A Computational Model of Cognitive Maps

Abstract

Why do some individuals pick up arms as opposed to others who live under the same conditions? Environmental and group theories fail to differentiate between these individuals. In response, we apply the cognitive mapping approach and model violence as decisions based on chains of beliefs about various types of factors, including state aggression, access to violent groups, religion, and personal characteristics. Based on a double-paired comparison, data are constructed from ethnographic interviews with Muslim and non-Muslim individuals engaging in violent and nonviolent activity in authoritarian and democratic states—Egypt and Germany. The analysis develops a computational model formalizing the cognitive maps into Bayesian networks. In 477,604 runs, the model (1) identifies the beliefs connected to decisions, (2) traces inference chains antecedent to decisions, and (3) explores counterfactuals. This suggests that both violent and nonviolent activities are responses to state aggression, and not to Islam, group access, or personal characteristics.

Keywords

belief structure conflict international security terrorism

On September 3, Sadat arrested 1,536 of our people

This event made us all very angry

We wanted revenge

Individual Involved in Planning the Assassination of the Egyptian President Sadat (1981)

Why do some individuals pick up arms against their states as opposed to others who live under the same conditions? Investigating groups rather than individuals, many existing analyses fail to address this question (e.g., Dudouet 2013; Nemeth 2014)—yet, understanding political violence also depends on understanding the behavior of individuals, who form or join violent groups, carry out violent attacks, or break away from violent groups. Consequently, this article focuses on the individuals who engage in political violence.

To study the mentioned question, we apply Axelrod’s cognitive mapping approach (CMA). To our knowledge, this is the first application of the CMA to systematically study political violence. Although a large body of literature has investigated belief systems and successfully applied the CMA (Young 1996, 395), most political scientists have abandoned the approach. This is due to the complexity of cognitive maps (CMs), which typically contain more than a dozen of beliefs and connections between beliefs. This abandonment is unfortunate for at least two reasons: (1) the CMA bridges the gap between human behavior and external factors by modeling behavior as decisions based on both external and internal factors, and (2) it promises to synthesize the existing literature of particular fields by offering a consistent method to explore various types of factors, so far examined by different, and sometimes contradictory theories.

Violence is one of the most studied topics in political science. Due to its devastating effects and continuous occurrence, its study remains a pressing need. The literature on violence can roughly be divided into four fields: (1) environmental theories emphasize the political, economic, or cultural environment (e.g., Fearon and Laitin 2003; Huntington 1993); (2) group theories emphasize the social environment (see previously); (3) rational choice theories emphasize calculations given certain goals and constraints (e.g., Enders and Sandler 2006, 11); and (4) psychological theories have often searched for a “terrorist personality” (e.g., Hacker 1976). To study our research question, these approaches suffer from four main liabilities: first, environmental theories fail to differentiate between violent and nonviolent individuals living under the same conditions. Second, group theories fail to explain why only certain individuals join violent groups as opposed to others who also have access. Third, rational choice theories cannot explain why, given the same conditions, individuals can make such different calculations that some pick up arms while others do not. Fourth, empirical findings indicate that there seems to be no “terrorist personality.” In the words of Crenshaw (1981, 390), “the outstanding common characteristic of terrorists is their normality.”

The CMA is an approach from the field of political psychology that can serve to overcome these liabilities. It belongs to a class of psychological approaches that offer models to analyze how beliefs and attitudes relate to human behavior, providing a theoretically basic and mathematically simple methodology.¹ Specifically, the CMA models human behavior, such as political violence, as decisions based on chains of beliefs about various types of factors. For example, one can hold beliefs about state aggression, or poverty (cf. environmental theories); access to violent groups or group dynamics (cf. group theories); goals and constraints (cf. rational choice theories); or personal characteristics like self-confidence (cf. psychological theories).

Since beliefs can have both direct and indirect connections with decisions to pick up arms, the CMA moreover provides a consistent methodology to systematically explore the mechanisms by which these factors motivate individuals to engage in violence. These mechanisms are logical, and modeled as chains of beliefs that are antecedent to decisions. Based on this, the CMA offers a rigorous analytical framework synthesizing existing approaches to violence: It adds to environmental and group theories focusing exclusively on external factors; and to the psychology and rational choice literature by systematically exploring the mechanisms underlying violence from the actors’ perspective.

To cope with the complexity of CMs, this article develops a formal model. Specifically, we formalize CMs into graphical models (Bayesian Networks) and conduct a computational analysis examining 477,604 belief combinations connected to violent as opposed to nonviolent activity. The model is applied to data constructed from about eighty ethnographic interviews in Arabic and German with twenty-seven violent and nonviolent individuals. Applying a double-paired comparison, we examine Muslims from an authoritarian state—Egypt—and non-Muslims from a democratic state—Germany.

The analysis provides surprising insight contradicting the existing literature: it shows that the belief systems underlying violent and nonviolent activity are surprisingly similar and that there are no significant differences between the beliefs of violent Muslims and non-Muslims. Both violent and nonviolent activity are found to be explained best by beliefs about aggressive state behavior, which confirms existing studies presenting violence as a form of self-defense (e.g., Pape 2005). It also confirms Tarrow’s (1998) and McAdam, Tarrow, and Tilly’s (2001) findings that the closing of political opportunities is often associated with an increase in the likelihood of violence and theoretical assumptions that the availability of official legitimate channels as opposed to state repression “acts as a brake on larger scale—especially violent—opposition” (Regan and Henderson 2002, 121). It moreover suggests that neither beliefs about violent groups nor personal characteristics explain violence, which contradicts group theories and theories attributing violence to the individuals’ personality.

The analysis identifies ten chains of beliefs underlying decisions to engage in violent versus nonviolent activity. These indicate that violence can be differentiated from nonviolent activity by (1) strongly versus weakly aggressive state behavior, (2) absence of acceptance of state structures (violence) versus acceptance of state structures (nonviolent activity), and (3) acceptance of consequences of violence (violence) versus absence of acceptance of consequences of violence (nonviolent activity). The chains also indicate that, as expected from theories of asymmetric warfare (cf. Mansdorf and Kedar 2008), individuals may decide to engage in violence based on power calculations that are unfavorable to themselves in comparison with the state; and that, unlike what is expected from rational choice theories, individuals may decide to engage in nonviolent activity without believing that this type of activity is effective to confront the state.

The article proceeds as follows: first, we introduce the CMA and its application to this study. Second, we present the research design and hypotheses. Third, we present the computational model, and analysis, including a description of the methodological steps. Finally, we elaborate on the mentioned results.

A Cognitive Mapping Approach to Political Violence

In this article, we draw on the CMA to study political violence, modeled as decisions based on chains of beliefs about various types of factors. The CMA was introduced by Axelrod (1976), and applied by numerous researchers in the field of foreign policy analysis. However, to our knowledge, it was never employed to systematically analyze violence: at best, it was used to explore war games by studying military officers (Klein and Cooper 1982), but not extended to study real events, or violent individuals. Other applications of the CMA have focused on fields like Chinese strategic culture (Johnston 1995), international negotiation (Bonham, Sergeev, and Parshin 1997), and foreign economic policy (Tapio 2003).

Today, most political scientists have abandoned the CMA. However, the approach has been spreading in other disciplines like economics (e.g., Zhang, Shen, and Jin 2011), engineering (e.g., Bhatia and Kapoor 2011), medical studies (e.g., Papageorgiou 2011), geography (Soler et al. 2011), or biology (e.g., Wills et al. 2010). In fact, the CMA itself has become a subject of research (e.g., Miao et al. 2010; Nadkarni and Shenoy 2001; Peng, Wu, and Yang 2011; Young 1996).

The CMA provides complex, in-depth mappings of the mechanisms connected to behavior, while bridging the gap between actors and structures. Therefore, its abandonment in political science is unfortunate. As mentioned, its abandonment is due to the complexity of CMs in general. In the particular field of violence, the CMA’s absence is moreover connected with the difficulty of obtaining information about the beliefs held by the actors. Those researchers who have gained access to violent individuals (e.g., Atran 2010; Post, Sprinzak, and Denny 2003; Stern 2003) have not applied the CMA, and, consequently, much remains unknown about the actors’ view of the complex micro-level mechanisms underlying their behavior.

