CASE STUDY / Fielding Graduate University · Doctor of Philosophy in Psychology / 14 min read

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

The FBI's Behavioral Analysis Unit has held since 2015 that every perpetrator of mass violence is best categorized as an instrumental offender; this dissertation takes the opposite position, and asks whether the language of a discovered pre-attack communication can sort attackers into two psychometrically distinct types that predict how the attack itself unfolds.

Read Time5 min read

Abstract

The FBI’s Behavioral Analysis Unit has held since 2015 that every perpetrator of mass violence is best categorized as an instrumental offender; this dissertation takes the opposite position. One hundred pre-attack communications were analyzed with LIWC2015 along four summary dimensions (Analytic, Clout, Authentic, Tone); K-means clustering separated the corpus into two psychometrically distinct types, Affective (n=53) and Instrumental (n=47), and linear discriminant analysis correctly classified 100% of cases. A binary logistic regression on four refined attack-behavior variables (lethality, demographic-based targeting, perpetrator’s home among attack locations, perpetrator killed by another person) recovered cluster membership at McFadden R² = 0.24, with the strongest single predictor, perpetrator killed, carrying an odds ratio of 14.11. What this offers is the first within-groups typology of mass murderers with statistically validated predictive ties to specific attack characteristics, meant as input to threat assessment and risk management once a pre-attack communication has been discovered.

The Prediction Problem

Theoretical Background

Existing typologies of mass violence tend to fall into three categories; those that distinguish mass murderers from the general population, those that distinguish them from other violent offenders, and those that distinguish kinds of mass murderers from each other. The first two have produced workable likelihood-of-attack screens (Knoll & Meloy, 2014; Langman, 2009; Verlinden, Hersen, & Thomas, 2000); the third has not. Theoretical proposals from Dietz (1986), Holmes and Holmes (1994), and Kelleher (1997) have not been empirically replicated, and recent attempts at empirical within-groups typology (Langman, 2009; Ioannou, Hammond, & Simpson, 2015) sort attackers ex post facto without predicting how an attack will unfold; what the literature has not yet had is a within-groups distinction with statistically validated ties to specific attack behaviors, and that gap is where this work sits.

In threat assessment, the standard frame separates threat (the probability an event will occur) from risk (the potential impacts if it does). A burglary of an unoccupied home is low-threat and low-risk; a school shooting is low-threat and high-risk, carrying 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. This dissertation leaves the threat side of the matrix alone, the question of whether an attack will occur, and works the risk side instead, the question of what kind of attack is more likely once a pre-attack communication has been discovered, by treating that communication as a measurable psychometric signal.

Threat (Probability of Event) →

Risk (Impact of Event) →

Low Threat / High Risk

A school shooting in the abstract. Dire consequences for any individual school, but low base rate.

High Threat / High Risk

An armed individual with credible communicated intent attacking an accessible location. The case that tends to call 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.

Separating attackers into affective and instrumental types lets the linguistic signals in a discovered pre-attack communication speak to what kind of attack is more likely, instead of waiting on the event itself to disclose its character.

Language as Psychological Evidence

Methodology

The methodological premise rests on a finding that as many as 67% of perpetrators of mass violence create meaningful pre-attack communications (Hempel, Meloy, & Richards, 1999; Meloy, Hempel, Mohandie, Shiva, & Gray, 2001; Langman, 2013; Silver, Horgan, & Gill, 2018); most perpetrators do not survive their attack, but the communications do, and computer-mediated linguistic analysis on those communications is the one route I could find to a standardized, posthumous psychological characterization of this population, which is the field’s central methodological constraint.

I also set aside the FBI’s definition of mass murder, four or more killed without significant cooling-off period, because that definition embeds a success component that lives outside the perpetrator’s psychology. Geddy Kramer, who at a FedEx facility in 2014 fired on six coworkers but killed only himself, was prepared and committed to a mass attack; his victims survived because of emergency response speed, hospital proximity, and his own choice of less-lethal ammunition, and the FBI definition excludes him on those grounds. The definition used here, “any perpetrator who presents as prepared and motivated to murder four or more individuals, with insubstantial time passing between violent acts,” includes him, and treats lack of success as a typology factor instead of a disqualifier.

100 pre-attack communications were collected from individuals who carried out or attempted mass violence between 1966 and 2019. Sources included SchoolShooters.info (curated by Peter Langman), the Stanford Mass Shootings of America database, the Mother Jones US mass shootings dataset, and the Gun Violence Archive. LIWC2015 (Pennebaker, Booth, Boyd, & Francis, 2015) generated standardized percentile scores along four summary variables (Analytic, Clout, Authentic, and Tone), and K-means clustering and binary logistic regression handled the small-sample, mixed-variable structure.

