Você está na página 1de 26

613828

research-article2015
HSXXXX10.1177/1088767915613828Homicide StudiesFujita et al.

Article
Homicide Studies
1­–26
A Multivariate Model for © 2015 SAGE Publications
Reprints and permissions:
Analyzing Crime Scene sagepub.com/journalsPermissions.nav
DOI: 10.1177/1088767915613828
Information: Predicting hsx.sagepub.com

Stranger, Solo Offender,


and Money-Oriented Motive
of Japanese Homicides

Goro Fujita1, Kazumi Watanabe1, Kaeko Yokota1,


Mamoru Suzuki1, Taeko Wachi1, Yusuke Otsuka1,
and Hiroki Kuraishi1

Abstract
This study examines the validity of a statistical offender profiling technique that predicts
the multi-dimensional classification of homicide offenders. Analyzing 539 Japanese
homicide cases, we constructed multivariate prediction models that infer classifications
defined by three dichotomous variables (stranger offender, solo offender, money-
oriented motive) on the basis of crime scene information. We evaluated the validity
of the models with a 10-fold cross-validation procedure and a receiver operating
characteristic (ROC) analysis, and found the models to have moderate accuracy (area
under the curve [AUC] = .73 to .82). We discussed the results from the perspective
of the offender’s rational choices in the crime scene and crime specialization.

Keywords
crime scene, policing, investigation, profiling, multivariate prediction model, predictive
validity, rational choice, crime specialization

Offender profiling or criminal profiling is a technique that infers characteristics of


offenders based on information about their actions at the crime scene. Most of the
research on offender profiling has focused on models based on typology or behavioral

1National Research Institute of Police Science, Kashiwa-shi, Japan

Corresponding Author:
Goro Fujita, National Research Institute of Police Science, 6-3-1 Kashiwanoha, Kashiwa-shi,
Chiba-ken 277-0882, Japan.
Email: fujita@nrips.go.jp

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


2 Homicide Studies 

themes. For example, research conducted by Ressler, Burgess, Douglas, Hartman, and
D’Agostino (1986) at the Federal Bureau of Investigation proposed that sexual homi-
cides could be classified as either organized or disorganized. The organized/disorga-
nized typology is based on offenders’ motives and is associated with offender mental
characteristics (e.g., psychopathy, neurodevelopmental insult, sadistic fantasy), which
have been suggested as developmental factors in homicide or violent behavior (Harris,
Rice, & Lalumière, 2001; Hickey, 2006; Sewall, Krupp, & Lalumière, 2013). Holmes
and Holmes (1998) and Keppel and Walter (1999) also proposed classifications of
sexual homicides based on offenders’ motives.
In contrast, psychological researchers at the University of Liverpool using psycho-
metrically measurable constructs or behavioral themes proposed an instrumental ver-
sus expressive distinction (e.g., Canter & Fritzon, 1998). This instrumental/expressive
behavioral theme is based on offenders’ behaviors in crime scenes and is associated
with a theory of aggressive behavior (e.g., Feshbach, 1964) as well as with the social
learning of criminal behavior (e.g., Akers, 1973). The instrumental/expressive behav-
ioral theme has been repeatedly confirmed empirically using quantitative homicide
data in diverse cultures and societies (Fox & Allen, 2014; Salfati, 2000; Salfati &
Bateman, 2005; Salfati & Canter, 1999; Salfati & Park, 2007; Santtila, Häkkänen,
Canter, & Elfgren, 2003).

Limitations of Offender Profiling Based on Typology or


Behavioral Themes
Typology and behavioral themes are practically useful for suspect prioritization, which
is one of the aims of offender profiling (e.g., Alison, Goodwill, Almond, van den
Heuvel, & Winter, 2010; Alison, West, & Goodwill, 2004; Canter, 1994). If police
have a reasonable basis for classifying a homicide case as one of several types, then
resources could be deployed with a smaller number of suspects. More specifically,
Canter (2000) identified offender profiling as the so-called A-C equation (actions-
characteristics equation), which is intended to identify which crime scene behaviors
(actions) predict offender characteristics or clusters of characteristics. Furthermore,
Alison, Bennell, Mokros, and Ormerod (2002) argued that offender profiling requires
an assumption of homology that the people who commit crimes in a similar style will
have similar background characteristics. However, most studies based on typology
only classify offenders’ motivations but do not attempt to relate these to offender char-
acteristics. Accordingly, most studies based on behavioral themes only classify the
offenders’ crime scene behavior but do not attempt to relate these to offender charac-
teristics. Only indirect or weak evidence has been provided by limited research that
attempted to show a statistical association between crime scene actions and offender
characteristics (e.g., Kocsis, Cooksey, & Irwin, 2002; Ressler et al., 1986; Salfati,
2000; Salfati & Park, 2007). For instance, Canter, Alison, Alison, and Wentink (2004)
tested the organized/disorganized typology using data from U.S. serial killers and
found no distinct subset of the dichotomy. Canter and Wentink (2004) tested the

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 3

Holmes and Holmes (1998) classification and found limited support. Moreover,
Goodwill, Alison, and Beech (2009) tested the accuracy of predictive models based on
the organized/disorganized typology and behavioral themes and found low accuracy.
Although many developed countries have adopted psychological or offender profiling
for police investigations, the validity of offender profiling has often been criticized for
lack of scientific evidence (Doan & Snook, 2008; Snook, Eastwood, Gendreau,
Goggin, & Cullen, 2007).

Multivariate Models Predicting a Single Dependent


Variable
Instead of trying to collate crime scene information into a psychological type or behav-
ioral theme, some recent research has focused on the direct association between
offender characteristics and the crime scene action. In crime investigations, police
need to predict offender characteristics that are of actual value to their investigation
rather than psychological and latent traits or behavioral theme (Alison et al., 2010).
Previous studies have tested the validity of multivariate regression analysis to predict
dependent variables of offender characteristics on the basis of crime scene information
(Davies, Wittebrood, & Jackson, 1997; Fujita et al., 2013; Goodwill et al., 2009;
Yokota, Fujita, Watanabe, Yoshimoto, & Wachi, 2007). Using receiver operating char-
acteristic (ROC) analysis, Goodwill et al. (2009) tested the accuracy of models pre-
dicting the criminal characteristics of British sex offenders and found that the
multivariate prediction model was more accurate than models based on typology or
behavioral theme. Similarly, using a ROC analysis and a cross-validation procedure,
Fujita et al. (2013) constructed multivariate prediction models that would infer various
characteristics of Japanese homicide offenders (e.g., sex, age, criminal record, residen-
tial location) on the basis of crime scene information and found that the values of the
area under the ROC curves (AUC) ranged between .70 and .87, suggesting medium
accuracy of their prediction models.
However, all these multivariate prediction models have the same shortcoming. The
previous studies indicated complex associations between offender characteristics and
crime scene actions, suggesting that the previous multivariate predictions are too sim-
plistic to infer these complex associations. For instance, Goodwill and Alison (2007)
examined prediction models that infer the age of the sexual offender based on the
victim’s age and found that the models were valid only if they controlled for the
offender’s degree of planning or aggression. According to the authors’ account, sex
offenders who conducted intentional targeting with cognitive effort (higher planning
tendency) tended to choose victims based on the offender’s age-specific focus, whereas
the characteristics of victims of offenders who perpetrated the crime in an unplanned
and impulsive way did not reflect the offender’s age-specific preferences. Santtila
et al. (2003) indicated that non-stranger offenders vary in their body disposal behavior
after the killing. They found that homicide offenders who murdered a family member
or relative in an expressive manner tended to hide the victim’s body, whereas

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


4 Homicide Studies 

offenders who murdered a friend or colleague in an expressive manner did not have
such a tendency in their body disposal behavior. The authors further indicated that the
former tended to be socially maladjusted and have psychiatric problems whereas the
latter expressive offenders tended to be socially successful (employed, having their
own home).

Multivariate Models Predicting Multi-Dimensional


Classification
One of the next research steps would be to construct models that infer more than one
characteristic of offenders simultaneously (e.g., solo and stranger offender). Such mod-
els would be expected not only to be as accurate as the previous multivariate models but
also to provide richer and more useful background information for subsequent research
as well as for police investigations. The validity of such models, however, might depend
on how the multi-dimensional offender characteristics are defined or how the depen-
dent variables in the equation are created. One method is to define the classification
operationally by using more than one dichotomous variable (0 = not present, 1 = pres-
ent). For example, in the case of focusing on offender characteristics of two dichoto-
mous variables (e.g., solo offender, stranger offender), a prediction model could be
considered that infers a classification or pattern defined by a 2×2 contingency table
(1-1, 1-0, 0-1, 0-0). In this study, we focus on such multivariate prediction models.
Although research has been conducted using classifications based on corresponding
analyses for sex crimes (Beauregard, Proulx, Rossmo, Leclerc, & Allaire, 2007), multi-
dimensional scaling for street robberies (Goodwill et al., 2013), stranger rapes (Mokros
& Alison, 2002), and homicides (Goodwill, Allen, & Kolarevic, 2014), as far as we
know, no research exists on homicides or violent crimes that has investigated multivari-
ate prediction models that infer multi-dimensional offender classifications defined by 2
× 2 contingency tables on the basis of crime scene information.
For constructing such a prediction model, choosing dependent variables is a matter
of importance not only for research but also in practice. More specifically, the vari-
ables should be chosen based on homicide trends, cultural characteristics, or a coun-
try’s police investigation needs. In addition, as mentioned above, we are interested in
constructing multivariate prediction models that provide richer and more useful back-
ground information for subsequent research as well as for police investigations.
Therefore, we focused on psychological and social characteristics of offenders.
First, despite the fact that the prevalence of sexual homicide is relatively low in
North America (Firestone, Bradford, Greenberg, & Larose, 1998; Meloy, 2000;
Roberts & Grossman, 1993), many studies on offender profiling have focused on sex-
ual homicide. In Japan, sexual assault is observed in just 5% of homicides, whereas a
motive to rob money from the victims is found in approximately one third of homicide
cases (Satsumi & Ikegami, 1997). Therefore, in Japan, police are more interested in
whether the victim was murdered due to a money-oriented motive. Although a money-
oriented motive is subjective and might not be able to be regarded as an objective
offender characteristic, we included money-oriented motive in the dependent variables
because of its importance for police investigation.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 5