The CMA involves two components: (1) beliefs and (2) connections between beliefs. Beliefs are “a person’s subjective probability that an object has a particular characteristic (e.g., how sure the person feels that “This book is interesting”…)” (Fishbein and Ajzen in Oskamp and Schultz 2005, 11). Unlike what might be assumed, beliefs are not by nature exclusively subjective, and a major task of the CMA is to identify shared beliefs (cf. Bar-Tal 2000) motivating more than one individual to make certain decisions. This is possible because many beliefs have propositional contents addressing observations accessible to anybody. As Nilsson (2013, 1) writes, “I believe I exist on a planet that we call Earth and that I share it with billions of other people.” Furthermore, there are beliefs whose propositional contents address “things” that are not necessarily observable; for example, beliefs about moral norms like “I believe that it is wrong to kill someone” or unverifiable events like “I believe that God will save the world.” These beliefs are also shared frequently (Bar-Tal 2000, xii).

Connections between beliefs indicate people’s subjective probability that an object has a particular characteristic in relation to the particular characteristic of another object, and the logical order of this relation. These connections are also called inferences. For example, the connection between B1 “I believe that it is raining” and B2 “I believe that the street is wet” indicates an order in which B1 is logically prior to B2. This inference can be expressed as B1 → B2, where B1 is the antecedent belief and B2 the consequent belief. Each inference contains at least one antecedent and one consequent belief. Each consequent belief of a particular inference can be the antecedent belief of another inference and vice versa. Based on this, it is possible to trace chains of beliefs, or chains of inferences, that consist of both direct and indirect connections between beliefs. For example, one can imagine a chain that contains a belief that the state is aggressive (B1), which is an antecedent of another belief that the government is a threat to its citizens (B2), which is in turn an antecedent of a belief that there is a necessity of doing something against the government (B3). This could be expressed as B1 → B2 → B3.

These chains represent the mechanisms by which people reason about the world. They allow the systematic tracing of the interconnections between factors underlying human behavior. Since human reasoning processes are not deterministic, these mechanisms are logical rather than causal. As Stenning and van Lambalgen (2008, 3) note, “the psychology of reasoning and logic are in a sense about the same subject.”

Belief chains are in turn antecedent to decisions, such as decisions to pick up arms. Since decisions are components of CMs representing belief systems, we define decisions as certain types of beliefs differing from other beliefs by semantics and structure.² Semantically, they express the subject’s intention to engage in certain behavior, for example, an individual’s decision to participate in the assassination of President Sadat. Structurally, the position of decisions in belief systems is located at the end of inference chains, so that they are logically preceded by all other beliefs of the chain (belief → belief → belief → decision). These chains represent the mechanisms underlying human behavior.

The connections between decisions and behavior are not causal, either. Once a decision occurs, it does not necessarily translate into behavior.³ Rather, there is a “gap,” indicating that “we do not normally experience the stages of our deliberation and voluntary actions as having causally sufficient conditions or as setting causally sufficient conditions for the next stage” (Searle 2001, 50, 61-96). The CMA can therefore serve the development of hypotheses for behavioral experiments.

Research Design and Hypotheses

The following analysis uses a double-paired comparison involving four opposite groups of individuals: First, it compares the beliefs of violent versus nonviolent individuals. Second, it compares Muslims from an authoritarian state—Egypt—with non-Muslims from a democracy—Germany. The Table 1 provides an overview.

Table 1.
A Double-paired Comparison of Violent and Nonviolent Individuals from Egypt and Germany.

Political violence Nonviolent activity

Egyptians (Muslims) 7 individuals 8 individuals

Germans (non-Muslims) 6 individuals 6 individuals

This design adds analytical rigor by controlling for the findings obtained for each group. For example, studying non-Muslims from a democracy (Germany) controls for findings obtained for Muslims from an authoritarian state (Egypt). Furthermore, the design investigates factors related to violence by the literature: It explores the relevance of (1) environmental theories by studying individuals from opposite states and cultures, that is, individuals from an authoritarian and democratic state who are Muslim and non-Muslim; (2) group theories by studying individuals rather than groups; (3) rational choice theories by investigating goals and constraints expressed by the individuals (rather than inferred by the researcher); and (4) psychological theories by focusing on beliefs in general and beliefs about personal characteristics in particular.

Although we noted that these theories do not immediately serve to study our research question, we used them to develop hypotheses that guide our analysis by providing a theoretical framework connected to the literature. Specifically, we reformulated the main argument of each theory as beliefs. This differentiated between violent and nonviolent individuals and allowed us to study various types of factors connected to violence (modeled as beliefs).

Environmental theories argue that violence is best explained by economic and political conditions. Fearon and Laitin (2003, 75), for instance, have explained civil wars by “conditions” including “poverty” and “political instability.” Similarly, Piazza (2011, 350) has shown that “countries that permit their minority communities to be afflicted by economic discrimination make themselves more vulnerable to domestic terrorism in a substantive way.” This implies that individuals who engage in violence are motivated by beliefs about environmental factors:
Hypothesis 1a: Violent individuals are motivated by beliefs about their state environment.

Although questioned by more recent findings (e.g., Canetti et al. 2010), some environmental theories have argued that violence is best explained by cultural factors like Islam. In the words of Lewis (1990), “Islam, like other religions has also known periods when it inspired in some of its followers a mood of hatred and violence.” Huntington (1993) extended this argument by asserting that there is a “clash of civilizations” between Western and non-Western civilizations. Following this argument, violent individuals are assumed to be motivated by Islam:
Hypothesis 1b: Violent individuals are motivated by beliefs in Islam.

Other researchers argue that individuals pick up arms in response to interacting with violent groups. Sageman (2004, vii), for instance, has observed, “terrorism is an emergent quality of the social networks formed by alienated young men who become transformed into fanatics yearning for martyrdom and eager to kill.” Similarly, McDoom (2013, 1) shows that violent individuals “are likely to live either in the same neighbourhood or in the same household as other participants.” “Specifically, as the number of violent to nonviolent individuals in an individual’s neighbourhood or household increases, the likelihood of this individual’s participation also increases.” This suggests that violent individuals are motivated by beliefs about interacting with violent groups:
Hypothesis 2: Violent individuals are motivated by beliefs about contact with violent groups.

Theories connecting violence to rational calculations investigate “the manner in which an agent responds to environmental and other constraints,” arguing that “[b]y responding in a sensible and predictable fashion to changing risks, terrorists are judged to be rational” (Enders and Sandler 2006, 11, italics in original). This suggests that violent individuals are motivated by rational calculations, including beliefs about goals, constraints—such as one’s comparative strength with the state—and the availability and consequences of violent and nonviolent means.
Hypothesis 3: Violent individuals are motivated by specific beliefs, namely, rational calculations including goals (a), their comparative strength to the state (b), and the availability (c) and consequences of violent (d) and nonviolent (e) means.

Although arguments connecting violence to the individuals’ personality have been discarded and it is unclear how much individuals are aware of their traits, we used the opportunity of having direct quotations by violent individuals to investigate the importance of personal characteristics. Specifically, we formulated a hypothesis investigating the relevance of the individuals’ self-perception:
Hypothesis 4: Violent individuals are motivated by beliefs about self-perception.

A Sample of Individuals from Egypt and Germany

We studied twenty-seven individuals from eight violent and nonviolent groups. A list of the individuals is available upon request, although some wish to remain unknown.⁴ Since the number of individuals is rather small, our sample’s representativeness is limited and not directly generalizable to different contexts. It does, however, provide an interesting basis of comparison of violent and nonviolent activity in two opposite states. It must also be noted that samples including formerly violent individuals and nonviolent individuals living in hiding are naturally small and that our interviews with only twenty-seven individuals still provided trillions of belief combinations, indicating that extraordinarily rich information can be obtained from small samples.

The Egyptians are from the violent groups al-Jihad (AJ) and al-Jama’at al-Islamiyya (JI), and the nonviolent Muslim Brotherhood (MB), whereas the Germans are from the Red Army Faction (RAF), Bewegung 2. Juni (B2 J), the Socialist German Student Union (SDS), Kommune 1 (K1), and Rote Zelle (RZ). These are anti- rather than prostate groups (e.g., the Loyalists in Northern Ireland), underlining that our findings are not immediately generalizable to other types of groups, either. Rather, our sample provides interesting insight into the motivations of violent and nonviolent individuals from radical groups opposing the state.

