Criminal Justice and Behavior
http://cjb.sagepub.com
Predictive Modeling in Hostage and Barricade Incidents
Patrick Michaud, Michel St-Yves and Jean-Pierre Guay
Criminal Justice and Behavior 2008; 35; 1136
DOI: 10.1177/0093854808319955
The online version of this article can be found at:
http://cjb.sagepub.com/cgi/content/abstract/35/9/1136
Published by:
http://www.sagepublications.com
On behalf of:
International Association for Correctional and Forensic Psychology
Additional services and information for Criminal Justice and Behavior can be found at:
Email Alerts: http://cjb.sagepub.com/cgi/alerts
Subscriptions: http://cjb.sagepub.com/subscriptions
Reprints: http://www.sagepub.com/journalsReprints.nav
Permissions: http://www.sagepub.com/journalsPermissions.nav
Citations (this article cites 31 articles hosted on the
SAGE Journals Online and HighWire Press platforms):
http://cjb.sagepub.com/cgi/content/refs/35/9/1136
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
PREDICTIVE MODELING IN HOSTAGE
AND BARRICADE INCIDENTS
PATRICK MICHAUD
University of Montreal
MICHEL ST-YVES
Sûreté du Québec
JEAN-PIERRE GUAY
University of Montreal
The study examined 534 hostage and barricade incidents. Its objective was to present an integrated and formalized vision of
current conceptualizations of critical incidents to optimize decision making. To this end, logistic regression and recursive partitioning models are presented and compared. In all, 18 distinct static and dynamic individual and situational factors were
used to predict various outputs during this type of situation. Results highlight the importance of considering the origin of the
emergency call and being barricaded in one’s home in the assessment of perpetrator’s propensity for violence, the fact that
time is not always on the police’s side during critical incidents, and the relatively minor importance of hostage taking in the
general comprehension of these high-risk events.
Keywords:
modeling
critical incidents; hostage situations; barricade situations; crisis negotiation; police decision making; predictive
S
ince its very beginnings, the study of hostage and barricade incidents has itself been
hostage to an intuitive and empirically fragmentary conceptual framework. This has not
only limited innovative thought and the development of new ideas but also handicapped
police decision making. In fact, the predictive abilities of law enforcement personnel managing critical incidents have been impaired by both the absence of a formal, integrated
vision of the factors that contribute to the conceptualization of these events and the current
exclusive reliance on the experience, judgement, and intuition of practitioners working in
the field. This lack of decision-making guidance is particularly problematic in the dynamic
context of hostage and barricade incidents, which are characterized by a high risk of violence, high-stakes outcomes, time pressure, and limited information. Consequently, there
seems to be a need for empirical assistance that can support police decision makers in their
efforts to peacefully resolve such complex and volatile situations. This aid can be ensured
by quantitatively modeling and subsequently predicting perpetrators’ behaviors during critical incidents.
AUTHORS’ NOTE: An earlier draft of this article was presented at the 2007 annual meeting of the American
Society of Criminology in Atlanta, Georgia. Please note that the content of this article does not necessarily
reflect the views and opinions of the Sûreté du Québec. Correspondence concerning this article should be
addressed to Patrick Michaud, School of Criminology, University of Montreal, C.P. 6128 succursale centreville, Montreal, Quebec, Canada, H3C 3J7; e-mail: patrick.michaud@umontreal.ca.
CRIMINAL JUSTICE AND BEHAVIOR, Vol. 35 No. 9, September 2008 1136-1155
DOI: 10.1177/0093854808319955
© 2008 International Association for Correctional and Forensic Psychology
1136
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1137
THE PERPETUAL RISK OF VIOLENT BEHAVIOR
The prediction of imminent violent behavior is of crucial importance to decision makers
managing hostage and barricade incidents, mainly because of the omnipresent risk of violence.
There are several reasons for this. First, the instigators of these crisis situations are disorganized, exhibit impaired judgement, and have, especially in the first 45 minutes of the incidents,
a limited ability to fully appreciate the consequences of their actions (Noesner & Dolan, 1992).
Second, the consumption of intoxicants—observed in two thirds of perpetrators (Michaud, StYves, & Guay, 2008)—stimulates irritability and aggressiveness, which accentuates this impairment in judgement and increases the propensity to violence (Fagan, 1990, 1993; Murphy &
Wetzel, 1990; Parker & Auerhahn, 1998). Third, the perpetrators are generally unprepared for
the arrival of the police, as the emergency call to authorities is usually made by a witness or
someone close to them (Michaud et al., 2008). The arrival of the police is therefore an additional emotional burden and a potential source of frustration. Fourth, the recognition by the
perpetrators of the only two probable—but not initially expected—outcomes of the incident
(surrender, with eventual criminal conviction, or death) is surely a significant cause of dissatisfaction that can promote violence. Finally, perpetrators may be driven to violence by negotiators’ incessant attempts to establish verbal contact, by their realization of their precarious
position (besieged, with a limited range of movement), or by their fear of being overcome by
force. Police experience indicates that the potential for violence is higher when the perpetrator
is barricaded in his or her home (Flood & Dalfonzo, 2005) or when he or she disposes of a
considerable arsenal, such as multiple weapons and ample ammunition (Strentz, 2005).
PREDICTING VIOLENCE IN THE ABSENCE OF VALID AND RELIABLE TOOLS
According to Meehl (1954), there are two approaches to risk assessment: application of
predictive actuarial rules based on statistical probability and clinical prognostics based on
practical experience. In critical incidents, each of these approaches has limitations that significantly alter its effectiveness in predicting violence.
The generic actuarial instruments currently used to predict violence are ill suited to the
operational realities of hostage and barricade incidents. First, none of them have been validated for the populations encountered in such situations. Second, although the goal in
hostage and barricade incidents is the prediction of imminent violent behavior (i.e., violence in the first few minutes and hours of the incident), the majority of these instruments
focus on the prediction of violence over the short (days, weeks) or long (months, years)
term. Examples of the former are the Broset Violence Checklist (Almvik, Woods, &
Rasmussen, 2000) and the Short-Term Assessment of Risk and Treatability (Webster,
Martin, Brink, Nicholls, & Middleton, 2004), whereas examples of the latter are Hare’s
Psychopathy Checklist (Hare, 1991, 2003), the Historical, Clinical and Risk Management
20-item scale (Webster, Douglas, Eaves, & Hart, 1997), and the Violence Risk Assessment
Guide (Quinsey, Harris, Rice, & Cormier, 1998). Third, this type of high-risk situation is
associated with decisional urgency, a context for which existing actuarial instruments have
not been developed. For example, some checklists may take considerable time to complete
or may require information that may not be available at the time of assessment. Both of
these requirements may lead to delays in assessing the risk.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1138
CRIMINAL JUSTICE AND BEHAVIOR
Given the current state of actuarial instruments, clinical judgement based on firsthand practical experience is currently the only basis for predicting violence in hostage and barricade incidents. Clinical prognostics not only have poorer predictive power than probabilistic methods
(Grove et al., 2000) and a tendency to overpredict violence (McAuley, 1993) but also are particularly complex to apply in such situations. There are two main reasons for this. First, the
instigators of critical incidents are characterized by a high potential for unpredictable behavior. By its very nature, this type of crisis situation is triggered by individuals “[in] a temporary
state of upset and disorganization, characterized chiefly by an individual’s inability to cope
with a particular situation using customary methods of problem solving” (Slaiku, 1990, p. 15).