Second, many studies on homicide offender profiling have focused on cases perpe-
trated by solo offenders (e.g., Salfati, 2000). However, a significant proportion of homi-
cide cases involve multiple offenders. For example, in the United States, 21% of
homicide cases are perpetrated by multiple offenders (Federal Bureau of Investigation,
2014); in England and Wales, 25% of homicide cases involve multiple offenders (Roscoe
et al., 2012); and in Japan, approximately 30% of homicide cases are perpetrated by
more than one offender (Fujita et al., 2013). The higher proportion in Japan might be due
to social attitudes of Asian people who place a greater value on groups than people in
America or Europe (e.g., Markus & Kitayama, 1991; Salfati & Park, 2007). For Japanese
police, in the early stage of police investigations, it is particularly important to infer
whether the case was perpetrated by a solo offender or by multiple offenders.
Finally, as pointed out in previous research (e.g., Salfati & Canter, 1999; Satsumi &
Ikegami, 1997), one of the most important pieces of information in the early stage of
police investigation is whether or not the offender is a stranger to the victim. Although
recent research argues that unsolved homicides are not always disproportionately per-
petrated by strangers (Quinet & Nunn, 2014), it is still important to infer whether a
homicide was perpetrated by a stranger or not because the suspect pool would be
entirely different between stranger homicides and non-stranger homicides. Furthermore,
research has shown associations between victim–offender relationships and crime
scene information such as wounding or weapon use, which suggests the possibility of
inferring victim–offender relationships on the basis of crime scene information (Last &
Fritzon, 2005; Trojan & Krull, 2012).
In conclusion, we have focused on multivariate prediction models which infer
homicide classifications that are defined by three dichotomous variables of offenders’
psychological and social characteristics: whether the crime was perpetrated by a
stranger, whether the crime was perpetrated by a solo offender, and whether the crime
involved a money-oriented motive.

Aim of This Study


The aim of this study was to examine the validity of a multivariate prediction model
that infers homicide classifications defined by the three dichotomous variables regard-
ing offenders’ psychological and social characteristics as a stranger offender, solo
offender, and a money-oriented motive. Analyzing Japanese homicide data, we con-
structed multivariate logistic regression models for predicting the homicide classifica-
tion on the basis of crime scene information, including victim characteristics, behavior
in the murder phase, and body disposal. We then were able to examine the accuracy of
inferences from these models using ROC and cross-validation procedures.

Method
Data Sample
We analyzed data concerning 539 Japanese homicide cases in which offenders were
arrested for murder from 1998 to 2002. In the sample, 64 offenders (12.0%) murdered

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


6 Homicide Studies 

more than one victim. In such cases, we analyzed only the information concerning the
victim who was murdered earliest in temporal order. As a result, in this study, we ana-
lyzed information about 539 victims who were murdered by 539 offenders or offender
groups. In the sample, 179 cases (33.2%) involved multiple offenders. In such cases,
basically we used only the information of the offender who was considered by the
police to have played a leading role in the crime. However, in recording offender char-
acteristics of stranger, residence (prefecture) and criminal record, we focused on com-
mon characteristics shared by all offenders. For example, the offender groups were
recorded as having no criminal record only if all of the offenders had no criminal record.
We obtained the data from the Murder Analysis Database (MAD) constructed for
academic research purposes by the National Research Institute of Police Science
(Watanabe, Suzuki, & Tamura, 1999). The MAD database stores data of police inves-
tigation case reports, which includes diverse information such as testimony of offend-
ers, testimony of witnesses, objective evidence, and autopsy reports. We considered
the MAD database as suitable for the current research for the following reasons. First,
MAD contains information about homicide cases that were investigated by the Special
Investigation Headquarters (SIH) of the prefectural police. Basically, local police
investigate homicide cases in Japan. However, the case is investigated by SIH if the
prefectural police headquarters considers that the case should be investigated thor-
oughly and cooperatively based on the provision of the Ordinance of the National
Public Safety Commission (Hanzai sousa kihan, 2012). SIH investigates approxi-
mately 8% of homicide cases (National Police Agency, 2013). SIH homicide cases
tend to be difficult for police to clear; the clearance rate of SIH homicide cases is
approximately 89% (National Police Agency, 2013), which is lower than that of all
Japanese homicide cases (96%). One of the reasons for the lower clearance rate may
be the higher percentage of stranger murders. The proportion of stranger murders in
SIH cases is almost double of that of all Japanese homicide cases (Satsumi & Ikegami,
1997). A second reason for using the MAD database was that a representative sample
could be obtained because the MAD database contains SIH cases occurring through-
out Japan. Finally, we used the MAD database because several other Japanese studies
on homicide have been conducted using the MAD database (Satsumi & Ikegami,
1997; Suzuki, Watanabe, & Tamura, 1999; Watanabe et al., 1999).

Homicide Classification
As explained previously, we focused on the Japanese homicide classifications that
were operationally defined by the three dichotomous variables: stranger offender, solo
offender, and money-oriented motive. In this study, the offender was defined as a
stranger if he or she was unknown to the victim, and was defined as a non-stranger if
he or she was a partner, other family member, or acquaintance of the victim. As
explained previously, for cases involving multiple offenders, the offender groups were
classified as stranger only if all of the offenders were strangers to the victim. In the
sample, 16.1% of the cases were perpetrated by a stranger offender or stranger offender
group, 33.2% of the cases were perpetrated by multiple offenders, and 51.2% of the

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 7

Table 1.  Offender Characteristics as a Percentage of the Multi-Dimensional Classification of


Japanese Homicides.

Offender Class I Class II Class III Class IV Class V Class VI Total


characteristics (n = 134) (n = 162) (n = 83) (n = 71) (n = 59) (n = 30) (N = 539)
Male 94.0* 87.0 80.7* 78.9** 98.3* 100.0* 88.7***
Age 50 or older 36.6* 32.1 30.1 14.1** 25.4 13.3 28.8**
Foreigner 7.5 9.9 4.8 19.7** 16.9 10.0 10.6*
Living alone 29.1 30.9 16.9* 9.9*** 32.2 33.3 25.8**
Unemployed 48.5 39.5* 43.4 46.5 67.8** 63.3 47.7**
Outside of 14.9 12.3** 28.9* 32.4** 18.6 20.0 19.3**
prefecture
Criminal record 51.5 36.4** 49.4 49.3 59.3* 50.0 47.1*
  Larceny offenses 23.9 16.0* 25.3 16.9 35.6** 16.7 21.7*
  Violent offenses 11.9 13.0 9.6 19.7 18.6 23.3 14.3
 Robbery 6.0 3.1 2.4 7.0 15.3*** 6.7 5.8*
 Homicide 2.2 6.8 2.4 2.8 3.4 6.7 4.1
  Rape or indecency 3.0 2.5 3.6 1.4 5.1 6.7 3.2
  Illegal drug 2.2 4.3 4.8 5.6 6.8 10.0 4.6
 Intellectual 8.2 3.7 6.0 5.6 11.9 3.3 6.3
offenses

Note. Class I = solo non-stranger offender with money-oriented motive; Class II = solo non-stranger
offender without money-oriented motive; Class III = multiple non-stranger offenders with money-oriented
motive; Class IV = multiple non-stranger offenders without money-oriented motive; Class V = stranger
offenders with money-oriented motive; Class VI = stranger offenders without money-oriented motive.
*p < .05. **p < .01. ***p < .001.

cases involved a money-oriented motive. Although the three dichotomous variables


produce eight classifications, we integrated the four classifications regarding stranger
murders into two classifications. Thus, we produced two classifications regarding
stranger murders without considering whether or not the crime was perpetrated by a
solo offender. This was necessary due to the relatively small number of stranger homi-
cide cases in the sample. Table 1 shows the homicide classifications that we used in
this study. The definition and statistical characteristics of the classifications are
described in the “Results” section.