AJ and JI developed during the 1970s and conducted various attacks on the state, including the assassination of President Sadat. They are considered the roots of al-Qaeda (cf. Sageman 2004). The MB was founded in 1928 and is the largest nonviolent opposition movement in the Middle East. The RAF was the most violent organization in Germany after the end of the Second World War, killing thirty-four people between 1971 and 1993 (Peters 2004, 17, 846). B2 J was active during the same time, but killing only two people, it received less attention by the state security (Reinders and Fritzsch 1995, 7-9). The SDS, K1, and RZ developed as part of the worldwide student revolts at the end of the 1960s. They are known as major drivers of the German protests.

Data Construction

To construct CMs, we conducted about eighty ethnographic interviews in Arabic and German. In Egypt, we met each individual two to ten times, as it was impossible to immediately conduct interviews, given the sensitivity of the topic; in Germany, we met each individual once. Interviews in Egypt lasted between twenty minutes and entire days, interrupted by families and friends, while interviews in Germany lasted between two to five hours. Nevertheless, the information gathered in Egypt and Germany was comparable, as shown by our coding scheme abstracting the individuals’ quotes into more general categories (see the following).

In Germany, we contacted violent individuals through their lawyers (identified from the news), and nonviolent individuals via public phone entries. Violent individuals were highly reluctant to be interviewed: of the sixty requests to RAF members, only four were answered. Two responses were positive and led to two additional individuals. In Egypt, both violent and nonviolent individuals were extremely difficult to locate and convince to meet with us. Drawing on prior connections, we identified individuals through a wide network of friends of friends, including students, journalists, and bloggers. We often held several meetings allowing the individuals to interview us before we began interviewing them. In total, we undertook field research over two years.

Under these circumstances, only certain information was accessible. We know that all violent Germans actually participated in violence, although they were sometimes unwilling to specify participation in particular attacks. However, not all violent Egyptians were willing to say that they engaged in violence—some only expressed support for violence. Nevertheless, their membership in a violent group, which they confirmed, posed high risks and indicates a strong commitment to violence.

Providing postfact explanations, one might assume that our interviews cannot provide complete representations of what motivated the individuals’ decisions. In response, it is important to note that decisions to engage in violence involve high risks and mark turning points in the individuals’ lives: spending the following decades in prison, the individuals had reconsidered their decisions extensively when we interviewed them, recalling related factors (e.g., the 1977 bread riots in Egypt) in extraordinary detail. This confirms that repetition can help recall the organization and structure of things, which are “essential for memory retrieval” (Mandler 2006). Indeed, the individuals’ explanations, complex and usually historically verifiable, by far exceed our own memory of decisions we made only last week.

One might further object that the individuals’ explanations suffer from post-event rationalization where current attitudes and past behaviors are aligned (cf. Albarracin and McNatt 2005, 719). We cannot exclude this possibility and acknowledge that parts of the individuals’ explanations may differ from their reasoning processes at the time when they made decisions. However, we emphasize that postfact explanations are the best way to obtain insight into the reasoning processes of violent individuals, as it is impossible to interview people as they are engaging in violence. We also emphasize that, like all explanations, postfact explanations have to be intelligible and structured according to the rules specified previously (see the A Cognitive Mapping Approach to Political Violence section). Therefore, they naturally indicate possible mechanisms underlying violence, regardless of whether these are the real mechanisms in particular cases.

To construct CMs from the interviews, we identified beliefs and inferences from the individuals’ quotes.⁵ Beliefs were identified from (1) content words, (2) subclauses, (3) main clauses, and (4) entire assertions. Inferences were identified from (1) linguistic connectors indicating (a) causal connections like “because” or “therefore,” and (b) condition connections like “if…then” and (2) the order in which certain assertions were made. For instance, our analysis of an interview with an MB identified the following five beliefs (right-hand side of Table 2) from the following quote (left-hand side of Table 2).

Table 2.
Identifying Beliefs.

Quote Identification of beliefs

Line 1 When I was in prison Belief 1 I believe that I was in prison

Line 2 All the people were talking about me Belief 2 I believe that all the people were talking about me

Line 3 They supported me Belief 3 I believe that they supported me

Line 4 I was on TV Belief 4 I believe that I was on TV

Line 5 I won Belief 5 I believe that I won

Moreover, our analysis of an interview with another MB identified the following two inferences (causal connectors are indicated by “>”; Table 3).

Table 3.
Identifying Inferences.

Quote Identification of inference

Line 1 The governments’ stance differs from the Arab peoples

Line 2 >Peaceful effort Inference

Line 3 I think that stopping violence needs pressure from the whole world

Line 4 Alone it is too weak

Line 5 It needs the whole world

Line 4 It needs the agreement of the peoples

Line 5 >It is a kind of jihad Inference

Based on this identification of beliefs and inferences, the CMs represented the individuals’ direct speech and were not immediately comparable. In response, we applied Spradley’s (1979) theme analysis to create a coding scheme. Theme analysis is a qualitative method exploring the semantics of belief systems by identifying “meaning that is integrated into some kind of larger pattern” (Spradley 1979, 140-141). It enabled us to systematically abstract beliefs according to the main theme they addressed and to assess what types of factors can motivate individuals to pick up arms. Our coding scheme was theoretically motivated (cf. Narayanan and Armstrong 2005, 32, “theory-driven coding procedures”), aiming at presenting categories that serve the systematic comparison of the beliefs underlying violent versus nonviolent activity, while speaking to the literature on violence.

To ensure that the scheme was both abstract enough to capture the larger differences between violent and nonviolent activity, and specific enough to identify their underlying micro-level mechanisms, it contains three levels of abstractions: (1) belief classes, (2) belief superclasses, and (3) belief super-superclasses. The beliefs identified directly from the interviews are called instances.

While creating these categories, we assigned positive and negative values to belief classes within superclasses. We moreover assigned numbers to belief classes, indicating their comparative strength inside their superclasses. These values and numbers helped us interpret the results of our analysis by adding more detailed information about the micro-level mechanisms underlying violent and nonviolent activity. Table 4 shows an excerpt from the super-superclass state environment.

Table 4.
Excerpt from Super-Superclass State Environment.

Super-superclass Superclass Belief class (BC) Rank Example of an instance

State environment Negative nature of home state Continuous aggression by home state +1 There was a circle of violence conducted by the state

Aggression by home state +2 Sadat arrested the brother of Khalid Islambouli

Domination by home state +3 There was no real party, only Sadat’s party, there were no possibilities

Strained living conditions +4 There was social poverty

Unreligious state +5 My country wasn’t Islamic

Aggression by previous state +6 Hitler waged the Second World War

Domination by previous state +7 People had lived under the terrible Nazis

Change of Home State +8 Sadat changed everything

Ignorance of fellow citizens +9 The people don’t care about politics, they are interested in the African cup

Support by home state −1 The state agency gave me an apprenticeship as a librarian

Religious government −2 The government is Muslim

Education of the people −3 The people become educated

Absence of domination by home state −4 Sadat had given us freedom

Absence of aggression by home state −5 The police stood in front of the building for days and didn’t do anything

Absence of state structures −6 If there were no state structures

Graphical Modeling

Since CMs contain large numbers of beliefs and inferences, it is impossible to systematically analyze them by hand. However, it is possible to formalize them. As shown by Axelrod (1984) and others (e.g., Bhavnani, Miodownik, and Choi 2011; Schrodt 1995), formal models make traceable political processes that would otherwise not be analyzable, or only be analyzable on a much smaller scale. Specifically, we formalized CMs into graphical models and developed a computational model for their systematic analysis.

A graph (or network) is a mathematical structure combining a set of nodes with a set of edges. An edge represents a connection between two nodes. Edges may be directed or undirected. A directed edge indicates that information or influence only flows in one direction. A graphical model associates variables with nodes. Edges arise from functions that express relationships between variables.

In our formalization, nodes correspond to belief classes (BCs) and decision classes (DCs), as identified from the interviews. Edges represent the connections between all classes, as indicated by the inferences derived from the interviews. An inference is a logical statement connecting beliefs. For examples that follow, uppercase letters (A, B, C,…) will represent belief classes with X, Y, and Z reserved for decision classes. An example of a basic inference comes from the statement “if A then B,” which translates to an edge pointing from A to B. All the edges in this model are directed.