This state of disorganization, often compounded by intoxication and/or psychiatric problems
(Michaud et al., 2008), considerably increases the chances that the perpetrator will react irrationally and with no warning. The clinical assessment of the risk of violence—already complicated by unstructured intuitive prognostics and a context in which there is an urgent need
for results—is thus rendered even more difficult. Furthermore, the relatively low prevalence of
hostage and barricade incidents (Hatcher, Mohandie, Turner, & Gelles, 1998; Michaud et al.,
2008) may hinder the predictive ability of decision makers managing critical incidents. This is
because of, among other things, the fact that the effectiveness of clinical prediction is a function of several assessor-specific factors (knowledge, experience, motivation; Garb, 1998).
AN INTUITIVE AND ANECDOTAL PREDICTIVE APPROACH
Currently, the only guidelines for decision making during hostage and barricade incidents
are derived from police experience. Several authors (Fuselier, 1981; Noesner, 1999; Soskis &
Van Zandt, 1986; St-Yves & Tanguay, 2007) have proposed predictors of a positive outcome:
(a) a relationship of trust between the negotiator and the perpetrator, (b) a perpetrator who is
more rational than emotional, (c) a perpetrator who opens up and talks at length, (d) a perpetrator who has stopped making threats and has expressed a desire to live, (e) the absence of
gunshots and of inflicted injury since the beginning of the incident, (f) the release of hostages
(where applicable), (g) the absence of repercussions for having failed to meet ultimatums
issued by the perpetrator, and (h) a perpetrator who talks about his or her immediate future and
the probable consequences of his or her actions and/or surrender.
Other authors (Call, 2003; Fuselier, Van Zandt, & Lanceley, 1991; Soskis & Van Zandt,
1986; Strentz, 2005) have identified predictors of imminent danger during critical incidents: (a) the absence of verbal contact or satisfactory relationships with the perpetrator,
even after several hours of negotiation; (b) the absence of explicit demands from the perpetrator; (c) insistence on face-to-face negotiations; (d) a perpetrator who insists on speaking to a specific person with whom he or she is in conflict; (e) an increasingly expressive
and emotional individual in crisis; (f) a perpetrator who is isolated, has a limited social network, and has recently had to cope with several stress factors; (g) hostages who were deliberately chosen by, were known to, and had a history of conflict with the perpetrator; and
(h) increasingly frequent and convincing threats of violence and/or suicide. The above situational factors are currently the main foundations of the intuitive and anecdotal predictive
approach used by police officers during critical incidents.
The literature on hostage and barricade incidents has not really considered the personal factors associated with violent behavior. Research in psychiatry, psychology, and criminology has
abundantly ploughed this field, however. For instance, significant links have been established
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1139
between heteroaggressive behaviors (those directed against others) and male gender (Cross &
Madson, 1997), youth (Farrington, 1998), poverty (Brownfield, 1986), poor education
(Kenway & Fitzclarence, 1997), unemployment (Rodriguez, Lasch, Chandra, & Lee, 2001),
low IQ (Loeber & Hay, 1997), prior criminal convictions (Salekin, Rogers, & Sewell, 1996),
and mental disorders (Link & Stueve, 1994; Monahan, 1981; Monahan & Steadman, 1994). In
the same vein, other variables have been proved to be valuable predictors of autoaggressive
behaviors (those directed against oneself). These include subject history of suicide attempts
(Kessler, Borges, & Walters, 1999), history of suicide attempts in the subject’s family (Roy,
1983), and depression, often triggered by a significant loss in the subject’s life (Isometsa et al.,
1994). Although these personal predictors of violent behavior are well established in the psychiatric, psychological, and criminological literature, they have not been validated in the particular context of hostage and barricade incidents.
In summary, the ability of police authorities to make predictions in hostage and barricade incidents is limited by several factors. Essentially, decision makers have no instruments adapted to their need to predict imminent violent behavior. Their decisions are thus
exclusively based on professional judgement that is derived from police experience and that
generates empirically untested anecdotal and intuitive predictions. As experience in other
areas of criminology has shown, exclusively speculative prediction is bound to be inadequate (Grove et al., 2000). Given the complex, high-risk, and high-stakes situations with
which law enforcement personnel managing critical incidents are confronted, a structured
decision-making process might yield better results than does the current reliance on intuitive strategies.
The present study was therefore undertaken to improve and formalize decision making
in hostage and barricade incidents, through an empirical approach. To this end, two distinct
predictive methods were used: logistic regression and recursive partitioning modeling. The
initial analyses sought to identify various predictors of the recourse to autoaggressive and
heteroaggressive behaviors by the perpetrator during the critical incident, using several criteria that are easily assessed at the outset of these crisis situations. The more fully decision
makers understand the risk factors present in hostage and barricade incidents, the better
equipped they are to detect high-risk situations and avoid irreparable acts. Subsequent
analyses using the same criteria attempted to distinguish between predictors of three turning
points in critical incident: (a) complete deployment of the specialized crisis-intervention structure, (b) successful verbal contact with the perpetrator, and (c) voluntary and nonviolent
surrender of the perpetrator. By identifying the criteria that best—and most poorly—characterize the principal structural elements of critical incidents, the results of this study will
allow decision makers to make more judicious decisions, which will improve their chances
of nonviolently resolving these high-risk situations.
METHOD
POPULATION AND SAMPLING
We examined 534 hostage and barricade incidents that occurred in Quebec between
1990 and 2004 and were managed by the province’s specialized crisis-intervention structure, the Sûreté du Québec’s Groupe Tactique d’Intervention (GTI; tactical intervention
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1140
CRIMINAL JUSTICE AND BEHAVIOR
group).1 In each incident, the municipal police force, who were the first responders,
assessed the situation and established an exterior perimeter on arriving on the scene. The
situation then either was resolved or deteriorated to the point that they were obliged to
retreat and request the assistance of the Sûreté du Québec’s GTI. The GTI was deployed
whenever the municipal police force was faced with an individual who refused to surrender to the police and at least one of the following conditions was met: (a) Hostages had
been taken, (b) gunshots had been fired, or (c) the lives of the perpetrator or of others was
in imminent danger.
The scope of the present study was therefore the total population of hostage and barricade incidents that were managed by an established intervention structure (Sûreté du
Québec’s GTI) during a defined period (1990 to 2004). However, the events studied are not
representative of all the hostage and barricade incidents that occurred in Quebec between
1990 and 2004. There are two main reasons for this. First, the less “problematic” critical
incidents—with no hostage, no gunshot, and no imminent danger—were usually managed
by municipal police forces without the participation of the Sûreté du Québec. Second, the
hostage and barricade incidents occurring in Montreal, the province’s biggest city, were not
included in the study, as that city has its own specialized crisis-intervention structure, which
manages all critical incidents in its territory, without the intervention of the Sûreté du
Québec (there are some rare exceptions). Thus, this study relies on a biased sample of
hostage and barricade incidents in Quebec which underrepresents less “problematic” critical incidents (i.e., events managed by municipal police forces) as well as crisis situations
occurring in highly urban settings.
