Implicit Power Drives in the Manifestos Preceding Autogenic Massacres

Abstract

I scored 23 manifestos written by perpetrators of autogenic massacre on one LIWC2015 dimension, power language. I tested the manifesto mean of 3.00 against three published LIWC baseline rates, for personal blogs, expressive writing, and a multi-source composite. In one-sample t-tests, I found t(22) = 3.47, p = .002 against blogs; t(22) = 3.66, p = .001 against expressive writing; and t(22) = 2.43, p = .024 against the composite. The manifestos scored higher than all three genres, and the first two differences remain significant after a Bonferroni correction for the three tests. I presented the study as a poster at Fielding Graduate University in 2017 to set direction for my dissertation.

Autogenic Massacres

Mullen (2004) coined the term autogenic massacre for the killing of multiple people, without a significant lapse of time between the killings, motivated by the perpetrator’s personal problems or psychopathology; the violence is self-generated and often planned in isolation. The perpetrator often leaves a written or recorded artifact meant to be read afterward, and as many as 67% of perpetrators of mass violence create meaningful pre-attack communications (Hempel, Meloy, & Richards, 1999; Silver, Horgan, & Gill, 2018). Researchers consistently find that many of these perpetrators leave documents; they agree less on what the documents share linguistically. Neuman, Assaf, Cohen, and Knoll (2015) used computer-aided text analysis to find themes of humiliation, revenge, and narcissism among people who complete these attacks, and researchers studying these crimes have repeatedly placed social status, identity, and control at the center (Harrison & Bowers, 2010; Meloy et al., 2004; Thompson & Kyle, 2005).

Mohandie (2013) proposes four categories of violent offenders within this broader population. Of the 23 writers, 8 fit Mohandie’s Type I and 14 fit Type II; I could not categorize one document. The offenses occurred between 1966 and 2016, and the authors wrote most of the texts (n = 18) between 2004 and 2016. I set aside content, theme, political affiliation, demographic background, and reported motive and looked at one LIWC dimension, power, which Pennebaker, Boyd, Jordan, and Blackburn (2015) describe as tracking words relevant to status, dominance, and social hierarchy, and which Pennebaker (2011) reads as concern with who is in control and with the social status of others. I asked whether the manifestos use power language at a different rate than baseline writing genres.

Method

I processed each manifesto with LIWC2015, scoring each document as the percent of its words in each of roughly ninety psychometric categories, and analyzed one category, power. The corpus mean was 3.00 (SD = 1.28, SE = 0.268). I treat the LIWC power category as a word-rate marker of power concern; Schultheiss (2013) found that LIWC word counts converge only modestly with content-coded implicit motives, and least for the power motive, so I read the elevation as a property of the texts.

I compared the manifestos with LIWC2015’s published baseline rates. Pennebaker and colleagues publish baseline rates for several large reference genres (Pennebaker, Boyd, Jordan, & Blackburn, 2015), and I used three of those baselines, for personal blogs (M = 2.07), expressive writing (M = 2.02), and a multi-source composite of blogs, expressive writing, natural speech, novels, newspaper text, and Twitter postings (M = 2.35). I used each as a fixed test value in a one-sample t-test of the manifesto mean, with df = 22 in every case. Because a published norm has no variance and no matching on length or genre, I read the one-sample tests as a description of how far the manifestos sit from typical writing.

Results

The manifestos scored higher than all three published baselines, by 0.65 to 0.98 percentage points. The two largest differences, against expressive writing and blogs, were significant at p = .001 and p = .002, and the narrower difference against the multi-source composite was significant at p = .024. In all three tests the 95% confidence interval for the mean difference excluded zero; after a Bonferroni correction for the three comparisons, the composite result (adjusted p = .07) is no longer significant at .05, and the other two remain significant.

Bar chart comparing mean LIWC power scores: Expressive Writing 2.02, Blogs 2.07, Multi-source Composite 2.35, and Manifestos 3.00. The manifesto bar is highlighted.

Mean LIWC power score by source. The manifestos (M = 3.00) scored higher than all three published LIWC baseline rates, expressive writing (M = 2.02, t = 3.66, p = .001), personal blogs (M = 2.07, t = 3.47, p = .002), and the multi-source composite (M = 2.35, t = 2.43, p = .024). One-sample t-tests, df = 22. Asterisks denote significance level (** p < .01, * p < .05), uncorrected.

