Reading the Behavior Signature: Predicting Paul J. Sticha , Elise A. Weaver

Reading the Behavior Signature: Predicting
Leader Personality from Individual and Group Actions
Paul J. Sticha1, Elise A. Weaver1, Joseph A. Tatman2, Suzanne M. Mahoney2, and Dennis M. Buede2
1
Human Resources Research Organization (HumRRO)
66 Canal Center Plaza, Suite 400, Alexandria, VA 22314
2
Innovative Decisions, Inc. (IDI)
1945 Old Gallows Rd, Suite 207, Vienna, VA 22182
{psticha | eweaver}@humrro.org, {jatatman | smahoney | dbuede}@innovativedecisions.com
Abstract1
correctly (Kelley 1967), with limited information and when
outcomes are negative, people tend to attribute their own
behavior to situational constraints but the behavior of
others to personality (Jones and Nisbett 1971; Choi and
Nisbett 1998; Malle 2006). Because individuals
underestimate the impact of situational demands when
inferring personality from behavior, i.e. they fall prey to
the fundamental attribution error (Ross 1977), analysts
wanting to understand a leader’s personality face a difficult
task. This task is often further complicated by the need to
infer leader personality indirectly from group behavior.
The interaction among, situational variables, personal
characteristics, and leader action suggests the need for two
modeling tools: (a) a diagnostic tool that aids inferences
about leader characteristics based on leader and group
actions and the situational context in which they occur, and
(b) a predictive tool that uses inferred leader characteristics
to estimate the likelihood of future actions. The second of
these tools was developed by Sticha, Buede, and Rees
(2005; 2006), who used Bayesian networks to predict
leader actions, based on personality and situational
variables. This paper focuses on the diagnostic component.
The overall goal of the effort was to develop a method to
infer the personality of a leader from information about
leader and group actions and about the situational context
in which these actions occur. A number of scholars
(Mecham 2006; Post, Ruby, and Shaw 2002a; 2002b)
regard a radical group’s use of violence as a strategic
choice, subject to leader personality traits, interpretive
tendencies, and behavioral proclivities, i.e. the “behavior
signature,” along with group and situational factors. To
provide a testbed for model development, we focused on
those leader characteristics related to group violence.
The personality of a leader can be used to predict that
leader’s actions as well as those of the group that he or she
leads. However, except for a small number of well-known
leaders, the personality of leaders must be inferred from
actions and other evidence. We have developed a Bayesian
network to infer leader personality variables related to
violence from evidence of leader and group actions and the
situational demands and context in which the actions occur.
The network was applied to a historical situation, and its
ability to distinguish extreme personalities was established.
Introduction
The term “behavior signature” describes the behavioral
consistency of an individual, group, organization, society,
or nation-state. In the case of a group leader, the behavior
signature facilitates the analysis of intelligence data to
identify and interpret indicators of leader characteristics
and behaviors. Derived from information about the leader,
the signatures could be used to describe the leader’s
methods of operation and to predict the actions that the
leader and the group that he or she leads might take.
A leader’s traits, tendencies, and proclivities making up
his or her personality are an important determiner of the
behavior signature. These personal tendencies interact with
the leader’s knowledge and experience and with the
specific demands of the situation to determine what actions
will be taken in a particular circumstance.
While personal characteristics are well understood for a
small number of well-known national leaders, they must be
inferred for the vast majority of leaders, including leaders
of terrorist cells, adversary military commanders, and
leaders of local organizations or allied forces. Because of
the complex relationship between personal characteristics,
situational demand, and leader action, it is difficult to
attribute a particular leader action to personal or situational
causes. Although people often make these attributions
Model Description
A Bayesian network (Pearl 1988) is a graphical
representation of relationships among a set of random
variables. Probability distributions associated with each
random variable augment the graph. An inference
algorithm uses these distributions to calculate revised
Approved for public release; distribution is unlimited. 88ABW-20080524. Copyright © 2009, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
130
beliefs about the random variables given evidence about
some of them.
knowing the state of Situation d-separates Situation Report
from Behavior.
Finally, complete lack of knowledge about a variable dseparates the variables upon which it depends. For
example, in Figure 1, not knowing anything about
Behavior or Observation of Talking d-separates Situation
from Personality. However, when we observe the behavior
of the individual, the variables Situation and Personality
become dependent.
