Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Evidence for a new, five-class typology of male sexual offender and offence characteristics Mundy, Crystal Lea 2016

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2016_september_mundy_crystal.pdf [ 3.2MB ]
Metadata
JSON: 24-1.0305065.json
JSON-LD: 24-1.0305065-ld.json
RDF/XML (Pretty): 24-1.0305065-rdf.xml
RDF/JSON: 24-1.0305065-rdf.json
Turtle: 24-1.0305065-turtle.txt
N-Triples: 24-1.0305065-rdf-ntriples.txt
Original Record: 24-1.0305065-source.json
Full Text
24-1.0305065-fulltext.txt
Citation
24-1.0305065.ris

Full Text

 EVIDENCE FOR A NEW, FIVE-CLASS TYPOLOGY OF MALE SEXUAL OFFENDER AND OFFENCE CHARACTERISTICS by   Crystal Lea Mundy  B.Sc., The University of British Columbia, 2012   A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF  THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF ARTS  in  THE COLLEGE OF GRADUATE STUDIES  (Psychology)THE UNIVERSITY OF BRITISH COLUMBIA (Okanagan) June 2016 © Crystal Lea Mundy, 2016   	 ii	The undersigned certify that they have read, and recommend to the College of Graduate Studies for acceptance, a thesis entitled:   Evidence for a New, Five-Class Typology of Male Sexual Offender and Offence Characteristics submitted by  Crystal Lea Mundy  in partial fulfillment of the requirements of the degree of Master of Arts .    Dr. Jan Cioe, Psychology  Supervisor, Associate Professor  Dr. Brian O’Connor, Psychology  Supervisor, Professor  Dr. Susan Wells, Psychology/Social Work  Supervisory Committee Member, Professor  Dr. Bonar Buffam, Sociology  University Examiner, Assistant Professor  June 16th, 2016   (Date Submitted to Grad Studies)     Additional Committee Members include:   Dr. Michael Woodworth, Psychology  Supervisory Committee Member, Professor    Dr. David Morgan, Faculty of Medicine, Psychiatry  Supervisory Committee Member, Adjunct Professor      iii Abstract The present study sought to replicate existing research in order to evaluate the similarities between distinct samples of sexual offenders. The research also sought to use a more advanced statistical procedure than previous research to derive a typology and see how it fit within the framework of existing literature. Data were gathered from 106 pre-sentence reports assessing adult male sexual offenders. Prior research examining typologies of adolescent male sexual offenders and adult female sexual offenders were replicated and compared to the collected data. Some similarities between the adolescent male and adult male sexual offenders were found, such as a relationship between type of offender and recidivism risk. The adult female and adult male sexual offenders had a differing number of sub-types, but offender age and victim age differentiated between the groups for both samples. Following replication, latent class analysis was used to uncover naturally occurring groups in the collected data using offender and offence characteristics. The analyses revealed five classes of offenders, which were labelled Mixed Victim Assaulters, Non-Pedophilic Mixed Gender Offenders, Preferential Pedophiles, Non-Aggressive Incest Offenders, and Non-Aggressive Non-Pedophilic Child Molesters. Of the eight variables selected for the analyses, the classes varied mainly in the areas of presence of pedophilia, presence of a substance/alcohol disorder, use of aggression, victim gender, victim age, and victim relationship. Sexual victimization history and offender age did not provide much class differentiation. External analyses revealed that among the classes, ethnicity, presence of a mood/anxiety disorder, and presence of a personality disorder differed. The classes involving child-oriented offences supported existing typologies in the literature, whereas the class involving adult-oriented offences did not align with existing typologies. A typology rooted in empirical findings, such as the typology developed for this research, can provide significant insight into the heterogeneous nature of those who commit sexual offences. This insight can allow researchers and clinicians to best assess, treat, and reintegrate sexual offenders, as the     iv offenders may require different preventive techniques and have different treatment needs based on their class characteristics.        v Preface The following research was reviewed by the University of British Columbia – Okanagan’s Behavioural Research and Ethics Board, certification number H07-01114. An ethics application was also approved by the Forensic Psychiatric Services Commission. The data used in the current research were collected by Crystal Lea Mundy in Northern British Columbia.         vi Table of Contents Abstract………………………………………………………………….....……………… iii Preface……………………………………………………………………............... ……... v Table of Contents…………………………………………………………......................... vi List of Tables……………………………………………………………………………… ix List of Figures…………………………………………………………….......................... x Acknowledgements……………………………………………………………………….. xi Dedication…………………………………………………………………………………. xii Chapter 1 Introduction…………………………………………………………............... 1 1.1 Sexual Offending and Typology Research………………………………… 1 1.1.1   Sexual offending…………………………………………………… 1 1.1.2   Offender typologies………………………………………………... 2 1.1.3   Sexual offender typologies………………………………………… 4     1.1.3.1   Typologies and sexual offending…………………………. 5    1.1.3.2   Methodology of sexual offender typologies……………… 10    1.1.4   Latent class analysis………………………………………………... 11   1.1.4.1   Latent class analysis and sexual offending……………….. 12 1.2 Purpose of the Present Study………………………………………………. 17 1.2.1   Background typology: Adolescent sexual offenders………………. 17 1.2.2   Background typology: Female sexual offenders…………………... 19 1.2.3 Deriving an adult male sexual offender typology………………….. 22 Chapter 2 Methodology…………………………………………………………………... 24 2.1  Procedure Preceding Site Visit…………………………………………….. 24   2.1.1   Selection of files……………………………………………………. 24   2.1.2   Identifying file content for coding………………………………….. 24     vii   2.1.3   Coding scheme……………………………………………………… 25   2.1.4   Development of file review criteria………………………………… 26 2.2  Procedure During Site Visit………………………………………………... 27   2.2.1   File coding on-site…………………………………………………... 27   2.2.2   Inter-rater reliability assessments…………………………………... 27 2.3  Procedure Following Site Visit…………………………………………….. 28   2.3.1   Variable re-coding………………………………………………….. 28 2.4 Sample Characteristics……………………………………………………... 29 Chapter 3 Results…………………………………………………………………………. 31 3.1   Replication of Prior Analyses……………………………………………… 31 3.1.1 Replicated analyses from Chu and Thomas (2010)………………... 31 3.1.2  Adolescent sexual offender typology versus male sexual offender  typology……………………………………………………………………. 33 3.1.3  Replicated analyses from Vandiver and Kercher (2004)…………... 34 3.1.4  Female sexual offender typology versus male sexual offender  typology……………………………………………………………………. 45 3.2 Latent Class Analysis of Adult Male Sexual Offender Data………………. 46   3.2.1 Selection of variables for analyses…………………………………. 47   3.2.2 Determining the number of classes………………………………… 48   3.2.3 Reviewing the model and naming the classes……………………… 51   3.2.4 Class differences on external variables…………………………….. 56   3.2.5 Class differences in case file textual information………………….. 61 Chapter 4 Discussion……………………………………………………………………... 65 4.1 Summary of the Typology Replications…………………………………… 65 4.2 Summary of the Derived Adult Male Sexual Offender Typology…………. 66     viii 4.3 Comparison of the Derived Typology and Existing Literature……………. 68 4.4 Limitations and Future Research…………………………………………... 70 4.5  Research Contributions and Future Directions…………………………….. 72 Chapter 5 Conclusion…………………………………………………………………….. 75 References…………………………………………………………………………………. 76 Appendices………………………………………………………………………………… 87 Appendix A: SSA Plot from Lundrigan and Mueller-Johnson (2013)…………….. 87 Appendix B: LCA Profiles from Busina (2014)…………………………………… 88 Appendix C: HLM Results from Vandiver and Kercher (2004)…………………... 89 Appendix D: Coding Guide – Site Visit…………………………………………… 90 Appendix E: Inter-rater Reliability of Select Coded Files………………………..... 95 Appendix F: Demographic Characteristics……………………………………….... 96 Appendix G: Replicated HLM Results – K-Way and Higher-Order Effects……… 99        ix List of Tables Table 1 Coding Scheme…………………………………………………………….. 26 Table 2 Contingency Table and Adjusted Residuals for Generalist/Specialist  Grouping and Relationship to Victim……………………………………… 32 Table 3 Contingency Table and Adjusted Residuals for Generalist/Specialist  Grouping and Sexual Recidivism Risk.......................................................... 32 Table 4 Statistical Indices for Latent Class Models Tested………………………… 50 Table 5 Average Latent Class Probabilities for Most Likely Class Memberships…. 52 Table 6 Probability for Each Item Category for Latent Classes……………………. 53 Table 7 Cell Frequencies and Residuals of External Variable Categories for the  Latent Classes……………………………………………………………… 58      x List of Figures Figure 1 Within Groups Sum-of-square Graph……………………………………… 38 Figure 2 2D Representation of the Two-Cluster Solution…………………………… 39 Figure 3 2D Representation of the Three-Cluster Solution………………………….. 40 Figure 4 2D Representation of the Four-Cluster Solution……………………………41 Figure 5 Dendrogram of a Three-Cluster Solution…………………………………...42 Figure 6 Dendrogram of a Four-Cluster Solution…………………………………….43 Figure 7 BIC Values for the Different Number of Clusters…………………………. 44 Figure 8 Profile of the Five Latent Classes………………………………………….. 56      xi Acknowledgements I would like to express sincere gratitude to my supervisors, Drs. Jan Cioe and Brian O’Connor, for their continued guidance throughout my graduate education. Without their support, this work would not have been successfully completed. I also thank Dr. David Morgan for his ongoing effort and allowing me to use his data for this study. Although an exceptionally busy man, he took time to mentor me and direct me in my data collection efforts; this allowed me an opportunity I would not have had otherwise. I would also like to thank my committee members, Dr. Susan Wells and Dr. Michael Woodworth, for their time and input, which were essential in order to conceptualize and have this project come to fruition. I cannot begin to express the appreciation I have for all of you, and the support you have provided me in my academic career.      xii Dedication I would like to dedicate this work to my husband, Jarrett Mundy, whose commitment to our family and continued love and support has allowed me to continue pursuing my dreams and deal with the challenges of graduate school on a daily basis.   	 1 Chapter 1 Introduction 1.1   Sexual Offending and Typology Research 1.1.1  Sexual offending   Sexual offending is a complex and controversial social issue. Sexual offences, as defined by the Criminal Code of Canada, encompass a wide range of criminal acts. These acts include crimes such as sexual interference, invitation to sexual touching, sexual exploitation, and incest (see R.S.C., c. C-46, s. 150-162; Criminal Code, 1985). Sexual assault is not included with the sexual offences section of the Criminal Code of Canada. Instead, it is included in the assault section and there are specific punishments provided for the different levels sexual assault (see R.S.C., c. C-46, s. 271-273; Criminal Code, 1985). Due to the nature of these crimes, some of the sexual offences and sexual assault make an offender eligible for long term offender or dangerous offender status (see R.S.C., c. C-46, s. 752-761; Criminal Code, 1985). Within Canada, sexual offences account for roughly 1% of overall offending and 9% of violent offences (Department of Justice, 2015a). In recent years the overall violent and non-violent crime severity indices have decreased in Canada (Statistics Canada, 2014). However, reports of sexual violations against children, possession of child pornography, and aggravated sexual assault were among the few crimes that had increased (Statistics Canada, 2014). It is also important to note that sexual offending is believed to be under-reported. This occurs for a variety of reasons, such as not wanting the police involved, finding the offence too personal to disclose, considering the offence unimportant, or fearing retribution for reporting (Department of Justice, 2015b; Statistics Canada, 2013).   Although under-reported, sexual victimization has a large impact on victims, and has been linked to many detrimental outcomes. A meta-analysis conducted by Madigan, Wade, Tarabulsy, Jenkins, and Shouldice (2014) reported an association between sexual (and physical) victimization and an increased risk of adolescent pregnancy. Jordan, Combs, and Smith (2014)     2 conducted research assessing the impact of sexual victimization on high school and college-aged women. Their study found that women who had been sexually victimized during high school entered college with lower GPA scores and earned lower grades than those who had not been victimized. They also found that if women were sexually victimized during their first semester of college, they went on to have lower GPAs than those who had not been sexually victimized. In terms of suicide, research has shown sexual victimization to be related to both non-fatal suicidal behavior (Sadeh & McNeil, 2013; Tripodi, Onifade, & Pettus-Davis, 2014; Tripodi & Pettus-Davis, 2013) and completed suicide (Gradus et al., 2012). Sexual victimization also has been linked with substance use/abuse issues (Tripodi & Pettus-Davis, 2013) and accidental overdoses (Cutajar et al., 2010). Overall, research demonstrates that individuals who have been sexually victimized, in either adulthood or childhood, are at increased risk for a number of problematic, and potentially fatal, behaviors.   Although sexual offending continues to be a serious concern within Canada and other countries, much about sexual offending remains unknown or uncertain. Factors such as intimacy/attachment issues (Ogilvie, Newman, Todd, & Peck, 2014), maladaptive schemas (Carvalho & Nobre, 2014), social anxiety (Nunes, McPhail, & Babchishin, 2012), pornography use (Babchishin, Hanson, & VanZuylen, 2015), and childhood abuse (Mallie, Viljoen, Mordell, Spice, & Roesch, 2011) have been identified as potentially having a role in sexual offending behaviors. However, those who sexually offend are a heterogeneous group, and different factors are likely to impact subsets of offenders in different ways (Seto, Kingston, & Stephen, 2015). Attaining a full understanding of those who sexually offend is a necessary step before moving on to determine how to best assess, treat, and reintegrate such individuals. 1.1.2  Offender typologies   The development of offender typologies began with the work of Cesare Lombroso in the late 1800s (Ellwood, 1912; Mannheim, 1960; Wolfgang, 1961). Lombroso’s typology, based on     3 descriptions and clinical observations, posited that criminality had a biological basis and could be identified by physical differences (e.g., skull measurements). The typology suggested that the physical anomalies present in criminals were associated with them being more closely related (i.e., throwbacks) to evolutionary ancestors than non-criminals. The typology included subtypes such as the epileptic criminal, the insane criminal, and the born criminal. Lombroso was criticized for not confirming the presence of the typology with adequate statistical methods. Lombroso’s criminal typology has been met with further cynicism due to the direct comparisons between criminals and specific races, leading to popularization of atavistic theories (Moyer, 2001). Although discredited, Lombroso’s work led to further interest in creating and evaluating offender typologies. Recent investigations into typologies have included evaluating crimes such as serial homicide (Horning, Salfati, & Labuschagne, 2014), street robbery (Goodwill, Stephens, Oziel, Yapp, & Bowes, 2012), and driving under the influence (Okamura, Kosuge, Kihira, & Fujita, 2014).  The typological approach needs to be distinguished from the taxonomic approach, as both can be used in psychological research (Bailey, 1994; Richey, Holm-Denoma, Kotov, Schmidt, & Joiner Jr., 2008). Both approaches focus on distinguishing between categories present in data. However, taxonomic approaches result in the categories being mutually exclusive. This means that once an individual is a member of a certain category, that person cannot be the member of another category. Typological approaches are not restricted in this manner, and instead an individual may have characteristics that align with more than one category. This is beneficial when considering psychological phenomena, as individuals are unlikely to fit discretely into categories. Problems may arise if the taxonomic approach is used because individuals may be seen as incapable of change, or defined by the characteristics of the group within which they fit. The typological approach allows for some fluidity, and to approach     4 psychological phenomena as a continuum in which individuals may move between categories based on the characteristics being considered. Byrne and Roberts (2007) suggest that there are three themes that need to be considered by researchers when attempting to develop a typology of offending behavior: design, development, and implementation. First, in terms of design, Byrne and Roberts state that the purpose of a typology needs to be clear, as different purposes will require different typologies. For example, a typology with the purpose of determining risk level of certain behaviors will be different than a typology with the purpose of determining institutional control levels. Second, in terms of development, they state that that the reliability, validity, and false positives/negatives of the typology need to be considered rather than ignored. Finally, in terms of implementation, Byrne and Roberts state that researchers must pay attention to, and address, the typologies’ ability to actually classify and predict the behaviors of interest. Further, they state that typologies should be considered at the community level, as well as the individual level, as both levels play a role in an individual’s outcome. 1.1.3  Sexual offender typologies   Sexual offending typologies have traditionally been developed in an attempt to provide a complete understanding of sexual offending behaviors (National Criminal Justice Association, 2014). As stated by Robertiello and Terry (2007), motivations and characteristics of sexual offenders can be used as a framework to derive groupings that allow for the distinction between subsets of offenders (i.e., typologies). The hope is that this comprehensive understanding will help researchers and clinicians make informed decisions about treatment, intervention, and supervision of offenders (National Criminal Justice Association, 2014).   However, developing accurate and useful typologies has proven to be a difficult task. As mentioned previously, sexual offenders are a highly heterogeneous group (Seto et al., 2015). Sexual offenders may have committed the same crime and present similarly, yet they can have a     5 wide range of differing background characteristics, attitudes and beliefs, and clinical and criminogenic needs (National Criminal Justice Association, 2014; Robertiello & Terry, 2007). In consideration of these individual differences and the complexity of sexual offending behavior, it is important to develop multidimensional typologies (Robertiello & Terry, 2007). Establishing accurate and useful typologies also has been difficult because typologies are often created and used in inconsistent manners (National Criminal Justice Association, 2014). Further, most existing typologies fail to address treatment issues and to predict recidivism, which limits their utility in forensic and clinical contexts (National Criminal Justice Association, 2014). The following describes three existing sexual offender typologies, as well as some of the methods, aims, and limitations of the research involved in deriving the typologies.  