PROCEDURE AND CODING
After every crisis situation managed by the GTI, a postincident report of approximately
15 pages was written by the investigators involved. The information contained in these
reports was primarily drawn from testimony (by police officers, victims, witnesses, family
members, acquaintances) and official documents (police and/or correctional service databases, medical and psychiatric files). The use of multiple and cross-checked sources probably contributes to the quality of the information obtained. The document included a
summary of the incident, the operation log, the negotiation log, a list of police personnel
deployed, and descriptive information on the perpetrator and the characteristics of the incident. Of the 534 postincident reports studied, 25 were rejected because they were damaged
or illegible.
To quantitatively analyze the qualitative data appearing in these postincident reports, an
analytical checklist was developed from a sample of 50 randomly selected reports. A
researcher meticulously analyzed the content of the reports and drew up a list of all the
information that was available and efficiently codable. In all, 90 variables were created, and
mutually exclusive categories corresponding to their possible values were clearly defined
on the basis of the data collected. Variables were then coded for each of the 509 postincident reports, a task that required an average of 30 minutes per report. Conferring the coding to a single researcher ensured systematic and standardized measurement of the
variables but had the disadvantage of precluding the validation of the measurements by
analyses of interobserver agreement. To minimize the disadvantage of using one coder, a
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1141
second coder later rescored a random sample of reports, as is described in the results section. The same researcher listened to audio recordings of the negotiations, when these were
available (11.0% of cases), in an attempt to obtain data missing from postincident reports.
Finally, to get more accurate criminal histories of the perpetrators, a restricted-access police
database (Centre des Renseignements Policiers du Québec, Quebec Police Information
Centre) was systematically consulted. It should be noted that this database was also available to decision makers during hostage and barricade incidents.
CONCEPTUALIZATION AND OPERATIONALIZATION
Independent variables. In all, 18 predictors were used in the analyses. These were chosen on the basis of three main criteria: their theoretical relevance in the literature, their practical qualities for those working in the field (good ecological validity, information that is
readily observable and promptly assessable on-scene), and their methodological attributes
(no redundancy, few missing values, sufficient base rate, absence of obvious multicollinearity). The variables were then divided into three distinct domains: static personal
factors, static situational factors, and dynamic situational factors. There were six static personal factors: (a) sex (0 = female, 1 = male), (b) age, (c) presence of known psychiatric
problems, (d) presence of known suicide attempts, (e) number of convictions for nonviolent crimes, including all crimes against property and all other nonviolent crimes not
involving firearms, and (f) number of convictions for violent crimes, including all crimes
against persons and all crimes involving firearms. The six static situational factors were (a)
whether at least one hostage was taken, (b) whether the perpetrator was intoxicated with
alcohol and/or drugs and/or medication, (c) whether the emergency call was made by the
perpetrator, (d) whether the incident was precipitated by conjugal or familial problems, (e)
whether the perpetrator was barricaded in a residential home, and (f) the total number
of weapons (blunt objects, bladed weapons, explosives, firearms, and other offensive
weapons, such as longbows and crossbows) in the perpetrator’s possession at the time of
the incident. Finally, the six dynamic situational factors were (a) the time elapsed before
the perpetrator spoke to a negotiator (i.e., the number of hours between the emergency call
and the beginning of negotiation attempts), (b) whether gunshots were fired at any time, (c)
whether the perpetrator immediately responded to contact (i.e., whether successful verbal
contact was made within the first 5 minutes of attempts to begin negotiations), (d) whether
the perpetrator made explicit demands, (e) the duration of the first verbal contact, as a proportion of the total duration of the negotiation period (the negotiation period was determined from a precise negotiation log that recorded the exact nature of the conversations
between the negotiators and the perpetrator), and (f) the total duration of the negotiation
period, as a proportion of the total duration of the operation (the total duration of the operation corresponds to the time elapsed between the beginning of negotiation attempts and
the end of the operation—i.e., the point at which the perpetrator no longer represents a
threat and/or is under the physical control of the police).
Dependent variables. A total of five dichotomous dependent variables were used for the
analyses. These were grouped into two domains: violent behaviors (two variables) and turning points of the incident (three variables). The first step was to determine the relationship
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1142
CRIMINAL JUSTICE AND BEHAVIOR
between the various predictors and the probability of autoaggressive and heteroaggressive
behaviors by the perpetrator during the police intervention. Accordingly, the first dichotomous variable was the presence of autoaggressive behaviors during the incident. These
behaviors include explicit suicidal threats, writing of a suicide letter or note, self-mutilative
behaviors, dangerous behaviors such as advancing on the police while pointing a firearm at
them, attempted suicide, and suicide. “Suicide by cop” situations—where an individual provokes the police to be killed—were also considered autoaggressive behaviors. The second
dichotomous variable was the presence of heteroaggressive behaviors—toward either
hostages or the police—during the incident. Such behaviors include explicit death threats,
physical assault, sexual assault, deliberate gunshot in the direction of another person for the
purposes of wounding or killing him or her, attempted murder, and murder. These two
dependent variables are thus not limited to the strictest definition of violent behavior but
encompass the entire range of aggressive behaviors, regardless of their gravity. In other
words, a manifestation of aggressiveness (a score of 1) was considered to have existed if the
perpetrator had violent intent and clearly verbalized it, regardless of whether the violence
was actually realized. The next step was to identify the relationship between various predictors and three turning points in the incident. A first dichotomous variable measured whether
the specialized crisis-intervention structure (GTI) had been completely deployed. Complete
deployment, it should be noted, coincided with the beginning of negotiators’ attempts to
make verbal contact. This variable was in some ways a proxy measure of the persistence of
the perpetrator’s disorganized crisis state. A second dichotomous variable indicated the success of negotiators in establishing verbal contact with the perpetrator once the intervention
structure had been fully deployed. A final dichotomous variable reflected whether the fully
deployed intervention structure was able to achieve a voluntary and nonviolent surrender of
the perpetrator. Involuntary and violent surrender (a score of 0) was considered to have
occurred if the perpetrator was killed by the police, committed suicide, or necessitated the
use of force by the tactical intervention unit (commonly known as the SWAT team).
STATISTICAL ANALYSES
To improve predictive power during critical incidents, two distinct but complementary
analyses were conducted. First, we performed logistic regression, an analysis that follows
a logic essentially intended to assess the relative weight of various predictors to quantify
their main effect on the occurrence of a dependent variable. In this study, our logistic
regression analyses had 80% power to detect an odds ratio (OR) of 2.4 for risk factors with
a prevalence of 20% or greater or an OR of 2.0 for risk factors with a prevalence of 40%
or greater, at a two-tailed alpha of .05. Risk factors with ORs of less than 2.0 were unlikely
to be detected by the analyses. Second, we performed recursive partitioning modeling, also
known as classification and regression tree (CART) analysis (Breiman, Friedman, Olshen,
& Stone, 1984). This predictive approach has several advantages over traditional analyses.