In the model, BCs can be on (asserted) or off (not asserted). Inferences may contain several BCs. An example inference is “If A and B then C.” In this case, both A and B must be on to activate C. The “or” relationship is represented by the presence of (at least) two inferences in an inference set. For example, the combination of inferences “if A then C” and “if B then C” is equivalent to the inference “if A or B then C.” BCs that appear in the “if” portion of an inference are called predicates while those in the “then” portion are called consequents. DCs only show up as consequent. Each individual CM has only one DC.

Consider the following set of inferences: “if A and B then C,” “if C then X,” and “if D then X.” This set may be viewed as the graph visualized by the upper part of Figure 1. As shown by this visualization, graphs identify both direct and indirect connections between certain classes. An indirect connection is called a path. It is possible to have more than one path between BCs. For example, the inference set “if A then B,” “if A then C,” “if B then D,” and “if C then D” has two paths between A and D. This is shown by the lower part of Figure 1.

Figure 1.
Two sets of inferences.

Cycles in a graph (“if A then B,” “if B then C,” and “if C then A”) are not allowed in this model because they would represent circular logic. Care was taken to represent the interviews as sets of inferences without cycles.⁶ The resulting structure is known in many science and engineering disciplines as a directed acyclic graph (DAG). We call an individual instance of our model a Cognitive Map Graph (CMG).

The CM graphical model is a special case of a Bayesian network (BN). In a BN, each node corresponds to a conditional probability distribution. Nodes in a CMG correspond to logical functions, which may be viewed as special probability distributions, taking the values of 1 and 0 to indicate the presence and absence of inferences between classes. BNs are often used in computer science to study structures of variables that are directed and limited (cf. Koller and Friedman 2009; Pearl 2009). Since the CMs’ chains of beliefs are also directed by involving antecedent and consequent beliefs, and limited by involving traceable chains of beliefs that end in decisions, their formalization into a graphical model offers a convenient basis for an automated analysis.

Our model was developed in MATLAB.⁷ It accesses CMs encoded in a simple tabular format and stored in Microsoft Excel files. To represent an or inference, our input format simply uses two inference statements. The statements “if A then C” and “if B then C” are combined to form “if A or B then C.” The and inference requires some extra notation in the spreadsheet format, as shown by the subsequent example: each row of the table records a single inference. There are two columns. The left column holds the antecedent belief or “and” combination of beliefs. The right column holds the consequent belief. BCs are represented by a string identifier such as “political-goal.” If an inference uses an “and” combination, the strings are placed in the same field and separated by a semicolon. Logical “or” combinations are represented by two independent inferences.

Excerpt of a CM in Tabular Form

Predicate Consequent

Political-goal; political-interest Peaceful-activity-against-home-state

Peaceful-activity-against-home-state Aggression-by-home-state

Domination-by-home-state Aggression-by-home-state

The model reads the files encoding the CMs and translates the data into a form required by the BNet toolbox. It first creates cell arrays that store the predicates (P) and consequents (C). The cells corresponding to the previous example are as follows:

MATLAB Cell Array

P = {“political-goal; political-interest,” “peaceful-activity-against-home-state,” “domination-by-home-state”} C = {“peaceful-activity-against-home-state, aggression-by-home-state, aggression-by-home-state”}

From these cell arrays, the program constructs an inference graph and checks for cycles. Finally, it constructs logic tables for all classes and forms the structure used by BNet. MATLAB can now process the CMs.

We used the model to systematically apply inferences given asserted input beliefs to determine decision possibilities (1); systematically trace inference chains connecting certain combinations of beliefs and decisions to engage in violent or nonviolent activity (2); and systematically modify the CMs to investigate counterfactuals (3).

Searching for decision possibilities (1) relies on functions that take beliefs as input and deliver decisions as output. Specifically, this proceeds in three main stages.

Stage 1: Input→ Stage 2: Processing→ Stage 3: Output

Assert certain beliefs in all CMs Search CMs for inferences associated with asserted beliefs. Assert the consequent beliefs and repeat the process until no change is detected Display names of the individuals whose CMs display decisions based on activated belief chains

To trace inference chains (2), the program applies a visualization tool (based on Graphviz; Gansner and North 2000) that produces figures displaying the inference chains between the asserted beliefs and decisions. To modify CMs (3), the program additionally denies (turns off) certain beliefs during stage 1 (input) and then tests if this denial interrupts the belief chains connecting the asserted beliefs with decisions (indicated in stage 3, output). Figure 2 illustrates a simple example.

Figure 2.
Processing without denial of belief classes. Processing including denial of a belief class.

Computational Analysis

The computational analysis systematically identified and evaluated the individuals’ reasoning processes, represented by the CMs. It proceeded in five main steps:
We tested how many individuals decide to engage in violent or nonviolent activity based on the assertion of belief classes, superclasses, and random belief combinations.

We created a sample of experimental belief classes based upon whose combined assertion all individuals reach decisions.

We tested how many individuals make decisions, given all possible combinations of belief classes of the sample.

We produced and compared visualizations of the belief chains between belief combinations and decisions.

We modified the CMs to test when individuals would not have made decisions.

Significance of the State Environment (Step 1)

The analysis suggested that beliefs about the state environment (Hypothesis 1a) matter most to decisions to engage in violent and nonviolent activity. It also indicated that beliefs about the consequences of the activity (Hypotheses 3d and 3e) matter to both types of decisions, while beliefs about religion (Hypothesis 1b), violent groups (Hypothesis 2), and self-perception (Hypothesis 4) do not appear to matter significantly. This offered strong support for Hypothesis 1a, and support for Hypotheses 3d and 3e, but not for the remaining hypotheses.

We obtained these results from different types of analyses. First, we conducted a nonparametric analysis of random belief combinations to investigate differences in decisions to engage in violent versus nonviolent activity (Monte Carlo sampling).⁸ We observed that the number of violent and nonviolent individuals making decisions increases proportionally to the number of asserted belief classes (see Figure 3 for an excerpt of this analysis). We then conducted a Kolmogorov–Smirnov (KS) test to check whether the distributions of the decisions to engage in violent versus nonviolent activity are the same. We found that, while the distributions differ (KS statistics = .79 and p = 0 for a significance level of .05), they have similar structure (see the Online Supplement).

Figure 3.
Distribution of decisions by violent and nonviolent individuals motivated by random belief classes in comparison with belief classes representing hypotheses.¹¹

This similarity was confirmed by investigating correlations between individuals making decisions to engage in violent versus nonviolent activity, given random selection of belief classes: all of these were positive or zero (see the Online Supplement). It was furthermore confirmed by a comparison of the paths lengths (numbers of inferences) between each belief class and decisions to engage in violent versus nonviolent activity. This showed that there is a correlation of .58 between violent and nonviolent individuals for belief classes held by at least one violent and one nonviolent individual. For classes held by either violent or nonviolent individuals, the correlation coefficient is positive but reduced (.40).

Second, we compared the analysis of random belief classes with an analysis of specific belief classes and superclasses representing the hypotheses. The results were obtained from 215,460 runs and are illustrated subsequently.⁹ They show that only the beliefs about the state environment and consequences of violent and nonviolent activity encourage more than the average number of individuals to make decisions, given the same number of random belief classes (Figure 3 and Table 5). Specifically, they show that the highest number of individuals motivated to make decisions by belief classes representing the hypotheses was nineteen—a number that is observably higher than the average obtained for the same number of randomly selected belief classes. This number was obtained for the superclass state environment (Hypothesis 1a). The p-values associated with each hypothesis also indicated that the state environment matters (Table 5): regarding Hypothesis 1a, there is a 21.6 percent chance that a random selection of classes would motivate at least as many individuals to make decisions. By contrast, with the exception of Hypotheses 3d and 3e, the chances for the remaining hypotheses range between 71.2 and 99.7 percent.

Table 5.
p-Values Associated with Hypotheses Relative to the Empirical Distributions Obtained from Monte Carlo Sampling.

