Politician hate speech and domestic terrorism

ABSTRACT

Does hate speech – rhetoric that targets, vilifies or is intended to intimidate minorities and other groups in society – fuel domestic terrorism? This question is, unfortunately, relevant given the convergence of the use of hate speech by political figures and domestic terrorist incidents in a variety of countries, including the United States. In this study I theorize that hate speech by politicians deepens political polarization and that this, in turn, produces conditions under which domestic terrorism increases. I test this proposition using terrorism and hate speech data for 135 to 163 countries for the period 2000 to 2017. I produce two findings. First, hate speech by political figures boosts domestic terrorism. Second, the impact of political hate speech on domestic terrorism is mediated through increased political polarization.

KEYWORDS:

• Terrorism
• Hate Speech
• Political Polarization
• Empirical Analysis
• Hate Speech
• polarization

Supplementary Material

Supplemental data for this article can be accessed on the publisher’s website.

Notes

¹ Terrorism is defined in this study as the deliberate, premeditated use of violence, or the threat of violence, by nonstate actors that is politically motivated and is intended to influence a wider audience beyond individuals affected directly by the attack.

² Source: frequency distributions calculated by the author using data from Mechkova et al. (2019) and the Global Terrorism Database.

³ Fiorina and Abrams (2008) enumerate significant nuances in the conceptualization and measurement of political polarization. In this study, my conception of political polarization aligns with what Iyengar, Sood, and Lelkes (2012) term “affective polarization” or what (Mason 2015) refers to as “social polarization.” That is, it is both a mass and elite phenomenon involving deep divisions of opinion about a range of political and social topics. This conception of political polarization conforms to the operationalization of the indicator for political polarization used in the analysis (see Mechkova et al. 2019, 21).

⁴ Terrorism differs from these other forms of political violence because it is low-intensity, unconventional and intended to influence an audience rather than to score a battlefield victory – unlike violence that typically is used by insurgent groups in civil wars – and is premeditated rather than impulsive or spontaneous – unlike hate crime.

⁵ Note, in the studies cited here, political polarization is linked to increased domestic terrorism of all ideological types. Political polarization is associated with societal ideological clashes that prompt individuals and groups to engage in terrorism against political opponents, outgroups and other demonized targets. As a consequence, my expectation is that political polarization boosts domestic terrorist incidents generally. To empirically check this, I ran naïve estimations regressing the measure of political polarization within countries to counts of right-wing, left-wing, nationalist-separatist and religious-Islamist terrorism. I found that political polarization is a significant, positive predictor of all of these types. (Results available from the author.) This gives me further assurance that political polarization is associated with increases of terrorism in general, and of all types.

⁶ The temporal limitations of the study are set by data availability. Data on terrorist attacks, the dependent variable, are available only through 2017. Data on hate speech, the independent variable, is available only from 2000 on. A list of countries included in the analysis is presented in Appendix Table 1.

⁷ A domestic terrorist attack is defined as a terrorist attack occurring with the boundaries of one country that is perpetrated by a citizen or resident of the country against domestic targets or victims.

⁸ Data and codebook for the GTD can be found at: https://www.start.umd.edu/gtd/.

⁹ To check for outlier effects, I collapsed the counts of terrorist attacks into a dichotomous measure coded 1 for observations containing one or more attack and zero for observations containing no terrorist attacks. I then reran the main model (model 2) using a logistical regression technique. This produced the same substantive results as the main model. Results summarized in Appendix Table 2.

¹⁰ Variable: “v2smpolhate_ord” from (Mechkova et al. 2019, 26). The question used to develop this variable reads, “How often do major political parties use hate speech as part of their rhetoric?” Note, it would be useful to compare the Political Parties Hate Speech indicator – a measure developed by country expert ratings, as explained in the next footnote – with some sort of count or content measure of actual hate speech messages from politicians to improve confidence in the findings. However, to my knowledge, no cross-national database of hate speech messages yet exists. This would be a very useful tool for future study of the relationship between hate speech and terrorism.

¹¹ The hate speech measure developed by Mechkova et al. (2019) is produced using the V-Dem expert panel coding system. Coppedge, et al. (2019) explain this process in detail. Evaluative indicators in V-Dem are produced using multiple ratings by groups of country experts. Country experts are recruited by V-Dem staff for each country in the database through a rigorous process that involves competitive selection from a larger pool based upon the experts’ backgrounds, levels of expertise, diversity of experience, and assessment of their impartiality. Around 60 percent of country experts are nationals or permanent residents of the country they evaluate. For each indicator, a minimum of five country experts provide an independent rating. Country experts provide a rating for each country-year observation independently. For ordinal-level indicators like those used in this study, experts are provided detailed vignettes (similar to rubrics) describing each ordinal level to allow the experts to match their rating decision to the situation in the country for that year. Moreover, to ensure that expert ratings are consistent across countries and years, V-Dem employs “bridge coding” whereby some experts produce ratings across multiple countries, or multiple years, for an indicator. These bridge codes are then used to check the ratings of experts who have coded an indicator for only one country or one year at a time. Once experts submit their ratings, V-Dem combines them using a an IRT (item response theory) model that takes into account patterns of cross-expert agreement and disagreement. Specifically, the expert ratings are combined into an, “ … integerized median ordinal highest posterior probability category over measurement model output (Coppedge et al. 2019, 143).” Also, V-Dem has experts fill out a post-survey questionnaire which is used to identify sources of attitudinal or other biases.