These restrictions on inference in Bayesian networks
correspond to common-sense inference rules. For example,
knowing the situation does not give information about an
individual’s personality, while observing behavior does.
However, knowledge of both behavior and situation allows
a more precise and accurate estimate of personality.
We were interested in modeling situations in which a
leader’s personality could be indicated by the actions of the
group that he or she leads. Previous research indicates that
a variety of variables are associated with group risk for
violence (Post, Ruby, and Shaw 2002a; 2002b). We
wanted to model the variables that affect group action,
including triggering events, the situational context, culture
and group type, as well as the leader action.
Figure 2 shows a model structure incorporating both
multiple situational variables and group actions. The model
shows that group action probabilities depend on the leader
action and the group characteristics or type. Knowing the
group type does not provide information about the leader
action (or personality) unless something is known about
the group action. The triggering event, culture, and
immediate context are all types of situational variables, so
the change from the basic model merely increases the
specificity with which these variables are represented.
Structure of the Model
The Bayesian network in Figure 1 gives a simplified view
of the relationship between the situation, personality, and
behavior. This network shows that behavior is determined
jointly by personality and the situation. Both the situation
and behavior may be observed, but the personality must be
inferred from other variables.
Situation
Situation Report
Personality
Behavior
Observations of Talking
Figure 1. Basic personality and behavior model.
For example, we might want to know whether an
individual is an introvert or an extravert. In general, we
expect an extravert to talk with others more often than an
introvert does. However, the situation makes a difference
in an individual’s gregarious behavior. In some situations,
such as at a party, we expect to observe most people
talking with others. In other situations, such as in a library,
we expect to observe most people behaving quietly. In both
situations, we expect to observe an introvert talking less
frequently than an extravert. The combination of
situational demands and the individual’s personality cause
the individual to behave in a particular way. We learn
about the situation and behavior through observations or
reports. Given such information, we can make inferences
about the individual’s personality.
Note that in Figure 1, we created nodes specifically to
record observations of behavior and reports of the
situation. Since observations or reports are not always
accurate, we include these nodes to capture our uncertainty
about the information in the reports.
In general, inference flows both with and against the
directions of the arrows in a Bayesian network. However,
there are specific limits to the flow of inference in a
Bayesian network. This limits are characterized by a
property called d-separation (Pearl 1988), which can occur
in three ways.
First, if three variables are linked together in a chain,
then certain knowledge regarding the intermediate variable
halts the flow of information from one end of the chain to
the other end (unless there is another path connecting these
variables). For example, in Figure 1, certain knowledge of
Behavior halts the flow of inference between the variables
Situation and Observations of Talking.
Second, knowledge of a variable d-separates the other
variables that depend on it. For example, in Figure 1,
Triggering Event
Leader Personality
Culture
Leader Action
Immediate Context
Group Action
Group Type
Figure 2. Personality and behavior model showing situational
and group variables.
Representing Personality Traits
In earlier work (Sticha, Buede, and Rees 2005; 2006), we
developed a Bayesian network representing personality
that could be used to predict leader behavior. The
personality variables used in this model were drawn from
the Five Factor Model (Costa and McCrae 1992), one of
the dominant models in personality psychology, and a set
of traits specifically developed to characterize political
leaders (Hermann 1980). The model was subsequently
modified, using a wider variety of personality variables,
incorporating the judgments of over 40 prominent
131
leadership researchers, and employing a formal statistical
analysis to provide model parameters (Sticha et al. 2008).
The modified network was further revised to reverse
arrow directions to make it consistent with the general
model structure shown in Figure 2. The resulting
personality network is shown in Figure 3. The variables in
the left column represent general personality variables
derived from the judgments of the leadership researchers.
Correlations between these variables are represented by the
two variables named DAF Factor and SAS Factor. The
variables in the right column represent classes of leader
actions, with the correlations between these action
categories represented by the variables LF1 and LF2. The
links between the columns represent the relationship
between personality and action categories. For example,
leaders who are high in Drive are more likely to engage in
actions that initiate change and require follow through.
Similarly, leaders who have a high Power Motive are less
likely to act with integrity and have a more conflictual
style for managing conflict.
