CASE STUDY / Fielding Graduate University · Doctor of Philosophy in Psychology · 2019 / Predictive Typology

Instrumental and Affective Mass Murder: Establishing a Predictive Typology with Computer-Mediated Linguistic Analysis

In its 2015 report, the FBI Behavioral Analysis Unit classifies perpetrators of mass violence as instrumental offenders; I tested whether the language of a discovered pre-attack communication separates an affective type from an instrumental type, and whether the two types attack differently.

Read Time10 min read

Abstract

In its 2015 report, the FBI Behavioral Analysis Unit classifies perpetrators of mass violence as instrumental offenders (National Center for the Analysis of Violent Crime, 2015); I tested whether the language of a discovered pre-attack communication separates an affective type from an instrumental type, and whether the two types attack differently. I scored 100 pre-attack communications with LIWC2015 on four summary dimensions (Analytic, Clout, Authentic, and Tone) and clustered them with K-means into two types, Affective (n = 53) and Instrumental (n = 47). I confirmed the partition with a cross-validated linear discriminant analysis, tested cluster membership against how each attack unfolded, and regressed membership on four attack-behavior variables (victims killed, demographic-based targeting, the perpetrator’s home among the attack locations, and the perpetrator being killed by another person). The refined model reached McFadden R² = 0.24; for the strongest single predictor, the perpetrator being killed by another person, the odds ratio was 14.11 toward the Instrumental type. The result is a within-groups typology of mass murderers with statistically tested ties to specific attack characteristics, built for the moment a threat assessment team has a pre-attack communication in hand and has to decide what kind of attack to plan against.

The full dissertation (81 pages, Fielding Graduate University, 2019) is available as a PDF.

The Prediction Problem

Researchers have built three kinds of typology; some distinguish mass murderers from the general population, some from other violent offenders, and some distinguish kinds of mass murderers from each other. Researchers have drawn some predictive value from typologies of the first kind for whether a person is at risk of attacking (Knoll & Meloy, 2014; Langman, 2009; Verlinden, Hersen, & Thomas, 2000), though only for attack likelihood in general terms. Researchers first proposed within-groups types from theory; Dietz (1986) named three types, Holmes and Holmes (1994) added two, and Kelleher (1997) proposed seven categories organized by motive. Researchers building more recent empirical within-groups typologies (Langman, 2009; Ioannou, Hammond, & Simpson, 2015) sorted attackers after the fact without linking type to how an attack unfolds. I set out to build a within-groups distinction with statistically tested ties to specific attack behaviors.

In threat assessment, practitioners separate threat (the probability an event will occur) from risk (the potential impact if it does). A burglary of an unoccupied home is low-threat and low-risk; a school shooting is low-threat and high-risk, with low probability for any given school and catastrophic impact when it happens; an armed individual with credible communicated intent at an accessible location is high-threat and high-risk. I left the threat side of the matrix, whether an attack will occur, to other methods and worked the risk side, what kind of attack is more likely once a team has discovered a pre-attack communication, by treating that communication as a measurable psychometric signal.

Risk (impact of event) ↑
Low threat / high risk
A school shooting in the abstract; dire consequences for any one school, but a low base rate.
High threat / high risk
An armed individual with credible communicated intent at an accessible location; usually calls for immediate tactical intervention.
Low threat / low risk
Burglary of an unoccupied home; low probability for any given home, limited impact.
High threat / low risk
High-probability events with minimal destructive impact; procedural deterrence is the standard response.
Threat (probability of event) →

With the two types defined, an assessor can read a discovered communication for the kind of attack that is more likely.

Language as Psychological Evidence

As many as 67% of perpetrators of mass violence create meaningful pre-attack communications (Hempel, Meloy, & Richards, 1999; Meloy, Hempel, Mohandie, Shiva, & Gray, 2001; Silver, Horgan, & Gill, 2018). Most perpetrators do not survive their attack, and their communications remain; by scoring those communications with computer-mediated linguistic analysis, I could characterize this population psychologically, in a standardized way, after the perpetrator’s death.