1.1.3.1   Typologies and sexual offending   Male stranger rape. Typologies relating specifically to sexual offending have been an area of interest for researchers, and the developed typologies are often related to specific types of sexual offending. Lundrigan and Mueller-Johnson (2013) attempted to derive a behavioral typology for male (victim and perpetrator) stranger rape. The aim of developing this typology was to distinguish between offender types based on crime scene behaviors such as offender-victim interactions. They believed this enhanced understanding could increase knowledge about the psychological processes involved in the offending behaviors, as well as contribute to criminal investigations. They analyzed 209 case files involving male stranger rape, and they included variables that assessed physical behaviors, verbal behaviors, sexual behaviors, and precautionary behaviors displayed in the offence. The case files were analyzed using smallest space analysis (SSA; Barkus, & Yavorsky, & Foster, 2006; Guttman, 1968). SSA is a form of nonmetric multidimensional scaling, which essentially is a visual representation of data that achieves the best fit with the least number of dimensions. SSA uses matrices of similarity/dissimilarity between the included variables (Kruskal & Wish, 1978).     6 For their research, each variable representing a specific offence behavior was shown by a point on the SSA plot (see Appendix A). The proximity of the points to one another indicated the likelihood that they occurred together. If two points were close together, they were more likely to co-occur in the same offence. The analysis indicated there were two offender types, involvement and hostility. These types represented distinct styles of offender-victim interactions. The region of involvement was further split into the sub-types of involvement intimate and involvement exploit (see Appendix A). The hostility type was characterized by aggressive interactions between the offender and the victim during the offence. Example characteristics of this type include the following: violence was used to control the victim, language during the offence was demeaning towards the victim, and degrading behaviors were used (e.g., anal/object penetration). The involvement intimate sub-type had a lack of aggressive behaviors. It instead was characterized by behaviors that reflected the offender trying to act as if in a relationship with the victim. This may include conversational interaction attempts, forced reciprocation by victim during sexual acts, or personal details being disclosed by the offender.  The involvement exploit sub-type was characterized by a blend of behaviors associated with control, criminality, and involvement. Examples of these behaviors are bringing a weapon, using the weapon to control the victim, and the offender providing justification of behaviors. After establishing the types, Lundrigan and Mueller-Johnson tested the model by examining whether each offence could be assigned to one theme based on crime scene behaviors.  The behaviors present were used to assign a score to each offence for each of the three themes. A dominant theme was present for 80% of the cases. The remaining 20% were categorized into the broader involvement sub-type, as the offence behaviors existed across both of the involvement sub-types.  Child sexual abuse. Wortley and Smallbone (2014) aimed to develop a typology of child sexual abusers based on the criminal careers model. The criminal careers model looks at abuse in     7 terms of the persistence and versatility of the offenders, and differences in these two factors is assumed to differentiate between types of abusers. Wortley and Smallbone analyzed records from 349 male offenders, 177 of whom filled out a self-report questionnaire. The records were initially analyzed using cross-tabulations.   Using the information from the cross-tabulations, four offender types were derived based on offence information. These four types were limited/specialized, limited/versatile, persistent/specialized, and persistent/versatile. Persistence was defined as the presence or absence of prior sexual offence convictions (persistent/limited). Versatility was defined as the presence or absence of prior non-sexual offence convictions (versatile/specialized). The researchers then utilized multivariate analysis of variance (MANOVA; Field et al., 2012) to assess the relationships between the types and other characteristics, such as age at first conviction. Significant multivariate interactions were found and post hoc analyses revealed a number of specific relationships. Limited/specialized offenders were significantly older than the other three types. Persistent/specialized offenders were significantly older at first conviction than the limited/versatile or persistent/versatile offender types and also had significantly more convictions than the other three types. Main effects were also found. Offenders with no prior sexual convictions received significantly shorter sentences and had fewer current nonsexual convictions than offenders with prior sexual convictions. No differences were found with regards to age at present conviction and offender category.   The self-report data were analyzed using cross-tabulations, and the same four offender types were found as described above. Multivariate analysis of covariance (MANCOVA; Field et al., 2012) was then used to investigate how sexual history variables were related to the offender types. The dependent variables consisted of 10 sexual history variables, and the independent variables were the prior sexual offence and nonsexual offence convictions. The covariate used was age at current conviction, as it was expected that the sexual history variables might be     8 influenced by how long the offenders have been able to offend. Significant multivariate interactions were found, and post hoc analyses were used to describe the specific relationships. Persistent/specialized offenders were exclusively heterosexual at a significantly lower rate than the other three types and abused more victims than the limited/specialized and limited/versatile types. Main effects were also found. Offenders with prior non-sexual convictions abused over shorter periods of time and had fewer sexual contacts with their initial victim. Offenders with prior sexual convictions were more likely to have suffered sexual victimization as a child, were younger when they abused their initial victim, and their initial victim was more likely to be an unrelated male. The multivariate analyses indicated that the covariate relationship was significantly associated with age at first sexual contact.   Wortley and Smallworth recognized a number of limitations associated with the development of their typology, but suggested that it can be used to help conceptualize the heterogeneity among child sexual abusers. By recognizing these differences, their hope is that treatment and prevention strategies can be better targeted and administered.   Sexual homicide. Sewall, Krupp, and Lalumière (2013) tested two models of sexual homicide. The first model proposed by Sewall et al. is based on a more general anti-sociality typology than what is in the existing literature. The general model was expected to have three types: psychopathic, sadistic, and competitively disadvantaged. These types were based heavily on perpetrator characteristics. The psychopathic type included engaging in crimes such as sexual homicide, criminal careers beginning at a young age, and criminal careers persisting through adulthood. The sadistic type included fantasizing about torturing others, steady employment, and a lack of a persistent criminal career. A lifelong criminal career, low degree of intelligence, and low socioeconomic status characterized the competitively disadvantaged type. The second model tested an organized/disorganized typology that had been previously presented in the literature. The organized type was characterized by order and planning, as evident through crime scene     9 behaviors. This type included signs of premeditation, such as removing a murder weapon or restraining the victim. The disorganized type was instead characterized by a lack of order and planning. This type reflects a lack of premeditation, such as leaving evidence/weapons behind or using excessive violence/mutilation.   Sewall and colleagues analyzed data from 82 sexual homicide offenders that had full biographic information on TruTV. Information was also pulled from Wikipedia (www.wikipedia.org) and the Encyclopedia of Serial Killers (Newton, 2006). The data were then analyzed using a principal components analysis (PCA; Field et al., 2012). PCA works by analyzing the interrelationships among a set of variables. It uses these relationships to then explain a variable set using a smaller number of variables, called principal components, with a minimal loss of information. The PCA initially resulted in nine components, but after examining scree plots, only five components were selected for further analyses. The five components were slashing, sadistic, instrumentality/order, poor school performance, and antisociality. Slashing included high loadings on variables such as disembowelment and ritualism. Sadistic included high loadings on variables such as victim torture and physical restraints. Instrumentality included high loadings on variables such as order and instrumentality. Poor school performance included high loadings on variables such as the Child and Adolescent Taxon Scale score and low education levels. Finally, antisociality included high loadings on items such as psychopathy score and antisocial personality disorder score. A cluster analysis was then used to analyze the five components to see whether a typology was present. Model-based cluster analysis aims to estimate the number of clusters that are present in a given set of data (Fraley & Raftery, 1998). Different models with different assumptions are tested, and a fit index is calculated for each model. The goodness of fit index that is calculated is used to decide what the best model is. Sewall and colleagues analyzed models ranging from two to five clusters. Based on the cluster analysis, a four-cluster solution     10 yielded the best fit to the data. High scores on the instrumentality/order and sadistic components characterized Cluster 1. High scores on the poor school performance and antisociality components characterized Cluster 2. Low scores on the antisociality and instrumentality/order components and high scores on the slashing component characterized Cluster 3. Low scores on the slashing, sadistic, and poor school performance components, and average scores on the instrumentality/order and antisociality components characterized Cluster 4. However, follow-up analyses using the adjusted Rand index (ARI; Hubert & Arabie, 1985) indicated that a one-cluster solution was more stable than a four-cluster solution. Sewall and colleagues tested the two potential typologies against one another, and found mixed support for both typologies. They suggested that this is due to the heterogeneous nature of sexual offenders but that testing the validity of models is critical in this area of research.  1.1.3.2   Methodology of sexual offender typologies As demonstrated in the above examples, the development of sexual offender typologies varies in terms of content, analyses, and results. However, a general pattern of investigating sexual offender typologies can be drawn from these examples. First, a statistical technique is used to analyze the data and derive a typology. This may occur through different analytic techniques such as SSA (Barkus et al., 2006; Guttman, 1968) or latent class analysis (Collins & Lanza, 2010). The type of data a researcher has often plays an important role in deciding what type of analysis will be used. Second, statistical techniques are then used to assess the relationships between external variables and the typology. These variables were not used in the initial derivation of the typology, but are used to see what may be associated with the offender types. This allows for a more comprehensive understanding of the typology and what variables are likely to be related to the offending behaviors being considered.   Another application of typology research was described in Sewall et al. (2013). Rather than seeking to find or confirm a typology that exists in a given sample, these researchers sought     11 to test two competing models of sexual homicide against one another. Although their research did not come to a decisive conclusion, it did show an alternative use to typology-based research. Although this approach is not often used in the sexual offending literature, it has been used in other areas such as knowledge management (Denford & Chan, 2011), consumer emotions (Havlena & Holbrook, 1986; Havlena, Holbrook, & Lehmann, 2006), and consumer strategies (Galbraith & Schendel, 1983). However, the statistical techniques used to compare the typologies in these studies are often not as advanced as can be seen in more recent examples. One such advanced statistical technique that can be used in typology research is latent class analysis. The following section will describe latent class analysis, and outline research that has applied this type of analysis to sexual offender typologies. 1.1.4  Latent class analysis   Latent class analysis is a particularly useful method for capturing heterogeneity both among and within groups (Rosato & Baer, 2012). Paul Lazarsfeld and colleagues introduced latent class analysis in the 1950s and 1960s, and since then it has continued to be improved upon (Lazarsfeld, 1950; Lazarsfeld; 1959; Lazarsfeld & Henry, 1968; Vermunt & Magidson, 2003). Latent class analysis uses maximum likelihood estimates to identify latent groups in a set of observations and to classify objects based on model probabilities. This is an improvement upon alternative methods that derive groupings based solely on distance, such as multidimensional scaling (Kruskal & Wish, 1978; Vermunt & Magidson, 2002). Latent class analysis is also favored, as it does not rely on the assumptions of traditional modeling techniques (Magidson & Vermunt, 2005). Latent class analysis is used when the variables within a data set are categorical. This set of observed categorical variables are assumed to be the result of latent groupings, also called classes (Hadzi-Pavlovic, 2010). These classes are what a latent class analysis is designed to find, if present. Generally, a few models with a plausible number of classes are tested. Comparing the     12 observed classification frequencies to the expected frequencies predicted by the model assesses model fit. Two common measures used to assess model fit are the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC; Hagenaars & McCutcheon, 2002). The AIC takes into account the number of model parameters, whereas the BIC takes into account both the number of model parameters and the number of observations. When testing multiple models, the model with the lowest AIC or BIC is often considered the model with the best fit. Although these two information criteria are commonly used, Nylund, Asparouhov, and Muthén (2007) suggest that the bootstrap likelihood ratio test may be a more consistent indicator of classes when assessing multiple models. Overall, latent class analysis is considered superior to other techniques such as cluster analysis, because not only does it identify groupings (like cluster analysis), it provides the probability of group membership for each case and informs researchers about the probability of each group endorsing different variable categories (Hadzi-Pavlovic, 2006). Following group identification, external variables that were not used in the initial latent class analysis can be mapped over the solution to help researchers more comprehensively understand the solution by seeing which variables are related or unrelated to the groupings.  1.1.4.1   Latent class analysis and sexual offending Latent class analysis has been used to develop typologies that have both theoretical and practical importance. The following sections outline recent research using latent class analysis that focused on sexual offending behaviors.   Sexual offending and crime site selection. Deslauriers-Varin and Beauregard (2014) used latent class analysis to investigate the stability of crime sites used by serial sex offenders based on environmental choices. They analyzed information from 361 sex offences, committed by 72 serial sex offenders. The data were collected via a semi-structured interview with the offenders, guided by a questionnaire they created for the purposes of the study. Deslauriers-    13 Varin and Beauregard used latent class analysis to investigate victim encounter and release, and then used cross-tabulations to investigate the crime sites across the offenders’ crime series (chronological progression of their crimes). Two latent class analyses were initially used to investigate eight environmental indicators related to the location of victim encounter and victim release. For example, residential area and offender familiarity with site were two of these indicators. Victim encounter and release were selected for evaluation because police commonly have knowledge of these locations during investigations. Multiple information criterion statistics were reviewed to assess model fit. It was determined that a four-class model was the optimal solution for both the victim encounter and victim release analyses.  The final classes for the victim encounter profiles were labeled neighborhood, shopping center, victim’s home, and offender’s home. The final victim release profiles were labeled home, neighborhood, shopping center, and unfamiliar site. Following the latent class analyses, cross-tabulations were used to analyze the associations that existed between the victim encounter profiles and the victim release profiles.  A significant relationship was found between the two sets of classes, meaning that victim encounter profiles were related to victim release profiles. For example, it was found that 94% of the crimes associated with the victim’s home during the encounter were associated with the home indicator of victim release. Their findings suggest that offenders have limited environments within which they will commit their crime. This, and similar research, utilizing latent class analysis to derive and confirm typologies using external variables may prove useful. For instance, knowing whether offenders consistently use the same crime site has self-evident value during a criminal investigation.   Sexual offending and interviewing/interrogation techniques. Recent Master’s research completed by Busina (2014) applied latent class analysis to sexual offending in an attempt to find characteristics that may help with investigative interviewing and the interrogation process.     14 Busina created both offender and victim profiles, assessed significant profile combinations, and then used the profile combinations to yield practical applications in terms of getting offenders to confess during interrogations. Busina analyzed information from 624 semi-structured interviews with offenders. The data were analyzed using latent class analysis, and models consisting of one to six classes were tested. Multiple information criterion statistics were reviewed to assess model fit, and it was determined that a five-class model was the optimal solution.  