive findings show that 47 of all petitions are accompanied by a highly controversial debate, 6 of the petitions are associated with a scandal in news media,Fig 1. Observed amount of online aggression per comment. doi:10.1371/journal.pone.0155923.gPLOS ONE | DOI:10.1371/journal.pone.0155923 June 17,11 /Digital Norm Enforcement in Online Firestormsand 26 of the commenters are motivated by fairness concerns. Social media thus indeed seem to offer an environment in which the second-order public good dilemma of norm enforcement can be overcome. Whether these conditions indeed contribute to norm enforcement is tested in Tables 1 and 2. The random-effects model in Table 1, Model 1, confirms that situations that offer selective incentives, i.e., a petition is accompanied by a highly controversial XL880 biological activity debate or is connected with a scandal in news media, significantly encourage online aggression in comments. This preliminarily supports Hypothesis 2 (for the size of the effects see Figs 2 and 3). The fixed-effect model in Table 2 entails no results for selective incentives because petition-invariant effects are dropped. Further, the random-effects as well as the fixed-effects models in Tables 1 and 2,Table 1. Predicted amount of online aggression dependent on the anonymity of aggressors (random-effects regression). Model 1 Y: Amount of online aggression (log) Anonymity Controversy of accusation Accusation is connected to a scandal Intrinsic motivation (log) Anonymity x Controversy Anonymity x Scandal Anonymity x Intrinsic motivation Length of comment in words Time of comment after petition opening Number of protest participants (log) Scope of protest Success of the petition Status of the accused (log) Accused is a natural person (vs. legal entity) Anonymity of social environment of aggressors (log) Motives: Income/minimization of costs Motive: Se.Regression, Poisson regression, or negative binomial regression for panel data, as our dependent variable is (if not transformed) a count variable, or can be transformed into a binary variable that indicates whether a person is an aggressor or not. The results are similar with the results that follow and will therefore not be presented here.ResultsIn accordance with Hypothesis 1, the data substantiate that online aggression in social media is a more frequent phenomenon than in the non-digital context. In the analyzed online petition platform we find 197,410 aggressions according to our definition. 20.62 of all comments entail a minimum of one aggressive expression (Fig 1). In 9 of all comments we find two, up to fifteen, aggressive expressions. On the petition level, only 11 of all petitions include no aggressions. 34 include a negligible amount of aggressions from 1, up to 10. 37 include 11 up to 100 aggressions. 16 include 101 up to 1,000 aggressions. 2 include 1,001, up to 25,360, aggressions. Even if the prevailing majority of commenters make no use of aggressive language in social media, the numbers demonstrate that online aggression occurs not only in a vanishing minority of comments or petitions (compared to the observed vanishing minority of max 4 of bystanders aggressively sanctioning in the non-digital context [49]). This supports the claim that in social media, aggressive sanctioning behavior is a relatively frequent phenomenon because it takes place in low-cost situations. We now move to the presence of selective incentives and intrinsically motivated actors in social media. The descriptive findings show that 47 of all petitions are accompanied by a highly controversial debate, 6 of the petitions are associated with a scandal in news media,Fig 1. Observed amount of online aggression per comment. doi:10.1371/journal.pone.0155923.gPLOS ONE | DOI:10.1371/journal.pone.0155923 June 17,11 /Digital Norm Enforcement in Online Firestormsand 26 of the commenters are motivated by fairness concerns. Social media thus indeed seem to offer an environment in which the second-order public good dilemma of norm enforcement can be overcome. Whether these conditions indeed contribute to norm enforcement is tested in Tables 1 and 2. The random-effects model in Table 1, Model 1, confirms that situations that offer selective incentives, i.e., a petition is accompanied by a highly controversial debate or is connected with a scandal in news media, significantly encourage online aggression in comments. This preliminarily supports Hypothesis 2 (for the size of the effects see Figs 2 and 3). The fixed-effect model in Table 2 entails no results for selective incentives because petition-invariant effects are dropped. Further, the random-effects as well as the fixed-effects models in Tables 1 and 2,Table 1. Predicted amount of online aggression dependent on the anonymity of aggressors (random-effects regression). Model 1 Y: Amount of online aggression (log) Anonymity Controversy of accusation Accusation is connected to a scandal Intrinsic motivation (log) Anonymity x Controversy Anonymity x Scandal Anonymity x Intrinsic motivation Length of comment in words Time of comment after petition opening Number of protest participants (log) Scope of protest Success of the petition Status of the accused (log) Accused is a natural person (vs. legal entity) Anonymity of social environment of aggressors (log) Motives: Income/minimization of costs Motive: Se.