.878
Drive
6.19
Very Low
23.8
Low
39.9
Average
24.0
High
Very High 6.12
50 ± 11
Follow Through
5.89
Very Low
23.9
Low
40.3
Average
24.0
High
Very High 5.97
50 ± 11
.18
.17
.915
Risk and Adjustment
6.17
Very Low
24.1
Low
39.6
Average
24.0
High
Very High 6.12
50 ± 11
.782
.539
SAS Factor
.95
.836
Flexibility
6.09
Very Low
24.0
Low
40.0
Average
23.7
High
Very High 6.15
50 ± 11
Social Responsibility
5.92
Very Low
24.0
Low
40.4
Average
23.7
High
Very High 5.92
50 ± 11
Power Motive
5.23
Very Low
24.1
Low
41.2
Average
24.2
High
Very High 5.30
50 ± 11
Affiliation Motive
6.03
Very Low
24.1
Low
40.0
Average
23.9
High
Very High 6.06
50 ± 11
Social Perceptiveness
6.15
Very Low
24.0
Low
39.9
Average
23.9
High
Very High 6.09
50 ± 11
.716
LF1
Initiate Change
6.86
Very Low
23.9
Low
38.4
Average
24.0
High
Very High 6.85
50 ± 11
.26
DAF Factor
.863
of the model. Each node represents a leader characteristic
or personality trait. Currently, we can only make inferences
about two characteristics, a leader’s affiliation motive and
social responsibility, because these two variables are
correlated with violent action. Including a wider range of
leader or group behaviors in the model would support
inferences about a greater number of personality variables.
The acts component of the model consists of the
violence of the leader recommended action and of the
group action. The violence of the leader action is
influenced by the situational demand for violence and the
leader propensity to violence. The violence of the group
action is influenced by the violence of leader action, leader
power, and the group’s threat of violence.
The links component consists of variables that link
personality to action. The main linking variables in the
model are the leader’s propensity to violence and social
dominance orientation. The leader’s propensity to violence
is a function of experience with violence. In addition,
propensity to violence and social dominance orientation
are functions of the personality characteristics of affiliation
motive and social responsibility. These linking variables
need to be modeled anew with each action scenario, while
the personality component is stable.
On the bottom right of the model is the group threat of
violence component, a variable that is a function of a
group’s readiness and propensity for violence. For a nonviolent group to engage in violence, the leader must both
be recommending violence and have power over the group.
Each of these requirements indicates something about the
leader’s personality.
Finally, the model includes a situational demand
component. This variable, representing multiple pressures
for violence, influences the violence of the action
recommended by the leader and taken by the group. It has
the strongest influence on a group’s actions when the
leader is low in power.
Situational demand for violence is made up of four
predictors: (a) trigger demand for violence, a measure of
the extent to which past and recent events might trigger a
violent response; (b) expected value of violence, a costbenefit planner’s view of the likely outcome of a violent
act (e.g., future funding and other support may be
enhanced by a successful strike or lost with a failed one);
(c) the culture’s history of violence; and (d) the violence of
the local context. Each of these pressures is felt by both the
leader and the group members.
The model shown in Figure 4 includes the five
components and variables described above and the more
specific indicators that an analyst would provide.
Resourcefulness
6.39
Very Low
24.0
Low
39.4
Average
23.8
High
Very High 6.35
50 ± 11
.31
.36
-.17
Act with Integrity
6.28
Very Low
24.0
Low
39.5
Average
24.0
High
Very High 6.26
50 ± 11
.725
.799
.677
LF2
-.05
.3
.03
.19
Manage Conflict
6.48
Very Low
23.9
Low
39.3
Average
23.9
High
Very High 6.50
50 ± 11
.833
.868
Consult with Others
6.43
Very Low
23.9
Low
39.4
Average
23.8
High
Very High 6.39
50 ± 11
Figure 3. Bayesian network representing personality variables.
Estimating Model Conditional Probabilities
Details of the Personality Prediction Model
We used the following strategies to estimate conditional
probability tables to link model variables, depending on the
source and quality of the information on which the
conditional probabilities were based. (a) In some cases
(e.g., in interpreting the results of Post et al., 2002a, b),
relationships were estimated by subject-matter experts
The structure of the prediction model can be seen in Figure
4. On the left is the personality component of the model
with the same variables that are in Figure 3. Since the goal
of behavioral signature modeling is to make inferences to
personality, the variables in this component are the output
132
Situational Demand
Perceived Benefit due to Success of Viole...