Offender Characteristics
In addition to the offender characteristic variables that we used for the definitions of
the homicide classifications, we obtained another seven variables of offender charac-
teristic from the MAD database to describe the statistical characteristics of the homi-
cide classifications. These included male offender, offender of age 50 or older, foreign
offender, offender who was living alone, offender who was living outside of the pre-
fecture where the body was disposed, unemployed offender, and offender who had a
criminal record.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


8 Homicide Studies 

Crime Scene Information


Previous homicide research on offender profiling has focused on the association
between offender characteristics and crime scene information in pre-murder, murder,
and post-murder phases. Based on these findings, we obtained 19 crime scene infor-
mation variables from the MAD database. First, previous homicide studies have
focused on victim characteristics to infer offender characteristics, considering that vic-
tim characteristics reflect offenders’ specific preferences and their hunting process
(e.g., Ressler et al., 1986; Salfati & Bateman, 2005; Salfati & Canter, 1999). In our
study, seven independent variables regarding victim characteristics were chosen.
These included foreign victim, female victim age 29 or younger (young woman), vic-
tim age 60 or older, employed victim, female victim who was working in the sex
industry, and victim who was living alone. Second, previous homicide studies have
indicated that violence or the use of a weapon in the crime scene is associated with
offender characteristics (Keppel, 2000; Salfati, 2000; Trojan & Krull, 2012), suggest-
ing that violent behavior reflects offenders’ cognitive processes, such as planning, or
their aggressiveness. In our study, seven dependent variables were chosen from the
crime scene information in the murder phase. These included stabbing, shooting,
strangling, beating (by object, hand, or foot), sexual abuse, and stealing money or
valuables (robbery). Finally, previous homicide studies have indicated that body dis-
posal behavior or the body’s location is associated with offender characteristics
(Lundrigan & Canter, 2001; Salfati, 2000; Santtila et al., 2003). Regarding action in
the post-murder phase, five variables were chosen. These included disposing of the
body in the victim’s house, disposing of the body in a public space (e.g., on a road, in
a parking lot or square), disposing of the body in a mountain area (including riverbank
or sandy beach), burying the body underground or in water (body buried), and mutilat-
ing or setting fire to the body (destruction of corpse).

Variable Coding
Variables in the MAD database were coded by researchers or trained research assis-
tants from the National Research Institute of Police Science. To estimate the interrater
reliability, Cohen’s kappa was calculated for 54 cases (10% of the sample) between
data in the MAD database and data that were coded independently by the first author.
In the current study, approximately half of the variables (52%) were in the almost
perfect range (e.g., solo offenders, money-oriented motive, criminal record, young
female victims) and one third of the variables (30%) were in the substantial agreement
range. The remaining 19% of variables were in the moderate agreement range (e.g.,
sexual abuse, foreigner victims).
It should be mentioned that using police investigation records for research involves
several limitations. First, available independent and dependent variables are limited.
Second, in this study, all the variables are dichotomous. Coding variables dichoto-
mously often have been adopted for quantitative research analyzing police investiga-
tion records (e.g., Canter et al., 2004; Salfati, 2000). Finally, it is possible that even if

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 9

the offender acted out a behavior, the behavior might not be recorded in the original
archives (Alison, Snook, & Stein, 2001; Canter & Alison, 2003).

Data Analysis
First, we conducted a descriptive analysis to determine the statistical characteristics of
Japanese homicide classifications. Second, we constructed multivariate logistic mod-
els to examine which crime scene information was associated with which Japanese
homicide classification using the 539 cases. Because some independent variables were
regarded as mutually correlated, to minimize the multi-collinearity, we performed a
forward stepwise regression (p < .05 to enter, p > .10 to remove). To examine the
robustness of the estimates, we calculated bootstrap 95% confidence intervals (95%
CI) of the adjusted odds ratio (OR). Finally, we evaluated the accuracy of the multi-
variate logistic models. To avoid an overly optimistic evaluation of model validity, we
used a 10-fold cross-validation procedure. We calculated the probabilities of the six
dependent variables (classification) in each trial set, exclusively based on the coeffi-
cients and the constants of the multivariate logistic regression models that were con-
structed for each learning data subset. We calculated AUC and bootstrap 95% CIs for
the final multivariate models. We also calculated sensitivity, specificity, positive pre-
dictive value, and the negative predictive value using the cutoff point as the point of
the ROC curves closest to the upper left corner (0, 1). We performed the statistical
analyses using the Japanese version of PASW Statistics 17 (IBM Japan Ltd, Tokyo,
Japan) and Stata 12 (College Station, TX, USA).

Results
Descriptive Characteristics of Japanese Homicide Classification
Table 1 shows the offenders’ descriptive characteristics and Table 2 shows the crime
scene information for the Japanese homicide classifications. Class I homicide classifi-
cation was defined as solo offenders who target a non-stranger due to a money-
oriented motive. In this study, a money-oriented motive included not only murder-
robbery but also insurance scam murder or murder due to financial troubles with the
victim. Class I homicide consisted of 134 cases that constituted a quarter of all
the samples. Comparing Class I cases with all the cases in this study, Class I homicide
cases included a greater percentage of male offenders (94.0% vs. 88.7%), offenders
age 50 or older (36.6% vs. 28.8%), victims age 60 or older (48.5% vs. 28.2%), and
victims who were living alone (50.0% vs. 31.0%). Class I homicide cases included a
smaller percentage of young female victims (7.5% vs. 13.4%). Regarding behavior in
the murder phase, Class I homicide cases included a greater percentage of strangula-
tion (42.5% vs. 35.1%). Robbery was recorded in 55.2% of Class I homicide cases,
suggesting that offenders with money motives do not always steal money or valuables
at the crime scene. Regarding body disposal behavior, Class I homicide cases included
a greater percentage of disposing of the body in the victim’s house (40.3% vs. 28.9%),

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


10 Homicide Studies 

Table 2.  Crime Scene Information as a Percentage of Multi-Dimensional Classification of


Japanese Homicides.

Crime scene Class I Class II Class III Class IV Class V Class VI Total
information (n = 134) (n = 162) (n = 83) (n = 71) (n = 59) (n = 30) (N = 539)
Victim characteristics
 Foreigner 6.7 12.3 3.6 15.5* 6.8 3.3 8.9*
  Young woman 7.5* 21.6*** 2.4** 5.6* 22.0* 26.7* 13.4***
  Age 60 or older 48.5*** 21.6* 24.1 4.2*** 37.3 23.3 28.2***
 Employed 56.7 59.3 65.1 62.0 64.4 40.0 59.4
  Working in sex 11.2 12.3 4.8 4.2 16.9 3.3 9.8*
industry
  Living alone 50.0*** 28.4 38.6 18.3* 13.6** 3.3*** 31.0***
Behavior in murder phase
 Stabbing 33.6 36.4 12.0*** 29.6 40.7 50.0* 32.3***
 Strangling 42.5* 40.1 36.1 26.8 25.4 10.0** 35.1**
 Beating 11.9 9.9* 25.3** 23.9* 11.9 6.7 14.7**
 Shooting 1.5 1.9 2.4 8.5 3.4 0.0 2.8
  Sexual abuse 2.2 4.3 1.2 0.0* 18.6*** 23.3*** 5.4***
 Robbery 55.2*** 0.6*** 33.7 0.0*** 74.6*** 0.0*** 27.3***
Body disposal
  Victim’s house 40.3*** 34.6 18.1* 11.3*** 33.9 10.0* 28.9***
  Public space 23.1 27.8 18.1* 19.7 49.2*** 56.7*** 28.0***
  Mountain area 24.6** 29.0 60.2*** 53.5*** 13.6*** 23.3 34.0***
  Body buried 9.0** 13.6 32.5*** 39.4*** 5.1** 6.7 17.4***
  Destruction of 9.7 11.1* 4.8 7.0 1.7 0.0 7.6*
corpse

Note. Class I = solo non-stranger offender with money-oriented motive; Class II = solo non-stranger
offender without money-oriented motive; Class III = multiple non-stranger offenders with money-oriented
motive; Class IV = multiple non-stranger offenders without money-oriented motive; Class V = stranger
offenders with money-oriented motive; Class VI = stranger offenders without money-oriented motive.
*p < .05. **p < .01. ***p < .001.

whereas Class I homicide cases included a smaller percentage of disposing of the body
in a mountain area (24.6% vs. 34.0%) or burying the body underground or in water
(9.0% vs. 17.4%). Although these associations were statistically significant, their
magnitude was relatively small in some comparisons because of the relatively large
sample size of this study.
Class II homicide classification was defined as solo offenders who targeted a non-
stranger due to a motive other than money-oriented (e.g., trouble between lovers, spite,
revenge). Therefore, most of Class II homicide cases have an expressive nature. Class
II homicides had the largest number of cases (162 cases) and constituted one third of
all the samples. Comparing Class II cases with all the cases in this study, Class II
homicide cases included a greater percentage of young female victims (21.6% vs.
13.4%), whereas Class II cases included a smaller percentage of victims age 60 or