0Communications Analyzed

0Affective Type

0Instrumental Type

0.00McFadden R²

LIWC2015 Summary Variables

Variable	High Score	Low Score
Analytic	Formal, logical, hierarchical thinking	Personal, narrative, present-focused
Clout	Confident, expert, authoritative	Tentative, humble, anxious
Authentic	Honest, personal, disclosing	Guarded, distanced, evasive
Tone	Positive, upbeat emotional valence	Anxious, sad, or hostile

A Detectable Divide

Cluster Analysis

K-means cluster analysis of the four LIWC summary variables produced two groups with no fuzzy boundary between them; linear discriminant analysis correctly classified 100% of the 100 cases under cross-validation, with Analytic (r = .64), Clout (r = .88), and Authentic (r = -.80) carrying the strongest contributions to component separation and Tone (r = .48) carrying a moderate one.

Cluster 1, the Affective Type (n = 53), shows the lower Analytic, lower Clout, lower Tone, and higher Authentic profile; their writing reads personal, present-tense, narrative, tentative, often saturated with hostility and despair. Cluster 2, the Instrumental Type (n = 47), shows the inverse profile, higher Analytic, higher Clout, higher Tone, lower Authentic; their writing reads formal, logical, hierarchical, confident, guarded, closer to ideological treatise than to personal grievance. The split aligns with the theoretical distinction in the violence literature between emotionally-driven and goal-driven offending (Dodge, 1991; Glenn & Raine, 2009; McEllistrem, 2004; Meloy, 1988), and applies it within the mass-murder population for the first time.

Relative Profile Absolute Scale

Analytic

Tone

Clout

Authentic

Affective Type (n=53)

Instrumental Type (n=47)

Analytic

Tone

Clout

Authentic

Affective (LIWC percentile)

Instrumental (LIWC percentile)

Formulating the Six Hypotheses

Behavioral Predictions

Six hypotheses mapped the linguistic clusters onto physical attack outcomes, each one asking whether the language in a pre-attack communication predicts how the event unfolds.

The text will naturally cluster into two distinct behavioral typologies. (Confirmed via K-means analysis).

The cluster an attacker belongs to will predict the number of victims killed. (True; Instrumental attackers are more lethal. Mann-Whitney U, p = .019).

Membership predicts whether an attacker will target a location they have a personal connection to. (True; Affective attackers are 270% more likely to strike schools or workplaces they know. Fisher’s exact, p = .032).

The cluster will dictate whether specific victim demographics are targeted. (True; systematic demographic targeting correlates almost exclusively with the Instrumental group).

The cluster predicts how the attack ends. (True; Affective attackers are substantially more likely to end the attack via suicide or apprehension. Fisher’s exact, p = .006).

We can build a unified logistical regression model correlating text clusters with physical lethality, targeting, location, and termination states.

Key Odds Ratios Test Statistics

Perpetrator Suicide or Apprehension

Odds-ratio gauge for the Perpetrator Killed predictor, showing OR 7.65. The arc is a relative visual indicator, not a calibrated scale.

Perpetrator Killed (OR 7.65, p = .006), the strongest single behavioral predictor of Affective type membership. Affective attackers are more than seven times as likely to be killed during or at the conclusion of the attack. The arc is a relative visual indicator, not a calibrated scale.

p = .006 · Affective vs Instrumental

SCALE 0 TO 10

Personal Connection to Attack Location

Personal Connection to Attack Location (OR 2.7 inverse, p = .032). Affective attackers are approximately 2.7 times more likely than Instrumental attackers to strike a location (such as a school or workplace) with which they have a personal history. The arc is a relative visual indicator, not a calibrated scale.

p = .032 · Affective vs Instrumental

SCALE 0 TO 10

Hypothesis	Test	Statistic	Result
H2, Victims Killed	Mann-Whitney U	U=908, z=-2.35	p = .019
H3, Location Connection	Fisher’s Exact	OR = 0.37 (inv = 2.7)	p = .032
H4, Initial Targeting	Fisher’s Exact	OR = 2.09	p = .075 (n.s.)
H5, Perpetrator Killed	Fisher’s Exact	OR = 7.65	p = .006