¹² Major political parties use hate speech: 4 = extremely often; 3 = often; 2 = sometimes; 1 = rarely; 0 never or almost never.

¹³ Variable “v2smpolsoc.ord” from (Mechkova et al. 2019, 21). Question: “How would you characterize the differences of opinion on major political issues in society?.”

¹⁴ 4 = serious polarization; 3 = moderate polarization; 2 = medium polarization; 1 = limited polarization; 0 = no polarization.

¹⁵ Derived from “Government Censorship Effort – Media” (v2mecenefm) from Permstein et al. (2019, 21). To ease interpretation, I invert the ordinal sale for government censorship of media so that 4 = government “attempts to censor are direct and routine,” 3 = attempts are “indirect but nevertheless routine,” 2 = attempts are “direct but limited to especially sensitive issues,” 1 = attempts are “indirect and limited to sensitive issues,” and 0 = “government rarely attempts to censor major media in any way.”

¹⁶ Source: United Nations National Accounts. Available online at: https://unstats.un.org/unsd/snaama/.

¹⁷ Source: Alesina et al. (2003). Data obtained from the Quality of Government Database, which provided updated data for entire timeseries. Variable=”al_ethnic.”

¹⁸ Source: UCDP/PRIO. Variable = “ucdp_type3, Internal Armed Conflict.”

¹⁹ Note, as a check I reran all models excluding terrorist attacks from the previous year. These produce the same results as those in the main analysis. Results available from the author.

²⁰ As a check, I also ran the estimations without lagging the independent variables and with lagging them by two years instead of one. These reproduced the main findings of the study. I have published the results for the main model (model 2) in Appendix Table 3. (Full results available from author.)

Table 3. Endogeneity Tests.

²¹ Calculated using post-estimation marginal effects simulations of a one-unit increase of hate speech while holding all other covariates constant at their mean values. Note, the substantive impact of a one-unit increase in hate speech in model 4, which excludes the lagged dependent variable as a predictor, is 3.071 more domestic terrorist attacks.

²² Note, Figure 1 presents the marginal effects simulations for the main estimation: model 2.

²³ It is also possible that hate speech and polarization are endogenous to one another. Statistically, they are correlates (ρ =.590). In the paper, I theorize that hate speech increases polarization to produce conditions under which domestic terrorism is more likely. However, it is possible that hate speech by politicians is more common, and acute, in environments that are already highly polarized, and that this might complicate the relationship between hate speech and terrorism. I empirically check for this and do not find it to be the case. I split the sample into sets of observations characterized by “low” and “high” polarized environments. The former consists of observations where polarization is completely absent or is “limited.” The latter consists of observations where polarization is “medium,” “moderate” or “serious.” I then reran the main estimations on the two sets. I found that hate speech drives domestic terrorism in both “low” and “high” polarized environments. To check further, I split the sample again and examined the impact of hate speech on terrorism in countries with the highest two ordinal categories of polarization: “moderate” and “serious.” I found, again, that in these highly polarized environments, hate speech increased domestic terrorism. These results are summarized in Appendix Table 5. The results give me confidence that though hate speech and polarization are possibly related, the impact of hate speech on terrorism is not due to a wider environment of elevated polarization.

²⁴ Statistical checks provide support for this logical assumption. While weak anti-defamation or hate speech laws in countries statistically predict subsequent hate speech, such laws do not statistically predict subsequent domestic terrorist attacks. This is demonstrated in Appendix Table 4.

²⁵ Mechkova (2019, 21) “Defamation protection” (v2smlawpr).

²⁶ To ease interpretation, I inverted the scale so that 4 = No. The law provides no protection against Internet defamation and hate speech; 3 = Not really; 2 = Somewhat; 1 = Mostly; 0 = Yes.

²⁷ To produce the results in Table 4, I used a Stata package (“medeff”) developed by Hicks et al. (2011).

²⁸ “medeff” (Hicks and Tingley 2011), which is based upon the technique for mediation developed in Imai, Keele and Tingley (2010) and Imai, Keele, Tingley and Yamamoto (2010).

²⁹ Moreover, as previously noted, to check for outlier effects I ran a logit analysis using a dichotomous measure of terrorism and found it to produce the same substantive results as the main estimations in the study. See Appendix Table 2.

³⁰ Results available from author.

³¹ β =.684***.

³² β =.486***.

³³ β =.387*** presented in parentheses in Figure 3.

³⁴ β =.064 presented not in parentheses in Figure 3.

³⁵ Average mediation =.071***.

³⁶ Percentage of Total Effect Mediated =.839411.

³⁷ Stata command “medsens”.

³⁸ Rho at which ACME = 0.

³⁹ All of these results are available from the author.

⁴⁰ Unfortunately, scholars lack reliable cross-national, time-series data on lone perpetrators. The latest iteration of the GTD 2019 does include a variable that identifies attacks by individuals unaffiliated with terrorist groups (“individual”). However, the GTD codebook includes a disclaimer stating that the individual variable reliably identifies only a portion of lone perpetrators and is affected by significant missingness. Preliminary descriptive statistics suggests that data missingness for the “individual” variable is not random. For example, there are only 223 lone perpetrator attacks in the GTD (comprising around 2.14 percent of attacks since 2000) and around 21.5 percentage of these occurred in just one country: the United States. This suggests to me that due to coding challenges and missingness, the distribution of lone perpetrator attacks in GTD is highly skewed.