Hypothesis 1a 0.216433

Hypothesis 1b 0.741767

Hypothesis 2 0.711538

Hypothesis 3a 0.7854

Hypothesis 3b 0.968

Hypothesis 3c 0.710567

Hypothesis 3d 0.0153667

Hypothesis 3e 0.144233

Hypothesis 4 0.9967

Note: The p-value is the probability that the belief classes representing the hypotheses encourage at least as many decisions as random belief combinations including the same number of classes. Mathematically, P(decisions from random selection ≥ decisions made by hypotheses).

Based on our analysis of the individuals’ quotes, the superclass state environment also contained significantly more beliefs than most other superclasses, which further underlines its relevance. Moreover, the analysis of individual belief classes showed that the highest number of individuals (nine) make decisions based on a belief class addressing the state environment: aggression by Home State.

Our analysis furthermore showed that beliefs about the consequences of violent and nonviolent activity (Hypotheses 3d and 3e) encourage observably more individuals to make decisions than the average, given the same number of random belief classes (Figure 3). The p-values associated with Hypotheses 3d and 3e indicated that there is only a 1.5 and 14.4 percent chance that a random selection of classes would motivate at least as many individuals to make decisions (Table 5). This might lead to the conclusion that beliefs about the consequences of violent and nonviolent activity matter more than beliefs about the state environment. However, as we show subsequently, the mechanisms underlying violent and nonviolent activity instead suggest that beliefs about the consequences of violent and nonviolent activity are in turn based on beliefs about the state environment.

The analysis also showed that no other beliefs representing hypotheses motivated more individuals to make decisions than the average, given the same number of random belief classes. Specifically, four individuals make decisions based on the superclasses strength of state and strength of resistance (Hypothesis 3b); four based on the superclasses Personality and Personal Life (Hypothesis 4); five based on superclass Goals (Hypothesis 3a); and fifteen based on the superclass availability of means (Hypothesis 3c).

Religion (Hypothesis 1b) and violent groups (Hypothesis 2) were the only hypotheses not addressed by belief superclasses, indicating that these factors matter the least. Religion is addressed by five belief classes, namely, Religious State, Unreligious State, Disunity of Muslims, Obedience to God, and God’s Might. These motivated only three individuals to make decisions: disunity of Muslims motivated a decision to engage in nonviolent activity, Unreligious State a decision to engage in violence, and Unreligious State in combination with Obedience to God a decision to engage in nonviolent activity. Violent groups are addressed by only two belief classes: support for Violence in Direct Environment and Efficient Structure of Violent Group. They motivated only one individual to make a decision: support for Violence in Direct Environment motivated an individual to decide to engage in violence.

Experimental Belief Classes (Step 2)

The following analysis explored belief combinations (steps 2 and 3), and chains of inferences between belief combinations and decisions (step 4). This provided information about the mechanisms underlying violent and nonviolent activity. To conduct this analysis, we developed a sample of experimental belief classes. Sampling was necessary, because there were too many belief classes to run the program on all possible belief combinations. Specifically, there were trillions of combinations; for example, the number of all possible combinations for only twenty-four belief classes is 16,777,216.

The goal of our selection process was to create a set of the most significant belief classes based on whose combined assertion all individuals reach decisions (so that the range of output is between 0 and 27; not 0 and 19 individuals as in the previous step). To achieve this, we created a set of (1) main classes and (2) complementary classes. Main classes include the beliefs about the state environment found to matter most in the first step of the analysis. Complementary classes include beliefs upon whose assertion by themselves nobody reaches a decision; but upon whose combined assertion with the main classes, all individuals reach decisions. Complementary classes were selected by the manual analysis of the CMs of eight individuals who do not reach decisions based on the superclass State Environment. The sample is shown in Table 6.

Table 6.
Experimental Classes.

Main set of experimental classes Complementary classes

Super-superclass Superclass Belief Class Rank Super-superclass Superclass Belief Class Rank

State environment Negative nature of home state Continued aggression by home state +1 State perception Strength of state Weakness of home state −1

Aggression by home state +2 Absence of strength of home state −2

Domination by home state +3 Strength of resistance Strength of international resistance +1

Strained living conditions +4 Flexible structure of violent group +2

Unreligious state +5 Goals Political goal +2

Aggression by previous home state +6 Private issues Personal life Experience of violence −1

Domination by previous home state +7 Personality Self-conviction +1

Change of home state +8

Ignorance of fellow citizens +9

Support by home state −1

Religious state −2

Education of the people −3

Absence of domination by home state −4

Absence of aggression by home state −5

Absence of state structures −6

Negative nature of foreign state Imperialism +1

Absence of domination by foreign state −1

Beliefs Encouraging Violent and Nonviolent Activity (Step 3)

The following analysis asserted all possible belief combinations of the sample (131,072; one combination per run). We represented the results by a two-dimensional matrix (see the Online Supplement), which confirms that the belief systems underlying violent and nonviolent activity are surprisingly similar: the largest difference between violent (13) and nonviolent individuals (14) making decisions based on all possible belief combinations is only six individuals.

We then applied filters to identify the beliefs constituting the combinations shown by the matrix. This identified four types of belief combinations:
Combinations encouraging no individual to decide to engage in violent or nonviolent activity (absence of decisions)

Combinations encouraging a small number of individuals to decide to engage in nonviolent activity but not in violence (minimum of nonviolent activity)

Combinations encouraging a small number of individuals deciding to engage in violent and in nonviolent activity (minimum of violence)

Combinations encouraging all individuals to decide to engage in violent and nonviolent activity (maximum of violent and nonviolent activity)

Specifically, the filters showed that type 1 combinations only include positive classes about the state environment, such as Support by Home State. Type 2 combinations additionally include a moderately negative class about the state environment: Strained Living Conditions. Type 3 combinations include a more negative class about the state environment: Domination by Home State. Type 4 combinations include two of the most negative classes: Aggression by Home State and Domination by Home State.

This suggests that both violent and nonviolent activity are responses to beliefs about threatening state behavior and that violence is a response to particularly negative beliefs about the state. This confirms Hypothesis 1a. The results furthermore imply that, with the exceptions of the complementary classes of the sample, none of the remaining beliefs is necessary for all individuals to make decisions. This indicates that neither beliefs about Islam (Hypothesis 1b) nor beliefs about violent groups (Hypothesis 2) play a significant role in decisions to engage in violent or nonviolent activity.

Figure 4 (upper part) illustrates the results referring to the class Aggression by Home State. It shows that of all 131,072 belief combinations (y-axis), belief combinations including Aggression by Home State (blue bar) motivate observably more individuals to decide to engage in violent (x-axis, section V) and nonviolent activity (x-axis, section NV) than belief combinations not including Aggression by Home State (red bar). The lower part of the figure shows that such a difference is not observable concerning Unreligious State, confirming that Islam does not play a major role. These findings were supported by KS tests: KS statistics¹⁰ were .77 (Aggression by Home State) versus .16 (Unreligious State) for decisions to engage in violence, and .53 (Aggression by Home State) versus .08 (Unreligious State) for decisions to engage in nonviolent activity.

Figure 4.
Increase in decisions for violent and nonviolent activity based on aggression by home state. Note: Absence of increase of decisions for violent and nonviolent activity based on unreligious state.

Mechanisms of Violent and Nonviolent Activity (Step 4)

The following analysis traced chains of beliefs connecting the four combination types with decisions. From the 131,072 runs, it identified ten chains that are initialized by the combinations and activate decisions to engage in violent and nonviolent activity.

We constructed these chains from automated inference counts identifying shared inferences and manual comparison between inference chains. All chains obtained from this analysis contain at least one, and mostly several shared inferences. However, most chains, involving between five and twenty beliefs, are unique in their entirety. Although beliefs can by nature be shared (see the A Cognitive Mapping Approach to Political Violence section), that they are shared is not a necessary requirement for our analysis: once observed, an inference indicates the possibility for anybody to make it. This is the case because, as explained, inferences indicate the logical connections between what is addressed by beliefs.

Political violence: The analysis identified five chains of beliefs connected to decisions to pick up arms:
1. Violence as last resort of defense against increasingly threatening state environments

2. Violence as response to state aggression

3. Violence as response to state aggression and acceptance of negative consequences

4. Violence as means to improve increasingly threatening state environments

5. Violence as means to reach goals whose implementation is absolutely necessary

All chains show that individuals decide to pick up arms based on beliefs about threatening state environments (confirmation of Hypothesis 1a). They moreover show that decisions may include additional beliefs about the consequences and goals of violence (confirmation of Hypotheses 3a and 3d) and that these additional beliefs are in turn mostly based on beliefs about the state environment (confirmation of Hypothesis 1a). This is illustrated by the top visualization of Figure 5 (chain 1): it shows how beliefs about state domination and aggression (Hypothesis 1a) can encourage the belief that there are no peaceful means (Hypothesis 3c), which can in turn motivate a decision to engage in violence.