Among other things, CART analysis provides a simple graphical decision tree representation of relations, analyzes the hierarchical structure of the prediction, allows for simultaneous analysis of a large number of categorical and continuous predictors, and takes into
account interaction effects. In the present study, CART analysis was performed with the
following criteria: (a) The maximum tree depth was of five levels below root, (b) the minimum number of cases was 50 for parent nodes and 25 for child nodes, and (c) the impurity measure for categorical targets was a Gini coefficient.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1143
Subsequently, the predictive power of each model—obtained from logistic regression or
CART analysis—was estimated by an area under the curve (AUC) coefficient based on
receiver operating characteristic (ROC) analyses. The AUC coefficient varies from .5 (random prediction, accurate in half the cases) to 1.0 (prediction accurate in all cases) and
allows models to be compared to each other to identify the ones with the best predictive
power. AUCs were compared using the totally nonparametric method of DeLong, DeLong,
and Clarke-Pearson (1988), which uses the Mann-Whitney U statistic to estimate the AUC
and its standard error (SE) for each of the correlated curves, followed by the covariance and
correlation matrices of the curves. This procedure uses structural components analysis to
contrast the curves and generates a chi-square statistic with a two-tailed p value.
RESULTS
INTERRATER RELIABILITY ANALYSES
Before proceeding with further statistical analyses, it was crucial to assess the reliability
of our measurements. Consequently, a second person was asked to rescore 18 variables
(i.e., all the predictors used in our analyses) in a random sample of 150 hostage and barricade postincident reports (which represents approximately 30% of the total sample), using
the same analytical checklist used by the initial coder. Overall, results of the interrater reliability analyses show substantial agreement in our measurements (Landis & Koch, 1977;
Shrout & Fleiss, 1979). In fact, the Kappa value (κ) for the 11 dichotomous independent
variables ranged from 0.56 to 1.00 (M = 0.73, SD = 0.15), and the intraclass correlation
coefficient (ρI) for the 7 continuous independent variables ranged from 0.73 to 1.00 (M =
0.89, SD = 0.11). The interrater reliability value of each predictor is presented in Table 1.
DESCRIPTIVE ANALYSES
Descriptive statistics for all of the 18 independent variables used in our analyses are also
presented in Table 1. As for the dependent variables, autoaggressive and heteroaggressive
behaviors were observed in 76.2% and 38.2% of perpetrators, respectively. Complete deployment of the crisis-intervention structure (GTI) was achieved in 74.9% of cases; in 80.6% of
these cases, verbal contact was successfully established with the perpetrator. Finally, 85.8%
of incidents in which the intervention structure was completely deployed concluded with the
voluntary and nonviolent surrender of the perpetrator. For a detailed analysis of the characteristics of the sample use in the present study, see Michaud et al. (2008).
PREDICTION OF PERPETRATOR’S VIOLENT BEHAVIOR
To predict the recourse to autoaggressive and heteroaggressive behaviors during hostage
and barricade incidents, two logistic regression analyses were performed, using the six static personal factors and the six static situational factors. Results are shown in Table 2.
The first logistic regression model allows effective prediction of the recourse to autoaggressive behavior by the perpetrator during the incident (χ2 = 66.57, p < .001; AUC = .75, p <
.001). The 12 predictors making up the model explain 19.6% of the variance of the dependent
variable. The five best predictors were (a) a history of known suicide attempts (OR = 4.72,
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1144
CRIMINAL JUSTICE AND BEHAVIOR
TABLE 1: Descriptive Statistics of the 18 Independent Variables Used in the Analyses
Dichotomous Independent Variables
Sex
Female
Male
Presence of known psychiatric problems
No
Yes
History of known suicide attempts
No
Yes
Holding of at least one hostage
No
Yes
Being intoxicated
No
Yes
Emergency call made by the perpetrator
No
Yes
Incident triggered by conjugal or familial problems
No
Yes
Being barricaded in a residential home
No
Yes
Firing of gunshots since the beginning of the incident
No
Yes
Immediate response to contact
No
Yes
Made explicit demands
No
Yes
Continuous Independent Variables
Age
Number of convictions for nonviolent crimes
Number of convictions for violent crimes
Number of weapons in the perpetrator’s possession
Time elapsed before talking to a negotiator (hours)
First contact–total negotiation proportion (%)
Negotiation–total operation proportion (%)
κ
n
%
1.00
25
484
4.9
95.1
270
239
53.0
47.0
399
110
78.4
21.6
441
68
86.6
13.4
141
281
33.4
66.6
411
81
83.5
16.5
189
255
42.6
57.4
116
393
22.8
77.2
338
165
67.2
32.8
182
191
48.8
51.2
259
119
68.5
31.5
.80
.56
.76
.77
.76
.59
.94
.63
.60
.57
ρI
M
1.00
1.00
.99
.73
.88
.83
.78
37.3
4.7
1.2
1.6
3.9
56.4
42.1
SD
11.6
10.9
3.4
1.4
3.0
43.6
36.5
p < .001), (b) the fact that the call to emergency services was made by the perpetrator himself
or herself (OR = 3.45, p < .01), (c) the presence of known psychiatric problems (OR = 1.99,
p < .01), (d) the triggering of the incident by conjugal or familial problems (OR = 1.86, p < .05),
and (e) the taking of at least one hostage (OR = 0.48, p < .05). The CART model presented in
Figure 1 has a slightly better predictive power than its logistic regression model (AUC = .78,
p < .001) but does not identify the presence of known psychiatric problems as a predictive factor for autoaggressive behaviors. Nevertheless, it indicates that being armed has a significant
impact on the risk of autoaggressive behaviors by the perpetrator, but only if the individual has
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1145
TABLE 2: Logistic Regression Models of Perpetrator’s Violent Behavior
Recourse to
Autoaggressive
Behavior by the
Perpetrator During
Critical Incident
(n = 478)
Predictors
Being a man
Age
Presence of known psychiatric problems
History of known suicide attempts
Number of convictions for nonviolent crimes
Number of convictions for violent crimes
Holding of at least one hostage
Being intoxicated
Emergency call made by the perpetrator
Incident triggered by conjugal or familial problems
Being barricaded in a residential home
Number of weapons in the perpetrator’s possession
Recourse to
Heteroaggressive
Behavior by the
Perpetrator During
Critical Incident
(n = 478)
B
SE
OR
B
SE
OR
–0.32
–0.01
0.69
1.55
–0.01
–0.01
–0.74
0.24
1.24
0.62
–0.55
0.11
0.69
0.01
0.26
0.45
0.01
0.03
0.31
0.24
0.41
0.25
0.32
0.10
0.73
0.99
1.99**
4.72***
0.99
0.99
0.48*
1.27
3.45**
1.86*
0.58
1.11
–0.29
0.01
0.29
0.26
0.02
0.07
1.38
0.86
0.64
–0.10
1.35
0.15
0.57
0.01
0.23
0.27
0.01
0.03
0.32
0.22
0.28
0.22
0.33
0.08
0.75
1.01
1.34
1.30
1.02
1.07*
3.97***
2.35***
1.90*
0.91
3.84***
1.16*
*p < .05. **p < .01. ***p < .001.
no known suicide attempt history. In addition, the decision tree indicates that age influences
the propensity to autoaggressive behaviors in individuals who are armed, who have not contacted emergency services, who have no known suicide attempt history, and whose incident
was not triggered by conjugal or familial problems.