The final classes for the offender profiles were labeled introverted specialists, versatile specialists, immature specialists, unemployed specialists, and aggressive introverts. The offender’s decision to confess during interrogation was used as a covariate within the clusters. The results indicated that greater than 50% of the immature specialists (58.5%), unemployed introverts (96.1%), and aggressive introverts (73.1%) made the decision to confess during the interrogation. Less than 50% of the introverted specialists (9.6%) and versatile extraverts (23%) made the decision to confess during interrogation. The final classes for the victim profiles were labeled familiar female children, adult female strangers, dysfunctional adult females, dysfunctional male children, and resistant male children. The offender’s decision to confess during interrogation was again used as a covariate within the clusters. The results indicated that greater than 50% of those with the adult female strangers profile (59.2%) and the resistant male children profile (99%) made the decision to confess during the interrogation. Less than 50% of those with the familiar female children profile (40.5%), dysfunctional adult females profile (19.8%), and dysfunctional male children profile (42.7%) made the decision to confess during interrogation. The final analysis was used to link the offender and victim-related profiles with a chi-square test.  The results indicated that there was a moderate relationship between the offender and victim-related profiles. These results can be viewed in Appendix B. Overall the outcomes suggested that factors such as offender and victim characteristics might play a role in the     15 decision of offenders to confess during an interrogation. Findings such as these have practical value. Understanding how these factors do and do not influence interrogation outcomes can help target interrogation strategies. However, Busina noted that such targeting is not the only factor having an influence on the outcome.   Residential sexual burglary. Pedneault, Harris, and Knight (2012) investigated the development of a residential sexual burglary typology. Pedneault and colleagues analyzed 224 archival files involving incidents of burglary that included a sexual component. The sexual component had to be evident, such as stealing underwear, engaging in voyeurism, or committing rape during the burglary. The same perpetrator may have committed multiple incidents; 104 offenders committed the 224 incidents. The archival files were analyzed using latent class analysis (Collins & Lanza, 2010). The latent class analysis included nine variables that consisted of offender behaviors and situational characteristics involved in the crime. The models that were tested ranged from one to five classes. Multiple information criterion statistics were reviewed to assess model fit, and it was determined that a three-class model was the optimal solution. The final classes from the solution were labeled fetishistic noncontact, versatile contact, and sexually oriented contact. The fetishistic noncontact burglary profile consisted of offences that included accessing property that led to something with sexually arousing value (e.g., stealing underwear). The versatile contact burglary profile consisted of offences that included accessing property for a variety of motivations. Although some offences may have sexually arousing value, that was not the only motivating factor. In the sexually oriented contact burglary profile the offences were strictly of a sexual nature (e.g., rape), and accessing the property was just required to commit the sexual crime. An analysis of variance (ANOVA; Field et al., 2012) was used to evaluate the relationship between the three classes and criminal charges. The independent variable consisted     16 of the three classes derived from the latent class analysis, and the dependent variables were the criminal charges prior to and following the incident.  The results indicated that the profiles were associated with differing criminal histories. The versatile contact burglary profile was associated with a more serious criminal history. Few violent and sexual charges, both prior to and following the incident, were associated with the fetishistic noncontact burglary profile. Lastly, the sexually oriented contact burglary fell in between the other two profiles. It was associated with fewer sexual charges than the versatile contact burglary profile, but more sexual charges than the fetishistic noncontact burglary profile.   The development of this typology served to increase understanding of sexual burglaries. It linked the behaviors that occur during a sexual burglary with three distinct classes, and then related those classes to prior and future criminal offences. These associations can provide important information in criminal investigations. Overall, research evaluating sexual offender typologies has sought to elucidate the relationship among variables such as offence characteristics, offender characteristics, recidivism, and demographic characteristics. The typologies were developed or tested in an attempt to provide a more complete understanding of sexual offending behaviors (National Criminal Justice Association, 2014). The hope is that this comprehensive understanding will help researchers and clinicians make informed decisions about treatment, intervention, and supervision of offenders (National Criminal Justice Association, 2014). In many cases, the findings of these studies yielded useful information for practical settings, such as police investigations. However, the studies focused on subsets of sexual offenders, which may have limited their ability to derive a comprehensive conclusion. The present research will seek to address this limitation by looking at male sexual offender data as a whole using the typological approach, rather than focusing on any specific subset of sexual offender data (e.g., rapists).      17 1.2  Purpose of the Present Study Psychological research is often split into variable- or case-centered approaches (Mandara, 2003). Mandara suggests that the typological approach can provide a blend of these two approaches, and therefore can allow for a fuller understanding of psychological phenomena. The typological approach is rooted in providing researchers and clinicians with the ability to effectively classify the people they are studying. In the proposed project, the group of interest is those who have sexually offended. The ability to effectively classify sexual offenders can lead to researchers and clinicians being better able to prevent, identify, and treat sexual offenders. This ideally can lead to a reduction in future sexual offences through a variety of pathways (e.g., prior to offending, after offending).   The present project used the typological approach to look at (a) whether naturally occurring groups existed within offender and offence characteristics data of a sample of male sexual offenders and (b) whether the naturally occurring groupings were consistent with selected offender typologies in the existing literature.   As any potential typology is based on the sample itself, it was unknown whether a similar typology will be found or what the level of agreement with existing typologies will be. This is why the typology derived was compared to existing typologies. Further, more advanced analytical techniques were available than those used in previous research. The newer method, latent class analysis, allowed for the refinement of previously established typologies. The following describes the two typologies that were used for comparative purposes. 1.2.1  Background typology: Adolescent sexual offenders Chu and Thomas (2010) developed a typology of adolescent-perpetrated sexual offences. Data on 156 adolescent males were pulled from psychological evaluation information provided by the Clinical and Forensic Psychology Branch of the Ministry of Community Development, Youth, and Sports in Singapore. These offenders had completed their evaluations between the     18 years of 1996 and 2007. The offenders were classified into two types by Chu and Thomas.  The first type was referred to as specialists; these were offenders who committed only sexual offences. The second type was referred to as generalists and were those offenders who committed both sexual and non-sexual offences. There were four categories of variables that were coded for: sociodemographic characteristics, offender characteristics, offence characteristics, and recidivism information. Sociodemographic characteristics included information on age at first offence, education, ethnicity, and structure of family. Offender characteristics included information on past offences (non-sexual), intellectual deficits, psychiatric conditions, sexual victimization history, and exposure to pornography. Offence characteristics included information on victim age, victim preference, relationship between offender and victim, use of aggression, use of weapon, and the victim’s gender. Recidivism information included presence of sexual recidivism, violent recidivism, nonviolent recidivism, or recidivism generally. The information on recidivism was provided through criminal record checks that were completed in 2008, one year after the last included data file. Univariate analyses were used to assess the characteristics of the pre-specified groupings, specialists and generalists, and recidivists and nonrecidivists. Relationship with the victim was found to be significantly related to the generalists and specialists typology. No difference was found between the generalists and specialists in regards to sexual recidivism. However, the generalists were found to be more likely to engage in all other types of recidivism. A forward stepwise logistic regression model was then used to model the significant univariate associations. Logistic regression is similar to multiple regression, but with a categorical dependent variable (Field et al., 2012).  This type of analysis evaluates whether the independent variable (continuous or categorical) can predict which categories of the dependent variable the person belongs to. In the context of this research, the significant univariate     19 associations were the predictor variables, and the dependent variable was whether the offender was a recidivist or nonrecidivist. Classification accuracy of the model was assessed by plotting the area under the curve of the receiver operating characteristics. The Hosmer-Lemeshow test was used to assess the fit of the model. The logistic regression indicated that relationship to victim remained significant even after total offence number and age at first sexual offence were accounted for as covariates. The model was better as classifying generalists than specialists, and overall correctly classified 74% of the sample. The results of the Hosmer-Lemeshow test did not indicate a poor model fit.   Finally, the recidivistic outcomes of the generalists and specialists groupings were assessed using Cox regression models. Cox regression models are a form of survival analyses that can be used to incorporate more than one independent variable (Wright, 2000). No difference was found between the generalists and specialists in regards to sexual recidivism after total offence number and age at first sexual offence were accounted for as covariates. However, the generalists were found to be more likely to engage in other types of recidivism. The research by Chu and Thomas (2010) replicated previous findings in regards to adolescent offenders.  There was support for their hypothesized typology of specialist adolescent sexual offenders and generalist adolescent sexual offenders. Although their model was better at classifying generalist offenders, it still correctly classified 74% of the offenders. Findings such as these can help to make decisions about treatment/supervision, and further our understanding of adolescent offenders. 1.2.2  Background typology: Female sexual offenders  Vandiver and Kercher (2004) developed a typology of female sexual offenders using offender and victim characteristics. Data on 471 female sexual offenders were pulled from the Texas Department of Public Safety’s sex offender registry website. This website included demographic information, offence (leading to registration) information, and victim information.     20 Demographic information included variables such as sex, race, registration date, and date of conviction. Offence information included the type of offence. Victim information included age of victim, sex of victim, and relationship to victim. The Texas Department of Public Safety also provided documentation on the criminal histories of each offender. The criminal histories included variables such as total arrests, date of arrests, and sentence imposed.  Vandiver and Kercher chose these variables, as they thought that certain sets of offender characteristics would be associated with sets of victim characteristics. The relationship between offender and victim characteristics was assessed using hierarchical loglinear modeling (HLM). The HLM analysis allowed for the investigation of higher order interactions amongst variables. This means that Vandiver and Kercher were not constrained to only look at two-way interactions. They instead could look at higher order interactions such as three- or four-way interactions. The variables included in the HLM were whether the arrest was for sexual assault, the offender’s age at arrest, the age of the victim, the sex of the victim, and whether the offender and victim were related. The HLM analysis indicated that there were two significant three-way relationships between offender and victim characteristics. The first relationship that existed was between sexual assault, relationship to victim, and victim’s age (see Appendix C). Roughly 70% of offenders who were unrelated to the victim and were arrested for sexual assault perpetrated against children in the 12-17 age range. In general, the 12-17 age range was offended against more than the other age ranges. The second relationship that existed was between offender’s age, relationship between offender and victim, and age of the victim. The victims of young offenders (18-25 age range) were most likely to be in the 12-17 age range, regardless of relationship to the victim. However, older offenders (33-78) who were unrelated to the victim were the most likely offender group to have victims that were younger than 6 years old.  Vandiver and Kercher also identified groupings of female sexual offenders using cluster analysis. Model-based cluster analysis aims to estimate the number of clusters (sub-populations)     21 that are present in a given set of data (Norman & Streiner, 2003). The variables included in the cluster analysis were sexual offence type, age of offender at time of the sexual offence arrest, age of victim, sex of victim, the relationship between the victim and offender, whether the offender was arrested after the sexual offence arrest, and total number of arrests. The analysis indicated that there were six groupings of female sexual offenders. These were labeled heterosexual nurturers, noncriminal homosexual offenders, female sexual predators, young adult child exploiters, homosexual criminals, and aggressive homosexual offenders.  Heterosexual nurturers were characterized by victimizing males who were, on average, age 12. These were the typical media cases in which women in an authority role (e.g., teacher) had sex with boys, but the relationship appeared to be nonabusive. Noncriminal homosexual offenders tended to have female victims who were an average age of 13. They had the lowest rates of re-arrests, and Vandiver and Kercher (2004) suggested these offenders may have been acting with male co-offenders. Female sexual predators tended to have male victims who were an average age of 11. These offenders were the most likely to have subsequent arrests, and had the highest number of total arrests compared to other groupings. Young adult child exploiters were the youngest offenders at the time of arrest, and tended to offend against victims with an average age of 7 years. Half of the victims were related to the offender. Homosexual criminals tended to be arrested for offences other than sexual assault (e.g., indecency). They had victims that were an average age of 11 and their motivations appeared to be monetary based. Finally, the aggressive homosexual offenders included offenders who offended against adult females. They were the most likely group to have committed a sexual assault. The results of Vandiver and Kercher’s (2004) study exhibit a complex relationship between offender and victim characteristics in sexual offences committed by female offenders. A study conducted by Sandler and Freeman (2007) sought to replicate the findings in a     22 geographically separate sample of adult female sexual offenders. That study found similar findings to the original study. 1.2.3  Deriving an adult male sexual offender typology Both Chu and Thomas’ (2010) typology and Vandiver and Kercher’s (2004) typology were useful to this research project due to the scope and coding scheme of the research. Both of the typologies focus on over-arching typologies of sexual offenders, rather than focusing on sub-sets of offenders, such as those who commit child sexual abuse. As demonstrated by the earlier examples, this type of specificity is typical of psychological research into sexual offenders. The coding schemes within both of the research articles are similar to the coding scheme that was used in the present project. The sources of data in those projects are similar to the files that were available for this project, and variables similar to those used for their analyses were readily available. The coding scheme for this project consisted of four variable categories: current offence characteristics, previous offence characteristics, offender characteristics, and sexual recidivism measures. The first three categories consisted of variables that are commonly coded for in research looking at sexual offender files. For example, variables such as age at first offence, victim age, victim gender, and absence or presence of a DSM diagnosis were included. However, depending on the goal of the research, often times only some of these variables were coded for. Chu and Thomas (2010) focused largely on offence characteristics and different types of recidivism (sexual, violent, nonviolent), but also coded for some offender characteristics and sociodemographic information. Vandiver and Kercher (2004) focused largely on offender and offence characteristics. In terms of sexual recidivism variables, many research projects use the Static-99R score as their only measure of sexual recidivism. Although we considered the Static-99R score, we also included a structured professional judgment measure that is also used to assess sexual recidivism. This included more qualitative information for coding. The present     23 research coded for variables that are seen across these studies, rather than focusing on a subset of the variables. The coding was limited to what was present in the files, but overall, drawing on the additional information allowed for rich analyses.   Developing over-arching typologies of sexual offenders is a necessity before deriving more specific typologies based on assumed sub-types. If such sub-types do not actually exist, then using the derived typologies may be misguided. For example, there have been many suggested typologies of child molesters. However, all those who have committed child molestation may not fit into one cluster if an over-arching typology of sexual offenders is derived. Offender and offence characteristics may differentiate those who commit child molestation into multiple sub-types that overlap with other types of sexual offenders. If these sub-types from an over-arching typology are analyzed, a typology may arise that classifies and predicts at a higher rate than existing typologies.   The present research sought to replicate existing research in order to evaluate the similarities between distinct samples of sexual offenders and to use a more advanced statistical procedure (i.e., latent class analysis) than previous research to derive a typology and see how it fit within the framework of existing literature. It was expected that differences would be present among the distinct samples of sexual offenders, as adolescents male, adult females, and adult males tend to have differing risk factors and criminal patterns (e.g., Cortoni & Hanson, 2005; Gal & Hoge, 2015). The latent class analysis was exploratory in nature, and driven by the data, so there were no explicit expectations. However, it was assumed that the classes would be somewhat similar to typologies proposed in the past. A typology rooted in empirical findings, such as the typology developed for this research, was investigated because it may provide significant insight into the heterogeneous nature of sexual offenders. This insight could allow researchers and clinicians to better target and administer treatment and prevention strategies.       