Figure 5.
Violence as last resort of defense (excerpt from chain 1). Violence as means to overcome threatening state behavior (excerpt from chain 4). Violence as response to state aggression and acceptance of negative consequences (excerpt from chain 3).

The middle visualization (chain 4) shows that beliefs about state domination and aggression (Hypothesis 1a) may encourage power calculations favoring the individual (Hypothesis 3b). Combined with beliefs in transformatory goals (Hypothesis 3a), these may encourage the belief that one’s goals can be reached by violence, which can in turn motivate a decision to pick up arms.

The bottom visualization (chain 3) shows that beliefs about state aggression (Hypothesis 1a) may motivate individuals to pick up arms in spite of power calculations favoring the state and beliefs about negative consequences of violence (contradiction to Hypotheses 3b and 3d). Specifically, it shows that believing in state aggression can encourage individuals to accept the negative consequences of violence and decide to pick up arms.
Nonviolent activity: the analysis identified five chains of beliefs connected to decisions to engage in nonviolent activity:

6. Nonviolent activity as means to improve strained living conditions

7. Nonviolent activity as acceptance of state structures in spite of strained living conditions

8. Nonviolent activity as response to unacceptable consequences of violence

9. Nonviolent activity as response to threatening state environment and support by the people

10. Nonviolent activity as response to threatening state environment and impossibility of reaching goals by violence

All chains show that nonviolent individuals also make decisions based on negative beliefs about the state environment, confirming that the belief systems underlying violent and nonviolent activity are rather similar. Furthermore, nonviolent individuals also make decisions based on additional beliefs about consequences and goals and confirm that these additional beliefs are mostly based on beliefs about the state environment. This confirms that the state environment matters most and that goals and consequences play a complementary role.

Specifically, some chains include additional beliefs about positive consequences of nonviolent activity (chains 6), and others about negative consequences of violence (chains 8 and 10). The majority does not include beliefs about the positive consequences of nonviolent activity (chains 7–10). This suggests that nonviolent activity is often not believed to be an effective means responding to threatening state behavior. This is illustrated by the top visualization of Figure 6.

Figure 6.
Nonviolent activity as response to threatening state environment and impossibility of reaching goals by violence, including unreligious state (excerpt from chain 10). Nonviolent activity as means to improve strained living conditions (excerpt from chain 6).

Some decisions to engage in nonviolent activity include only moderately negative beliefs about the state environment, namely, about strained living conditions (chains 6 and 7). This stands in contrast to decisions to pick up arms, supporting the earlier finding that nonviolent activity can occur based on beliefs about moderately rather than strongly threatening state environments. It is illustrated by the bottom visualization of Figure 6.

The analysis furthermore confirms that Islam does not appear to motivate violence, because the belief class Unreligious State can also encourage individuals to engage in nonviolent activity. An example is illustrated by the top visualization of Figure 6.

Counterfactual Analysis (Step 5)

We furthermore conducted a counterfactual analysis, testing whether the individuals would not have decided to engage in violent or nonviolent activity, had they not held certain beliefs. This analysis involved the same sample and the same 131,072 belief combinations. The only difference was that it denied the nonasserted belief classes during each run. Confirming the results that the belief systems underlying violent and nonviolent activity are surprisingly similar, the analysis showed that no individual decides to pick up arms and significantly less individuals decide to engage in nonviolent activity based on the denial of only three belief classes: Continuous Aggression by Home State, Aggression by Home State, and Domination by Home State.

The analysis also identified two mechanisms by which the belief chains underlying decisions can be interrupted: the first involves the denial of entire inference chains representing state environments becoming increasingly aggressive (Domination by Home State → Aggression by Home State → Continuation of Aggression by Home State), or of beliefs addressing state aggression and domination which are not directly connected. The second shows that decisions to engage in nonviolent activity are additionally reduced by only denying beliefs about state domination rather than aggression. This confirms that nonviolent activity may occur based on beliefs about less threatening state environments.

Conclusion

This article has applied the CMA to study the question why certain individuals pick up arms against their states as opposed to others who live under the same conditions. The analysis developed a computational model formalizing CMs into graphical models (Bayesian networks). It examined 477,604 of beliefs held by twenty-seven violent and nonviolent individuals from Egypt and Germany.

We find that the belief systems underlying violent and nonviolent activity are surprisingly similar, and that, contrary to explanations attributing violence to Islam, there are no significant differences between violent Muslims and non-Muslims. The analysis moreover puts into perspective group theories and theories focusing on personality by showing that violence cannot be explained by beliefs about the violent groups or personal characteristics. Furthermore, it identifies ten chains of beliefs antecedent to decisions to engage in violent versus nonviolent activity. These show that, as expected from the literature on asymmetric warfare (Mansdorf and Kedar 2008), decisions to engage in violence may involve power calculations that are unfavorable to the individual in comparison with the state. They also show that decisions to engage in nonviolent activity may be made without believing that this activity can be effective to confront the state, contradicting what is expected from rational choice theories.

The analysis confirms environmental theories by showing that violent and nonviolent activity are explained best by beliefs about threatening state environments. While decisions to engage in violence are always motivated by beliefs about rather threatening state environments (Domination by Home State is the least threatening belief underlying violence), decisions to engage in nonviolent activity may be motivated by beliefs about less threatening environments (Strained Living Conditions). Both types of decisions occur most frequently based on the most negative beliefs about the state environment (Continuous Aggression by Home State, Aggression by Home State, and Domination by Home State). The counterfactual analysis shows that upon the denial of the three most threatening beliefs about the state (see previously), no individuals decide pick up arms and significantly less decide to engage in nonviolent activity.

These findings confirm existing studies arguing that violence is a form of self-defense. Pape’s (2005) investigation of suicide attacks, for instance, also shows that violence is a reaction to threatening state behavior (foreign occupation) and not based on Islam. Similarly, Kepel (2006, viii), drawing on examples from the Muslim world, writes that “jihadism…must be seen as a primordially political, rather than religious movement,” and Beggan’s (2006) examination of violence in Northern Ireland relates violence to three factors including repressive state policies.

Our counterfactual findings imply that individuals do not pick up arms against their states in the absence of threatening state behavior—an implication supported by various studies addressing state aggression indirectly. Ashour (2007), for instance, emphasizes that government interaction is among the “common causes” of de-radicalization; Horgan and Braddock (2010) note the importance of “rehabilitation” of violent individuals into society (cf. Horgan and Altier 2012), and Kruglanski, Gelfand, and Gunaratna (2010) observe that trying to de-radicalize individuals by a religious dialogue has been unsuccessful.

Our results raise the question why some individuals may believe the state to be more threatening than others. Following Altier (2011), this may be related to the individuals’ socialization by others who have experienced violence. It could also be related to personal experience—although our interviews suggest that beliefs that the state is aggressive can also be motivated by news reports or observations on the street. Future research could provide more specific knowledge about this important aspect.