It is also clear that the second logistic regression model also effectively predicts the
recourse to heteroaggressive behavior by the perpetrator during the incident (χ2 = 97.78,
p < .001; AUC = .75, p < .001). Thus, the 12 predictors of the model explain 25.1% of the
variance of the dependent variable. The six best predictors were (a) at least one hostage
having been taken (OR = 3.97, p < .001), (b) a perpetrator barricaded in a residential home
(OR = 3.84, p < .001), (c) intoxication by alcohol and/or drugs and/or medication (OR =
2.35, p < .001), (d) the fact that the call to emergency services was made by the perpetrator himself or herself (OR = 1.90, p < .05), (e) the number of convictions for violent crimes
(OR = 1.07, p < .05), and (f) the number of weapons in the perpetrator’s possession (OR = 1.16,
p < .05). The corresponding CART model, presented in Figure 2, has a predictive power
identical to its logistic regression model (AUC = .75, p < .001). However, the decision tree
shows that the number of convictions for nonviolent crimes is also a predictor of heteroaggressive behaviors, although only for individuals who are armed, not intoxicated, have no
more than two convictions for violent crimes, and are barricaded in a residential home.
PREDICTION OF THE OCCURRENCE OF TURNING POINTS OF THE INCIDENT
To predict the occurrence of three turning points of the incident, three other logistic
regression analyses were performed using the same 12 independent factors. Furthermore,
six other dynamic situational factors with appropriate temporal relationships (the independent variables had to have been measured before the dependent variables) were progressively integrated into the models. Results are presented in Table 3.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1146
CRIMINAL JUSTICE AND BEHAVIOR
Figure 1: Classification and Regression Tree of the Recourse to Autoaggressive Behavior by the
Perpetrator During Critical Incident
Note. AUC = area under the curve.
The first logistic regression model effectively predicts the complete deployment of the
specialized crisis-intervention structure (χ2 = 48.88, p < .001; AUC = .72, p < .001), the 12
predictors explaining 14.5% of the variance of this dependent variable. The four best predictors were (a) the presence of known psychiatric problems (OR = 2.00, p < .01), (b) the
triggering of the incident by conjugal or familial problems (OR = 1.91, p < .01), (c)
age (OR = 1.02, p < .05), and (d) the number of weapons in the perpetrator’s possession
(OR = 1.26, p < .05). The CART model, presented in Figure 3, has a slightly poorer predictive power than does the corresponding logistic regression model (AUC = .69, p < .001).
In addition to the four preceding predictors, the decision tree also identifies the physical
location of the incident as a predictor of the complete deployment of the crisis-intervention
structure, but only in cases involving individuals with known psychiatric problems. The
number of convictions for violent crimes is a predictive factor in the CART model, but only
for individuals with no known psychiatric problems, armed with two weapons or less, and
whose incident was triggered by conjugal or familial problems.
The second logistic regression model effectively predicts the success of attempts to make
verbal contact with the perpetrator following the deployment of the crisis-intervention
structure (χ2 = 30.39, p < .01; AUC = .72, p < .001). The 14 predictors explain 13.3% of
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1147
Figure 2: Classification and Regression Tree of the Recourse to Heteroaggressive Behavior by the
Perpetrator During Critical Incident
Note. AUC = area under the curve.
the variance of the dependent variable. The three best predictors were (a) at least one gunshot having been fired since the beginning of the incident (OR = 0.34, p < .001), (b) age
(OR = 0.97, p < .05), and (c) the number of hours elapsed before the perpetrator talked to
a negotiator (OR = 0.90, p < .05). The CART model, presented in Figure 4, has a slightly
poorer predictive power than does the corresponding logistic regression model (AUC = .67,
p < .001). The CART model does not identify the presence of gunshots since the beginning
of the incident as a significant predictor but does suggest that the triggering event of the
incident is a predictor of a successful verbal contact with perpetrators who waited less than
4.3 hours before talking to a negotiator and were no older than 49.
Finally, the last logistic regression model effectively predicts the voluntary and nonviolent
surrender of the perpetrator once the crisis-intervention structure has been deployed (χ2 =
31.50, p < .05; AUC = .73, p < .001). The 18 predictors explain 16.3% of the variance of the
dependent variable. The three most significant predictors were (a) the total duration of the
negotiation period, as a proportion of the total duration of the operation (OR = 1.02, p < .01),
(b) the taking of at least one hostage (OR = 0.36, p < .05), and (c) the firing of at least one gunshot since the beginning of the incident (OR = 0.50, p < .05). The CART model, presented in
Figure 5, has a poorer predictive power than the corresponding logistic regression model (AUC =
.65, p < .001). The decision tree identifies only one significant predictor, namely the total
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1148
CRIMINAL JUSTICE AND BEHAVIOR
TABLE 3:
Logistic Regression Models of Incident Turning Points
Specialized
Crisis-Intervention
Structure Was
Completely
Deployed (n = 479)
Predictors
B
Being a man
–0.82
Age
0.02
Presence of known psychiatric problems
0.69
History of known suicide attempts
0.28
Number of convictions for nonviolent crimes
0.01
Number of convictions for violent crimes
0.11
Holding of at least one hostage
–0.25
Being intoxicated
0.27
Emergency call made by the perpetrator
–0.32
Incident triggered by conjugal or
0.65
familial problems
Being barricaded in a residential home
0.14
Number of weapons in the perpetrator’s
0.23
possession
Time elapsed before talking to a
—
negotiator (hours)
Firing of gunshots since the beginning
—
of the incident
Immediate response to contact
—
Made explicit demands
—
First contact–total negotiation proportion (%)
—
Negotiation–total operation proportion (%)
—
Successful Verbal
Contact Established
With the Perpetrator
(n = 358)
Voluntary and
Nonviolent
Surrender
of the Perpetrator
(n = 330)
SE
OR
B
SE
OR
B
SE
OR
0.66
0.01
0.26
0.31
0.01
0.06
0.33
0.24
0.30
0.24
0.44
1.02*
2.00**
1.33
1.01
1.11
0.78
1.31
0.73
1.91**
–0.50
–0.03
–0.15
0.00
0.00
0.03
0.17
0.11
0.54
–0.35
0.81
0.01
0.32
0.37
0.02
0.05
0.46
0.32
0.48
0.32
0.60
0.97*
0.86
1.00
1.00
1.02
1.18
1.12
1.72
0.71
–1.12
–0.02
0.00
0.30
–0.01
0.09
–1.03
–0.05
–0.18
–0.11
1.09
0.02
0.37
0.45
0.02
0.08
0.49
0.39
0.48
0.37
0.33
0.98
1.00
1.34
0.99
1.09
0.36*
0.95
0.83
0.90
0.29
0.11
1.15
1.26*
–0.01
0.18
0.41
0.12
0.99
1.19
0.46 0.46 1.59
0.08 0.13 1.08
—
—
–0.10
0.05
0.90*
0.00 0.06 1.00
—
—
–1.08
0.31
0.34*** –0.70 0.35 0.50*
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
0.06
–0.29
0.00
0.02
0.46
0.40
0.00
0.01
1.06
0.75
1.00
1.02**
*p < .05. **p < .01. ***p < .001.
Figure 3:Classification and Regression Tree of the Complete Deployment of the Specialized CrisisIntervention Structure
Note. AUC = area under the curve.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1149
Figure 4: Classification and Regression Tree of the Successful Verbal Contact Established With the
Perpetrator
Note. AUC = area under the curve.