Comparison	Test value	Mean diff	95% CI	t (df=22)	p
vs Blogs	2.07	+0.93	[0.37, 1.48]	3.47	.002
vs Expressive writing	2.02	+0.98	[0.42, 1.53]	3.66	.001
vs Multi-source composite	2.35	+0.65	[0.09, 1.20]	2.43	.024

Vocabulary in the Highest-Power Manifesto

For the poster, I built a word cloud of the most frequent words in the manifesto that scored highest on power, using Voyant Tools (Sinclair & Rockwell, 2016); word size corresponds to frequency. The first-person plural we was the most frequent word, a pattern consistent with the identification warning behavior Meloy, Hoffmann, Guldimann, and James (2012) describe, and Pennebaker (2011) discusses we as a word that can signal a speaker’s identification with a group. Around we sit words of abstract civic stakes (tyranny, tyrants, freedom, liberty, free, truth), in-group anchors (children, men), adversaries (enemy, enemies), and action (fight, stand, face, know, live, fail, fear); the singular I also appears, at a smaller size.

Word cloud from the manifesto with the highest LIWC power score. We is the largest word, followed by children, know, free, tyrants, enemy, fight, men, tyranny, liberty, and freedom.

Most frequent words in the manifesto with the highest power score; word size corresponds to frequency. I made the cloud with Voyant Tools (Sinclair & Rockwell, 2016) for the 2017 poster.

When I read the scores against the vocabulary of the highest-scoring manifesto, the elevation looks like collective positioning; the author speaks for a group, sets it against an adversary, and claims moral standing over a stake he has decided is at risk.

Findings and Limits

Power language ran at a higher rate in this manifesto corpus than in three baseline writing genres, including a multi-source composite that itself averages over genres, and there is a lesson in the design for anyone screening text against published norms; a difference that appears against several reference distributions is harder to attribute to an artifact of any one of them, so it pays to test against more than one baseline before trusting an elevation.

The study was built to find a signal worth testing at scale, and I took that test up in my dissertation with 100 communications, using this result to choose the variables. With 23 documents whose authors are already known, the result describes a known class of texts; identifying future perpetrators from their writing is a prospective question, and it needs a prospective design. In keeping with Ethical Standard 9.01(b) of the American Psychological Association (2002), the finding stays at the level of the texts, since a conclusion about an individual writer requires an assessment none of them received.

References

American Psychological Association. (2002). Ethical principles of psychologists and code of conduct. Author.

Harrison, M. A., & Bowers, T. G. (2010). Autogenic massacre as a maladaptive response to status threat. The Journal of Forensic Psychiatry & Psychology, 21(6), 916-932.

Hempel, A. G., Meloy, J. R., & Richards, T. C. (1999). Offender and offense characteristics of a nonrandom sample of mass murderers. Journal of the American Academy of Psychiatry and the Law, 27(2), 213-225.

Meloy, J. R., Hempel, A. G., Gray, B. T., Mohandie, K., Shiva, A., & Richards, T. C. (2004). A comparative analysis of North American adolescent and adult mass murderers. Behavioral Sciences & the Law, 22(3), 291-309.

Meloy, J. R., Hoffmann, J., Guldimann, A., & James, D. (2012). The role of warning behaviors in threat assessment: An exploration and suggested typology. Behavioral Sciences & the Law, 30(3), 256-279.

Mohandie, K. (2013). Threat assessment in schools. In J. R. Meloy & J. Hoffmann (Eds.), International handbook of threat assessment (pp. 126-141). Oxford University Press.

Mullen, P. E. (2004). The autogenic (self-generated) massacre. Behavioral Sciences & the Law, 22(3), 311-323.

Neuman, Y., Assaf, D., Cohen, Y., & Knoll, J. L. (2015). Profiling school shooters: Automatic text-based analysis. Frontiers in Psychiatry, 6, 86. https://doi.org/10.3389/fpsyt.2015.00086

Pennebaker, J. W. (2011). The secret life of pronouns: What our words say about us. Bloomsbury Press.

Pennebaker, J. W., Boyd, R. L., Jordan, K., & Blackburn, K. (2015). The development and psychometric properties of LIWC2015. University of Texas at Austin.

Schultheiss, O. C. (2013). Are implicit motives revealed in mere words? Testing the marker-word hypothesis with computer-based text analysis. Frontiers in Psychology, 4, 748. https://doi.org/10.3389/fpsyg.2013.00748

Silver, J., Horgan, J., & Gill, P. (2018). Foreshadowing targeted violence: Assessing leakage of intent by public mass murderers. Aggression and Violent Behavior, 38, 94-100.

Sinclair, S., & Rockwell, G. (2016). Voyant Tools [Computer software]. http://voyant-tools.org/

Thompson, S., & Kyle, K. (2005). Understanding mass school shootings: Links between personhood and power in the competitive school environment. Journal of Primary Prevention, 26(5), 419-438.