	Political violence	Nonviolent activity
Egyptians (Muslims)	7 individuals	8 individuals
Germans (non-Muslims)	6 individuals	6 individuals

Quote	Identification of beliefs
Line 1	When I was in prison	Belief 1	I believe that I was in prison
Line 2	All the people were talking about me	Belief 2	I believe that all the people were talking about me
Line 3	They supported me	Belief 3	I believe that they supported me
Line 4	I was on TV	Belief 4	I believe that I was on TV
Line 5	I won	Belief 5	I believe that I won

Quote	Identification of inference
Line 1	The governments’ stance differs from the Arab peoples
Line 2	>Peaceful effort	Inference
Line 3	I think that stopping violence needs pressure from the whole world
Line 4	Alone it is too weak
Line 5	It needs the whole world
Line 4	It needs the agreement of the peoples
Line 5	>It is a kind of jihad	Inference

Super-superclass	Superclass	Belief class (BC)	Rank	Example of an instance
State environment	Negative nature of home state	Continuous aggression by home state	+1	There was a circle of violence conducted by the state
Aggression by home state	+2	Sadat arrested the brother of Khalid Islambouli
Domination by home state	+3	There was no real party, only Sadat’s party, there were no possibilities
Strained living conditions	+4	There was social poverty
Unreligious state	+5	My country wasn’t Islamic
Aggression by previous state	+6	Hitler waged the Second World War
Domination by previous state	+7	People had lived under the terrible Nazis
Change of Home State	+8	Sadat changed everything
Ignorance of fellow citizens	+9	The people don’t care about politics, they are interested in the African cup
Support by home state	−1	The state agency gave me an apprenticeship as a librarian
Religious government	−2	The government is Muslim
Education of the people	−3	The people become educated
Absence of domination by home state	−4	Sadat had given us freedom
Absence of aggression by home state	−5	The police stood in front of the building for days and didn’t do anything
Absence of state structures	−6	If there were no state structures

Predicate	Consequent
Political-goal; political-interest	Peaceful-activity-against-home-state
Peaceful-activity-against-home-state	Aggression-by-home-state
Domination-by-home-state	Aggression-by-home-state

Stage 1: Input→	Stage 2: Processing→	Stage 3: Output
Assert certain beliefs in all CMs	Search CMs for inferences associated with asserted beliefs. Assert the consequent beliefs and repeat the process until no change is detected	Display names of the individuals whose CMs display decisions based on activated belief chains

Main set of experimental classes	Complementary classes
State environment	Negative nature of home state	Continued aggression by home state	+1	State perception	Strength of state	Weakness of home state	−1
Aggression by home state	+2	Absence of strength of home state	−2
Domination by home state	+3	Strength of resistance	Strength of international resistance	+1
Strained living conditions	+4	Flexible structure of violent group	+2
Unreligious state	+5	Goals		Political goal	+2
Aggression by previous home state	+6	Private issues	Personal life	Experience of violence	−1
Domination by previous home state	+7	Personality	Self-conviction	+1
Change of home state	+8
Ignorance of fellow citizens	+9
Support by home state	−1
Religious state	−2
Education of the people	−3
Absence of domination by home state	−4
Absence of aggression by home state	−5
Absence of state structures	−6
Negative nature of foreign state	Imperialism	+1
Absence of domination by foreign state	−1

Footnotes

Acknowledgment

We thank the Institute for Computational and Mathematical Engineering at Stanford University, the Swiss National Fund and the German Academic Exchange Service for their generous financial support for this research. We thank Ravi Bhavnani, Miles Hewstone, David Sylvan, and the anonymous reviewers for their comments on earlier drafts of this article. We are especially indebted to David Sylvan for his encouragement and close reading of multiple drafts. We thank Riccardo Bocco, Margot Gerritsen, Martha Crenshaw, Marwa Daoudy, Bahgat Korany, Clark McCauley, Walter Murray, Marc Sageman, Philip Schrodt, and all other professors with whom we had discussions at the Graduate Institute of International and Development Studies and Stanford University for their advice related to this research. This study is based on interviews. We thank those who helped us arrange the interviews (who cannot be named here), and the individuals themselves for giving us their trust and having the courage to speak with us. When we completed this article, several of our interviewees had been imprisoned. It was with them and their families in mind that we completed this article. All mistakes are our own.

Authors’ Note

Stephanie Dornschneider conducted the interviews. Nick Henderson wrote the code of the computational model. The program developed for this research and instructions on how to replicate the results are available online. Data are available as .csv files. Notes of the interviews are not included to protect the identity of our interviewees.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by the Swiss National Fund (BPGEP1-134217 and PBGEP1-139842) and the German Academic Exchange Service (D/10/47115).

Supplementary Materials

The online [appendices/data supplements/etc] are available at .

Notes

References

Albarracin

Dolores

McNatt

Penny S.

. 2005. “Maintenance and Decay of Past Behavior Influences: Anchoring Attitudes on Beliefs Following Inconsistent Actions.” Personality and Social Psychology Bulletin 31 (6): 719–33.

Altier

Mary Beth Ehrhardt

. 2011. Voting for Violence. Dissertation submitted to the Department of Politics, Princeton, NJ: Princeton University.

Ashour

Omar

. 2007. “Lions Tamed? An Inquiry into the Causes of De-radicalization of Armed Islamist Movements: The Case of the Egyptian Islamic Group.” The Middle East Journal 6 (4): 596–625.

Atran

Scott

. 2010. Talking to the Enemy. Faith, Brotherhood, and the (Un)Making of Terrorists. New York: HarperCollins.

Axelrod

Robert M.

1976. Structure of Decision: The Cognitive Maps of Political Elites. Princeton, NJ: Princeton University Press.

Axelrod

Robert M.

1984. The Evolution of Cooperation. New York: Basic Books.

Bar-Tal

Daniel

. 2000. Shared Beliefs in a Society: Social Psychological Analysis. Thousand Oaks, CA: Sage.

Beggan

Dominic M.

2006. “State Repression and Political Violence: Insurgency in Northern Ireland.” International Journal on World Peace 23 (4): 61–90.

Bhatia

Nitin

Kapoor

Namarta

. 2011. “Fuzzy Cognitive Map based Approach for Software Quality Risk Analysis.” ACM SIGSOFT Software Engineering Notes 36 (6): 1–9.

10.

Bhavnani

Ravi

Miodownik

Dan

Choi

Hyun J.

. 2011. “Three Two Tango: Control and Violence in Israel, the West Bank, and Gaza.” Journal of Conflict Resolution 55 (1): 133–58.

11.

Bonham

Matthew G.

Sergeev

Victor M.

Parshin

Pavel B.

. 1997. “The Limited Test-ban Agreement: Emergence of New Knowledge Structures in International Negotiation.” International Studies Quarterly 41 (2): 215–40.

12.

Canetti

Daphna

Hobfoll

Stevan E.

Pedahzur

Ami

Zaidise

Eran

. 2010. “Much Ado about Religion: Religiosity, Resource Support, and Support for Political Violence.” Journal of Peace Research 47 (5): 575–87.

13.

Crenshaw

Martha

. 1981. “The Causes of Terrorism.” Comparative Politics 13 (4): 379–99.

14.

Dudouet

Veronique

. 2013. “Dynamics and Factors of Transition from Armed Struggle to Nonviolent Resistance.” Journal of Peace Research 50 (3): 401–13.

15.

Eagly

Alice H.

Chaiken

Shelly

. 1993. The Psychology of Attitudes. Orlando, FL: Harcourt Brace College Publishers.

16.

Enders

Walter

Sandler

Todd

. 2006. The Political Economy of Terrorism. Cambridge: Cambridge University Press.

17.

Fearon

James

Laitin

David

. 2003. “Ethnicity, Insurgency, War.” American Political Science Review 97 (1): 75–89.

18.

Gansner

Emden R.

North

Stephen C.

. 2000. “An Open Graph Visualization System and Its Applications to Software Engineering.” Software Practice and Experience 30 (11): 1203–33.

19.

Hacker

Frederick

. 1976. Crusaders, Criminals, Crazies: Terror and Terrorism in our Time. New York: W. W. Norton.

20.

Horgan

John

Altier

Mary Beth

. 2012. “The Future of Terrorist De-radicalization Programs.” Georgetown Journal of International Affairs 13:83–90.

21.

Horgan

John

Braddock

Kurt

. 2010. “Rehabilitating the Terrorists? Challenges in Assessing the Effectiveness of De-radicalization Programs.” Terrorism and Political Violence 22 (2): 267–91.

22.

Huntington

Samuel P.

1993. “The Clash of Civilization?” Foreign Affairs 72 (3): 22–49.

23.

Johnston

Alastair Iain

. 1995. Cultural Realism: Strategic Culture and Grand Strategy in Chinese History. Princeton, NJ: Princeton University Press.

24.

Kepel

Gilles

. 2006. Jihad: The Trail of Political Islam. London: I. B. Tauris.

25.

Klein

Jonathan H.

Cooper

Dale F.

. 1982. “Cognitive Maps of Decision-makers in a Complex Game.” The Journal of the Operational Research Society 33 (1): 63–71.

26.

Koller

Daphne

Friedman

Nir

. 2009. Probabilistic Graphical Models: Principles and Techniques. Cambridge, MA: MIT Press.