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 11

older (21.6% vs. 28.2%). Class II cases included smaller percentages of unemployed
offenders (39.5% vs. 47.7%) and offenders who had criminal records (36.4% vs.
47.1%). Specifically, Class II cases included a smaller percentage of offenders who
had criminal records of larceny offenses (16.0% vs. 21.7%). Regarding behavior in the
murder phase, Class II cases included a smaller percentage of beatings (9.9% vs.
14.7%). It might be noteworthy that Class II cases included a relatively small number
involving sexual abuse (4.3%). Regarding body-disposing behavior, Class II cases
included a greater percentage of destruction of corpses (11.1% vs. 7.6%), whereas
Class II cases included a smaller percentage of offenders who were living outside the
prefecture where the body was disposed (12.3% vs. 19.3%).
The Class III homicide classification was defined as multiple offenders who tar-
geted a non-stranger due to a money-oriented motive. Class III homicides consisted of
83 cases that constituted 15% of all the samples. Comparing Class III cases with all the
cases in this study, Class III homicide included a smaller percentage of male offenders
(80.7% vs. 88.7%), offenders who were living alone (16.9% vs. 25.8%), and young
female victims (2.4% vs. 13.4%). Regarding behavior in the murder phase, Class III
cases included a greater percentage of beatings (25.3% vs. 14.7%), whereas Class III
included a smaller percentage of stabbings (12.0% vs. 32.3%). Of the Class III cases,
robbery was recorded in 33.7% of the cases, suggesting that the Class III offenders
may have been more likely to have had a dispute with victims over financial troubles
or have targeted insurance money rather than targeting money or valuables at the
crime scene. Regarding body disposal, Class III cases included a greater percentage of
disposing of the body in a mountain area (60.2% vs. 34.0%), disposing of the body
underground or in water (32.5% vs. 17.4%), and offenders who were living outside the
prefecture where the body was disposed (28.9% vs. 19.3%), whereas Class III cases
included a smaller percentage of disposing of the body in the victim’s house (18.1%
vs. 28.9%) or in a public space (18.1% vs. 28.0%).
The Class IV homicide classification was defined as multiple offenders who tar-
geted a non-stranger due to a motive other than money-oriented. One of the typical
Class IV cases is group lynching. It is noteworthy that group lynching is typically
observed in Japanese criminal groups such as yakuza (criminal organization), bosozoku
(motorcycle gang), or delinquent group (Kaplan & Dubro, 2012; Morisue, 2000;
Osuka, 2003). Class IV homicide consisted of 71 cases that constituted 13% of all the
samples. Comparing Class IV cases with all the cases in this study, Class IV homicide
cases included a smaller percentage of male offenders (78.9% vs. 88.7%), offenders
age 50 or older (14.1% vs. 28.8%), and offenders who were living alone (9.9% vs.
25.8%). Class IV cases included a greater percentage of foreign offenders (19.7% vs.
10.6%) and foreign victims (15.5% vs. 8.9%). Class IV included a smaller percentage
of victims age 60 or older (4.2% vs. 28.2%) and victims living alone (18.3% vs.
31.0%). Regarding behavior in the murder phase, Class IV included a greater percent-
age of beatings (23.9% vs. 14.7%), whereas Class IV included a smaller percentage of
sexual abuse (0.0% vs. 5.4%) and robbery (0.0% vs. 27.3%). Characteristics of body
disposal of Class IV cases were similar to those of Class III cases. For instance, Class
IV cases included a greater percentage of disposing of the body in a mountain area

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


12 Homicide Studies 

(53.3% vs. 34.0%), a greater percentage of burying the body underground or in water
(39.4% vs. 17.4%), a greater percentage of offenders who were living outside the pre-
fecture where the body was disposed (32.4% vs. 19.3%), and a smaller percentage of
disposing of the body in the victim’s house (11.3% vs. 28.9%). It should be noted that
in the sample of this study, all offenders and victims were living in Japan at the time
of the crime.
Class V homicide classification was defined as stranger offenders with a money-
oriented motive. One of the typical cases of Class V was armed robbery. Class V
homicide cases consisted of 59 cases that constituted 11% of all the samples.
Comparing Class V cases with all the cases in this study, Class V homicide cases
included a greater percentage of male offenders (98.3% vs. 88.7%), unemployed
offenders (67.8% vs. 47.7%), and offenders who had criminal records (59.3% vs.
47.1%). More specifically, Class V cases included a greater percentage of offenders
who had a criminal record of larceny offenses (35.6% vs. 21.7%) and robbery (15.3%
vs. 5.8%). Regarding the victims’ characteristics, Class V cases included a greater
percentage of young female victims (22.0% vs. 13.4%), whereas Class V included a
smaller percentage of victims who were living alone (13.6% vs. 31.0%). Regarding
behavior in the murder phase, Class V cases included a greater percentage of sexual
abuse (18.6% vs. 5.4%). Of the Class V cases, 74.6% of cases included robberies.
Regarding body disposal, Class V cases included a greater percentage of disposing of
the body in a public space (49.2% vs. 28.0%), whereas Class V cases included a
smaller percentage of disposing of the body in a mountain area (13.6% vs. 34.0%) or
burying the body underground or in water (5.1% vs. 17.4%).
Class VI homicide classification was defined as stranger offenders without a
money-oriented motive. It included homicide cases perpetrated by a sexual murderer
or a random attacker. Class VI homicide had the smallest number of cases (30 cases),
constituting 6% of all the samples. All offenders of Class VI cases were male.
Comparing Class VI homicide cases with all the cases in this study, Class VI cases
included a greater percentage of young female victims (26.7% vs. 13.4%). Class VI
cases included a smaller percentages of victims who were living alone (3.3% vs.
31.0%). Regarding behavior in the murder phase, Class VI cases included a greater
percentage of stabbing (50.0% vs. 32.3%) and sexual abuse (23.3% vs. 5.4%), whereas
Class VI included a smaller percentage of strangulation (10.0% vs. 35.1%). Regarding
body disposal, Class VI cases included a greater percentage of disposing of the body
in a public space (56.7% vs. 28.0%) and a smaller percentage of disposing of the body
in the victim’s house (10.0% vs. 28.9%).

Multivariate Logistic Regression Models


Table 3 shows the estimates produced by logistic regression for predicting the six
Japanese homicide classifications using the 539 cases. To evaluate the goodness-of-fit
of the predicting models, a Hosmer–Lemeshow test was conducted. The test accepted
the null hypothesis for all our six prediction models, indicating that the prediction by
the models fit the data. In addition, to evaluate multi-collinearity, we examine the

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 13

Table 3.  Estimates of Multivariate Logistic Regression Models Predicting Multi-Dimensional


Classification of Japanese Homicides (N = 539).
Estimates Class I Class II Class III Class IV Class V Class VI

Adjusted odds ratioa


  Victim characteristics
  Young woman 2.4 0.1 0.2 4.4 2.9
[1.3, 4.3] [0.0, 0.3] [0.1, 0.7] [1.8, 10.7] [1.1, 7.6]
   Age 60 or older 2.0 0.1  
[1.3, 3.3] [0.0, 0.4]
  Living alone 2.7 0.2 0.1
[1.7, 4.3] [0.1, 0.5] [0.0, 0.3]
  Behavior in murder phase
  Stabbing 0.3  
[0.1, 0.7]
  Strangling 1.7 0.2
[1.0, 2.8] [0.1, 0.6]
  Beating 0.5  
[0.3, 0.9]
  Robbery 4.5 0.0 1.8 0.0 b 16.4 0.0b
[2.8, 7.3] [0.0, 0.0] [1.0, 3.3] [8.0, 33.4]
  Body disposal
  Victim’s house 3.0  
[1.7, 5.0]
  Public space 2.9
[1.3, 6.5]
  Mountain area 3.6 0.3  
[2.1, 6.3] [0.1, 0.9]
  Body buried 0.3 3.5  
[0.2, 0.8] [1.9, 6.3]
   Destruction of corpse 2.7  
[1.2, 5.9]
Constant 0.1 0.5 0.1 0.2 0.0 0.1
p value (Hosmer–Lemeshow test) .626 .587 .147 .601 .907 .852
Nagelkerke R2 .28 .35 .19 .30 .35 .28

Note. Class I = solo non-stranger offender with money-oriented motive; Class II = solo non-stranger offender without
money-oriented motive; Class III = multiple non-stranger offenders with money-oriented motive; Class IV = multiple
non-stranger offenders without money-oriented motive; Class V = stranger offenders with money-oriented motive;
Class VI = stranger offenders without money-oriented motive.
aFigures in parenthesis are 95% confidence interval.
bConfidence interval could not be calculated due to no cases of robbery.

variance inflation factor (VIF) of the six prediction models. The VIF ranged between
1.0 and 1.2, which indicated no need for concern about multi-collinearity.
Stepwise selection of a multivariate logistic model for predicting the Class I homi-
cide classification (solo non-stranger offenders with a money-oriented motive) identi-
fied five crime scene variables. Homicide cases whose victims were age 60 or older
and cases whose victims were living alone were more likely to have been committed
by Class I offenders (OR = 2.0, OR = 2.7). Homicide cases whose victims were mur-
dered by strangulation and cases in which money or valuables were stolen at the crime