I also set aside the FBI’s definition of mass murder, four or more killed without a significant cooling-off period, because the FBI built a success component into that definition, and success depends on factors outside the perpetrator’s psychology. In 2014 Geddy Kramer shot six coworkers at a FedEx facility and killed only himself; he was prepared and committed to a mass attack, and his victims survived because of fast emergency response, rapid access to medical care, and his own choice of less-lethal ammunition. Under the FBI definition, he does not count. I defined a mass murderer as “any perpetrator who presents as prepared and motivated to murder four or more individuals, with insubstantial time passing between violent acts,” a definition that includes Kramer, and I coded attack success as a typology variable.

I collected 100 pre-attack communications from people who carried out or attempted mass violence between 1966 and 2019. I drew the communications from SchoolShooters.info, curated by Peter Langman, and from social media, and I gathered incident details from the Stanford Mass Shootings of America database (Stanford University, n.d.), the Mother Jones dataset of US mass shootings (Follman, Aronsen, & Pan, 2019), and the Gun Violence Archive (2019). I scored each communication with LIWC2015 (Pennebaker, Booth, Boyd, & Francis, 2015) on four summary variables (Analytic, Clout, Authentic, and Tone), and I used K-means clustering, nonparametric tests, and binary logistic regression to handle the small sample and the mix of categorical and continuous variables.

100 Communications analyzed
53 Affective type
47 Instrumental type
0.24 McFadden R² (refined model)
VariableHigh ScoreLow Score
AnalyticFormal, logical, hierarchical thinkingPersonal, narrative, present-focused
CloutConfident, expert, authoritativeTentative, humble, anxious
AuthenticHonest, personal, disclosingGuarded, distanced, evasive
TonePositive, upbeat emotional valenceAnxious, sad, or hostile

A Detectable Divide

I ran K-means cluster analysis on the four LIWC summary variables with k = 2, a number I set in advance from the two-type distinction in the violence literature. With a linear discriminant analysis on the same four variables, I classified all 100 cases into their K-means cluster under leave-one-out cross-validation. The structure correlations were largest for Clout (r = .88), Authentic (r = -.80), and Analytic (r = .64), and moderate for Tone (r = .48). Because I ran the discriminant analysis on the variables that formed the clusters, I read the 100% agreement as confirmation of a clean partition; I test the types against external criteria in the attack-outcome analyses that follow.

Writers in Cluster 1, the Affective Type (n = 53), scored lower on Analytic, Clout, and Tone and higher on Authentic; they wrote in a personal, present-focused, narrative, and tentative voice, often saturated with hostility and despair. Writers in Cluster 2, the Instrumental Type (n = 47), showed the inverse profile; they wrote in a formal, logical, hierarchical, confident, and guarded voice, closer to an ideological treatise than to a personal grievance. I drew the two labels from the distinction in the violence literature between emotionally driven and goal-driven offending (Dodge, 1991; Glenn & Raine, 2009; McEllistrem, 2004; Meloy, 1988) and applied that distinction within the mass-murder population.

Affective type (n = 53)Instrumental type (n = 47)
33.2 (SD 19.0)
Analytic
59.3 (SD 22.1)
15.7 (SD 14.9)
Tone
33.6 (SD 24.8)
33.4 (SD 15.6)
Clout
68.3 (SD 15.9)
73.9 (SD 17.5)
Authentic
38.5 (SD 21.4)
Mean LIWC2015 summary score by cluster, bars on a common 0 to 100 scale (dissertation Table 1). The within-cluster SDs are large, so individual communications overlap even where the cluster means differ.

Formulating the Six Hypotheses

I tested six hypotheses linking the clusters to attack outcomes; in each, I asked whether the language in a pre-attack communication relates to how the event unfolds.