24 Chapter 2 Methodology 2.1  Procedure Preceding Site Visit 2.1.1  Selection of files The present project focused on information collected from pre-sentence reports (PSRs). Judges presiding over criminal matters, defense attorneys, or prosecutors can request a PSR. A PSR often is completed when the court does not have sufficient expertise in the matter at hand (Bonta, Bourgon, Jesseman, & Yessine, 2005). The PSRs assist the court with sentencing decisions, and may include information on re-offence risk and treatment needs of the offender. The ethics application for the present project initially identified 331 PSRs for offenders who committed either sexual offences or violent offences. Within the PSRs the offenders were given an opportunity to provide personal information during a psychiatric assessment; this information was then used to corroborate the data provided in medical volumes held at a forensic clinic. The focus of these specific PSRs was to estimate the likelihood of re-offence, identify treatment needs, and suggest clinical interventions that may help lower the offender’s re-offence risk. The files included offences that took place between January 2011 and April 2015. Once on-site, the number of files was reduced to 285, as some files were off-site and some files had erroneously been selected and included non-violent, non-sexual crimes. Of the 285 files, 109 were sexual offenders (three were female) and 176 were violent offenders. This research project focused on the data from the male sexual offenders only. The violent offender data and female sexual offender data will be used in future research. 2.1.2  Identifying file content for coding    Three redacted files were reviewed to identify the relevant variables for the coding scheme. Some of the identified variables were demographic characteristics, offence characteristics, and DSM diagnoses (e.g., Antisocial Personality Disorder). Also included in the data were ratings on two sexual risk assessment measures. The Risk for Sexual Violence     25 Protocol (RSVP) is a structured professional judgment measure designed to identify potential risk factors, determine risk factors importance to future offending, and provide explicit guidelines for risk formulation (Hart, Kropp, Laws, Klaver, Logan, & Watt, 2003). The RSVP was developed using items from the Sexual Violence Risk – 20 (SVR-20) checklist, but was modified to allow for the determination of risk factor importance and include risk formulation such as risk scenarios, risk management strategies, and summary judgments based on the complete assessment. Over half of the files were completed using the RSVP (n = 59), and the rest were completed using the SVR-20 (n = 47). The Static-99R is an actuarial measure designed for use with adult males who have been charged or convicted with at least one sexual offence. The Static-99R estimates relative risk of sexual and violent recidivism, as well as the likelihood of reoffending in 5 or 10 years (Helmus, Thornton, Hanson, & Babchishin, 2012). Nearly two-thirds of the sample had Static-99R data in their files (n = 63).  2.1.3  Coding scheme   The coding scheme developed for extracting data from the files is provided in Table 1. The coding scheme was based on the information available within the redacted files, as well as similar research investigating sexual offender typologies. This research included Chu and Thomas’ (2010) study and Vandiver and Kercher’s (2004) study.        26 Table 1  Coding Scheme  Background Characteristics Offender Characteristics Previous/Index Offence Information Static-99R Risk for Sexual Violence Protocol Parental Drug/Alcohol Use Presence of Drug Arrests Number Age Sexual Violence History Sibling Drug/Alcohol Use Presence of Intellectual Disability Victim Gender Ever Lived With Psychological Adjustment Self-Drug/Alcohol Use DSM Diagnoses Victim Age Index Non-Sexual Violent Mental Disorder Occupation Presence of Psychopathy Victim Relationship Prior Non-Sexual Violence Social Adjustment Marital Status Sexual Victimization History  Use of Aggression  Prior Sex Offences Manageability Number of Children Pornography History Use of Weapon Prior Sentencing Dates Risk Scenarios Previous Hospitalizations  Sexual/Non-sexual Non-Contact Sex Offences Risk Rating Suicidal Ideation  Type of sexual-offence Unrelated Victims Management Suggestions Ethnicity   Stranger Victims  Education Level   Male Victims     Total Score     Risk Category   2.1.4  Development of file review criteria   File review criteria were developed to operationalize the variables listed in Table 1 and minimize the likelihood of subjective bias in the coding process. The file review criteria document was split into five sections: Background Characteristics, Offender Characteristics, Previous/Index Offence Information, Static-99R, and Risk for Sexual Violence Protocol. The document listed each variable within one section, listed the data entry options for each variable     27 (e.g., Yes/No/Other), and outlined the information that should be referred to when inputting the data. For example, the criteria stated that coders should code Yes for the presence of problematic drug/alcohol use by primary caregiver(s) only if alcohol/drug use related to neglect, abuse, etc. of the offender before he was 18. The document also provided the location within the PSRs that the variable information could usually be located. See Appendix D for the full file review criteria document. 2.2  Procedure During Site Visit 2.2.1  File coding on-site   As redaction of the files by an external researcher or reading of redacted files off-site was not approved by the forensic agency, the project was conducted on-site at a forensic clinic in Northern British Columbia. The researcher travelled to the clinic and spent 2 weeks reviewing and coding the un-redacted files. A doctoral student working with the researcher travelled to the clinic for the first 3 days. Although it would have been ideal to have a second coder present for the entirety of the coding process, this was not possible due to financial and time constraints. 2.2.2  Inter-rater reliability assessments   Inter-rater reliability assessments were conducted to check whether the two researchers were drawing the same information when reviewing the files. The primary researcher completed coding of all the files, while the secondary researcher completed coding a portion of the files. In consideration of these factors, and assuming that the researchers as coders are interchangeable (Fleiss, 1978) for the purposes of this research, Krippendorff’s α was used for the reliability check (Krippendorff, 2004). After reviewing the files on-site, it was decided that the best approach would be to calculate the inter-rater reliability check on all file sections except the risk assessment sections. The risk assessment sections introduced the least amount of subjectivity, as the selections were clearly marked in the files. Further, this approach allowed the researchers to review and check the inter-rater reliability of an increased number of files.     28   The researchers initially coded one file together to establish the general coding pattern. Following that, the researchers coded files separately and conducted an inter-rater reliability check on each file coded. The researchers were able to conduct inter-rater reliability checks on 46 of the files, representing roughly 16% of the total files. Any difficulties encountered during the coding process were addressed, recorded, and rectified through in-depth discussions between the researchers. The alpha values ranged from .70 to 1.0, with only seven values under .80. The average alpha value was .88. See Appendix E for the complete list of alpha values for all files coded by the researchers. 2.3  Procedure Following Site Visit 2.3.1  Variable re-coding    Following the site visit, re-coding of certain variables took place to allow for specific analyses or to account for information that was initially overlooked in the coding scheme. This was possible because many of the variables had an additional text option where the researchers could record pertinent information or issues that came up while coding. For example, the ethnicity question was re-formatted to include Caucasian as an option. The term Caucasian was not originally included in the variable coding categories for Ethnicity, but once the files were being reviewed and coded it was noted that the author of the pre-sentence reports often specified whether the offender was Caucasian. This information was included in the additional text box for the variable and used in the re-coding process. The DSM diagnoses variable was originally formatted as a text box where researchers could input all of the DSM diagnoses recorded in the files. This was re-coded to a present/absent format for the following groupings of diagnoses: pedophilia, mood/anxiety, personality disorder, substance/alcohol disorders, and psychotic/schizophrenia. The victim type variable originally had the options of stranger, acquaintance, and relative. The options were re-formatted to split the relative option into     29 romantic partner/ex-partner and non-romantic relative, as this yielded more information for the analyses and made interpretation of the results easier. 2.4  Sample Characteristics   Data from 106 male sexual offenders were included in the study. All identifying information was removed when the offender files were reviewed and coded. The majority (44.3% and 37.7%) of the offenders fell into the 18 to 34.9 age category and the 40 to 59.9 age category. Many offenders were single (40.6%), married (18.9%), or divorced/separated (19.8%), whereas a smaller number of offenders were co-habiting (13.2%) or non co-habiting (4.7%); a small minority of the offenders (2.8%) had no marital status information in their files. Most of the offenders (66%) had children. The majority of the offenders were Aboriginal (51.9%), and a smaller portion were Caucasian (25.5%). A sizable portion (21.7%) of the files did not have sufficient information to ascertain the ethnicity of the offender. A small portion (0.9%) identified other ethnicities (e.g., Indo-Canadian). The majority of the offenders had not completed high school (61.3%). In terms of mental health, most offenders (65.1%) had currently or previously been diagnosed with a DSM disorder. A small portion (17.0%) had previous hospitalizations for mental health concerns. A small portion (27.4%) also had evidence of an intellectual disability.  In terms of offence characteristics, most of the offenders committed their crimes against female victims (79.2%). A small portion committed their crimes against male victims (7.5%) or both female and male victims (10.4%). A few of the files (2.8%) did not have information on victim gender. The majority (42.5% and 33.0%) of the offenders’ victims were non-romantic relatives or acquaintances. Only 21.7% of the victims were adult age (19+). The majority (28.3% and 24.5%) of the offenders’ victims were aged 6 – 11 and 12 – 15. There were a number of victims (15.1%) aged 0 – 5, and a small number of victims aged 16 – 18 (5.7%). Information on victim age was not available in five of the files (4.7%). The majority of offenders did not use     30 aggression (80.2%) or weapons (95.3%) during the commission of their offences. Complete demographic characteristics information is provided in Appendix F.       31 Chapter 3 Results   The analytic plan focused on using a modern profile analysis procedure (i.e., latent class analysis) to identify naturally occurring groups in the data set based on offender and offence characteristics. As latent class analysis allows for the analysis of categorical data, it was well suited to this project. However, prior to the modern profile analysis, the analyses of Chu and Thomas (2010) and Vandiver and Kercher (2004) were replicated. 3.1  Replication of Prior Analyses 3.1.1  Replicated analyses from Chu and Thomas (2010)   The data were initially split into generalists and specialists. The criterion in the original paper for categorizing offenders was whether offenders had both sexual and non-sexual offences (in the past or currently). Since past offence information was not available, the categorization for the present project was based only on current offence information. Descriptive statistics showed that 30.2% of the offenders were categorized as generalists and 69.8% of the offenders were categorized as specialists. This was a different distribution than Chu and Thomas’ adolescent sample, which had an nearly even split of 52.0% generalist offenders and 48.0% specialist offenders.   After categorizing the offenders, Chu and Thomas (2010) used univariate analyses to compare the characteristics generalists versus specialists, as well as recidivists versus non-recidivists. As the collected data did not have follow-up recidivism information, the risk rating from the RSVP/SVR-20 was used instead. This risk rating was selected for the replicated analyses over the Static-99R because the data were nearly complete. However, it should be noted that both risk ratings were used for the latent class analysis. Chi-square analyses were used to evaluate the relationship between relationship to victim and grouping, as well as the relationship between sexual recidivism risk and grouping. Some of the sexual recidivism risk cells were less than five, so the levels were collapsed from five levels into three levels. The Low and Low to     32 Moderate levels were collapsed into Low to Moderate (Revised), Moderate was left as Moderate, and Moderate to High and High were collapsed into Moderate to High (Revised).   The chi-square score for relationship to victim and generalist/specialist grouping was significant, c2(2) = 4.53, p = .03, as was the chi-square score for sexual recidivism risk and generalist/specialist grouping, c2(2) = 7.67, p = .02. The contingency tables for the chi-squares analyses are provided in Tables 2 and 3.  Table 2  Contingency Table and Adjusted Residuals for Generalist/Specialist Grouping and Relationship to Victim   Relationship to Victim Grouping  No Yes Total N Generalist Frequency 11 20 31  Adjusted Residual -2.1 2.1  Specialist Frequency 42 30 72  Adjusted Residual 2.1 -2.1  Total N  53 50 103  Table 3  Contingency Table and Adjusted Residuals for Generalist/Specialist Grouping and Sexual Recidivism Risk   Risk Rating (Collapsed)  Grouping  Low to Moderate Moderate Moderate to High Total N Generalist Frequency 12 6 10 28  Adjusted Residual -2.7 0.9 2.4  Specialist Frequency 46 9 9 64  Adjusted Residual 2.7 -0.9 -2.4  Total N  58 15 19 92      33   In the original paper, Chu and Thomas followed up their initial analyses with a forward stepwise logistic regression and Cox Regression model to evaluate variables that may be related to the variables, such as total number of convicted offences (across time) and age at first offence. As the present data did not have that information, no further analyses were conducted. 3.1.2  Adolescent sexual offender typology versus male sexual offender typology Chu and Thomas (2010) classified adolescent offenders into two types. The first type was referred to as specialists; these were offenders who committed only sexual offences. The second type was referred to as generalists and were those offenders who committed both sexual and non-sexual offences. Relationship with the victim was found to be significantly related to the generalists and specialists typology. No difference was found between the generalists and specialists in regards to sexual recidivism. However, the generalists were found to be more likely to engage in all other types of recidivism. A logistic regression indicated that the relationship to victim remained significant even after total offence number and age at first sexual offence were accounted for as covariates. The model was better at classifying generalists than specialists, and overall correctly classified 74% of the sample. The results of the Hosmer-Lemeshow test did not indicate a poor model fit. Cox regression models indicated that no difference was found between the generalists and specialists in regards to sexual recidivism after total offence number and age at first sexual offence were accounted for as covariates. However, the generalists were still found to be more likely to engage in other all types of recidivism. In Chu and Thomas’ research, a significant relationship between the generalists/specialist grouping and relationship to victim was found, c2(1) = 8.68, exact p = .004. In the present study, a significant relationship also was found, c2(2) = 4.53, p = .03. The present data showed the opposite pattern of Chu and Thomas’ research, with generalists offending more often against familial victims than non-familial (64.5% vs. 35.5%), and specialists offending more often against non-familial victims than familial (58.3% vs. 41.7%). This may be related to the coding     34 of the data, as the current project included romantic partners in the familial coding and acquaintances in the non-familial coding. Chu and Thomas had the basic categories of familial and strangers in their coding. Chu and Thomas’ research also found a significant relationship between the generalists/specialists grouping and recidivism. In the present study, a significant relationship also was found, c2(2) = 7.67, p = .02, between the generalists/specialists grouping and sexual recidivism risk. It should be noted that recidivism risk was not the variable Chu and Thomas used in their study, as they had recidivism outcomes (e.g., violent, sexual, nonviolent) for the offenders. Chu and Thomas’ interpretation focused on percentages, and in their article they found that generalists were more likely to engage in any type of criminal recidivism than specialists (45.5 % vs. 23.9%). If percentages are considered for the adult male sexual offenders, it appears that a larger proportion of the specialists were deemed low to moderate risk than the generalist group (62.0% vs. 38.0%). The increased proportion of the generalists grouping being rated moderate to high risk would seem to agree with the recidivism results of Chu and Thomas’ research. 3.1.3  Replicated analyses from Vandiver and Kercher (2004) The first step for the replication was coding a sexual assault variable. Vandiver and Kercher (2004) used a dichotomous yes/no variable, but it was unclear whether multiple offences were present in their coded files. For the present data the offence information was used to code the sexual assault variable, and Yes was selected if any of the current offence types included a sexual assault. Descriptive statistics showed that 30.2% of the offenders had not been charged with sexual assault and 69.8% of the offenders had been charged with at least one sexual assault. Vandiver and Kercher initially used hierarchical loglinear modeling (HLM) to evaluate the relationship between offender and victim characteristics. HLM is an analysis that is used to evaluate the relationship between three or more categorical variables (Meyers, Gamst, &     35 Guarino, 2013). It functions similarly to a chi-square analysis, in which the focus is on differences between the expected frequencies and observed frequencies in the cross-category cells. Due to the fact there are three or more variables, higher-level interactions can be included and analyzed (e.g., five-way interaction). HLM produces a goodness of fit statistic (loglikelihood ratio) to evaluate the hierarchical model, and a non-significant value indicates the model fits the data well. The variables included in Vandiver and Kercher’s were whether the arrest was for sexual assault, the offender’s age at arrest, the age of the victim, the sex of the victim, and whether the offender and victim were related. They reported that the HLM interactions in their data were complex, and used these findings to provide justification for their subsequent cluster analysis. An attempt to replicate the HLM was made. The variables used for the present analyses were the same as those of the original analysis, with the exception of the modified sexual assault arrest variable mentioned above. Following Vandiver and Kercher’s methods, a saturated model was used to explore the interactions within the data. A saturated model contains all possible effects (Meyers, Gamst, & Guarino, 2013). This means that three-, four-, and five-way interactions were modelled in the analysis. The statistics indicated that there were no significant three-, four-, or five-way interactions. There were significant two-way interactions and main effects (see Appendix G). The goodness of fit test indicated a good model fit for the model including only two-way interactions and main effects (likelihood ratio, c2 = 57.09, p = 1.00). Following the HLM the analytical focus was redirected to the cluster analysis, which was the main analysis within Vandiver and Kercher’s (2010) article.    