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


14 Homicide Studies 

scene were more likely to have been committed by Class I offenders (OR = 1.7, OR =
4.5). In contrast, homicide cases in which bodies were buried in the ground or in water
were less likely to have been committed by Class I offenders (OR = 0.3).
Stepwise selection of a multivariate logistic model for predicting the Class II homi-
cide classification (solo non-stranger offenders without a money-oriented motive)
identified five crime scene variables. Homicide cases in which young women were
murdered (OR = 2.4), cases in which bodies were disposed of in the victim’s house
(OR = 3.0), and cases in which the bodies were mutilated or set fire (OR = 2.7) were
more likely to have been committed by Class II offenders. In contrast, homicide cases
in which money or valuables were stolen (OR = 0.0) and cases whose victims were
murdered by beating (OR = 0.5) were less likely to have been committed by Class II
offenders.
Stepwise selection of a multivariate logistic model for predicting the Class III
homicide classification (multiple offenders of non-stranger with a money-oriented
motive) identified four crime scene variables. Homicide cases in which money or
valuables were stolen (OR = 1.8) and cases in which bodies were disposed of in a
mountain area (OR = 3.6) were more likely to have been committed by Class III
offenders. In contrast, homicide cases in which young female victims were murdered
(OR = 0.1) and cases in which victims were murdered by stabbing (OR = 0.3) were
less likely to have been committed by Class III offenders.
Stepwise selection of a multivariate logistic model for predicting the Class IV
homicide classification (multiple offenders of a non-stranger without a money-
oriented motive) identified four crime scene variables. Homicide cases in which bod-
ies were buried in the ground or in water were more likely to have been committed by
Class IV offenders (OR = 3.5). In contrast, cases in which young female victims were
murdered (OR = 0.2) and cases in which older victims were murdered (OR = 0.1) were
less likely to have been committed by Class IV offenders. We could not estimate the
CI of OR of the crime scene variable of robbery because there were no robberies
among the Class IV cases in the study sample. Nevertheless, we constructed the pre-
dictive model for Class IV homicide cases assigning 0 to the OR of that independent
variable because we were interested in constructing a statistical prediction model of
offender profiling rather than estimating the degree of association between offender
characteristics and crime scene information.
Stepwise selection of a multivariate logistic model for predicting Class V classifi-
cation (stranger offenders with a money-oriented motive) identified four crime scene
variables. Homicide cases in which young female victims were murdered and cases in
which money or valuables were stolen at the crime scene were more likely to have
been committed by Class V offenders (OR = 4.4, OR = 16.4). In contrast, homicide
cases in which victims living alone were murdered and cases in which the body was
disposed of in a mountain area (OR = 0.2, OR = 0.3) were less likely to have been
committed by Class V offenders.
Stepwise selection of a multivariate logistic model for predicting the Class VI clas-
sification (stranger offenders without a money-oriented motive) identified five crime
scene variables. Homicide cases in which young female victims were murdered and

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 15

cases in which bodies were disposed of in public spaces were more likely to have been
committed by Class VI offenders (OR = 2.9, OR = 2.9). In contrast, cases in which
victims living alone were murdered and cases in which victims were murdered by
strangulation were less likely to have been committed by Class VI offenders (OR =
0.1, OR = 0.2). Again, we could not estimate the 95% CI of the OR of the crime scene
variable of robbery because there were no robberies among Class VI cases in the study
sample. Nevertheless, we constructed the predictive model for Class VI homicide
cases by assigning a 0 to the OR of the independent variable.

Accuracy of the Multivariate Prediction Model


Table 4 indicates accuracy of the multivariate prediction models that we examined
in this study. The first row of the Table 4 indicates AUC values of the optimum
model (estimates are reported in Table 3), the second row of the Table 4 indicates
AUC values of the robbery model that includes the single independent variable of
robbery (estimates are not reported in this article). The higher the AUC, the better
the prediction power the model has, and the AUC is expected to be .50 if the model
is ineffective. AUCs of the optimum models were as follows: .77 (95% CI = [.72,
.82]) for Class I, .77 (95% CI = [.73, .81]) for Class II, .73 (95% CI = [.67, .78]) for
Class III, .79 (95% CI = [.74, .84]) for Class IV, .82 (95% CI = [.75, .88]) for Class
V, and .82 (95% CI = [.76, .89]) for Class VI. Swets (1988) suggested that AUC
between .70 and .90 shows medium accuracy. Therefore, based on AUC, all of the
six multivariate logistic models have medium accuracy. Moreover, we compared
the AUCs of the optimum models with those of the robbery models and found
that the AUCs of the optimum models were significantly better than those of the
robbery models. Therefore, the optimum model that includes independent variables
of victim characteristics, behavior in the murder phase, and body disposal has bet-
ter predictive power than the robbery model, even when homicide offender classi-
fications related to money-oriented motive were inferred.
The third and fourth rows of Table 4 show the sensitivity and specificity of the six
prediction models. The sensitivity and specificity of a model are expected to be approxi-
mately .50 if the model is ineffective. In the current study, sensitivity ranged from .63
(Class II) to .90 (Class IV), and specificity ranged from .60 (Class IV) to .76 (Class I and
Class V). Although the higher values indicate better performance, it is important that
sensitivity and specificity are well-balanced. Therefore, based on sensitivity and specific-
ity, the models for Class I and Class V performed better than the others. The fifth and sixth
rows of Table 4 show the positive predictive value and negative predictive value of the six
prediction models. For example, the positive predictive value is the probability that cases
are truly Class I when the Class I prediction model is positive, whereas the negative pre-
dictive value is the probability that cases are not truly Class I when the Class I prediction
model is negative. Although the higher values indicate better performance, the values
depend on the prevalence of dependent variables that are shown at the bottom of Table 4.
The values of the positive predictive value ranged from .14 (Class VI) to .52 (Class II),
and they were 1.7 to 2.6 times higher than the prevalence rates, indicating that the models

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


16 Homicide Studies 

Table 4.  Accuracy of Multivariate Logistic Regression Models Predicting Multi-Dimensional


Classification of Japanese Homicides.

Class I Class II Class III Class IV Class V Class VI


Area under the curve .77 .77 .73 .79 .82 .82
(Robbery modela) (.64) (.64) (.46) (.59) (.70) (.54)
Sensitivity .72 .63 .64 .90 .78 .87
Specificity .76 .75 .71 .60 .76 .67
Positive predictive value .49 .52 .29 .25 .28 .14
Negative predictive value .89 .83 .92 .98 .97 .99
Predictive accuracy .75 .71 .70 .64 .76 .68
Cutoff probabilityb .25 .34 .18 .17 .09 .06
Prevalence ratec .25 .30 .15 .13 .11 .06

Note. Class I = solo non-stranger offender with money-oriented motive; Class II = solo non-stranger
offender without money-oriented motive; Class III = multiple non-stranger offenders with money-oriented
motive; Class IV = multiple non-stranger offenders without money-oriented motive; Class V = stranger
offenders with money-oriented motive; Class VI = stranger offenders without money-oriented motive.
aThe models include the single independent variable of robbery.
bValue at the nearest point to upper left corner (0, 1) on receiver operating characteristic curve.
cProportion of 539 total sample.

are useful for prediction. The values of the negative predictive value rates were high,
ranging from .83 (Class I) to .99 (Class VI). The seventh row of Table 4 shows the predic-
tive accuracy. Although the predictive accuracy of the models for Class IV and Class VI
were under .70 (64%, 68%), the predictive accuracy of the other models were over .70.
Table 5 shows the results of simulated predictions that were conducted using the cutoff
probabilities shown in Table 4. To evaluate the degree of saving of police investigative
resources, the prioritized rate was calculated based on the prevalence rate in the sample.
Moreover, to evaluate the accuracy of the predictions, the hit rate was calculated based on
the proportion of truly predicted cases related to the total observation. For example, in the
cases of the No. 2 prediction (presented in the second row of Table 5), where the models
suggested the probability of a Class VI offender, the prioritized rate was calculated as 6%,
which is equivalent to the prevalence rate of Class VI in the sample. The hit rate of the No.
2 prediction was 31%, which indicates that four of the 13 observed cases that suggested a
Class VI offender were actually perpetrated by a Class VI offender. If the prediction was
conducted without any models or any statistical information, the expected values of the
prioritized rate and the hit rate would both be 17%. It is noteworthy that because the mod-
els in this study predict each class independently, the prediction using the models does not
always predict only a single class, but often predicts more than one class. Of the 539
simulated prediction cases, a single class was predicted in 81 predictions, and their means
of prioritized rates and hit rates were 18% and 43%, respectively. The most frequently
observed cases were that the models predicted two of the six classes, and their means of
prioritized rates and hit rates were 33% and 71%, respectively. In the total of 539 cases,
the means of prioritized rates and hit rates were 37% and 72%, respectively.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 17

Table 5.  Results of Simulated Prediction by the Multivariate Model (N = 539).