Perceived Benefit due to Success of Pe...
4.56
24.1
42.6
24.1
4.56
Very Low Benefit
Low Benefit
Moderate Benefit
High Benefit
Very High Benefit
4.56
24.1
42.6
24.1
4.56
Very Low Benefit
Low Benefit
Moderate Benefit
High Benefit
Very High Benefit
50 ± 10
Personality Variables
50 ± 10
Perceived Loss due to Failure of Pe...
Perceived Loss due to Failure of Viole...
Drive
.878
Very Low
Low
Average
High
Very High
6.30
24.0
39.0
24.2
6.52
Perceived Likelihood of Success of Viole...
Follow Through
.18
Very Low
Low
Average
High
Very High
50.1 ± 11
5.91
23.9
40.1
24.2
5.84
Expected Value of Violent Solut...
20.0
20.0
20.0
20.0
20.0
Very Low
Low
Moderate
High
Very High
Historical Culture of Violen...
50 ± 10
50 ± 10
Leader Tendency
4.56
24.1
42.6
24.1
4.56
Very Low Loss
Low Loss
Moderate Loss
High Loss
Very High Loss
4.56
24.1
42.6
24.1
4.56
Very Low Loss
Low Loss
Moderate Loss
High Loss
Very High Loss
0.5 ± 0.29
Expected Value of Peaceful Solut...
6.45
28.3
30.7
28.1
6.55
Very Low
Low
Moderate
High
Very High
50 ± 22
20.0
20.0
20.0
20.0
20.0
Very Low
Low
Moderate
High
Very High
50 ± 22
0.4
0.5 ± 0.29
.716
50 ± 11
.50 Notional
Risk and Adjustment
.915
6.75
23.3
38.6
24.9
6.47
Initiate Change
.26
50.1 ± 11
Very Low
Low
Average
High
Very High
Current Communal Conflict
.50 Notional
6.94
23.7
38.2
24.2
7.05
Recent Triggering Event Violen...
.725
LF1
Relative Expected Value of Viole...
5.34
24.0
41.1
24.2
5.34
Very Low
Low
Moderate
High
Very High
Very_Low
Low
Moderate
High
Very_High
50.1 ± 11
50 ± 11
Cultural Demand for Violen...
2.37
25.1
46.4
23.6
2.43
Very Low
Low
Moderate
High
Very High
4.56
24.1
42.6
24.1
4.56
Very Low
Low
Moderate
High
Very High
49.7 ± 18
Flexibility
Very Low
Low
Average
High
Very High
6.01
24.4
40.3
23.3
5.99
49.9 ± 11
Very Low
Low
Average
High
Very High
6.20
23.6
39.8
24.1
6.26
Past Triggering Event Violen...
.782
5.73
23.4
40.1
24.2
6.50
50.3 ± 11
Power Motive
.539
Very Low
Low
Average
High
Very High
5.31
24.0
41.1
24.2
5.31
50.1 ± 11
Affiliation Motive
.95
Very Low
Low
Average
High
Very High
5.79
24.9
39.3
23.9
6.11
50 ± 11
Social Perceptiveness
.836
Very Low
Low
Average
High
Very High
6.01
24.4
40.3
23.3
5.91
0.4
4.56
24.1
42.6
24.1
4.56
.35
Violence of Context
4.56
24.1
42.6
24.1
4.56
Very Low
Low
Moderate
High
Very High
.08
49.8 ± 13
5.54
24.1
40.7
24.1
5.52
Attacks by Ideological Oppone...
Very Low
Low
Moderate
High
Very High
0.4
Situational Demand for Viole...
.34.34
10.8
23.8
32.1
23.0
10.3
Very Peaceful
Peaceful
Neutral
Violent
Very Violent
50 ± 10
Supporter Pressure for Violen...
Very Low
Low
Moderate
High
Very High
50 ± 11
.51
Trigger Demand for Viole...
Very Low
Low
Moderate
High
Very High
50.2 ± 10
0.4
5.00
24.2
41.6
24.2
5.05
Very Low
Low
Moderate
High
Very High
0.4
50 ± 10
4.94
24.1
41.8
24.2
4.98
50 ± 10
Successful Violence by Similar Gro...