Following the complex results from their HLM, Vandiver and Kercher (2010) used a cluster analysis to investigate the presence of a typology. Prior to the analysis they created a variable for the offenders’ total number of offences. Since past arrest information was not available for the replication, the total number of offences was calculated using the current     36 offence information. The number of current offences offenders were charged with ranged from 1 to 5. Most offenders had only been charged with one crime (52.8%), followed by two crimes (28.3%), three crimes (10.4%), four crimes (6.6%), and five crimes (1.9%). Vandiver and Kercher also had a variable for type of sexual offence arrested for. This was not coded for in the present data, as it was too difficult to code because there were multiple offences recorded for each offender.   The main step in the replication was determining the number of clusters that were present in the present data set. Vandiver and Kercher (2010) mainly used traditional cluster analysis methods to evaluate whether there were offender categories in their data set. Specifically, they used visual inspection of dendrogram graphs and a within sum-of-squares graph. Dendrograms are tree like diagrams that arise through methods such as calculating the Euclidean distance (Nicol & Pexman, 2010). These diagrams illustrate hierarchical clusters. Branches that are closer to one another represent similarity between cases or clusters of cases. Hierarchical clustering assumes there is a hierarchy within the data (Norman & Streiner, 2003). This means that although a large set of data can be split into sub-groups, there is an over-arching group still present. Indices of similarity or dissimilarity are used to decide whether certain cases should be joined (agglomerative clustering) or divided (divisive clustering). The present data were also investigated using a within sum-of-squares graph and 2D graphical representations of the clusters. Both of these methods are derived using k-means, which are calculated using a form of cluster analysis known as partitioning. Unlike hierarchical methods, partitioning methods do not assume that an over-arching group is present in the data (Norman & Streiner, 2003). It instead assumes that each group derived is its own distinct entity. Although different criteria are used in partitioning methods than those used in hierarchical methods, they both attempt to make homogeneous groups that can be differentiated from other groups.     37   In addition to traditional cluster analytic methods, model-based cluster analysis was used. Both methods were used, as determining the number of clusters can be a subjective process. The researcher wanted to evaluate as many indicators as possible of what the optimal number of clusters may be. Further, Vandiver and Kercher (2010) did not use model-based cluster analysis methods. Model-based cluster analysis aims to estimate the number of clusters that are present in a given set of data (Fraley & Raftery, 1998). Different models with different assumptions are tested, and a fit index is calculated for each model. The goodness of fit index that is calculated is used to decide what the best model is. This represents an advance over traditional cluster analytic methods (e.g., Ward’s method, k-means), because selecting the best-fitting model does not rely as heavily on the researchers determining the number of clusters or deciding on the procedure. Following the traditional cluster analysis methods, inspection of a Bayesian Information Criterion (BIC) graph was used. BIC graphs an index derived using model-based cluster analysis. The BIC is designed to balance model fit with model parsimony, so that the selected model maximizes fit while minimizing the number of parameters (Vrieze, 2012). Better-fitting models will have less negative values (Raftery, 1995). Data from 100 offenders (rather than 106) were used for the analyses due to missing values. The first graph inspected was the within sum-of-squares graph, which can be seen below in Figure 1. The larger sum of squares decreases from one to four clusters (with smaller decreases after) suggests a two-, three-, or four-cluster solution.     38  Figure 1. Within groups sum-of-square graph. This figure depicts the relationship between the within groups sum of squares and the number of clusters extracted. The second graphs inspected were the 2D graphical representations of the clusters (see Figures 2, 3, and 4) based on k means clustering. As the within groups sum-of-squares graph indicated that a two-, three-, or four-cluster solution may be optimal, those three graphs were examined. All of the graphs had some areas of overlap, but the three- and four-cluster solutions appeared to provide better differentiation of the data. 1 2 3 4 5 6200250300350400450500Number of ClustersWithin groups sum of squares    39  Figure 2. 2D representation of the two-cluster solution. -2 -1 0 1 2-3-2-1012342D Representation of the Two Cluster SolutionComponent 1Component 2These two components explain 58.85 % of the point variability.    40  Figure 3. 2D representation of the three-cluster solution. -2 -1 0 1 2 3-2-1012342D Representation of the Three Cluster SolutionComponent 1Component 2These two components explain 58.85 % of the point variability.    41  Figure 4. 2D representation of the four-cluster solution.   The third graphs to be inspected were the dendrograms (see Figures 5 and 6). The dendrograms were derived using Euclidean distance, the similarity index Vandiver and Kercher (2010) used in their article. As the three- and four-cluster solutions stood out in the second set of graphs, the dendrograms for these solutions were examined. The two dendrograms were similar. However, the four-cluster appeared to isolate one small group of data that was part of the largest data cluster in the three-cluster solution. -2 -1 0 1 2-2-1012342D Representation of the Four Cluster SolutionComponent 1Component 2These two components explain 58.85 % of the point variability.    42  Figure 5. Dendrogram outlining a three-cluster solution. This cluster dendrogram was derived using a Euclidean distance method. The numbers at the bottom of the figure are the participant (i.e., offender) number in the data set. 50 64 482410 5742 877813 6349 51 90 81 86 73 67 55 37 18 2972 12 58 35 9246 963698 2 9715106 599 3 663432 4089 56 8247 7075103 26 4 17 7 68 6574 833860 805488 25 61 31 919539 79 76 10019 416 93 23 713044 5328 69 62 84 105 59 1 21 94 85 52 11 2027 104338 77 9 14051015Cluster Dendrogramhclust (*, "ward.D2")DistanceHeight    43  Figure 6. Dendrogram outlining a four-cluster solution. This cluster dendrogram was derived using a Euclidean distance method. The numbers at the bottom of the figure are the participant (i.e., offender) number in the data set.   The final graph to be inspected was the BIC graph (see Figure 7). From the appearance of the BIC graph, a solution of three-, four-, or five-clusters would yield roughly the same average BIC value. The two-cluster solution has one very high BIC value that would likely bring the 50 64 482410 5742 877813 6349 51 90 81 86 73 67 55 37 18 2972 12 58 35 9246 963698 2 9715106 599 3 663432 4089 56 8247 7075103 26 4 17 7 68 6574 833860 805488 25 61 31 919539 79 76 10019 416 93 23 713044 5328 69 62 84 105 59 1 21 94 85 52 11 2027 104338 77 9 14051015Cluster Dendrogramhclust (*, "ward.D2")DistanceHeight    44 average of the solution up above the other cluster solutions. Figure 7. BIC values for one to eight clusters. Lower BIC values indicate better model fit.    After considering the information from all four sources, a four-cluster solution was decided upon for the additional analyses. K means clustering was used to determine how many offenders would fall into the four clusters (refer to Figure 4). The cluster sizes were determined to be 17, 34, 21, and 28.  As employed by Vandiver and Kercher, the Kruskal-Wallis test was then used to test for differences among the clusters. The Kruskal-Wallis test is similar to a one-way independent ANOVA, but is robust and based on ranked data (Field et al., 2012). The first Kruskal-Wallis -1500-1400-1300-1200-1100-1000-900-800Number of componentsBIC1 2 3 4 5 6 7 8 9EIIVIIEEIVEIEVIVVIEEEEVEVEEVVEEEVVEVEVVVVV    45 test showed a statistically significant difference in age of offender among the clusters, H(3) = 82.13, p < .001, with a mean rank offender age score of 71.18 for Cluster 1, 75.87 for Cluster 2, 25.43 for Cluster 3, and 25.95 for Cluster 4. Multiple comparisons indicated that Class 1 had significant differences in terms of offender age than Cluster 3 (difference = 45.75) and Cluster 4 (difference = 45.23), and that Cluster 2 had significant differences in terms of offender age than Cluster 3 (difference = 50.44) and Cluster 4 (difference = 49.92). Cluster 1 and Cluster 2 did not significantly differ in terms of offender age, nor did Cluster 3 and Cluster 4. From these results, it can be concluded that offenders in Cluster 1 and 2 tended to be older than offenders in Cluster 3 and 4.  The second Kruskal-Wallis test showed a statistically significant difference in age of victim among the clusters, H(3) = 68.82, p < .001, with a mean rank offender age score of 71.15 for Cluster 1, 29.49 for Cluster 2, 86.24 for Cluster 3, and 36.68 for Cluster 4. The multiple comparison results indicated that Class 1 had significant differences when compared to Cluster 2 (difference = 41.66) and Cluster 4 (difference = 34.47). Additionally, Cluster 2 was significantly different from Cluster 3 (difference = 56.75), and Cluster 3 was significantly different (difference = 49.56) from Cluster 4. From these results, it can be concluded that Class 1 and 3 tended to offend against older groups of victims, followed by Class 4, and then Class 3, who tended to offend against the youngest group of victims. 3.1.4  Female sexual offender typology versus male sexual offender typology   Vandiver and Kercher (2004) conducted an HLM and found significant three-way interactions (Sexual Assault × Related to Victim × Victim’s Age; Offender’s Age × Relationship Between Victim and Offender × Victim’s Age) in their sample of female sexual offenders. Unlike their results, only significant two-way interactions and main effects were found in the present data. However, similar to Vandiver and Kercher, the HLM results were complex and suggested than an alternative method such as cluster analysis would be beneficial in deriving the     46 typology. Vandiver and Kercher used dendrograms and within sum-of-squares plots to graphically evaluate the number of clusters for their data. These methods were used in addition to 2D graphical representations and BIC graphs to determine the number of clusters in the sample of adult male sexual offenders.   Vandiver and Kercher (2004) identified six clusters of female sexual offenders using cluster analysis, whereas the present data were found to have four clusters. As was found in the original article, the clusters were related to age of offender and age of victim. Multiple comparisons indicated which clusters differed from one another, which would allow for further description of the clusters. In the present data, Cluster 1 and 2 tended to be older than offenders in Cluster 3 and 4. In terms of victim age, Class 1 and 3 tended to offend against older groups of victims, followed by Class 4, and then Class 3, who tended to offend against the youngest group of victims. Further examination of the clusters was not completed because the analytical plan was to use latent class to derive the adult male sexual offender typology. Vandiver and Kercher noted in their discussion that external analyses would have been more useful in validating the clusters, which was planned to be included in the latent class analysis. 3.2  Latent Class Analysis of Adult Male Sexual Offender Data     As previously mentioned, latent class analysis is a particularly useful method for capturing heterogeneity both between and within groups (Rosato & Baer, 2012). Latent class analysis uses maximum likelihood estimates to identify latent groups in a set of observations and to classify objects based on model probabilities. There are many benefits to using latent class analysis. It also does not rely on the assumptions of traditional modeling techniques (Magidson & Vermunt, 2005), nor does it derive groupings based solely on distance (Kruskal & Wish, 1978; Vermunt & Magidson, 2002). Latent class analysis is superior to other techniques such as cluster analysis because it can identify groupings, provide a probability of group membership for each case, and inform on the proportion of individuals who are present in a given group (Hadzi-    47 Pavlovic, 2006). Following group identification, external variables not used in the initial analysis can be mapped over the solution to help researchers better understand the groupings by seeing which variables are related or unrelated. Given the benefits of latent class analysis, this was the chosen analysis for evaluating the adult male sexual offender data set for potential typologies. 3.2.1  Selection of variables for analyses   Excessive variables can make a latent class analysis extremely complex, so it was decided no more than eight variables would be selected and used for the models. The offender and offence characteristics were selected for use in the models, whereas the background characteristics and risk ratings of the offender were reserved to use in external analyses following the class identification. This is typical of research investigating sexual offender typologies (e.g., Chu & Thomas, 2010; Vandiver & Kercher, 2004; Wortley & Smallbone, 2014). Using these variables allowed the researcher to compare the derived typology to existing typologies of a similar nature. Nine variables were initially selected and the frequencies were examined to determine whether there was enough variation to include in the model. The variables were sexual victimization history, age of offender, presence/absence of pedophilic disorders, presence/absence of mood/anxiety disorders, presence/absence of psychopathy, victim gender, victim age, victim relationship, and use of weapon during offence. The psychopathy variable was dropped, as only two cases were recorded as Yes, one case was recorded as a No, and the remainder did not have the information included. The variable of whether a weapon had been used during the offence was also dropped, as only four cases had been recorded as Yes. The presence/absence of mood/anxiety disorders was also not included, as only seven cases had been recorded as Yes. The remainder of the offender and offence characteristic variables were then examined to determine what variables may be included. Nearly all of the variables had enough variability to be included in the models. These variables included drug arrests, intellectual disability, mental health hospitalizations, presence/absence of a personality disorder,     48 presence/absence of a substance/alcohol disorder, psychotic/schizophrenia, pornography exposure, and use of aggression. The two variables selected were presence/absence of substance/alcohol use disorders and use of aggression during offence. They were selected as both variables were varied in terms of categories and it was thought that that could lead to good differentiation between classes. Therefore, the final eight for the latent class analyses were sexual victimization history, age of offender, presence/absence of pedophilia, presence/absence of substance/alcohol disorder, use of aggression, victim gender, victim age, and victim relationship. 3.2.2  Determining the number of classes   The existing literature was reviewed to determine how latent class analyses have been used to identify the number of classes for similar datasets (e.g., Goncy, Sullivan, Farrell, Mehari, & Garthe, 2016). Following the review, MPLUS 7.4 (Muthén & Muthén, 2013) was used to conduct a series of latent class analyses to determine the optimal number of classes for the male sexual offender data. Models of increasing classes (1 through 6) were assessed using a variety of statistical indices, rather than relying on a particular index or assuming the number of classes was pre-determined by the replicated analyses. The observed classification frequencies were compared to the expected frequencies predicted by the model, which assesses model fit. One index used was the loglikelihood. When only consulting the loglikelihood, the best solution tends to be that with the largest loglikelihood. The Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC; Hagenaars & McCutcheon, 2002) statistical indices were also used. The AIC takes into account the number of model parameters, whereas the BIC takes into account both the number of model parameters and the number of observations. When testing multiple models, the model with the lowest AIC or BIC is often considered the model with the best fit (Nylund et al., 2007; Raftery, 1995). The Lo-Mendell-Rubin adjusted likelihood ratio test (LMR-LRT) was also used to determine whether adding one more class improved the fit of the     49 model. A non-significant value suggests that the model with one fewer class is a better explanation of the data (Nylund et al., 2007). The bootstrapped parametric likelihood ratio test (B-LRT) was included in the analysis, which gives similar information to the LMR-LRT, but with bootstrapped statistics. Finally, entropy statistics were also consulted. Entropy statistics assess whether or not the model is classifying well, with higher entropy statistics indicating that the model is doing a good job of classifying cases. When the entropy value of a model is greater than 0.80, it indicates that the latent classes are highly discriminating (Muthén & Muthén, 2013). The fit indices for the six models tested are presented in Table 4.	 50  Table 4  Statistical Indices for Latent Class Models Tested Number of Classes Loglikelihood Free Parameters AIC BIC Adjusted BIC LMR -LRT (p) B-LRT (p) Entropy  1 -704.27 16 1440.54 1483.16 1432.61 - - -  2 -660.94 33 1387.88 1475.77 1371.51 85.45 (p = 0.10) -704.27 (p < .001) 0.79  3 -641.97 50 1383.94 1517.12 1359.15 37.43 (p = 0.66) -660.93 (p = 0.01) 0.87  4 -625.14 67 1384.27 1562.73 1351.05 33.25  (p = 1.00) -641.97 (p = 0.14) 0.84  5 -608.85 84 1385.70 1609.43 1344.04 29.37  (p = 0.80) -623.72 (p = 0.25) 0.90  6 -594.34 101 1390.69 1659.69 1340.60 28.65  (p = 1.00) -608.85 (p = 0.11) 0.91  Note. The adjusted BIC is adjusted for sample size. 	 51   As emphasized by Muthén and Muthén (2013), the analyses were checked to ensure that global maximum (i.e., best loglikelihood) could be reached in each model. The global maximum provides the single best set of parameters, unlike local maxima. There can be multiple local maxima in a given dataset. Local maxima are related to the complexity of the model, and they become more common as the number of latent classes increases. As the models were increasingly complex for the present data, increasing numbers of random starts were used in the latent class analyses to allow for the global maximum for each model to be reached. For more information, refer to Muthén and Muthén (2013) and Jung and Wickrama (2008).  