Prediction typea

Class (Prevalence rate in the sample)

I II III IV V VI

No. (.25) (.30) (.15) (.13) (.11) (.06) Observation (%) Prioritized rateb Hit rate
Number of predicted classes = 0  
 1 0 0 0 0 0 0 1 (0) — .00
Number of predicted classes = 1  
 2 0 0 0 0 0 1 13 (2) .06 .31
 3 0 0 0 0 1 0 1 (0) .11 .00
 4 0 0 0 1 0 0 29 (5) .13 .31
 5 0 0 1 0 0 0 8 (1) .15 .50
 6 0 1 0 0 0 0 26 (5) .30 .62
 7 1 0 0 0 0 0 5 (1) .25 .40
Subtotal 82 (15) .18 .43
Number of predicted classes = 2  
 8 0 0 0 1 0 1 63 (12) .19 .44
 9 0 0 1 0 0 1 2 (0) .21 .50
10 0 0 1 0 1 0 4 (1) .26 .75
11 0 0 1 1 0 0 36 (7) .29 .67
12 1 0 0 0 0 1 1 (0) .30 .00
13 0 1 0 0 0 1 25 (5) .36 .76
14 1 0 0 0 1 0 63 (12) .36 .90
15 1 0 0 1 0 0 2 (0) .38 .00
16 1 0 1 0 0 0 16 (3) .40 .81
17 0 1 0 1 0 0 29 (5) .40 .69
18 0 1 0 0 1 0 7 (1) .41 .86
19 0 1 1 0 0 0 1 (0) .45 1.00
20 1 1 0 0 0 0 22 (4) .55 1.00
Subtotal 271 (50) .33 .72
Number of predicted classes = 3  
21 0 0 1 1 0 1 25 (5) .34 .80
22 0 1 0 0 1 1 20 (4) .47 .75
23 0 1 0 1 0 1 39 (7) .49 .72
24 0 1 1 0 0 1 1 (0) .51 1.00
25 1 0 1 0 1 0 68 (13) .51 .99
26 1 0 1 1 0 0 6 (1) .53 .83
27 0 1 1 1 0 0 11 (2) .59 .91
28 1 1 0 1 0 0 8 (1) .68 .88
29 1 1 1 0 0 0 3 (1) .70 1.00
Subtotal 181 (34) .49 .86
(continued)

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


18 Homicide Studies 

Table 5. (continued)
Prediction typea

Class (Prevalence rate in the sample)

I II III IV V VI

No. (.25) (.30) (.15) (.13) (.11) (.06) Observation (%) Prioritized rateb Hit rate
Number of predicted classes = 4  
30 0 1 1 1 0 1 3 (1) .64 .33
31 1 1 1 1 0 0 1 (0) .83 1.00
Subtotal 4 (1) .69 .50
  539 (100) .37 .72
Note. Prediction using cutoff probability presented in Table 4. Cases with no observations are not shown in the table.
a1 = predicted, 0 = non-predicted.
bEstimated based on prevalence rate in the sample.

Discussion
The aim of this study was to examine the validity of a multivariate prediction model that
infers multi-dimensional homicide classifications on the basis of crime scene informa-
tion. The findings have a number of academic and practical implications. First, based on
the ROC analysis and cross-validation procedure, the results indicated that all of the six
multivariate logistic regression models had medium accuracy. Therefore, in this study,
the A-C equations (Canter, 2000) were established and the homology hypothesis (Alison
et al., 2002) was supported using a sample of Japanese homicides. Although these results
are consistent with Fujita et al. (2013), the results are inconsistent with previous studies
that claim offender profiling is not possible without considering moderators such as situ-
ational factors (Alison et al., 2002; Alison et al., 2010). The results also are in contrast
with studies that claim offender profiling is non-scientific (Doan & Snook, 2008; Snook
et al., 2007). Why this study’s models obtained moderate accuracy without interaction
variables might be due to the type of data used. Previous studies that claimed offender
profiling was non-scientific have not involved homicides but rather crimes such as arson
(Doan & Snook, 2008) and burglary (Doan & Snook, 2008). It is possible that the dif-
ferentiation or distinctiveness of homicide offender behavior might be larger than that
for arson, burglary, or sex crimes. Specifically, it is noteworthy that this study’s sample
included any type of homicide rather than just a specific type (e.g., sexual homicide). It
is possible that the differentiation or distinctiveness of homicide offender behavior in our
sample that included any type of homicide might be even greater than in samples that
included only a specific type of homicide (Fujita et al., 2013).
Regarding differences in the predictive accuracy of the classes, the results indicated
that models that predict stranger homicide (Class V and Class VI) were better than the
others in AUC values, suggesting the potential usefulness of the multivariate approach
for stranger homicide investigations. Moreover, the results of our study indicated that
the optimum models that predict money-oriented motivated offenders (Class I and

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 19

Class V) performed better than those with only independent variables of robbery.
These results suggest that the multivariate approach using crime scene information,
including victim characteristics, behaviors in the murder phase, and body disposal, is
more valid than predictions based on only the existence of robbery in the crime scene
even when it used for inferring money-oriented motive. Furthermore, to understand
the associations between the independent variables and the classes properly, it is nec-
essary not only to examine the estimates of stepwise selection but also to examine the
descriptive statistics that are presented in Table 2. This is because the optimum multi-
variate models do not always include all the independent variables that are associated
with the classes to avoid multi-collinearity. For instance, the optimum prediction
model that predicted a Class I offender includes the independent variable of “body
buried (burying underground or in water)” but did not include “victim house (dispos-
ing of the body in the victim’s house)” despite the fact that both independent variables
were significantly associated with Class I. The stepwise selection excluded “victim
house” and included “body disposal” because the two variables were negatively asso-
ciated with each other.
Second, as implied in previous studies, the results of our study indicated that asso-
ciations between offenders’ psychological and motivational characteristics and crime
scene information are complex (e.g., Goodwill & Alison, 2007; Santtila et al., 2003).
In this respect, the results suggest that the current study’s models would produce more
valid and more useful information than conventional multivariate prediction models
that infer a single offender characteristic (Davies et al., 1997; Fujita et al., 2013;
Goodwill & Alison, 2007; Yokota et al., 2007). For example, the results of our study
indicated that homicide cases in which young women were murdered were more likely
to be committed by a solo non-stranger offender without money-oriented motive
(Class II) or stranger offenders (Class V and Class VI). On one hand, the results sug-
gest that investigators should keep in mind the diverse possibilities of motive and the
background of the crime when they investigate a crime in which young women were
murdered. Possible interpretations include the following. The young women might be
targeted by a Class II offender because of trouble between lovers. She might be tar-
geted by a Class V offender because of the easiness to control her to get money. Or she
might be targeted by a Class VI offender because of sexual desire. On the other hand,
if the models suggest the possibility of a Class II, Class V, and Class VI offender, then
police could prioritize 41% of a suspects list because Class II, Class V, and Class VI
offenders accounted for 41% of the total sample in this study.
Third, our study findings suggest that crime scene behavior should be understood
based on the rational choices of the offenders (e.g., Beauregard, Leclerc, & Lussier,
2012; Clarke, 1997). For example, the current study’s models indicated that homicide
cases whose victims were living alone were more likely to be committed by solo non-
stranger offenders who have money-oriented motives (Class I) and that Class I offend-
ers tended to be older males. It is thought that older male offenders who need money
tend to target older non-strangers (family member, relatives, or acquaintances), expect-
ing a more certain benefit (knowing that the victim has some money) and a lower cost
(believing that he could control the victim by himself). There is a further evidence that

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


20 Homicide Studies 

is consistent with the importance of offender rational choice. The current study’s mod-
els indicated that homicide cases in which the body was mutilated or burned after the
killing were more likely to have been committed by solo offenders who were non-
stranger to the victims and who did not have a money-oriented motive (Class II). The
models also indicated that homicide cases in which the body was found in a mountain
area or cases in which the body was buried in the ground or water were more likely to
be committed by multiple offenders who were non-strangers to the victims (Class III
or Class IV). These body disposal behaviors may imply that Japanese homicide offend-
ers are sensitive to the cost of leaving the body at the crime scene, and that moving the
body to a remote area without being witnessed requires too many resources for solo
offenders. It is thought that solo offenders choose mutilating or burning the body to
delay police from learning about the case (Watanabe & Tamura, 1998). It should be
noted that because Japan is populated 10 times more densely than the United States,
the body would soon be found by the police if offenders did not hide the body.
Fourth, in 47.1% of the Japanese homicide cases in our study sample, offenders or
offender groups had prior criminal records. Specifically, the results indicated that
offenders or offender groups with money-oriented motive (Class I, Class III, and Class
IV) had more records of larceny offenses than violent offenses. Offenders or offender
groups who perpetrated homicide of an expressive nature (Class II, Class IV, and Class
VI) had more records of violent offenses than larceny offenses. More specifically, the
results of the study indicated that Class V offenders or offender groups (stranger
offenders with money-oriented motives) had 2.6 times more records of robbery and
had 1.6 times more records of larceny offenses when compared with the entire sample.
These results suggest a positive relationship between prior crime specialization and
homicide crime scene behavior (e.g., Trojan & Salfati, 2011). However, the results of
the study indicated that offenders or offender groups had no criminal record in more
than half of Japanese homicide cases. These results are consistent with recent research
that has been conducted in other countries (Ganpat, Liem, van der Leun, & Nieuwbeerta,
2014; Greenall & Richardson, 2015; Trojan & Salfati, 2011). As suggested in previous
research, it is possible that most homicide offenders had engaged in anti-social con-
duct or prior illegal activity even if they had no criminal record (e.g., Greenall &
Richardson, 2015).
Finally, the results of our study have implications for police investigators and
administrators. As suggested in previous research (Alison et al., 2010; Alison et al.,
2004; Canter, 1994), police could use the multivariate prediction models to prioritize
suspects and more efficiently deploy investigation resources. Figure 1 provides an
example of an investigation process using the multivariate models. First, investigators
would input crime scene information into the models. Second, they could calculate the
probabilities of each class with the six multivariate models using default cutoff prob-
abilities. Then, they could prioritize suspects and could deploy police resources based
on the results of the predictions by the models. For example, if the model suggests that
an offender is more likely to be a Class I offender, then police should divide investiga-
tion resources equally into two segments. One should focus on the victim’s family
members, relatives, or acquaintances needing money or having financial problems