50 ± 10
Pro-social Leadership
.36
-.17
.35
6.18
23.7
39.5
24.3
6.36
Very Low
Low
Average
High
Very High
.67 to .33
- .23
Leader Propensity for Viole...
50.1 ± 11
Very Low
Low
Moderate
High
Very High
.677
5.90
24.2
39.8
24.2
5.84
Violence of Leader Recommended Act...
Very Peaceful
Peaceful
Neutral
Violent
Very Violent
.55
50 ± 11
6.92
24.4
38.0
24.0
6.70
Violence of Group Action
Very Peaceful 10.7
Peaceful
23.5
Neutral
32.0
Violent
23.1
Very Violent
10.6
.33 to .67
49.9 ± 11
Group Risk
49.9 ± 13
Manage Conflict
-.05
50 ± 11
SAS Factor
4.78
23.8
41.7
24.6
5.16
Very Low
Low
Moderate
High
Very High
4.94
24.1
41.7
24.3
5.05
50 ± 10
0.6
.799
Social Responsibility
Very Low
Low
Average
High
Very High
Poli, Econ, and Soc Instabi...
Very Low
Low
Moderate
High
Very High
0.6
0.6
Resourcefulness
.31
5.44
24.5
41.1
23.7
5.29
49.9 ± 11
50 ± 10
DAF Factor
.863
4.94
24.0
41.7
24.3
5.05
50.1 ± 10
.17
Very Low
Low
Average
High
Very High
Very Low
Low
Moderate
High
Very High
Perceived Likelihood of Success of Pe...
6.51
28.2
30.6
28.3
6.46
Very Low
Low
Moderate
High
Very High
.3
0.6
6.35
23.7
39.8
23.7
6.39
Very Low
Low
Average
High
Very High
.833
LF2
- .26
.33
Acts
Leader Experience with Viol...
6.14
24.3
39.4
24.1
5.99
Very Low
Low
Moderate
High
Very High
50 ± 11
Threat of Violence from Gro...
49.9 ± 11
Consult with Others
.03
.19
Very Low
Low
Average
High
Very High
Perceived Threat and Opponent Dehumani...
6.29
23.8
39.6
23.7
6.58
Very Low
Low
Moderate
High
Very High
Very Low
Low
Moderate
High
Very High
6.03
24.1
39.7
24.1
6.11
49.9 ± 11
.50 Notional
- .24
0.4
Very Low
Low
Moderate
High
Very High
Very Low
Low
Moderate
High
Very High
50 ± 11
0.6
Very Low
Low
Moderate
High
Very High
0.6
5.37
24.1
40.7
24.2
5.54
50 ± 11
0.6
0.6
Formal Power
Very Low
Low
Moderate
High
Very High
0.6
Group Pressure to be Viole...
4.56
24.1
42.6
24.1
4.56
50 ± 10
5.32
23.5
41.1
24.6
5.50
50.1 ± 11
0.6
.50 Notional
5.84
24.2
39.9
24.2
5.79
Group Experience with Viole...
Very Low
Low
Moderate
High
Very High
Group Propensity for Violen...
4.56
24.1
42.6
24.1
4.56
50 ± 10
Social Dominance Orientati...
4.84
23.8
41.6
24.6
5.15
50.1 ± 10
0.6
Group Readiness for Violen...
Leader Power
Very Low
Low
Moderate
High
Very High
9.06e-5 ± 0.081
- .24
.85
.21
50 ± 11
Very Low
Low
Moderate
High
Very High
50.1 ± 11
.868
6.34
23.8
39.5
24.0
6.33
5.26
23.9
41.1
24.3
5.47
Very Low
Low
Moderate
High
Very High
50.1 ± 11
Group Pressure to Conform
Internal Capability
4.56
24.1
42.6
24.1
4.56
Very Low
Low
Moderate
High
Very High
50 ± 10
5.51
23.9
41.0
24.2
5.40
50 ± 11
External Support
Very Low
Low
Moderate
High
Very High
5.71
25.1
40.8
23.2
5.20
49.7 ± 11
Violence of Group Ideol...
Very Low
Low
Moderate
High
Very High
5.43
23.8
40.9
24.4
5.45
50.1 ± 11
Organizational Incentives for Viole...