Figure 5: Classification and Regression Tree of the Voluntary and Nonviolent Surrender of the Perpetrator
Note. AUC = area under the curve.
duration of the negotiation period, as a proportion of the total duration of the operation. The
CART model indicates that 94.5% of the incidents in which this proportion exceeded 50%
ended in the voluntary and nonviolent surrender of the perpetrator, compared to 79.8% of the
incidents in which this proportion was less than this critical threshold of 50%.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1150
CRIMINAL JUSTICE AND BEHAVIOR
DISCUSSION
The nature, frequently described as unpredictable, of hostage and barricade incidents
naturally calls into question the ability to effectively predict the outcome of these high-risk
situations. The results of this study indicate that it is possible not only to predict the
sequence of events in these incidents but also to do so with considerable precision. The
AUC coefficients of the logistic regression models (.72 to .75) and CART models (.65 to
.78) indicate that the models have fair predictive power (Swets, 1988). This predictive
power compares well to that obtained in other studies in which the prediction of violence
is a preoccupation (Quinsey et al., 1998). Although results from both models varied, AUC
comparisons using the method of DeLong et al. (1988) revealed no significant differences
between the logistic regression models and their corresponding CART models.
One of the main criticisms noted in the introduction was that prediction during hostage
and barricade incidents has been based exclusively on clinical criteria drawn from police
experience; for this reason, it appeared relevant to develop probabilistic criteria that would
improve and organize this prediction. The logistic regression analyses in this study demonstrate that several risk factors discussed in the introduction and identified by practitioners
who have studied hostage and barricade incidents are empirically well founded. However,
the problems with so-called clinical predictions are related not to the poor choice of predictors but to their ineffective combination (Meehl, 1954). Practitioners’ difficulty in
making accurate predictions is partly related to the nonoptimal weighting and organization
of predictors. Just as logistic regression permits the attribution of relative weight to various
predictors, other statistical techniques such as recursive partitioning modeling allow predictors to be reorganized in mechanical or algorithmic structures (Grove & Meehl, 1996).
The CART models are much better suited to concrete application, for several reasons.
First, the sequential nature of these hierarchical prediction models—and their use of conditional
probabilities—seems to mirror the decision-making logic of police forces managing critical
incidents (Zimmerman, 2006). Second, the CART models allow the identification of subtle
interaction effects that are difficult to detect with traditional main-effect models. The generation and testing of such interaction effects by logistic regression requires a preexisting
theoretical organization, which is problematic in a field of study such as hostage and barricade incidents, in which there are currently no theoretical models. Finally, the CART
models’ simple representation of relationships is readily transposable to the field and easily understood and interpreted by nonspecialists.
THE DUAL NATURE OF THE TIME FACTOR
Several authors have underscored the importance of negotiators saving time during hostage
and barricade incidents (Fuselier, 1981; McMains & Mullins, 2001; Mullins, 2003; Strentz,
2005; St-Yves, Michaud, & Tanguay, 2007; St-Yves & Tanguay, 2007; Van Zandt, 1991;
Vecchi, Van Hasselt, & Romano, 2005). Letting time pass has decided advantages during
hostage and barricade incidents, regardless of whether it serves to (a) reduce the anxiety or
emotional burden of the individual in crisis, (b) completely and optimally deploy a crisisintervention structure, (c) intensively gather information, (d) create a relationship of trust
between the perpetrator and negotiators, (e) increase the chances of establishing the Stockholm
syndrome (if there are hostages), (f) reduce the undesirable effects of intoxicants, and/or (g)
foster the desire for certain basic, but negotiable, needs (eating, sleeping, smoking). However,
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1151
two observations in this study contradict the widespread idea that time is always on the police’s
side during critical incidents. First, from the moment that the emergency call is made, every
hour that passes without contact with a negotiator diminishes the chances that a successful verbal contact will be established with the perpetrator. Our results indicate that no more than 4.3
hours should pass before initiating contact attempts with a perpetrator; after this time, the
chances of making verbal contact diminish. This is a significant drawback that limits the possibilities of beginning effective negotiations intended to increase the chances of voluntary and
nonviolent surrender of the perpetrator. Second, the greater the proportion of the operation
spent in negotiations, the greater the probabilities of voluntary and nonviolent surrender of the
perpetrator. This result underscores the crucial importance of not only negotiation per se but
also maximizing the duration of negotiation by minimizing the periods in which there is no
verbal contact between the perpetrator and negotiators. Maintaining a constant communication
channel between the police (emitter) and the individual in crisis (receptor) allows the latter to
vent negative emotions, establish a relationship of trust with negotiators, find alternatives to the
situation, and minimize the time available to brood over problems. Negotiators should make
the establishment of continual verbal contact with the perpetrator a priority, as it appears to be
one of the primary guarantees of his or her voluntary and nonviolent surrender. Our results
show that the threshold duration of negotiation for improving the chances of voluntary and
nonviolent surrender is 50%, which means that at least 50% of the total duration of the operation should be spent in negotiation. Because hostage and barricade incidents are laced with
imponderables (e.g., technical problems, interruption of communication by the perpetrator,
categorical refusal of the perpetrator to establish verbal contact), maintaining negotiations for
more than half the operation may prove particularly difficult. However, the incapability to
establish sustained communication is usually because of the whims of the receptor rather than
failures by the emitter.
THE PAST AS KEY TO THE FUTURE
The presence of a history of known suicide attempts in perpetrators multiplies the
chances of autoaggressive behaviors during the critical incident fivefold. The presence of a
suicidal history is an established risk factor for suicide (Kessler et al., 1999), reflecting the
idea that past autoaggressive behaviors effectively predict similar behaviors in the future.
If autoaggression is already part of an individual’s repertoire of solutions for overcoming
specific obstacles, the occurrence of a new, apparently insurmountable problem (i.e., the
triggering event of the critical incident) will predispose him or her to adopt this adaptive
operational mode. The reaction of a suicidal individual to his or her previous suicide
attempt is nevertheless an essential element to be taken into consideration. For instance,
research by Henriques, Wenzel, Brown, and Beck (2005) suggests that among individuals
who have attempted suicide, the risk of committing suicide in the future are 2.5 times greater
in individuals who regret not having died than in individuals who are ambivalent or happy
to be alive. Furthermore, the literature on violence provides many examples of previous
violent behavior being the best predictor of future violent behavior (Salekin et al., 1996).
This pattern can also be seen in critical incidents: The greater an individual’s history of
criminal violence, the greater the chances that he or she will resort to heteroaggressive
behaviors during the incident. Similarly, individuals who see violence as a conflict resolution strategy will be more inclined to be violent in the face of a new conflict.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1152
CRIMINAL JUSTICE AND BEHAVIOR
THE ORIGIN OF THE EMERGENCY CALL AS AN INDICATOR OF IMMINENT DANGER
When it is the perpetrator who makes the call to emergency services, the risks of autoaggressive and heteroaggressive behaviors during the incident significantly increase. Such an
act by a suicidal individual indicates an urgent situation and suggests that the perpetrator is
at the end of the suicidal process, alone, despairing, and in pain. The call to authorities is
thus part of a final attempt to obtain the aid, support, and audience that will perhaps help
one finally find an alternate solution that will revive his or her hope. The recourse to
autoaggressive behaviors is thus imminent, as individuals generally make the call at the
precise moment that they can no longer tolerate their suffering. For some, the emergency
call to authorities may reflect an explicit plan in which the presence of the police is a
required element (quest for an audience, desire for publicity, need for an intermediary to
convey a message, etc.). The incident is thus not circumstantial, but planned, with precise
objectives (suicide, murder, ransom, specific demands, etc.). As a result, violence toward
those who present an obstacle to the end in question is a possibility.
