Using a dataset containing the geo-tagged hate tweets for Italy in 2017, we have analyzed which are the local determinants of online hate.We have considered different types of variables: demographic, economic, geographic, political and institutional. With respect to institutional variables, alongside measures for trust and cooperation, we have introduced one of the most basic and persistent institutional forms -family types- (Duranton et al., 2010) and two measures for deviant behaviors: -offline hate behaviors and vaccine hesitancy.
Our dependent variable, the number of hate tweet per inhabitants in each LLMA- is a continuous variable with several zeros. Therefore, as suggested by the literature (i.a.Cameron and Trivedi, 2009; Woolridge, 2002), the most suitable model specifications are 2-Part-Model and Sample Selection Model. Considering the vigorous and ongoing debate opposing advocates of the two competing frameworks, we have decided to follow a balanced approach, consisting of estimation both models first and then applying several specification tests to get some gauge on the relative strength and weakness of each set-up.
Although our SSM specification, entailed with exclusion restriction and estimated according to two-step procedure, appears not to suffer from multicollinearity, in the end we get that the 2PM specification is characterized by a higher predictive power.
Our findings highlight that, under the 2PM specification:
- The set of regressors determining the predicted probability of online hate eventis partly different from the set of regressors influencing the expected number of online hate events;
- Human capital effects are consistent with self-interest theory: an increase in the share of educated people increases both the predicted probability of online hate event and theexpected number of online hate events; however, when it comes to the latter, the effect of human capital is moderated by income inequality: the more unequal the LLM, the more self-interest theory is at works.
- Family types influenceboth the predicted probability and theexpected number of online hate events, mainly through the transmission of the self-interest value: the more selfish the family, the more expected online hate is fueled. This finding fosters the role of self-interestin boosting online hate manifestations.
- Deviant behaviors exert a roleon online hate events: vax hesitancy has a positive influence on the predicted probability of online hate, the same holds for offline hate with respect to expected amount of online hate tweets.
- Geography matterson the expected number of online hate events: central areas reduce the expected amount of hate events, whereas the presence of a hub exerts the opposite effect.
- Online deviant behavior is triggered by situational features that are distinct from the one fueling offline hate, consistently with the literature.
Noteworthy, many findings remain consistent also under the SSM specification.
We have performed robustness checks and postestimation diagnostics for each model specification