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 21

Figure 1.  An example of an investigation process using the multivariate prediction models.
aDefault values are presented in Table 4.

with the victim. Because the results of this study indicated that positive predictive
value of a Class I offender was .49, this implies a 51% probability that the offender is
not a Class I offender even though the model suggested a Class I offender. The other
segment of police resources should focus on other types of offenders. In this case, the
remaining resources might primarily focus on a stranger armed robbery (Class V).
Because the results of the simulated prediction of the study showed that the models
predicted Class I in 195 prediction cases and that in 131 cases (67%) of them the mod-
els predicted Class V at the same time.
Moreover, as indicated in Figure 1, the investigator could adjust the cutoff proba-
bilities and re-predict the classes if the offender could not be found based on the first
trial prediction. For example, to broaden the range of the suspects, the investigator
could use lower cutoff probabilities. Conversely, to narrow the range of the suspects,
the investigator could use higher cutoff probabilities. It is noteworthy that the negative
predictive value of the multivariate models in this study ranged between .83 and .99.
The higher negative predictive values might indicate that the current study’s models
might be suitable for police investigation because the negative predictive value is the
probability that a negative result is truly negative. For instance, the results of this study
indicated that the negative predictive value of Class V (stranger robbery) was .97,
which means a 3% probability that the case was perpetrated by Class V offender if the
multivariate model suggests that an offender is less likely to be a Class V offender.
Therefore, in this case, police would have a reasonable basis for excluding Class V
offenders from the list of suspects. We based the multivariate prediction models in the
current study on widely used statistical diagnostic techniques in the field of medicine
(e.g., Zweig & Campbell, 1993). These techniques always produce false positives and

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


22 Homicide Studies 

false negatives, so it is necessary for users of these multivariate prediction models to


understand the statistical meanings of the false positive rate and the false negative rate.
Before concluding, we discuss the conceptual and practical limitations of our multi-
variate model. In our study, we tested offender classifications on the basis of only three
offender characteristics: stranger offender, solo offender, and money-oriented motive.
Research is needed to examine the validity of models predicting different offender clas-
sifications, such as classification related to criminal record. Moreover, previous studies
have shown that behavior at the crime scene as well as the motives and behavioral
themes could change among serial homicide cases involving the same offender, which
indicates the difficulty of criminal profiling for serial crimes (e.g., Salfati & Bateman,
2005). However, this is a limitation not only for our multivariate model but also for
other criminal profiling methods. Finally, this approach should not be a replacement for
investigations by experienced detectives but instead should give them new investiga-
tive tools for information analysis. As indicated in Figure 1, objective evidence should
be respected when available. A multivariate approach might be unnecessary when
police have strong evidence such as DNA or witnesses to identify the offender. However,
this approach could be a useful tool for information analysis when limited objective
evidence is available. Although future research should evaluate the usefulness of the
multivariate approach in actual police investigations, it might be that an experienced
detective who utilizes a multivariate approach would perform better.

Conclusion
In this study, we constructed multivariate logistic regression models to infer multi-
dimensional classification of Japanese homicide offenders on the basis of crime scene
information. We found that this approach had moderate accuracy. These multivariate
prediction models might produce more useful information for prioritizing suspects in
police investigations than previous models of offender profiling. Nevertheless, the
validity of the models might be improved if the moderators or interaction variables of
situational factors were included (Alison et al., 2002; Alison et al., 2010; Goodwill &
Alison, 2007). In addition, future research should explore the relationship between the
models and theories of psychology, psychiatry, and criminology in terms of offenders’
decision-making processes in crime scenes, motives for violent crime, and the devel-
opment of criminal careers. Finally, to test the validity of statistical and behavioral
offender profiling more fully, future research should apply the multivariate prediction
models to homicides in other countries.

Acknowledgments
The authors are grateful for the support of the First Investigation Division, Criminal Investigation
Bureau, National Police Agency.

Declaration of Conflicting Interests


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

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 23

Funding
The author(s) received no financial support for the research, authorship, and/or publication of
this article.

References
Akers, R. L. (1973). Deviant behavior: A social learning approach. Belmont, CA: Wadsworth.
Alison, L. J., Bennell, C., Mokros, A., & Ormerod, D. (2002). The personality paradox in
offender profiling: A theoretical review of the processes involved in deriving back-
ground characteristics from crime scene actions. Psychology, Public Policy, and Law,
8, 115-135.
Alison, L. J., Goodwill, A., Almond, L., van den Heuvel, C., & Winter, J. (2010). Pragmatic solu-
tions to offender profiling and behavioural investigative advice. Legal and Criminological
Psychology, 15, 115-132.
Alison, L. J., Snook, B., & Stein, K. L. (2001). Unobtrusive measurement: Using police infor-
mation for forensic research. Qualitative Research, 1, 241-254.
Alison, L. J., West, A., & Goodwill, A. (2004). The academic and the practitioner: Pragmatists’
views of offender profiling. Psychology, Public Policy, and Law, 10, 71-101.
Beauregard, E., Leclerc, B., & Lussier, P. (2012). Decision making in the crime commission
process: Comparing rapists, child molesters, and victim-crossover sex offenders. Criminal
Justice and Behavior, 39, 1275-1295.
Beauregard, E., Proulx, J., Rossmo, K., Leclerc, B., & Allaire, J. F. (2007). Script analysis of
the hunting process of serial sex offenders. Criminal Justice and Behavior, 34, 1069-1084.
Canter, D. V. (1994). Criminal shadows. London, England: HarperCollins.
Canter, D. V. (2000). Offender profiling and criminal differentiation. Legal and Criminological
Psychology, 5, 23-46.
Canter, D. V., & Fritzon, K. (1998). Differentiating arsonists: A model of firesetting actions and
characteristics. Journal of Criminal and Legal Psychology, 3, 73-96.
Canter, D. V., & Alison, L. J. (2003). Converting evidence into data: The use of law enforce-
ment archives as unobtrusive measurement. The Qualitative Report, 8, 151-176.
Canter, D. V., Alison, L. J., Alison, E., & Wentink, N. (2004). The organized/disorganized
typology of serial murder: Myth or model? Psychology, Public Policy, and Law, 10,
293-320.
Canter, D. V., & Wentink, N. (2004). An empirical test of Holmes and Holmes’s serial murder
typology. Criminal Justice and Behavior, 31, 489-515.
Clarke, R. V. (1997). Situational crime prevention: Successful case studies (2nd ed.).
Guilderland, NY: Harrow and Heston.
Davies, A., Wittebrood, K., & Jackson, J. L. (1997). Predicting the criminal antecedents of a
stranger rapist from his offence behaviour. Science & Justice, 37, 161-170.
Doan, B., & Snook, B. (2008). A failure to find empirical support for the homology assumption
in criminal profiling. Journal of Police and Criminal Psychology, 23, 61-70.
Federal Bureau of Investigation. (2014). Crime in the United States, 2013. Washington, DC:
U.S. Department of Justice.
Feshbach, S. (1964). The function of aggression and the regulation of aggressive drive.
Psychological Review, 71, 257-272.
Firestone, P., Bradford, J. M., Greenberg, D. M., & Larose, M. R. (1998). Homicidal sex offend-
ers: Psychological, phallometric, and diagnostic features. The Journal of the American
Academy of Psychiatry and the Law, 26, 537-552.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