KING OF THE CASTLE
In critical incidents, individuals barricaded in a residential home are 4 times more likely
to resort to some form of violence toward others. There are two main reasons for this. The
first reflects the “king of the castle” pattern of behavior (Flood & Dalfonzo, 2005): Most
barricaded individuals consider themselves safe, indeed untouchable, in a residential home
(which belongs to them 85% to 90% of the time, see Michaud et al., 2008). Therefore, some
may react very strongly to the possible loss of their last bastion, and their ferocious resistance may take the form of violence toward others. The second is derived from the advantages conferred on a perpetrator by being barricading in his or her own residence: a known
environment, effective protection by the walls of the building, availability of several highquality shooting positions (windows, doors, roof), ability, in many cases, to see the police
approach the location, and full control over everything that occurs within the environment
(particularly with regard to the treatment of hostages).
THE HOSTAGE AS “SUPPORTING CAST”
Research on critical incidents has often seen hostages as fundamental structural elements, occupying a central and preponderant place in the conceptualization of these highrisk events. However, the results of this study do not corroborate this widespread view. In
fact, our results suggest that the taking of hostages is of only minor importance for the general comprehension of this type of situation. Although the analyses indicate that the holding of a hostage increases the risk of heteroaggressive behaviors and of the violent
conclusion of the incident, its importance is far from dominant. Rather than being the foundation of critical incidents, the taking of hostages appears to be but one of a large number
of risk factors to be taken into account, with no exceptional contribution to the overall variance. It is certainly true that some situations are almost exclusively focused on the presence
of a hostage, particularly when the hostage is used as an intensely negotiated currency of
exchange. However, the prevalence of this type of event seems so low (McMains &
Mullins, 2001; Michaud et al., 2008) that its impact on the overall understanding of critical incidents is negligible. Therefore, our results do not support a dichotomous conception
of critical incidents based on the presence or absence of a hostage.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1153
CONCLUSION AND LIMITATIONS
To optimize decision making during hostage and barricade incidents, this study
attempted, through an empirical approach, to organize and improve prediction. Through
statistical analyses, the study demonstrated that it was possible to use a series of static and
dynamic individual and situational factors to predict the sequence of events of critical incidents effectively. Several valuable predictors that can be readily identified and evaluated by
decision makers were identified; each of these has its own predictive power and makes a
specific contribution to the overall understanding of this type of high-risk situation.
This study does, however, have methodological limitations that should be noted. First,
as indicated above, the sample underrepresents less “problematic” critical incidents as well
as those occurring in highly urban settings. Second, recommended methodological strategies inevitably underestimate the prevalence of psychiatric problems and suicidal history
among the instigators of this type of incident. Furthermore, the predictive analyses of perpetrators’ aggressive behaviors follow a dichotomous logic that disregards the continuous
nature of the spectrum of violent behavior. For example, verbal threats were given the same
weight as suicide or murder; this limits the scope and generalizability of the results
obtained. This situation is essentially because of the low prevalence of successful suicides
(8.9% of the total sample), murder of hostages (1.4% of the total sample), and murder of
police personnel (0.2% of the total sample). Finally, for methodological reasons, several
aspects of hostage and barricade incidents could not be coded (negotiator’s characteristics,
nature of negotiation, nature of interactions between the hostage takers and their captors,
etc.), which means that some of the models presented here could be underspecified.
The objective of this study was to develop an integrated and formalized conception of
critical incidents to optimize decision making. In addition to testing the validity of our
results, future studies should attempt to develop a more integrated view of the “lead actors”
and their motives.
NOTE
1. Located in eastern Canada, Quebec is the largest province in Canada, with an area of about 1,542,056 km2 (595,391
sq. mi.) and a population (2007) of approximately 7.7 million.
REFERENCES
Almvik, R., Woods, P., & Rasmussen, K. (2000). The Broset Violence Checklist. Journal of Interpersonal Violence, 15, 12841296.
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees. Belmont, CA:
Wadsworth.
Brownfield, D. (1986). Social class and violent behavior. Criminology, 24, 421-438.
Call, J. A. (2003). Negotiating crises: The evolution of hostage/barricade crisis negotiation. Journal of Threat Assessment,
2(3), 69-94.
Cross, S. E., & Madson, L. (1997). Models of the self: Self-construals and gender. Psychological Bulletin, 122, 5-37.
DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver
operating characteristic curves: A nonparametric approach. Biometrics, 44, 837-845.
Fagan, J. (1990). Intoxication and aggression. In M. Tonry & J. Q. Wilson (Eds.), Crime and justice: Vol. 13. Drugs and crime
(pp. 241-320). Chicago: University of Chicago Press.
Fagan, J. (1993). Interactions among drugs, alcohol, and violence. Health Affairs, 12(4), 65-79.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
1154
CRIMINAL JUSTICE AND BEHAVIOR
Farrington, D. P. (1998). Predictors, causes, and correlates of male youth violence. In M. Tonry & M. H. Moore (Eds.), Crime
and justice: Vol. 24. Youth violence (pp. 421-475). Chicago: University of Chicago Press.
Flood, J., & Dalfonzo, V. (2005, October). Critical incident. Paper presented at the 11th Crisis Negotiation Seminar, Calgary,
Alberta, Canada.
Fuselier, G. D. (1981). A practical overview of hostage negotiations [Part I and Conclusion]. FBI Law Enforcement Bulletin,
50(6-7), 2-6, 10-15.
Fuselier, G. D., Van Zandt, C. R., & Lanceley, F. J. (1991). High risk factors and the “action criteria” in hostage/barricade
situations. FBI Law Enforcement Bulletin, 60(1), 6-12.
Garb, H. N. (1998). Studying the clinician: Judgment research and psychological assessment. Washington, DC: American
Psychological Association.
Grove, W. M., & Meehl, P. E. (1996). Comparative efficiency of informal (subjective, impressionistic) and formal (mechanical,
algorithmic) prediction procedures: The clinical-statistical controversy. Psychology, Public Policy, and Law, 2, 293-323.
Grove, W. M., Zald, D. H., Hallberg, A. M., Lebow, B., Snitz, E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment, 12, 19-30.
Hare, R. D. (1991). The Hare Psychopathy Checklist–Revised. Toronto, Ontario, Canada: Multi-Health Systems.
Hare, R. D. (2003). Manual for the Revised Psychopathy Checklist (2nd ed.). Toronto, Ontario, Canada: Multi-Health
Systems.
Hatcher, C., Mohandie, K., Turner, J., & Gelles, M. G. (1998). The role of the psychologist in crisis/hostage negotiations.
Behavioral Sciences and the Law, 16, 455-472.