Odds-ratio gauge showing OR 7.65 for how the attack ends. The arc shows relative magnitude only.
H5, how the attack ends (OR 7.65, p = .006). Affective attackers had 7.65 times the odds of ending the attack by suicide or apprehension, compared with being killed by another person. The arc shows relative magnitude only and has no calibrated scale.
Odds-ratio gauge showing OR 2.7 (inverse of 0.37) for personal connection to the attack location. The arc shows relative magnitude only.
H3, personal connection to the attack location (OR 0.37, inverted 2.7, p = .032). Affective attackers had 2.7 times the odds of striking a location, such as a school or workplace, with which they had a personal history. The arc shows relative magnitude only and has no calibrated scale.
HypothesisTestStatisticResult
H2, Victims killedMann-Whitney UU = 908, z = -2.35p = .019
H3, Location connectionFisher’s exactOR = 0.37 (inv. 2.7)p = .032
H4, TargetingFisher’s exactOR = 2.09p = .075 (n.s.)
H5, How the attack endsFisher’s exactOR = 7.65p = .006

Affective type (n = 53) compared with Instrumental type (n = 47); alpha = .05, uncorrected. The Limitations section reports the corrected result.

Predictive Signal Matrix

My first binary logistic regression used the raw hypothesis variables (victims killed, connection, targeting, and perpetrator killed) and was significant (χ²(4) = 30.58, p < .001, McFadden R² = 0.22). I had coded targeting and connection too coarsely, so in a second model I refined both codings. In the first model I coded targeting as yes or no without distinguishing specific-individual targeting from demographic targeting, and I coded connection without distinguishing the perpetrator’s home from a less intimate site such as a school or workplace; in the refined model, targeting became demographic-based targeting and connection became the perpetrator’s home among the attack locations. The refined model fit better (χ²(4) = 32.52, p < .001, McFadden R² = 0.24), and each of the four predictors was significant when controlling for the others. Three predictors are associated with higher odds of the Instrumental type, and a home among the attack locations with higher odds of the Affective type.

Perpetrator killed by another person
OR 14.11+1,311% odds · p = .003
Perpetrator's home among attack locations
OR 0.20 (inv. 5.0)+400% odds · p = .022
Demographic-based targeting
OR 4.83+383% odds · p = .008
Each additional victim killed
OR 1.14+14% odds · p < .001
Raises the odds of the Instrumental type Raises the odds of the Affective type
Refined model with all four predictors entered together (χ²(4) = 32.52, p < .001, McFadden R² = 0.24), fit to the 100 cases that defined the clusters. Bar length is log-scaled.

Operational Manifestations

Matthew Murray and Anders Breivik sit at opposite poles of the two clusters. Murray wrote at the Affective limit; his pre-attack postings are personal and emotionally raw, dense with I and you. Breivik wrote a 1,500-page manifesto at the Instrumental limit, structured as a political treatise with no personal emotional disclosure; even in its introduction he avoided I and wrote of we, us, and our.

Affective type
Matthew Murray
Arvada and Colorado Springs, 2007
LIWC2015 summary scores
Analytic 18.81
Tone 1.00
Clout 46.47
Authentic 89.35
Attack outcomes
Victims killed
4
Victims injured
5
Personal connection to location
Yes
Attack ended in
Suicide
Instrumental type
Anders Breivik
Oslo and Utøya, 2011
LIWC2015 summary scores
Analytic 83.47
Tone 23.40
Clout 67.32
Authentic 26.71
Attack outcomes
Victims killed
77
Victims injured
319
Personal connection to location
No
Attack ended in
Apprehension

Threat Management Efficacy

The typology belongs on the risk side of threat assessment, the impact side; once your team has a discovered communication, an assessor can score it and use the type to estimate which kind of attack is more likely, and here is what that looks like in practice.

A university threat assessment team might receive what appears to be a pre-attack communication written by a student. If the team scores the text on the LIWC summary variables, applies the cluster model, and classifies the author as Affective, the team has evidence that the author is more likely to attack a location with a personal connection (potentially the campus itself), more likely to end the attack by suicide or apprehension, and more likely to be in an acute emotional state, so the team can harden the connected location and lead with mental-health de-escalation.

If the team classifies the author as Instrumental, it plans for a lower likelihood of a personal connection to the attack location, more victims killed, demographic-based targeting, and an attack more likely to end with the perpetrator killed by another person; teams staff the two postures with different resources, personnel, and timelines.