24 Homicide Studies 

Fox, K. A., & Allen, T. (2014). Examining the instrumental-expressive continuum of homi-
cides: Incorporating the effects of gender, victim-offender relationships, and weapon
choice. Homicide Studies, 18, 298-317.
Fujita, G., Watanabe, K., Yokota, K., Kuraishi, H., Suzuki, M., Wachi, T., & Otsuka, Y. (2013).
Multivariate models for behavioral offender profiling of Japanese homicide. Criminal
Justice and Behavior, 40, 214-230.
Ganpat, S. M., Liem, M., van der Leun, J., & Nieuwbeerta, P. (2014). The influence of criminal
history on the likelihood of committing lethal versus nonlethal violence. Homicide Studies,
18, 221-240.
Goodwill, A. M., & Alison, L. J. (2007). When is profiling possible? Offense planning and
aggression as moderators in predicting offender age from victim age in stranger rape.
Behavioral Sciences & the Law, 25, 823-840.
Goodwill, A. M., Alison, L. J., & Beech, A. R. (2009). What works in offender profiling? A
comparison of typological, thematic, and multivariate models. Behavioral Sciences & the
Law, 27, 507-529.
Goodwill, A. M., Allen, J. C., & Kolarevic, D. (2014). Improvement of thematic classification
in offender profiling: Classifying Serbian homicides using multiple correspondence, clus-
ter, and discriminant function analyses. Journal of Investigative Psychology and Offender
Profiling, 11, 221-236.
Goodwill, A. M., Stephens, S., Oziel, S., Sharma, S., Allen, J. C., Bowes, N., & Lehmann, R.
(2013). Advancement of criminal profiling methods in faceted multidimensional analysis.
Journal of Investigative Psychology and Offender Profiling, 10, 71-95.
Greenall, P. V., & Richardson, K. (2015). Adult male-on-female stranger sexual homicide: A
descriptive (baseline) study from Great Britain. Homicide Studies, 19, 237-256.
Hanzai sousa kihan [Regulation of criminal inquiry and investigation], Ordinance of the
National Public Safety Commission, No. 8 (2012). Retrieved from http://law.e-gov.go.jp/
htmldata/S32/S32F30301000002.html
Harris, G. T., Rice, M. E., & Lalumière, M. (2001). Criminal violence: The roles of psychopathy, neu-
rodevelopmental insults, and antisocial parenting. Criminal Justice and Behavior, 28, 402-426.
Hickey, E. V. (2006). Serial murderers and their victims (4th ed.). Belmont, CA: Wadsworth.
Holmes, R. M., & Holmes, S. T. (1998). Serial murder (2nd ed.). Thousand Oaks, CA: SAGE.
Kaplan, D. E., & Dubro, A. (2012). Yakuza: Japan’s criminal underworld. Berkeley: University
of California Press.
Keppel, R. D. (2000). Signature murders: A report of the 1984 Cranbrook, British Columbia
cases. Journal of Forensic Sciences, 45, 500-503.
Keppel, R. D., & Walter, R. (1999). Profiling killers: A revised classification model for under-
standing sexual murder. International Journal of Offender Therapy and Comparative
Criminology, 43, 417-437.
Kocsis, R. N., Cooksey, R. W., & Irwin, H. J. (2002). Psychological profiling of sexual mur-
ders: An empirical model. International Journal of Offender Therapy and Comparative
Criminology, 46, 532-553.
Last, S. K., & Fritzon, K. (2005). Investigating the nature of expressiveness in stranger, acquain-
tance and intrafamilial homicides. Journal of Investigative Psychology and Offender
Profiling, 2, 179-193.
Lundrigan, S., & Canter, D. (2001). A multivariate analysis of serial murderers’ disposal site
location choice. Journal of Environmental Psychology, 21, 423-432.
Markus, H. R., & Kitayama, S. (1991). Culture and the self: Implications for cognition, emotion,
and motivation. Psychological Review, 98, 224-253.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


Fujita et al. 25

Meloy, J. R. (2000). The nature and dynamics of sexual homicide: An integrative review.
Aggression and Violent Behavior, 5, 1-22.
Mokros, A., & Alison, L. J. (2002). Is offender profiling possible? Testing the predicted homol-
ogy of crime scene actions and background characteristics in a sample of rapists. Legal and
Criminological Psychology, 7, 25-43.
Morisue, O. (2000). Analysis of conflicts among Japanese gangs. Journal of Police Science,
53(6), 141-163.
National Police Agency. (2013). Heisei 24nen no hanzaizyousei [Crimes situation in 2012].
Retrieved from http://www.npa.go.jp/toukei/seianki/h24hanzaizyousei.pdf
Osuka, K. (2003). Aichi Pref. Police measures against delinquent groups such as motorcycle
gangs. Journal of Police Science, 56(12), 91-112.
Quinet, K., & Nunn, S. (2014). Establishing the victim–offender relationship of initially unsolved
homicides: Partner, family, acquaintance, or stranger? Homicide Studies, 18, 271-297.
Ressler, R. K., Burgess, A. W., Douglas, J. E., Hartman, C. R., & D’Agostino, R. B. (1986).
Sexual killers and their victims: Identifying patterns through crime scene analysis. Journal
of Interpersonal Violence, 1, 288-308.
Roberts, J. V., & Grossman, M. G. (1993). Sexual homicide in Canada: A descriptive analysis.
Annals of Sex Research, 6, 5-25.
Roscoe, A., Rahman, M. S., Mehta, H., While, D., Appleby, L., & Shaw, J. (2012). Comparison
of a national sample of homicides committed by lone and multiple perpetrators. The Journal
of Forensic Psychiatry & Psychology, 23, 510-521.
Salfati, C. G. (2000). The nature of expressiveness and instrumentality in homicide: Implications
for offender profiling. Homicide Studies, 4, 265-293.
Salfati, C. G., & Bateman, A. L. (2005). Serial homicide: An investigation of behavioral consis-
tency. Journal of Investigative Psychology and Offender Profiling, 2, 124-144.
Salfati, C. G., & Canter, D. V. (1999). Differentiating stranger murders: Profiling offender char-
acteristics from behavioral styles. Behavioral Sciences & the Law, 17, 391-406.
Salfati, C. G., & Park, J. (2007). An analysis of Korean homicide crime-scene actions. Journal
of Interpersonal Violence, 22, 1448-1470.
Santtila, P., Häkkänen, H., Canter, D., & Elfgren, T. (2003). Classifying homicide offenders
and predicting their characteristics from crime scene behavior. Scandinavian Journal of
Psychology, 44, 107-118.
Satsumi, Y., & Ikegami, S. (1997). Satsuzin sousa honbu ziken ni okeru kagaisya-higaisya
kan no menshiki kankei [Offender-victim relationship in murder cases solved by specific
investigative headquarters]. Reports of the National Research Institute of Police Science,
38, 115-123.
Sewall, L. A., Krupp, D. B., & Lalumière, M. (2013). A test of two typologies of sexual homi-
cide. Sexual Abuse: A Journal of Research and Treatment, 25, 82-100.
Snook, B., Eastwood, J., Gendreau, P., Goggin, C., & Cullen, R. M. (2007). Taking stock of
criminal profiling: A narrative review and meta-analysis. Criminal Justice and Behavior,
34, 437-453.
Suzuki, M., Watanabe, K., & Tamura, M. (1999). Kyouaku hanzai deitabeisu wo mochiita kin-
nen no satsuzinziken ni kansuru kenkyu: Chiriteki zyouhou ni kansuru bunseki [A study of
modern homicide using felonious offenses database: An analysis of geographic informa-
tion]. Japanese Journal of Criminal Psychology, 37(Suppl.), 10-11.
Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science, 240, 1285-1293.
Trojan, C., & Krull, A. C. (2012). Variations in wounding by relationship intimacy in homicide
cases. Journal of Interpersonal Violence, 27, 2869-2888.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016


26 Homicide Studies 

Trojan, C., & Salfati, C. G. (2011). Linking criminal history to crime scene behavior in single-
victim and serial homicide: Implications for offender profiling research. Homicide Studies,
15, 3-31.
Watanabe, K., Suzuki, M., & Tamura, M. (1999). Kyouaku hanzai deitabeisu wo mochiita kin-
nen no satsuzinziken ni kansuru kenkyu: Hankou genba no zyouhou ni kansuru bunseki [A
study of modern homicide using felonious offenses database: An analysis of crime scene
information]. Japanese Journal of Criminal Psychology, 37(Suppl.), 8-9.
Watanabe, K., & Tamura, M. (1998). Analysis on the characteristics of mutilation-murder cases
in the last 50 years. Reports of the National Research Institute of Police Science, 39, 52-65.
Yokota, K., Fujita, G., Watanabe, K., Yoshimoto, K., & Wachi, T. (2007). Application of the
behavioral investigative support system for profiling perpetrators of serial sexual assaults.
Behavioral Sciences & the Law, 25, 841-856.
Zweig, M. H., & Campbell, G. (1993). Receiver-operating characteristic (ROC) plots: A funda-
mental evaluation tool in clinical medicine. Clinical Chemistry, 39, 561-577.

Author Biographies
Goro Fujita, PhD, is the past head of the Investigation Support Section at the National Research
Institute of Police Science (NRIPS), Japan. He has focused on the use of quantitative research
in police psychology.
Kazumi Watanabe, PhD, is the current head of the Investigation Support Section at the NRIPS.
She received her PhD in medical science from Tokyo Medical and Dental University in forensic
psychiatry. She is currently investigating the behavioral characteristics of serial and mass mur-
derers and the investigative interviewing.
Kaeko Yokota, PhD, is a senior researcher in the Investigation Support Section at the NRIPS.
She earned her MSc and PhD degrees in investigative psychology from the University of
Liverpool. Her research interests include consistency of criminal behavior, case linkage of serial
crimes, and cognitive interviews.
Mamoru Suzuki, MA, is currently an associate professor in the Faculty Humanities and Social
Science at Iwate University, Morioka, Japan. He is interested in psychological research on
crime prevention and police investigations. His interests include the cognitive process of fraud
victimization, sexual arousal of sex offenders, and the geographic analysis of serial crimes.
Taeko Wachi, PhD, is a senior researcher in the Investigation Support Section at the NRIPS.
She obtained a PhD in psychology at the University of Cambridge. The main focus of her
current research is on police interviewing and interrogation techniques.
Yusuke Otuska, MA, is a researcher in the Investigation Support Section at the NRIPS. His
research interests include the modi operandi of sex and violent crimes, offender profiling of
extortion, and statistic modeling of offender’s characteristics.
Hiroki Kuraishi, MA, is currently a forensic psychologist in the Criminal Investigation
Department at the Shiga Prefectural Police Headquarters. His current research focus is on the
decision making process of residential burglars from the perspective of cognitive psychology.

Downloaded from hsx.sagepub.com at CLARKSON UNIV LIBRARY on May 9, 2016

Você também pode gostar