5.51
24.2
41.2
23.8
5.31
Very Low
Low
Moderate
High
Very High
49.9 ± 11
Links
Figure 4. Final personality inference model.
using broad categories, such as low, medium, and high. In
these cases, we assigned correlations to these categories,
and used the correlations to estimate the conditional
probability table. (b) When we could not find research that
linked model variables, we were sometimes able to infer
the correlations from their relationships with common
mediating variables. In these cases, we represented the
known correlations in a Bayesian network and generated
cases of the model variables that we then used to estimate
the correlation between model variables. (c) For some
variables, there were no known correlations between
parents and children, so we needed to elicit expert
judgments to determine them. First, we developed
plausible cases using our own estimates of intercorrelations
among the parent variables. Then, we conducted a
judgment analysis with an expert to determine appropriate
relative weights for each of the parent variables in
determining the value of the child variable. (d) Finally, we
created latent variables to absorb common variance in
correlated parent variables. The latent variables were
developed using a factor analysis or structural equation
model based on intercorrelations among the parent
variables.
for control of Gaza in June 2007. The partial timeline in
Table 1 describes the five key events in this battle that
were represented in the cases. Information about the cases
was obtained from open source historical record. We
estimated values for the model parameters from this record
and reviewed our assessments with an expert who was
familiar with the events. These cases provided information
for the model to incorporate to learn personality variables.
Table 1. Summary Timeline for Hamas Takeover of Gaza Strip
Feb 8, 2007 Palestinian rivals meeting in the Islamic holy city
of Mecca, Saudi Arabia reach an accord ensuring a
ceasefire and unity government. Haniyeh sworn in
as Prime Minister of new government. However,
isolated incidents continued through March and
April 2007. More than 90 people were killed in
these first months.
May 19,
Vicious circle of killings and revenge killings
2007
intensifies, resulting in ceasefire. Some violence
continues.
Jun 9, 2007 Ceasefire ends, substantial clashes and revenge
killings between Fatah and Hamas begin.
Jun 12,
Hamas launches full-scale attack after RPG attack
2007
on Haniyeh’s house, laying siege to the Fatahdominated security services, attempting decisive
victory. Hamas has likely made the decision at this
point to takeover Gaza.
Jun 24,
Conciliatory speech, efforts to rally Palestinians
2007
A Sample Application of the Model
In order to test the model with real data, we built five cases
based on the history of the battle between Hamas and Fatah
133
The results of executing this process are shown in Figure
5. Incorporating evidence that the leader’s experience with
violence is very high lowers the expected value of both
Affiliation Motive and Social Responsibility from their
prior means of 50 to the 44-45 range, a drop of
approximately one half of the standard deviation of the
prior distribution, which was set to 10. The two ceasefire
actions come next. Incorporating these actions raises the
expected value of both personality variables to reflect the
fact that these are peaceful actions taken in a violent
environment. The curves now bend down as evidence from
the breaking of the ceasefire and the decision to takeover
the Gaza Strip is incorporated. Finally, the conciliatory
speech given by Haniyeh (the Hamas leader) to rally all
Palestinians raises the expected value of the prosocial
personality variables slightly.
Remember that these curves are cumulative. Each new
case adds evidence to the evidence from all previous cases
read into the model. Note that the two personality variables
track together very closely, indicating that they are highly
correlated in the model.
These variables were set to their lowest category (with a
mean value of 27.5) to represent the violent leader, and to
their highest category (with a mean value of 72.5) to
represent the peaceful leader.
We then generated 1050 cases for each of these two
personalities simulating the values of 25 non-personality
variables included in the model. Since each case was
dependent only on the original personalities we specified,
but was independent of every other case, we split up the
cases into 5 samples of 210 cases each.
To test whether and how quickly the model converged to
the original personality values, we loaded 210 cases into
the model sequentially and plotted the series of personality
variable values that resulted with each added case. Figure 6
and Figure 7 display these results, for the two prosocial
personality variables, Affiliation Motive and Social
Responsibility. In each of these figures, the original values
set for the personality variables are indicated by dashed
lines, with the shorter dashes indicating the peaceful
personality and the longer dashes indicating the violent
personality.