27.

Kruglanski

Arie W.

Gelfand

Michele

Gunaratna

Rohan

. 2010. “Detainee Deradicalization: A Challenge for Psychological Science.” Observer 32 (1). Accessed June 9, 2014. http://www.psychologicalscience.org/index.php/publications/observer/2010/january-10/detainee-deradicalization.html.

28.

Lewis

Bernard

. 1990. “The Roots of Muslim Rage.” The Atlantic. Accessed September 20, 2012. http://www.theatlantic.com/magazine/archive/1990/09/the-roots-of-muslim-rage/304643/.

29.

Mandler

George

. 2006. “Remembering.” In The Oxford Companion to the Mind, edited by Gregory

Richard L.

. Oxford, UK: Oxford University Press. Accessed October 20, 2013 at http://www.oxfordreference.com/view/10.1093/acref/9780198662242.001.0001/acref-9780198662242-e-728.

30.

Mansdorf

Irwin J.

Kedar

Mordechai

. 2008. “The Psychological Asymmetry of Islamist Warfare.” Middle East Quarterly 15 (2): 37–44.

31.

MATLAB Release. 2011b. The MathWorks, Inc, 3 Apple Hill Drive, Natick, MA. http://www.mathworks.co.uk/company/newsroom/MathWorks-Announces-Release-2011b-of-the-matlab-and-Simulink-Product-Families.html.

32.

McAdam

Doug

Tarrow

Sidney

Tilly

Charles

. 2001. Dynamics of Contention. New York: Cambridge University Press.

33.

McDoom

Omar Shahbudin

. 2013. “Who Killed in Rwanda’s Genocide? Micro-space, Social Influence, and Individual Participation in Intergroup Violence.” Journal of Peace Research 11:1–15.

34.

Miao

Yuan

Miao

Chunyan

Tao

XueHong

Shen

Zhiqi

Liu

ZhiQiang

. 2010. “Transformation of Cognitive Maps.” Fuzzy Systems 18 (1): 114–24.

35.

Murphy

Kevin

. 2001. “Bayes Net Toolbox for MATLAB.” Computing Science and Statistics 33:351–370.

36.

Nadkarni

Sucheta

Shenoy

Prakash P.

. 2001. “A Bayesian Network Approach to Making Inferences in Causal Maps.” European Journal of Operational Research 128:479–98.

37.

Narayanan

V. K.

Armstrong

Deborah J.

. 2005. Causal Mapping for Research in Information Technology. London: Idea Group Inc.

38.

Nemeth

Stephen

. 2014. “The Effect of Competition on Terrorist Group Operations.” Journal of Conflict Resolution 58 (2): 336–62.

39.

Nilsson

Nils J

. 2013. “Understanding Beliefs.” Draft. Accessed October 3, 2013. http://ai.stanford.edu/˜nilsson/.

40.

Oskamp

Stuart

Wesley Schultz

. 2005. Attitudes and Opinions. 3rd ed. Mahwah, NJ: Lawrence Erlbaum Associates.

41.

Papageorgiou

Elpiniki I.

2011. “A New Methodology for Decisions in Medical Informatics using Fuzzy Cognitive Maps based on Fuzzy Rule-extraction Techniques.” Applied Soft Computing 11:500–513.

42.

Pape

Robert A.

2005. Dying to Win: The Strategic Logic of Suicide Terrorism. New York: Random House.

43.

Pearl

Judea

. 2009. Causality: Models, Reasoning, and Inference. Cambridge: Cambridge University Press.

44.

Peng

Zhen

Lifeng

Yang

Bingru

. 2011. “Fuzzy Cognitive Map Based on Multi-relationship.” Scientific Research and Essays 6 (13): 2746–51.

45.

Peters

Butz

. 2004. Toüdlicher Irrtum: die Geschichte der RAF. Berlin, Germany: Argon.

46.

Piazza

James A.

2011. “Poverty, Minority Ethnic Discrimination, and Domestic Terrorism.” Journal of Peace Research 48 (3): 339–53.

47.

Post

Jerrold

Sprinzak

Ehud

Denny

Laurita

. 2003. “The Terrorists in their Own Words: Interviews with Thirty-five Incarcerated Middle Eastern Terrorists.” Terrorism and Political Violence 15:171–84.

48.

Regan

Patrick M.

Henderson

Errol A.

. 2002. “Democracy, Threats and Political Repression in Developing Countries: Are Democracies Internally Less Violent?” Third World Quarterly 23 (1): 119–36.

49.

Reinders

Ralf

Fritzsch

Ronald

. 1995. Die Bewegung 2. Juni. Gespraeche ueber Haschrebellen Lorenzentfuehrung Knast. Amsterdam: Edition ID-Archiv.

50.

Sageman

Marc

. 2004. Understanding Terror Networks. Philadelphia: University of Pennsylvania Press.

51.

Schrodt

Philip

. 1995. Patterns, Rules and Learning: Computational Models of International Behavior. 2nd ed. Vinland, KS: Parus Analytical Systems.

52.

Searle

John R.

2001. Rationality in Action. Cambridge, MA: MIT Press.

53.

Soler

Luciana S.

Kok

Kasper

Camara

Gilberto

Veldkamp

Antoine

. 2011. “Using Fuzzy Cognitive Maps to Describe Current System Dynamics and Develop Land Cover Scenarios: A Case Study in the Brazilian Amazon.” Journal of Land Use Science 7 (2): 1–27.

54.

Spradley

James P.

1979. The Ethnographic Interview. New York: Holt, Rinehart and Winston.

55.

Stenning

Keith

van Lambalgen

Michiel

. 2008. Human Reasoning and Cognitive Science. Cambridge, MA: MIT Press.

56.

Stern

Jessica

. 2003. Terror in the Name of God. New York: HarperCollins Publishers.

57.

Tapio

Tuomas

. 2003. “The United States, European Union and the Development of Cooperation in Competition Policy. Measuring and Modeling Ideas in Foreign Economic Policy.” Doctoral thesis submitted to the Graduate Institute of International and Development Studies, Geneva.

58.

Tarrow

Sidney

. 1998. Power in Movement. Social Movements and Contentious Politics. 2nd ed. New York: Cambridge University Press.

59.

Wills

Tom J.

Cacucci

Francesca

Burgess

Neil

O’Keefe

John

. 2010. “Development of the Hippocampal Cognitive Map in Prewneanling Rats.” Science 328 (5985): 1573–76.

60.

Young

Michael

. 1996. “Cognitive Mapping Meets Semantic Networks.” Journal of Conflict Resolution 40 (3): 395–414.

61.

Zhang

Helen

Shen

Gang

Jin

David

. 2011. “A Research Methodology Based on Design Science for the Construction of Cognitive Maps: The Case of a Chinese Steel Company’s Strategic Decision Making.” Advanced Materials Research 204-10:2098–102.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.58 MB

Hypothesis 1a	0.216433
Hypothesis 1b	0.741767
Hypothesis 2	0.711538
Hypothesis 3a	0.7854
Hypothesis 3b	0.968
Hypothesis 3c	0.710567
Hypothesis 3d	0.0153667
Hypothesis 3e	0.144233
Hypothesis 4	0.9967

Main set of experimental classes				Complementary classes
Super-superclass	Superclass	Belief Class	Rank	Super-superclass	Superclass	Belief Class	Rank
State environment	Negative nature of home state	Continued aggression by home state	+1	State perception	Strength of state	Weakness of home state	−1
		Aggression by home state	+2		Strength of state	Absence of strength of home state	−2
		Domination by home state	+3		Strength of resistance	Strength of international resistance	+1
		Strained living conditions	+4		Strength of resistance	Flexible structure of violent group	+2
		Unreligious state	+5	Goals		Political goal	+2
		Aggression by previous home state	+6	Private issues	Personal life	Experience of violence	−1
		Domination by previous home state	+7	Private issues	Personality	Self-conviction	+1
		Change of home state	+8
		Ignorance of fellow citizens	+9
		Support by home state	−1
		Religious state	−2
		Education of the people	−3
		Absence of domination by home state	−4
		Absence of aggression by home state	−5
		Absence of state structures	−6
	Negative nature of foreign state	Imperialism	+1
	Negative nature of foreign state	Absence of domination by foreign state	−1