All of the statistical indices were consulted to determine the optimal number of classes for the dataset. When looking in a holistic manner, although the loglikelihood statistics were mid-range, three to five classes appeared to be possible solutions. The loglikelihood value for the five-class solution was the second highest of the models (both non-bootstrapped and bootstrapped), so it was evaluated further. AIC values for the three-, four-, and five-class solutions were all within a very close range. The BIC for the five-class solution was higher than the other two; however, the sample size adjusted BIC was lower for the five-class solution than the other two solutions. Further, the five-class solution had the highest entropy of the three solutions. For these reasons it was decided to move forward using a five-class latent class model. However, it should be noted that the three-class solution also seemed a strong possible solution for the dataset. 3.2.3 Reviewing the model and naming the classes    The average latent class probabilities (ACPs) for most likely latent class membership by latent class were then reviewed. In a well-fitted model, the probability for the most likely class should be considerably higher than the probabilities for the consecutive classes. In the present analysis the average latent class probabilities for the most likely class were very high, supporting the five-class model. The values for the five-class model are presented in Table 5.     52 Table 5  Average Latent Class Probabilities for Most Likely Class Membership  Class 1 Class 2 Class 3 Class 4 Class 5 Class 1 0.94 0.01 0.00 0.05 0.04 Class 2 0.00 1.00 0.00 0.00 0.00 Class 3 0.00 0.00 0.95 0.03 0.01 Class 4 0.03 0.00 0.00 0.91 0.06 Class 5 0.09 0.00 0.00 0.02 0.88   Each latent class was then examined to determine the probability of the class having endorsement for the different item categories (e.g., were the class members more likely to have the presence or absence of a pedophilic disorder?). The item probabilities for all items used in the latent class analysis are provided in Table 6. Some commonalities and differences emerged among the classes. For example, Class 1 and Class 5 had no or little presence of pedophilic disorders, whereas Classes 2, 3, and 4 had varying degrees of pedophilic disorder present. Substance and alcohol use disorders were endorsed within Classes 1, 2, and 4, but were not probable in Class 3 or Class 5. There were varying degrees of probability of sexual victimization history, with all classes showing some endorsement. Class 1, 2, and 3 had varying levels of endorsement of use of aggression, whereas Class 4 and 5 showed no endorsement. Victim gender, age, and relationship varied across all the classes. These variables assisted in preliminary naming of the groups, as there were specific combinations of these variables for each class. Offender age also varied across all five classes.            53 Table 6  Probability for Each Item Category for Latent Classes Item Class 1 (n = 35) Class 2 (n = 9) Class 3 (n = 7) Class 4 (n = 31) Class 5 (n = 24) Pedophilic Disorder    Absent    Present  1.00 0.00  0.78 0.22  0.00 1.00  0.51 0.49  0.94 0.06 Substance/Alcohol Use Disorder    Absent    Present   0.28 0.72   0.13 0.87   1.00 0.00   0.32 0.68   1.00 0.00 Sexual Victimization History    No    Yes  0.63 0.37  0.67 0.33  0.56 0.44  0.42 0.58  0.46 0.54 Victim Gender    Female    Male    Both  0.94 0.06 0.00  0.00 0.11 0.89  0.85 0.00 0.15  1.00 0.00 0.00  0.69 0.22 0.09 Victim Age    0 - 5    6 - 11    12 - 15    16 - 18    19 +  0.00 0.00 0.32 0.10 0.58  0.78 0.11 0.00 0.00 0.11  1.00 0.00 0.00 0.00 0.00  0.00 0.68 0.27 0.00 0.06  0.15 0.41 0.33 0.12 0.00 Victim Relationship    Romantic Partner/Ex-partner    Non-romantic Relative    Stranger    Acquaintance  0.08 0.09 0.19 0.64  0.22 0.22 0.56 0.00  0.00 0.60 0.27 0.14  0.00 0.85 0.15 0.00  0.00 0.48 0.00 0.53         54 Table 6  Probability for Each Item Category for Latent Classes Item Class 1 (n = 35) Class 2 (n = 9) Class 3 (n = 7) Class 4 (n = 31) Class 5 (n = 24) Offender Age    18 - 34.9    35 - 39.9    40 - 59.9    60 +  0.61 0.07 0.32 0.00  0.65 0.11 0.24 0.00  0.00 0.15 0.85 0.00  0.18 0.18 0.41 0.24  0.58 0.00 0.35 0.07 Use of Aggression    No    Yes  0.61 0.39  0.68 0.33   0.55 0.45   1.00 0.00  1.00 0.00    After reviewing the classes and their respective item probabilities, a tentative name for each class was generated. The names were preliminary and based solely on the latent class analysis. It was expected there could be slight changes in the names and descriptions once external analyses were completed.   Class 1 was labelled Mixed Victim Assaulters. This class was characterized by a lack of pedophilic disorder, a moderate-high probability of substance/alcohol use disorders, and a low-moderate probability of sexual victimization history. The offences were likely to be committed against females, primarily adults, with whom the offender had a limited or no relationship with prior to the offence. The offenders were most likely to be in the youngest age category, and there was a low-moderate probability of using aggression during the offence. Class 2 was labelled the Non-Pedophilic Mixed Gender Child Molesters. Class 2 had a low-moderate probability of having pedophilic disorder, a high probability of having a substance/alcohol use disorder, and a low-moderate probability of sexual victimization history. The victims tended to be both male and female, and were very likely to be pre-pubescent     55 children. The children were most likely to be strangers, but some were relatives. The offenders were most likely to be in the youngest age category, and there was a low-moderate probability of using aggression during the offence.  Class 3 was labelled the Preferential Pedophiles. Class 3 was characterized by the presence of pedophilic disorder. There was a no probability of substance/alcohol use disorders, and a moderate probability of sexual victimization history. The victims were primarily female, and were exclusively extremely young, pre-pubescent children. They were highly likely to be non-romantic relatives, but there was some probability of the victim being a stranger or acquaintances. This class had a higher probability of being older than the other classes, and was the most likely class to use aggression during the offence.   Class 4 was labelled Non-Aggressive Incest Offenders. Class 4 had a moderate probability of pedophilic disorders, with a moderate-high probability of substance/alcohol use disorders and sexual victimization history. The victims were exclusively female, tended to be non-romantic relatives, and were primarily pre-pubescent or pubescent children. The age of offender was variable, but included a higher probability of being in the older age range than other classes. The offences did not include aggression.  Finally, Class 5 was labelled Non-Aggressive Non-Pedophilic Child Molesters. Class 5 had a low probability of pedophilic disorders, no probability of substance/alcohol use disorders, and a moderate probability of sexual victimization history. The victims were most likely to be female, but there was a low probability of male or mixed gender victims. The victims were exclusively non-adults, with the probability dispersed across the age ranges. There was a moderate probability of the victims being acquaintances or non-romantic relatives. The offenders had a moderate probability of being in the younger age range, but there was some variability. The offences did not include aggression. A partial latent profile graph of the five classes is provided below in Figure 8.     56  Figure 8. Profile of the five latent classes. The figure shows the probability of the absence of an offender/offence characteristic based on class membership. Only four of the eight variables are shown here, as they were dichotomous. Only four variables are provided because MPLUS sums variable probabilities for non-dichotomous categorical variables. Therefore, it was simpler to focus on the dichotomous items for illustrative purposes. 3.2.4  Class differences on external variables   Following the establishment of the classes, differences among the classes were examined using variables that had not been included in the latent class analyses (i.e., external variable analyses). These variables consisted of demographic characteristics and the offender risk ratings. As the variables were categorical, bivariate correlations were not used to assess the relationships between variables. A multinomial logistic regression to examine the multivariate relationships     57 was considered, but it was not conducted due to small cell size. This was related both to some variables having a large number of categories, and to there being five classes to evaluate. Instead, chi-squares analyses were used to test for differences among the classes.   The chi-squares results were non-significant for problematic family drug/alcohol use, c2(4) = 3.80, p = .43, non-problematic family drug/alcohol use, c2(4) = 9.06, p = .06, marital status, c2(16) = 21.20, p = .17, number of children, c2(16) = 26.12, p = .05, education, c2(16) = 19.57, p = .24, previous drug arrests, c2(4) = 1.21, p = .88, intellectual disability, c2(4) = 2.69, p = .61, previous mental health hospitalizations, c2(4) = 4.90, p = .30, psychopathy, c2(8) = 7.55, p = .48, history of sexual victimization, c2(4) = 4.59, p = .33, and pornography exposure, c2(4) = 7.38, p = .12. The chi-square results also were non-significant for both the Static-99R risk ratings, c2(12) = 12.57, p = .40, and the RSVP risk ratings, c2(8) = 6.55, p = .59.   Although many results were non-significant, some of the chi-squares analyses used to examine demographic characteristics did yield significant results. This included ethnicity, c2(12) = 37.29, p < .001, and current/past DSM diagnoses, c2(4) = 33.63, p < .001. As noted earlier, the current past/DSM diagnoses was broken down into specific sub-categories. Given the significant chi-square score from the analysis of the DSM diagnoses, these sub-categories were then analyzed. The chi-square test was significant for presence/absence of a mood/anxiety disorder,  c2(4) = 11.74, p = .02, and presence/absence of a personality disorder, c2(4) = 20.61, p < .001. The chi-square result was non-significant for presence/absence of a psychotic/schizophrenic disorder, c2(4) = 2.44, p = .66. The presence/absence of a pedophilic disorder and presence/absence of a substance/alcohol use disorder were not included in the chi-square analyses because they were used as part of the latent class analysis. The observed frequencies, expected frequencies, and adjusted standardized residuals for the significant chi-square tests are provided in Table 7.	 58 Table 7  Cell Frequencies and Residuals of External Variable Categories for the Latent Classes Item Mixed Victim Assaulters (n = 35) Non-Pedophilic Mixed Gender Child Molesters (n = 9) Preferential Pedophiles (n = 7) Non-Aggressive Incest Offenders (n = 31) Non-Aggressive Pedophilic Child Molesters (n = 24) Ethnicity      Aboriginal           Count           Expected Count           Adjusted Residual      Caucasian           Count           Expected Count           Adjusted Residual      Other           Count           Expected Count           Adjusted Residual      Not Applicable           Count           Expected Count           Adjusted Residual   24 18.2 2.4  5 8.9 -1.9  1 0.3 1.4  5 7.6 -1.3   1 4.7 -2.6  5 2.3 2.2  0 0.1 -0.3  3 2.0 2.2   0 3.6 -2.8  6 1.8 3.8  0 0.1 -0.3  1 1.5 -0.5   22 16.1 2.5  5 7.9 -1.4  0 0.3 -0.6  4 6.7 -1.4   8 12.5 -2.1  6 6.1 -0.1  0 0.2 -0.5  10 5.2 2.7 Note. The adjusted residual indicates the standard deviations above or below the expected count the observed count is, while taking into account the sample size. Bolded values have reached the level of significance of +/- 1.96.           59 Table 7  Cell Frequencies and Residuals of External Variable Categories for the Latent Classes Item Mixed Victim Assaulters (n = 35) Non-Pedophilic Mixed Gender Child Molesters (n = 9) Preferential Pedophiles (n = 7) Non-Aggressive Incest Offenders (n = 31) Non-Aggressive Pedophilic Child Molesters (n = 24) Mood/Anxiety Disorder      No           Count           Expected Count           Adjusted Residual      Yes           Count           Expected Count           Adjusted Residual   35 32.7 1.9  0 2.3 -1.9   7 8.4 -2.0  2 0.6 2.0   5 6.5 -2.4  2 0.5 2.4   29 29.0 0.0  2 2.0 0.0   23 22.4 0.5  1 1.6 -0.5 Personality Disorder      No           Count           Expected Count           Adjusted Residual      Yes           Count           Expected Count           Adjusted Residual   25 29.1 -2.2  10 5.9 2.2   5 7.5 -2.3  4 1.5 2.3   4 5.8 -1.9  3 1.2 1.9   31 25.7 3.0  0 5.3 -3.0   23 19.9 1.9  1 4.1 -1.9 Note. The adjusted residual indicates the standard deviations above or below the expected count the observed count is, while taking into account the sample size. Bolded values have reached the level of significance of +/- 1.96.   	 60  The adjusted standardized residuals were examined for the three variables with significant chi-square results. The larger the absolute value of the residuals, the more contribution to the overall chi-square statistic the value has (Sharpe, 2015). The residuals are standardized z-values, and the level of significance is +/- 1.96 (Haberman, 1973).  Mixed Victim Assaulters and Non-Aggressive Incest Offenders had large, positive values for the Aboriginal category of ethnicity, while Non-Pedophilic Mixed Gender Child Molesters, Preferential Pedophiles, and Non-Aggressive Child Molesters, had large, negative values. This indicates that there were more Aboriginal offenders in the former classes than would be expected by chance, and fewer Aboriginal offenders in the latter classes than would be expected by chance. For the Caucasian category, Preferential Pedophiles and Non-Aggressive Incest Offenders had large, positive values, indicating more Caucasians than would be expected by chance in those classes. The remaining three classes had values under 1.96. Non-Pedophilic Mixed Gender Child Molesters and Non-Aggressive Child Molesters had large, positive values for the Not Applicable category of ethnicity, while the remaining three classes had values under 1.96. This indicates that there were more offenders that had Not Applicable in the two classes than would be expected by chance. All of the values for the Other category were under 1.96.  For the DSM diagnoses variables, the standardized adjusted residuals of the two categories mirrored each other because they were opposite (Yes/No). For the presence of a mood/anxiety disorder, Non-Pedophilic Mixed Gender Child Molesters and Preferential Pedophiles had large, negative values for the No category, and large, positive values for the Yes category. This indicates that for those classes there were more mood/anxiety disorders than would be expected by chance. The remaining classes had values under 1.96. For the presence of a personality disorder, Mixed Victim Assaulters and Non-Pedophilic Mixed Gender Child Molesters had large, negative values for the No category and large, positive values for the Yes category. This indicates that for those classes, there were more personality disorders than would     61 be expected by chance. Non-Aggressive Incest Offenders had a large, positive value for the No category and a large, negative value for the Yes category. This indicates that for that class there were fewer personality disorders than would be expected by chance. The remaining two classes had values under 1.96.   These findings were reviewed to see how they differentiated the classes. The Mixed Victim Assaulters had more Aboriginal offenders and more personality disorders than would be expected by chance. The Non-Pedophilic Mixed Gender Child Molesters had fewer Aboriginal offenders, more mood/anxiety disorders, and more personality disorders than would be expected by chance. The Preferential Pedophiles had fewer Aboriginal offenders, more Caucasian offenders, and more mood/anxiety disorders than would be expected by chance. The Non-Aggressive Incest Offenders had more Aboriginal offenders, more Caucasian offenders, and fewer personality disorders than would be expected by chance. Lastly, the Non-Aggressive Non-Pedophilic Child Molesters had fewer Aboriginal offenders than would be expected by chance.  3.2.5  Class differences in case file textual information   Following the chi-square analyses, textual offence information was reviewed to see whether qualitative information could help inform the class descriptions. As discussed by Braun and Clarke (2006), thematic analysis is a flexible method that can be used to provide a rich account of data, but does not necessarily have to be rooted in theory. In this project, thematic analysis was used in a simple manner to report on patterns within the class data. As part of the data collection efforts, brief descriptions of the offences were captured in a text box. These were generally one sentence, and included information such as who the victim of the offence was and what the context of the offence was. To assess the information the data were sorted by class, and then the data from each class were reviewed separately. The goal of reviewing these data was to see whether there were common types of victims or environments that could provide further class differentiation.     62   Class 1 (n = 35) was labelled Mixed Victim Assaulters, and was characterized by primarily adult female victims who were strangers or acquaintances. The offenders were also highly likely to have a substance/alcohol use disorder.  When examining the cases, victims included individuals such as neighbors, prison officers, fellow party goers, family friends, or in some cases, individuals they had not previously had any interactions with. Many of the cases involved alcohol or drug use at the time of assault, and included cases such as assaulting the victim at a party or during a camping trip. There were a number of younger males who committed assaults against strangers or acquaintances. Some of these offences included violence, which made sense as nearly all the offenders were charged with sexual assault. However, there was one case involving possession of child pornography, and a few cases involving sexual touching. This was the only class in which the victims were primarily adults. This may be related to the fact that only provincial data were gathered, restricting the number of violent adult-oriented assaults that were present.  Class 2 (n = 9) was initially labelled Non-Pedophilic Mixed Gender Child Molesters. This class was characterized by both male and female victims who were very likely to be pre-pubescent children, although there were some adult victims. The victims were most likely to be strangers, but some were relatives. The offenders were also highly likely to have a substance/alcohol use disorder. When examining the cases, it was noticeable that 4 of the 9 cases involved charges of possession of child pornography. One case involved charges of indecent exposure and exposing genitals for a sexual purpose. The remaining 4 cases involved sexual assault against partners, ex-partners, and their families, and in some cases severe violence was used (e.g., bodily harm, weapon use). In the sexual assault cases, there was often more than one victim. Given that some of the victims in this class were adults, the name of Class 2 was changed from Non-Pedophilic Mixed Gender Child Molesters to Non-Pedophilic Mixed Gender Offenders. From the case file information, Class 2 seemed like it could have been comprised of     63 two groupings, the child pornography users and then violent sexual assaulters. However, given the small sample size of these case types, it may have been difficult for the latent class analysis to identify and differentiate the groupings.  