Henriques, G., Wenzel, A., Brown, G. K., & Beck, A. T. (2005). Suicide attempters’ reaction to survival as a risk factor for
eventual suicide. American Journal of Psychiatry, 162, 2180-2182.
Isometsa, E. T., Henriksson, M. M., Aro, H. M., Heikkinen, M. E., Kuoppasalmi, K. I., & Lonnqvist, J. K. (1994). Suicide in
major depression. American Journal of Psychiatry, 151, 530-536.
Kenway, J., & Fitzclarence, L. (1997). Masculinity, violence and schooling: Challenging “poisonous pedagogies.” Gender
and Education, 9, 117-133.
Kessler, R. C., Borges, G., & Walters, E. E. (1999). Prevalence of and risk factors for lifetime suicide attempts in the National
Comorbidity Survey. Archives of General Psychiatry, 56, 617-626.
Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33, 159-174.
Link, B. G., & Stueve, A. (1994). Psychotic symptoms and the violent/illegal behavior of mental patients compared to community controls. In J. Monahan & H. J. Steadman (Eds.), Violence and mental disorder: Developments in risk assessment
(pp. 137-159). Chicago: University of Chicago Press.
Loeber, R., & Hay, D. (1997). Key issues in the development of aggression and violence from childhood to early adulthood.
Annual Review of Psychology, 48, 371-410.
McAuley, F. (1993). Insanity, psychiatry and criminal responsibility. Dublin, Ireland: Round Hall Press.
McMains, M. J., & Mullins, W. C. (2001). Crisis negotiations: Managing critical incidents and hostage situations in law
enforcement and corrections (2nd ed.). Cincinnati, OH: Anderson.
Meehl, P. E. (1954). Clinical versus statistical prediction. Minneapolis: University of Minnesota Press.
Michaud, P., St-Yves, M., & Guay, J. P. (2008). Analyse descriptive des situations de prise d’otages et de barricade au Québec
[A descriptive analysis of hostage and barricade incidents in Quebec]. Revue Internationale de Criminologie et de Police
Technique et Scientifique, 61, 133-150.
Monahan, J. (1981). Predicting violent behavior: An assessment of clinical techniques. Beverly Hills, CA: Sage.
Monahan, J., & Steadman, H. J. (Eds.). (1994). Violence and mental disorder: Developments in risk assessment. Chicago:
University of Chicago Press.
Mullins, W. C. (2003). The role of the first responder at a hostage situation. Journal of Police Crisis Negotiations: An
International Journal, 3(1), 31-49.
Murphy, G. E., & Wetzel, R. D. (1990). The lifetime risk of suicide in alcoholism. Archives of General Psychiatry, 47, 383-392.
Noesner, G. W. (1999). Negotiation concepts for commanders. FBI Law Enforcement Bulletin, 68(1), 6-14.
Noesner, G. W., & Dolan, J. T. (1992). First responder negotiation training. FBI Law Enforcement Bulletin, 61(8), 1-4.
Parker, R. N., & Auerhahn, K. (1998). Alcohol, drugs, and violence. Annual Review of Sociology, 24, 291-311.
Quinsey, V. L., Harris, G. T., Rice, M. E., & Cormier, C. A. (1998). Violent offenders: Appraising and managing risk.
Washington, DC: American Psychological Association.
Rodriguez, E., Lasch, K. E., Chandra, P., & Lee, J. (2001). Family violence, employment status, welfare benefits, and alcohol drinking in the United States: What is the relation? Journal of Epidemiology and Community Health, 55, 172-178.
Roy, A. (1983). Family history of suicide. Archives of General Psychiatry, 40, 971-974.
Salekin, R. T., Rogers, R., & Sewell, K. W. (1996). A review and meta-analysis of the Psychopathy Checklist and Psychopathy
Checklist–Revised: Predictive validity of dangerousness. Clinical Psychology: Science and Practice, 3, 203-215.
Shrout, P. E., & Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin, 86, 420-428.
Slaiku, K. A. (1990). Crisis intervention (2nd ed.). Boston: Allyn & Bacon.
Soskis, D. A., & Van Zandt, C. R. (1986). Hostage negotiation: Law enforcement’s most effective nonlethal weapon. Behavioral
Sciences and the Law, 4, 423-435.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008
Michaud et al. / PREDICTIVE MODELING IN CRITICAL INCIDENTS
1155
Strentz, T. (2005). Psychological aspects of crisis negotiation. Boca Raton, FL: CRC Press.
St-Yves, M., Michaud, P., & Tanguay, M. (2007). L’évaluation des risques en situation de crise [Risk assessment in crisis situation]. In M. St-Yves & M. Tanguay (Eds.), Psychologie de l’enquête criminelle: La recherche de la vérité (pp. 363376). Cowansville, Quebec, Canada: Éditions Yvon Blais.
St-Yves, M., & Tanguay, M. (2007). Psychologie de la négociation de crise [Psychology of crisis negotiation]. In M. St-Yves
& M. Tanguay (Eds.), Psychologie de l’enquête criminelle: La recherche de la vérité (pp. 327-362). Cowansville, Quebec,
Canada: Éditions Yvon Blais.
Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science, 240, 1285-1293.
Van Zandt, C. R. (1991). Hostage situations: Separating negotiation and command duties. FBI Law Enforcement Bulletin,
60(1), 18-19.
Vecchi, G. M., Van Hasselt, V. B., & Romano, S. J. (2005). Crisis (hostage) negotiation: Current strategies and issues in highrisk conflict resolution. Aggression and Violent Behavior, 10, 533-551.
Webster, C. D., Douglas, K. S., Eaves, D., & Hart, S. D. (1997). HCR-20: Assessing risk for violence (2nd ed.). Burnaby,
British Columbia, Canada: Simon Fraser University, Mental Health, Law, and Policy Institute.
Webster, C. D., Martin, M.-L., Brink, J., Nicholls, T. L., & Middleton, C. (2004). The Short Term Assessment of Risk and
Treatability (START). Hamilton and Port Coquitlam, British Columbia, Canada: St. Joseph’s Healthcare Hamilton and
Forensic Psychiatric Services Commission.
Zimmerman, L. A. (2006). Law enforcement decision making during critical incidents: A three-pronged approach to understanding and enhancing law enforcement decision processes. Dissertation Abstracts International, 67(04). (UMI No.
3214014)
Patrick Michaud, MSc, is completing his doctorate in criminology at the University of Montreal. His master’s thesis, conducted in conjunction with the Behavioural Science Unit of the Sûreté du Québec, investigated hostage and barricade incidents in Quebec. His current research interests focus on disjunctive escalation of aggression, policing, and sex offenders.
Michel St-Yves, MA, has been a forensic psychologist with the Behavioural Science Unit of the Sûreté du Québec since
2002. He works as a critical-incident specialist and is actively involved in criminal investigations as well as in preparing psychological profiles and interrogations. He has taught at Quebec’s provincial police academy and has been a lecturer at the
University of Montreal.
Jean-Pierre Guay, PhD, is assistant professor at the University of Montreal. He received his PhD in criminology and completed postdoctoral training in psychology at Brandeis University. His research interests lie primarily in assessing psychopathy, sex offenders, and street gang involvement.
Downloaded from http://cjb.sagepub.com by arno baltin on October 12, 2008