Limitations

With 100 cases, the study had low power to detect small and moderate effects, and the sample holds only perpetrators who left a communication to score, so Stephen Paddock (Las Vegas, 2017) and others who left none fall outside it. In keeping with Ethical Standard 9.01(b) of the American Psychological Association (2002), the typology characterizes writing by people who were never assessed directly, and it serves a team as a statement about the communication in front of them; a conclusion about the author is a separate act that requires an assessment.

The discriminant analysis, the attack-outcome tests, and both regressions were fit on the same 100 cases, and two predictor codings were refined after the first model; the next study is out-of-sample validation on new communications, and a team adopting the typology should expect the effect sizes to shrink there. The four hypothesis tests ran uncorrected; at a Bonferroni threshold of .0125 (.05 / 4), only H5 remains significant, and H4 did not reach significance at .05. The design grew out of a 2017 study in which I found elevated power language in the manifestos of mass murderers (Leffew, 2017).

Within those limits, attackers whose writing scores Affective and attackers whose writing scores Instrumental carry out attacks that differ in lethality, location, and ending.

References

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

Dietz, P. E. (1986). Mass, serial and sensational homicides. Bulletin of the New York Academy of Medicine, 62, 477-491.

Dodge, K. A. (1991). The structure and function of reactive and proactive aggression. In D. J. Pepler & K. H. Rubin (Eds.), The development and treatment of childhood aggression (pp. 201-218). Lawrence Erlbaum Associates.

Follman, M., Aronsen, G., & Pan, D. (2019, February 15). US mass shootings, 1982-2019: Data from Mother Jones’ investigation. Mother Jones. https://www.motherjones.com/politics/2012/12/mass-shootings-mother-jones-full-data/

Glenn, A. L., & Raine, A. (2009). Psychopathy and instrumental aggression: Evolutionary, neurobiological, and legal perspectives. International Journal of Law and Psychiatry, 32(4), 253-258.

Gun Violence Archive. (2019). Mass shootings. https://www.gunviolencearchive.org/mass-shooting

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.

Holmes, R. M., & Holmes, S. T. (1994). Murder in America. Sage.

Ioannou, M., Hammond, L., & Simpson, O. (2015). A model for differentiating school shooters characteristics. Journal of Criminal Psychology, 5(3), 188-200.

Kelleher, M. D. (1997). Flash point: The American mass murderer. Praeger.

Knoll, J. L., & Meloy, J. R. (2014). Mass murder and the violent paranoid spectrum. Psychiatric Annals, 44(5), 236-243.

Langman, P. (2009). Rampage school shooters: A typology. Aggression and Violent Behavior, 14(1), 79-86.

Leffew, H. T. (2017, June). Implicit power drives in the manifestos preceding autogenic massacres [Poster presentation]. Fielding Graduate University School of Psychology Summer Session, Alexandria, VA, United States.

McEllistrem, J. E. (2004). Affective and predatory violence: A bimodal classification system of human aggression and violence. Aggression and Violent Behavior, 10(1), 1-30.

Meloy, J. R. (1988). The psychopathic mind: Origins, dynamics, and treatment. Jason Aronson.

Meloy, J. R., Hempel, A. G., Mohandie, K., Shiva, A., & Gray, B. T. (2001). Offender and offense characteristics of a nonrandom sample of adolescent mass murderers. Journal of the American Academy of Child and Adolescent Psychiatry, 40, 719-728.

National Center for the Analysis of Violent Crime, Behavioral Analysis Unit. (2015). Making prevention a reality: Identifying, assessing, and managing the threat of targeted attacks. U.S. Department of Justice, Federal Bureau of Investigation.

Pennebaker, J. W., Booth, R. J., Boyd, R. L., & Francis, M. E. (2015). Linguistic Inquiry and Word Count: LIWC2015 [Computer software]. Pennebaker Conglomerates.

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.

Stanford University. (n.d.). Stanford Mass Shootings of America database. Stanford Geospatial Center. https://github.com/StanfordGeospatialCenter/MSA

Verlinden, S., Hersen, M., & Thomas, J. (2000). Risk factors in school shootings. Clinical Psychology Review, 20(1), 3-56.

Related Works

Next Publication
Loading...
→