Class 3 (n = 7) was labelled Preferential Pedophiles. This class was characterized by the presence of pedophilic disorder and primarily young, pre-pubescent female victims. The victims were highly likely to be non-romantic relatives, but there was some probability of the victims being strangers or acquaintances. This class was also the most probable class of using aggression during the offence. When examining the cases, it was noticeable that 2 of the 7 cases involved charges of possession, creating, and distributing child pornography, and a third case involved using a hidden camera to record victims. Some of the cases, and the remainder of the non-pornography cases, involved sexual touching of children, and/or sexual assault. In most cases, the victims were the offenders’ children. The establishing factor of these cases seemed to be the presence of the pedophilic interest in children. Rather than just possessing child pornography, such as the offenders in Class 2, these offenders actually became involved in making and distributing child pornography. These offenders mostly fell into the age range of 40 to 60, which is an older age range than the majority of the other classes.  Class 4 (n = 31) was labelled Non-Aggressive Incest Offenders, and was characterized by a moderate probability of pedophilic disorders and victims who were primarily pre-pubescent or pubescent related females. Although many of the offenders were charged with sexual assault and/or sexual touching, the offences did not include aggression. When examining the cases, many of the victims included children, grandchildren, and nieces. Unlike the victims of Class 3, Class 4 involved many “step” relatives that were related to through their spouse. Six of these cases were based on historical events.  Only one case involved possessing and distributing child pornography. The offenders in this class had a moderate-high probability of a substance/alcohol use disorder, and the age range spanned across all options.     64  Lastly, Class 5 (n = 24) was labelled Non-Aggressive Non-Pedophilic Child Molesters. This class was characterized by a low probability of pedophilic disorders, a lack of substance/alcohol use disorders, and a mix of pre-pubescent and pubescent male and female victims. The age of the victims was mainly in the later pre-pubescent or pubescent age range. The victims were likely to be acquaintances or non-romantic relatives. The cases included victims such as their children’s friends, a neighbor’s child, a student, an employee’s child, and family friends. These offenders appeared opportunistic, and used their positions to find and prey on their victims. As seen in Class 4, although the offenders were mainly charged with sexual assault and/or sexual touching, the offences did not include aggression.       65 Chapter 4 Discussion   The present project sought to use the typological approach to look at (a) whether naturally occurring groups exist within offender and offence characteristics data of a sample of male sexual offenders and (b) whether the naturally occurring groupings are consistent with selected offender typologies in the existing literature. Overall, both aims were successfully met. 4.1  Summary of the Typology Replications Chu and Thomas’ (2010) research classified adolescent sexual offenders into generalists and specialists. Generalists had committed a range of offences, whereas specialists had committed only sexual offences. Similar to the adolescent offender sample, significant relationships were found between the generalist/specialist grouping and relationship to victim and the generalist/specialist grouping and sexual recidivism risk in the sample of adult male sexual offenders. The first relationship showed the opposite pattern of Chu and Thomas’ research, with generalists offending more often against familial victims than non-familial and specialists offending more often against non-familial victims than familial. For the second relationship, even though the variable used in the present research was slightly different than that of Chu and Thomas’, the increased proportion of the generalists grouping being rated moderate to high risk agreed with the recidivism outcome results of Chu and Thomas’ research. Their research indicated that generalists were more likely to engage in any type of criminal recidivism than specialists. Vandiver and Kercher (2004) found significant three-way interactions in their sample of female sexual offenders. Unlike their results, only significant two-way interactions and main effects were found in the present data. However, similar to Vandiver and Kercher, the results were complex and suggested that an alternative method, such as cluster analysis, would be beneficial in deriving the typology. Vandiver and Kercher (2004) identified six clusters of female sexual offenders using cluster analysis, whereas the present data were found to have four     66 clusters. As was found in the original article, the clusters were related to age of offender and age of victim. Multiple comparisons indicated which clusters differed from one another. For example, Cluster 1 and 2 tended to be older than offenders in Cluster 3 and 4. Further examination of the clusters was not complete because the analytical plan was to use latent class to derive the adult male sexual offender typology. 4.2  Summary of the Derived Adult Male Sexual Offender Typology Sexual offending typologies have traditionally been developed in an attempt to provide a more complete understanding of sexual offending behaviors (National Criminal Justice Association, 2014). However, developing accurate and useful typologies has proven to be a difficult task. As mentioned previously, sexual offenders are a highly heterogeneous group (Seto et al., 2015). Sexual offenders may have committed the same crime and present similarly, yet they may have a wide range of differing background characteristics, attitudes and beliefs, and clinical and criminogenic needs (National Criminal Justice Association, 2014; Robertiello & Terry, 2007). The present research sought to established an adult male sexual offender typology using latent classes analyses. As noted earlier, latent class analysis is considered superior to other techniques such as cluster analysis, because not only does it identify groupings (like cluster analysis), it provides a probability of group membership for each case (Hadzi-Pavlovic, 2006). Latent class analysis also informs researchers about the proportion of individuals that are present in a given group (Hadzi-Pavlovic, 2006).  After reviewing the latent class solutions, it was determined that a five-class solution was optimal for the collected data. The item probabilities were reviewed to determine how each class was different and similar from one another. The items considered were sexual victimization history, age of offender, presence/absence of pedophilia, presence/absence of substance/alcohol disorder, use of aggression, victim gender, victim age, and victim relationship. The examination revealed that Class 1 and Class 5 had no or little presence of pedophilic disorders, whereas     67 Classes 2, 3, and 4 had varying degrees of pedophilic disorder present. Substance and alcohol use disorders were endorsed within Classes 1, 2, and 4, but were not probable in Class 3 or Class 5. When it came to a history of sexual victimization all classes exhibited some endorsement. Victim gender, age, and relationship varied across all the classes. Offender age also varied across all four classes. Use of aggression was the final variable, in which only Class 4 and Class 5 showed no probability of endorsing, while the remaining three classes had varying levels.  Given this information, the classes were named as followed (one through five, respectively): Mixed Victim Assaulters, Non-Pedophilic Mixed Gender Child Molesters, Preferential Pedophiles, Non-Aggressive Incest Offenders, and Non-Aggressive Non-Pedophilic Child Molesters. Following the latent class analysis, external analyses were completed. Chi-squares analyses were conducted to assess how demographic variables and risk ratings were related to the classes. The findings indicated that ethnicity, the presence of a mood/anxiety disorder, and the presence of a personality disorder were significantly related to the classes. Mixed Victim Assaulters and Non-Aggressive Incest Offenders had fewer Aboriginal offenders than expected by chance, whereas Non-Pedophilic Mixed Gender Child Molesters, Preferential Pedophiles, and Non-Aggressive Child Molesters had more Aboriginal offenders than expected. Preferential Pedophiles and Non-Aggressive Incest Offenders had more Caucasian offenders than expected. Non-Pedophilic Mixed Gender Child Molesters and Preferential Pedophiles had more mood/anxiety disorders than expected, whereas Mixed Victim Assaulters and Non-Pedophilic Mixed Gender Child Molesters had more personality disorders than expected. Finally, Non-Aggressive Incest Offenders had fewer personality disorders than expected.  A multivariate logistic regression was not used, as the high number of classes and small size of two of the classes would have made it difficult to derive information using this technique. Therefore, a review of textual offence information was conducted to provide qualitative information and further outline differences and similarities between classes. The review of the     68 cases led to Class 2 being renamed Non-Pedophilic Mixed Gender Offenders from Non-Pedophilic Mixed Gender Child Molesters. 4.3  Comparison of the Derived Typology and Existing Literature  Overall, the derived classes were mainly consistent with prior literature. Child molester typologies began with Groth, Hobson, and Gary’s (1982) classification system using a fixated-regressed dichotomy. According to the typology, fixated offenders tend to be sexually preferential towards children, whereas regressed offenders are externally stressed and the offences tend to be situational. The FBI expanded upon the fixated-regressed typology and developed seven sub-types, with four sub-types within the situational type and three sub-types within the preferential type (Holmes & Holmes, 1996). The differences among the sub-types are based around characteristics such as coping skills, psychosexual development, use of aggression, and self-esteem. Although these exact variables were not considered in the present research, four of the five classes that centered around child-oriented offences seemed to fit well within the existing typology. The Non-Pedophilic Mixed Gender Offenders class had a low-moderate probability of having pedophilic disorder, a high probability of having a substance/alcohol use disorder, and a low-moderate probability of sexual victimization history. This class seemed comprised of two groupings, those possessing and using child pornography and those committing violent sexual assaults.  The characteristics of this class appear to overlap with the characteristics of the Morally Indiscriminate and Sexually Indiscriminate sub-groups of the FBI typology (Holmes & Holmes, 1996), such as a lack of true sexual interest in children and the use of some force. As discussed in the limitations section, the small sample size of the data may have contributed to a lack of differentiation between two sub-groups within this class as would the source of offenders coming from the provincial corrections system. The Preferential Pedophiles class was characterized by the presence of pedophilic disorders. There was a no probability of     69 substance/alcohol use disorders, and a moderate probability of sexual victimization history. The victims were primarily female, and were exclusively extremely young, pre-pubescent children. They were highly likely to be non-romantic relatives, but there was some probability of the victim being a stranger or acquaintances. The characteristics of this class appear to overlap well with the characteristics of the Seduction and Sadistic sub-groups of the FBI typology, such as the primary sexual interest in children and some presence of violence. The Non-Aggressive Incest Offenders class had a moderate probability of pedophilic disorders, with a moderate-high probability of substance/alcohol use disorders and sexual victimization history. The victims were exclusively female, tended to be non-romantic relatives, and were primarily pre-pubescent or pubescent children. The characteristics of this class appear to overlap with some characteristics of the Regressed and Inadequate sub-groups of the FBI typology, although neither seemed to completely capture the class to a high degree. Finally, the Non-Aggressive Non-Pedophilic Child Molesters class had a very low probability of pedophilic disorders, no probability of substance/alcohol use disorders, and a moderate probability of sexual victimization history. The victims were most likely to be female, but there was a low probability of male or mixed gender victims. The victims were exclusively non-adults, with the probability dispersed across the age ranges. The victims were likely to be acquaintances or non-romantic relatives. The characteristics of this class appear to overlap with the characteristics of the Sexually Indiscriminate sub-group of the FBI typology, such as the lack of true sexual interest in children. However, the nature of the cases was not truly captured by this sub-type, as the type of victims selected tended to overlap more with the Seduction sub-group. Although the offences against children tended to distribute in expected classes according to previously developed typologies, it should be noted that the offences against adults aggregated into one class. The Mixed Victim Assaulters class was characterized by a lack of pedophilic disorder, moderate-high probability of substance/alcohol use disorders, and low-moderate     70 probability of sexual victimization history. The offences were likely to be committed against females, primarily adults, with whom the offender had a limited or no relationship with prior to the offence. The offenders were most likely to be in the youngest age category, and there was a low-moderate probability of using aggression during the offence. This was not necessarily as expected. For example, Groth’s (1979) well-known rapist typology asserts that there are four classifications of rapists. Other rapist typologies also assert multiple groupings of rapists, rather than just a single type (Barbaree, Seto, Serin, Amos, & Preston, 1994; Knight, 1999; Knight & Prentky, 1990). Often these typologies depend on more in-depth crime information, as well as motivation and interpersonal information of the offender. These pieces of information were not readily available in the data, which may have contributed to the lack of differentiation among adult-oriented offenders. The reliance on provincial offender data, rather than a combination of federal and provincial offender data, also may have impacted the groupings because there was an inherent selection bias. 4.4  Limitations and Future Research  As with any research, this project has limitations. The archival-based nature of the study design made it difficult for all variables to be coded with the same high level of accuracy. However, the inclusion of a second coder alongside the primary researcher allowed for inter-rater reliability to be established and therefore higher confidence in the coding of the variables for analyses. It would be beneficial for future research to have multiple coders available to code all of the data files to ensure any coding problems are identified and discussed. This can further increase the confidence in the coding system used, and the resulting analyses.   Related to the archival-based nature of the study design was the fact that some pre-sentence reports were based on file-only information, while others were based on file information and offender interviews. The depth of information varied between these types of files, which may have impacted the data that were gathered. This issue can be examined in future     71 research, to see whether the type of information included presents differences in the outcomes, holding other factors constant. These particular reports were written by a single psychiatrist, and future research will examine the impact of potential moderator variables of the risk assessment outcomes. The outcomes will be evaluated specifically for whether cooperation of the offender in a psychiatric assessment resulted in moderation of the final risk rating. Another study limitation present was in the breadth of the data that were collected. Although the data presented a unique opportunity to analyze pre-sentence reports, the information only consisted of reports written on provincial offenders. This limits the generalizability of the typology, and may have excluded additional classes of offenders in the resulting typology. This is due to the fact that the more aggressive and/or serious offences committed are not included in the data set since they are likely to be processed in the federal system. These cases may have provided pertinent information on how the offenders distribute based on the included variables (e.g., use of aggression). Future research should strive to collect data on a wide range of offences, which should include both provincial and federal level offences, so as not to become too focused on low- or high-risk crimes. This could allow for the development of a more generalizable typology across all sexual offences. As is often the case in psychological research, sample size posed a limitation to the project. Although 106 case files were included, only a small portion of the offences included violent and/or aggressive sexual crimes. This made it difficult to determine whether the classes were fully differentiated, or whether more nuanced patterns could be found using a larger set of offender files. Nearly all of the adult offenders presented in a single class, and the lack of a data set with a larger range of crimes may have been related to that finding. The researcher collected information on violent offenders during the data collection efforts, so future research will focus on investigating the differences that are present in terms of demographic and historical characteristics of sexual versus violent offenders. As noted previously, future research should     72 strive to collect data from a wide range offences, both at the provincial and federal level. This may help in determining whether classes are comprised on multiple sub-types, or whether the characteristics are truly linking different offences together (e.g., child pornography use and serious violent sexual assaults). This study also was unable to use certain variables for the chosen analysis, as there was a limited distribution across categories. A larger sample size should assist with this issue by generating a greater range across the different categories of each variable (e.g., use of a weapon). A larger sample size would also provide additional information on demographic characteristics such as ethnicity; these characteristics can provide compelling information on the composition of the classes. A larger data set would allow for the analyses to have more statistical power, which may lead to the discovery of additional relationships and allow researchers to further differentiate classes into meaningful the sub-types. 4.5 Research Contributions and Future Directions   Byrne and Roberts (2007) suggested that three themes need to be considered when developing typologies of offending behavior. This project tried to address these whenever possible given the context in which the research took place. The first theme was design, and Byrnes and Roberts suggested that researchers be clear on what the purpose of the typology is. This project sought to replicate existing research that derived sexual offender typologies in order to evaluate the similarities between distinct samples of sexual offenders (e.g., adolescent males and adult males). This project also aimed to use a more advanced procedure (i.e., latent class analysis) than previous research to derive a typology and to see how it fit within the framework of existing literature. The project aimed to conduct this research because although typologies have been theorized and commonly accepted (e.g., Groth, 1979; Groth et al., 1982), there has been a lack of empirical validation across multiple sample types. In this study, the classes involving child-oriented offences supported existing typologies in the literature, whereas the class involving adult-oriented offences did not clearly support existing typologies. Future     73 research can assess a variety of sample types to assess whether proposed typologies are empirically supported.   