AFFECTIVE TYPE (N=53) VS INSTRUMENTAL TYPE (N=47) · ALPHA = .05

Predictive Signal Matrix

Logistic Regression

The first binary logistic regression used the raw H1-H5 predictors (lethality, connection, targeting, perpetrator killed) and was significant (χ²(4) = 30.58, p < .001, McFadden R² = 0.22), though targeting and connection were too coarse to do their work. Targeting was coded as yes/no without distinguishing between specific-individual targeting and demographic targeting; connection was coded without distinguishing between the perpetrator’s home and a less intimate site like their school or workplace. A second iteration refined both, so that targeting became demographic-based targeting specifically, and connection became perpetrator’s home among the attack locations specifically. The refined model improved fit to McFadden R² = 0.24 (χ²(4) = 32.52, p < .001) and produced the odds ratios below; each of the four predictors holds up as significant when controlling for the others.

Odds Ratio % Change in Odds

Perpetrator KilledOR 14.11

Home as Attack LocationOR 5.0 inv.

Demographic TargetingOR 4.83

Per Victim KilledOR 1.14

RUST = PREDICTS AFFECTIVE TYPE · GLACIAL = PREDICTS INSTRUMENTAL TYPE

Perpetrator Killed+1,311%

Home as Attack Location+400%

Demographic Targeting+383%

Per Victim Killed+14%

CHI-SQUARED(4) = 32.52 · P < .001 · McFADDEN R² = 0.24

Operational Manifestations

Type Exemplars

Matthew Murray and Anders Breivik sit at opposite poles of the two clusters. Murray’s pre-attack writings represent the limit of the Affective Type, highly personal, emotionally raw, focused on his own perceived mistreatment; Breivik’s 1,500-page manifesto is the Instrumental limit, carrying no personal emotional disclosure, structured as a political treatise with chapter headings and citation conventions.

Affective Type

Matthew Murray

Arvada and Colorado Springs, 2007. Four killed, five injured. Murray attacked locations where he had personal history. His communications were saturated with first-person emotionality, immediate grievances, and a stated indifference to his own survival.

LIWC Profile Attack Outcomes

Analytic18.81

Tone1.00

Clout46.47

Authentic89.35

Victims Killed4

Victims Injured5

Location ConnectionYes

Attack ConclusionSuicide

Instrumental Type

Anders Breivik

Utoya Island and Oslo, 2011. Seventy-seven killed, 319 injured. Breivik generated a 1,500-page manifesto reading as a calculated scholarly treatise. It is structured entirely around rigid ideological argument with no personal grievance. He was ultimately apprehended alive.

LIWC Profile Attack Outcomes

Analytic83.47

Tone23.40

Clout67.32

Authentic26.71

Victims Killed77

Victims Injured319

Location ConnectionNo

Attack ConclusionApprehended

Threat Management Efficacy

Practical Constraints

The intended use of this typology is narrow; it does not predict whether an attack will occur, only what kind of attack is more likely, conditional on a pre-attack communication being discovered and submitted for analysis. The work belongs to risk management, the impact side of the threat-assessment equation, not to threat detection, the probability side.

Institutional Implementation Scenario

Consider a university threat assessment team presented with what appears to be a pre-attack communication composed by a student. Running the text through the LIWC summary variables and applying the cluster model classifies the author as Affective; the team then knows the author is more likely to attack a location with personal connection (potentially the campus itself), more likely to commit suicide at the conclusion, and likely to be in an acute emotional state, so the response can shift accordingly, toward target hardening of the connected location and mental-health-led de-escalation instead of a counter-terror posture.

The same model applied to an Instrumental classification points elsewhere, toward lower likelihood of personal connection to attack location, higher likelihood of demographic-based targeting, and substantially higher likelihood that the attack continues until the perpetrator is apprehended or killed by another person; the two response postures call for different resources, different personnel, and different timelines.

Limitations

n = 100 is small for a field with low base rates, and power analysis suggests significance may be underestimated for small and moderate effects. Stephen Paddock (Las Vegas, 2017) and other perpetrators who left no pre-attack communications are necessarily excluded, since the method has nothing to read where there is no communication. The within-sample classification accuracy of 100% is a property of the sample, not a claim about the world, so out-of-sample validation on new pre-attack communications is the next step worth taking; earlier work on increased communicated power drives in mass-murder manifestos (Leffew, 2017) was the seed for this dissertation’s variable selection. The language an attacker leaves behind, weighed this way, sorts cleanly into two types that predict how the attack itself unfolds.

Related Works

Next Publication

→