Original - Violent
Haniyeh and Hamas Case Results
80
Aff Motive Expected
Value
Expected Value
52
50
48
46
44
42
Feb 8
May 19
Affiliation Motive
Jun 9
Jun 12
Sample 3 - Violent
Sample 4 - Violent
Sample 5 - Violent
Original - Peaceful
60
Sample 6 - Peaceful
50
Sample 7 - Peaceful
40
Sample 8 - Peaceful
Sample 9 - Peaceful
30
Violent
20
0
Past
Violence
Sample 2 - Violent
Peaceful
70
40
Initial
Sample 1 - Violent
Affiliation Motive (Aff Motive)
50
Jun 24
100
150
Sample 10 - Peaceful
200
Number of Cases
Social Responsibility
Figure 6 Affiliation Motive values as a function of the number of
cases input to the model.
Figure 5. Expected value of personality variables given evidence
from cases.
Original - Violent
Sample 1 - Violent
Social Responsibility (SR)
Validation of the Model
80
SR Expected Value
To validate the model, we tested whether it could
discriminate well between highly violent and highly
peaceful personalities using only case findings that did not
include information about personality. First, we set up two
notional personalities: one highly violent and one highly
peaceful. Next, we generated a random set of cases based
on the leader personality to correspond to observable
events and actions by the leader and group in question.
Finally, we tested to see whether and how fast the model
could recover the original personalities we specified, using
only the cases. To test how reliable the model was at
discriminating between personalities, we ran this validation
for multiple samples of cases.
In the first step, we specified values for the personality
variables corresponding to a highly violent and a highly
peaceful individual. Two personality variables were
relevant, Social Responsibility and Affiliation Motive.
Peaceful
70
Sample 2 Sample 3 Sample 4 Sample 5 -
Violent
Violent
Violent
Violent
Original - Peaceful
Sample 6 - Peaceful
Sample 7 - Peaceful
Sample 8 - Peaceful
60
50
40
Sample 9 - Peaceful
Sample 10 - Peaceful
30
Violent
20
0
50
100
150
200
Number of Cases
Figure 7. Social Responsibility values as a function of the number
of cases input to the model.
Because they are independent, the five samples do not
follow exactly the same path, but all tend toward the
original value of each personality variable. The
convergence appears to be faster for the violent leader than
134
for the peaceful leader. Since the peaceful and violent
samples diverge from 50 in opposite directions, it is clear
that the model successfully discriminates. Furthermore, a
two sample t-tests assuming unequal variance showed
highly significant differences between the violent and
peaceful leaders at the end of the simulation for both
affiliation motive (t = -23.03, p < .0001) and social
responsibility (t = -28.79, p < .0001).
violence, restrictions on the amount of evaluation data, and
the limited validation analysis.
The focus of the model on group violence limits the
number of personality variables that can be predicted. A
more comprehensive personality prediction model would
consider a much wider range of leader and group actions,
categorized along several dimensions and would link to a
larger group of personality variables. In fact, the primary
benefit of the model comes when it incorporates many
personality variables. In that case, personalities estimated
by the model could be used to predict novel types of
behaviors that had never been observed, but that were
related to known personality variables. The results of the
current research suggest that the development of a more
comprehensive model might be both feasible and useful.
The model validation was limited because it relied on
simulated, rather than actual cases. The simulated cases are
guaranteed to be independent, while the independence of
actual cases must be established or assumed. Furthermore,
the use of simulated cases avoids the problem of coding
evidence according to the levels of the model variables.
Finally, the simulated data used in the model validation
assume that the model accurately reflects the relationship
between the variables included in it. Since models are
always imperfect representations of reality, there is a
possibility that the model validation gives an overly
optimistic view of the validity of the model predictions.
The limitations of the current results are typical of
research in an early state of development. The preliminary
results are promising, but they are still preliminary.
Overall, they support the belief that Bayesian networks
provide a good platform for making inferences about
leader personality from actions of the leader and of the
group that he or she leads.
Discussion and Conclusions
The primary goal of the effort was to develop methods to
infer leader personality from information about leader and
group actions and about the situational context in which
these actions occur. Central to this proof-of-concept are the
goals to demonstrate the feasibility of the method and the
validity of the inferences made using it. Although the
results are limited by the preliminary nature of this
research, we believe that they are positive regarding both
feasibility and validity.
Regarding feasibility, the modeling effort has overcome
several obstacles in building the predictive personality
model, including developing a model structure that
supports proper consideration of actions, situations, and
group characteristics to make inferences about leader
personality. In addition, to implement the model, it was
necessary to develop procedures to estimate conditional
probability distributions of variables using information
from a variety of sources, ranging from academic research
results to expert opinion. These procedures were
implemented so that analysts using the model will not need
to estimate large conditional probability tables.
The application of the model to the conflict between
Fatah and Hamas demonstrates the feasibility of using the
approach in a realistic situation. Although, it was
sometimes difficult to make some of the assessments
required by the model, these judgments were possible, and
the results reasonably reflect expectations about the
personality of the Hamas leader, Haniyeh.
The results of the validation for the model show that it
can recover personality information from independent
cases that describe leader and group actions, group
characteristics, and situational variables. Although this
result has limited application to realistic situations in which
the independence of cases cannot be ensured, it
demonstrates that the model is sensitive to the variables
that can predict leader personality. Although the model
requires a relatively large number of cases to converge on
the correct personality values, it can effectively distinguish
extreme personalities with a much smaller number of
cases. In addition, the model provides a logical way to
organize the many situational, leader, and group variables
that are associated with political violence. This feature of
the model provides a benefit to users in addition to the
accuracy of the predictions.
The generally positive results should be interpreted in
light of some of the limitations of the research, which
might come from the exclusive focus of the model on
Acknowledgments
This work was supported by the Air Force Research
Laboratory. The authors wish to thank Professor Fathali M.
Moghaddam for his help in revising the model and
defining the cases for the sample application.
References
Choi, I., and Nisbett, R.E. 1998. Situational salience and cultural
differences in the correspondence bias and in the actor-observer
bias. Personality and Social Psychology Bulletin 24: 949-960.
Costa, P.T., Jr., & McCrae, R.R. 1992. Revised NEO Personality
Inventory (NEO-PI-R) and NEO Five-Factor (NEO-FFI)
Inventory professional manual. Odessa, FL: Psychological
Assessment Resources.
Hermann, M. G. 1980. Explaining foreign policy behavior using
the personal characteristics of political leaders. International
Studies Quarterly, 24: 7-46.
135
Jones, E.E. and Nisbett, R. E. 1971. The actor and the observer:
Divergent perceptions of the causes of behavior. Morristown, NJ:
General Learning Press.
Kelley, H.H. 1967. Attribution theory in social psychology. In D.
Levine ed., Nebraska symposium of motivation 15: 192-240.
Lincoln: University of Nebraska Press.
Malle, B.F. 2006. The actor-observer asymmetry in attribution: A
surprising) meta-analysis. Psychological Bulletin 132: 895-919.
Mecham, Q. 2006. Why do Islamist groups become transnational
and violent? Audit of the Conventional Wisdom, Center for
International Studies, MIT. http://web.mit.edu/cis/acw.htmls
Pearl, J. 1988. Probabilistic Reasoning in Intelligent Systems. San
Francisco, CA: Morgan Kaufmann.
Post, J.M., Ruby, K.G., and Shaw, E.D. 2002a. The radical group
in context: 1. An integrated framework for the analysis of group
risk for terrorism. Studies in Conflict and Terrorism 25: 73-100.
Post, J.M., Ruby, K.G., and Shaw, E.D. 2002b. The radical group
in context: 2. Identification of critical elements in the analysis of
risk for terrorism by radical group type. Studies in Conflict and
Terrorism 25: 101-126.
Ross, L. 1977. The intuitive psychologist and his shortcomings.
In L. Berkowitz ed., Advances in Experimental Social Psychology
10: 173-220. New York: Academic.
Sticha, P.J., Buede D.M., and Rees, R.L. 2005. APOLLO: An
analytical tool for predicting a subject’s decision making. 2005
International Conference on Intelligence Analysis Proceedings.
Bedford, MA: the MITRE Corporation
Sticha, P.J., Buede D.M., and Rees, R.L. 2006. Bayesian model
of the effect of personality in predicting decisionmaker behavior.
Proceedings of the 22nd conference on uncertainty in artificial
intelligence.
Sticha, P.J., Ingerick, M.J., Handy, K., Ramli, M., Furr, D.,
Buede, D.M., and Mahoney, S.M. 2008. Apollo-DI Personality,
Cultural and Modeling Enhancements (FR-08-09). Alexandria,
VA: Human Resources Research Organization.
136