The second theme identified by Byrne and Roberts was development. The authors stated that there needs to be recognition of issues such as the reliability and validity of the typology. As addressed above, this project is not without limitations. Sample size, coding of the variables, and the absence of a wide range of data all posed issues in terms of the generalizability of the typology. The limitations of the study are fully recognized; however, this study utilized a rich data source, and also had many strengths. A multi-step process was used to develop a detailed coding scheme for data collection, and a high level of inter-rater reliability was established. Advanced statistical analyses yielded quality information on the probability of class membership, probabilities of item endorsements, and clear profiles for each class. Further, external analyses were used to describe the classes in further details, and identify avenues of future research. The findings were considered in context of existing literature, and have furthered our knowledge of sexual offender typologies.   The third and final theme identified by Byrne and Roberts was implementation: The authors suggest that researchers assess how well the typologies actually classify and/or predict the behaviors they set out. The typology derived within the present research did an excellent job classifying cases. The typology may change with additional information, and with that change the ability to classify the offenders may also change. Further research on the variables included in the study will strengthen the findings of the project and further elucidate differences between sub-types of sexual offenders.   A number of research studies have reported on the outcomes for sexual offender treatment programs, but the findings have been conflicting. In an early review, Wormith and Hanson (1992) collected data from 22 sexual offender treatment programs across Canada and found that there was no distinctive type of program, and the programs varied in factors such as     74 size, service duration, and setting. A more recent meta-analysis conducted by Lösel and Schmucker (2005) reported that treated offenders exhibited less sexual, violent, and general recidivism than controls. During the same time frame, Hanson and Morton-Bourgon (2005) conducted a meta-analysis that indicated many of the commonly targeted variables in sexual offender treatment programs were unrelated to violent or sexual recidivism.    The typological approach used in this study is useful because it provides researchers and clinicians with the ability to classify more accurately the people they are investigating or treating. This ability can allow researchers and clinicians to better target and administer preventive and treatment strategies. For example, in a sample of sexual offenders similar to ours, understanding the relationship between the classes and external variables may assist treatment providers. If the offender group being treated is likely to have an increased presence of Aboriginal offenders, culturally appropriate and relevant treatment options can be researched and provided. Similarly, if the offender group is likely to have an increased probability of personality disorders such as anti-social personality disorder, the treatment provider can include potential screening and treatment options that may be more relevant for such an individual. This would also be the case for mood and anxiety disorders (e.g., additional depression and anxiety screenings, cognitive behavioral therapy). Ideally, adaptations to treatment such as these can lead to a reduction in future sexual offences both prior to offending, as well as after offending. Differing background characteristics, attitudes and beliefs, and clinical and criminogenic needs may play an important role in successful treatment. As described above, understanding the sub-types of sexual offenders may help to target sexual offender treatment more appropriately, and thereby increase its effectiveness. Offender and offence characteristics, such as those used in the present research study, can help differentiate such sub-types. Overall, this research empirically derived a five-class typology of adult male sexual offence and offender characteristics, and articulated plans for future research into sexual offender typologies.     75 Chapter 5 Conclusion  Attaining a comprehensive understanding of those who sexually offend is a necessary step before researchers and clinicians can make informed decisions about treatment, intervention, and supervision of offenders (National Criminal Justice Association, 2014). Although typologies have historically been derived for different types of sexual offenders (e.g., Groth, 1979; Groth et al., 1982), there has been a lack of empirical research designed to determine the nature of over-arching typologies.   This research project sought to replicate existing research findings in order to evaluate the similarities between distinct samples of sexual offenders. A relationship between type of offender and recidivism risk was found for both the adolescent male and adult male sexual offender samples. The adult female and adult male sexual offender groupings were both differentiated by offender age and victim age. The research also used a more advanced statistical procedure than previous research to derive a typology of adult male sexual offence and offender characteristics, and to ascertain how well this typology fit within the framework of existing typologies. There was strong support of previously established child-oriented offender groupings, whereas there was limited support of differentiation in adult-oriented offenders. In conclusion, this study has provided an empirical investigation into over-arching typologies using distinct sexual offenders samples. The research presents a stepping stone on which to base future exploration into deriving empirically supported typologies of sexual offenders using advanced analytical techniques.       76 References Babchishin, K. M., Hanson, K. R., & VanZuylen, H. (2015). Online child pornography offenders are different: A meta-analysis of the characteristics of online and offline offenders against children. Archive of Sexual Behavior, 44, 45-66. doi:10.1007/s10508-014-0270-x Bailey, K. D. (1994). Typologies and taxonomies: An introduction to classification techniques. Thousand Oaks, CA: Sage Publications, Inc. Barbaree, H. E., Seto, M. C., Serin, R. C., Amos, N. L., & Preston, D. L. (1994). Comparison between sexual and nonsexual rapist subtypes. Criminal Justice and Behavior, 21, 95−114. doi:10.1177/0093854894021001007  Barkus, E., Yavorsky, C., & Foster, J. (2006). Understanding and using advanced statistics. Thousand Oaks, CA: Sage Publications, Inc. Bonta, J., Bourgon, G., Jesseman, R., & Yessine, A. K. (2005). Presentence reports in Canada. (User Report 2005-03). Ottawa: Public Safety Canada. Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research In Psychology, 3, 77-101. doi:10.1191/1478088706qp063oa Busina, I. (2014). Latent class analysis of sexual offending – Extracting offender and offense characteristics for investigative interviewing and interrogation process. Retrieved from http://summit.sfu.ca/item/14478 (etd8456) Byrne, J. M., & Roberts, A. R. (2007). New directions in offender typology design, development, and implementation: Can we balance risk, treatment and control? Aggression and Violent Behavior, 12, 483-492. doi:10.1016/j.avb.2007.02.009 Carvalho, J., & Nobre, P. J. (2014). Early maladaptive schemas in convicted sexual offenders: Preliminary findings. International Journal of Law and Psychiatry, 37, 210-216. doi:10.1016/j.ijlp.2013.11.007     77 Chu, C. M., & Thomas, S. M. (2010). Adolescent sexual offenders: The relationship between typology and recidivism. Sexual Abuse: Journal of Research and Treatment, 22, 218-233. doi:10.1177/1079063210369011 Collins, L., & Lanza, S. (2010). Latent class and latent transition analysis: With applications in the social, behavioral, and health sciences. Hoboken, NJ: John Wiley & Sons, Inc.  Cortoni, F., & Hanson, R. K. (2005). A review of the recidivism rates of adult female sexual offenders. Research Report No. R-169. Ottawa: Correctional Service of Canada. Criminal Code, R.S.C. 1985, c. C - 46. Cutajar, M. C., Mullen, P. E., Ogloff, J. R., Thomas, S. D., Wells, D. L., & Spataro, J. (2010). Suicide and fatal drug overdose in child sexual abuse victims: A historical cohort study. Medical Journal of Australia, 192, 184-187. Retrieved from https://www.mja.com.au Denford, C. S., & Chan, Y., E. (2011). Knowledge strategy typologies: Defining dimensions and relationships. Knowledge Management Research & Practice, 9, 102-119. doi:10.1057/kmrp.2011.7 Department of Justice. (2015a). Bill C-46: Records applications post-Mills, A case law review. Retrieved from http://www.justice.gc.ca/eng/rp-pr/csj-sjc/ccs-ajc/rr06_vic2/p3.html Department of Justice. (2015b). Statistics of sexual assault. Retrieved from http://www.justice.gc.ca/eng/rp-pr/csj-sjc/ccs-aj  Deslauriers-Varin, N., & Beauregard, E. (2014). Unravelling crime series patterns amongst serial sex offenders: Duration, frequency, and environmental consistency. Journal of Investigative Psychology and Offender Profiling, 11, 253-275. doi:10.1002/jip.1418 c/rr06_vic2/p3_4.html Ellwood, C. A. (1912). Lombroso’s theory of crime. Journal of Criminal Law and Criminology, 2, 716-723. Retrieved from http://scholarlycommons.law.northwestern.edu/jclc/     78 Field, A., Miles, J., & Field, Z. (2012). Discovering statistics using R. Thousand Oaks, CA: Sage Publications. Fleiss, J. L. (1978). Reply to Klaus Krippendorff’s “Reliability of binary attribute data.” Biometrics, 34, 144. Fraley, C. and Raftery, A.E. (1998). How many clusters? Which clustering methods? Answers via model-based cluster analysis. Computer Journal, 41, 578-588. Gal, M., & Hoge, R. D. (2015). A profile of the adolescent sex offender. Retrieved from http://www.csc-scc.gc.ca/research/forum/e112/e112b-eng.shtml Galbraith, C., & Schendel, D. (1983). An empirical analysis of strategy types. Strategic Management Journal, 4, 153-173. doi:10.1002/smj.4250040206 Goncy, E. A., Sullivan, T. N., Farrell, A. D., Mehari, K. R., & Garthe, R. C. (2016). Identification of patterns of dating aggression and victimization among urban early adolescents and their relations to mental health symptoms. Psychology of Violence. Advance online publication. doi:10.1037/vio0000039 Goodwill, A. M., Stephens, S., Oziel, S., Yapp, J., & Bowes, N. (2012). Multidimensional latent classification of ‘street robbery’ offences. Journal of Investigative Psychology and Offender Profiling, 9, 93-109. doi:10.1002/jip.1351 Gradus, J. L., Qin, P., Lincoln, A. K., Miller, M., Lawler, E., Sørensen, H. T., & Lash, T. L. (2012). Sexual victimization and completed suicide among Danish female adults. Violence Against Women, 18, 552-561. doi:10.1177/1077801212453141 Groth, A. N. (1979). Men who rape: The psychology of the offender. New York, NY: Plenum Press. Groth, A. N., Hobson, W. F., & Gary, T. S. (1982). The child molester: Clinical observations. In J. Conte & D.A. Shore (Eds.), Social work and child sexual abuse (pp. 129-144). New York, NY: Haworth.     79 Guttman, L. A. (1968). A general non-metric technique for finding the smallest co-ordinate space for a configuration of points. Psychometrika, 33, 495-506. doi:10.1007/BF02290164 Haberman, S. J. (1973). The analysis of residuals in cross-classified tables. Biometrics, 29, 205-220. doi.org/10.2307/2529686  Hadzi-Pavlovic, D. (2006). Finding patterns and groupings: I. Introduction to latent class analysis. Acta Neuropsychiatrica, 21, 312-313. doi:10.1111/j.1601-5215.2009.00429.x Hadzi‐Pavlovic, D. (2010). Finding patterns and groupings: II. Introduction to latent profile analysis and finite mixture models. Acta Neuropsychiatrica, 22, 40-42. doi:10.1111/j.1601-5215.2009.00442.x Hagenaars, J. A., & McCutcheon, A. L. (Eds.). (2002). Applied latent class analysis. Cambridge, UK: Cambridge University Press. Hanson, R. K., & Morton-Bourgon, K. E. (2005). The characteristics of persistent sexual offenders: A meta-analysis of recidivism studies. Journal of Consulting and Clinical Psychology, 73, 1154-1163. doi:10.1037/0022-006X.73.6.1154 Hart, S. D., Kropp, P. R., Laws, D. R., Klaver, J., Logan, C., & Watt, K. A. (2003). The Risk for Sexual Violence Protocol (RSVP): Structured professional guidelines for assessing risk of sexual violence. Burnaby, British Columbia, Canada: Mental Health, Law, and Policy Institute, Simon Fraser University. Havlena, W. J., & Holbrook, M. B. (1986). The varieties of consumption experience: Comparing two typologies of emotion in consumer behavior. Journal of Consumer Research, 13, 394-404. doi:10.1086/209078 Havlena, W. J., Holbrook, M. B., & Lehmann, D. R. (2006). Assessing the validity of emotional typologies. Psychology & Marketing, 6, 97-112. doi:10.1002/mar.4220060203     80 Helmus, L., Thornton, D., Hanson, R., & Babchishin, M. (2012). Improving the predictive accuracy of Static-99 and Static-2002 with older sex offenders: Revised age weights. Sexual Abuse: A Journal of Research and Treatment, 24, 64-101. doi:10.1177/1079063211409951 Holmes, R. M., & Holmes, S. T. (1996). Profiling violent crimes: An investigative tool. Thousand Oaks, CA; Sage Publications. Horning, A. M., Salfati, G. C., & Labuschagne, G. N. (2014). South African serial homicide: A victim-focused behavioural typology. Journal of Investigative Psychology and Offending Profiling, 12, 44-68. doi:10.1002/jip.1426 Hubert, L. J., & Arabie, P. (1985). Comparing partitions. Journal of Classification, 2, 193-218. doi:10.1007/BF01908075 Jordan, C. E., Combs, J. L., & Smith, G. T. (2014). An exploration of sexual victimization and academic performance among college women. Trauma, Violence, & Abuse, 15, 191-200. doi:10.1177/1524838014520637 Jung, T., & Wickrama, K. S. (2008). An introduction to latent class growth analysis and growth mixture modeling. Social and Personality Psychology Compass, 2, 302-317. doi:10.1111/j.1751-9004.2007.00054.x Knight, R. A. (1999). Validation of a typology for rapists. In W. L. Marshall & J. Frenken (Eds.), Sex offender research and treatment: State-of-the-art in North America and Europe (pp. 58-75). Beverly Hills, CA: Sage Publications. Knight, R. A., & Prentky, R. A. (1990). Classifying sex offenders: The development and corroboration of taxonomic models. In W. L. Marshall & H. E. Barbaree (Eds.), Handbook of sexual assault: Issues, theories, and treatment of the offenders (pp. 23-52). New York, NY: Plenum Press.     81 Krippendorff, K. (2004). Reliability in Content Analysis: Some common misconceptions and recommendations. Human Communication Research, 30, 411-433.  doi:10.1093/hcr/30.3.411 Kruskal, J. B., & Wish, M. (1978). Multidimensional scaling. Newbury Park, CA: Sage Publications. Lazarsfeld, P. F. (1950). The logical and mathematical foundation of latent structure analysis. In S. Stouffer (Ed.), Measurement and prediction (pp. 413-472). Princeton, NJ: Princeton University Press. Lazarsfeld, P. F. (1959). Latent structure analysis. In S. Koch (Ed.), Psychology: A study of a science: Conceptual and systematic. Vol. 3: Formulations of the person and the social context (pp. 476-543). New York, NY: McGraw Hill.  Lazarsfeld, P. F., & Henry, N. W. (1968). Latent structure analysis. Boston, MA: Houghton Mifflin.  Lösel, F., & Schmucker, M. (2005). The effectiveness of treatment for sexual offenders: A comprehensive meta-analysis. Journal of Experimental Criminology, 1, 117-146. doi:10.1007/s11292-004-6466-7 Lundrigan, S., & Mueller-Johnson, K. (2013). Male stranger rape: A behavioral model of victim-offender interaction. Criminal Justice and Behavior, 40, 763-783. doi:10.1177/0093854812474451 Magidson, J., & Vermunt, J. K. (2005). A nontechnical introduction to latent class models. DMA Research Council Journal. Retrieved from www.statisticalinnovations.com/articles/ lcmodels2.pdf Madigan, S., Wade, M., Tarabulsy, G., Jenkins, J. M., & Shouldice, M. (2014). Association between abuse history and adolescent pregnancy: A meta-analysis. Journal of Adolescent Health, 55, 151-159. doi:10.1016/j.jadohealth.2014.05.002     82 Mallie, A. L., Viljoen, J. L., Mordell, S., Spice, A., & Roesch, R. (2011). Childhood abuse and adolescent sexual re-offending: A meta-analysis. Child & Youth Care Forum, 40, 401-417. doi:10.1007/s10566-010-9136-0 Mandara, J. (2003). The typological approach in child and family psychology: A review of theory, methods, and research. Clinical Child and Family Psychology Review, 6, 129-146. doi:10.1023/A:1023734627624 Mannheim, H. (1960). Pioneers in criminology. Chicago, IL: Quadrangle Books. Meyers, L. S., Gamst, G. C., & Guarino, A. J. (2013). Hierarchical loglinear analysis. In L. S. Meyers, G. C. Gamst, & A. J. (Eds.), Performing data analysis using IBM SPSS (pp. 689-697). Hoboken, NJ: John Wiley & Sons. Moyer, I. (2001). Criminological theories: Traditional and non-traditional voices and themes. Thousand Oaks, CA: Sage Publications, Inc. Muthén, L. K., & Muthén, B. O. (2013). Mplus user’s guide (7th ed.). Los Angeles, CA: Author. National Criminal Justice Association. (2014). Sex offender management assessment and planning initiative. Retrieved from http://www.smart.gov/SOMAPI/ printerFriendlyPDF/complete-doc.pdf Newton, M. (2006). The encyclopedia of serial killers (2nd ed.). New York, NY: Checkmark.  Nicol, A. M., & Pexman, P. M. (2010). Dendrograms. In Displaying your findings: A practical guide for creating figures, posters, and presentations (6th ed.) (pp. 125-127). Washington, DC, US: American Psychological Association. Norman, G. R., & Streiner, D. L. (2003). Cluster analysis. In G. R. Norman & D. L. Streiner (Eds.), PDQ Statistics (pp. 177-187). Hamilton, Canada: BC Decker. Nylund, K. L., Asparouhov, T., & Muthén, B. O. (2007). Deciding on the number of classes in latent class analysis and growth mixture modeling: A Monte Carlo simulation study. Structural equation modeling, 14, 535-569. doi:10.1080/10705510701575396     83 Nunes, K. L., McPhail, I. V., & Babchishin, K. M. (2012). Social anxiety and sexual offending against children: A cumulative meta-analysis. Journal of Sexual Aggression, 18, 284-293. doi:10.1080/13552600.2010.549243 Ogilvie, C. A., Newman, E., Todd, L., & Peck, D. (2014). Attachment & violent offending: A meta-analysis. Aggression and Violent Behavior, 19, 322-339. doi:10.1016/j.avb.2014.04.007 Okamura, K., Kosuge, R., Kihira, M., & Fujita, G. (2014). Typology of driving-under-the-influence (DUI) offenders revisited: Inclusion of DUI-specific attitudes. Addictive Behaviors, 39, 1779-1783. doi:10.1016/j.addbeh.2014.07.007 Pedneault, A., Harris, D., & Knight, R. (2012). Toward a typology of sexual burglary: Latent class findings. Journal of Criminal Justice, 40, 278-284. doi:10.1016/j.jcrimjus.2012.05.004 Raftery, A. E. (1995). Bayesian model selection in social research. Sociological Methodology, 25, 111-163. Richey, J. A., Holm-Denoma, J. M., Kotov, R., Schmidt, N. B., & Joiner Jr., T. E. (2008). Using taxometrics in psychopathology research. In D. McKay (Ed.), Handbook of research methods in abnormal and clinical psychology (pp. 293-308). Thousand Oaks, CA: Sage Publications, Inc. Robertiello, G., & Terry, K. J. (2007). Can we profile sex offenders? A review of sex offender typologies. Aggression and Violent Behavior, 12, 508-518. doi:10.1016/j.avb.2007.02.010 Rosato, N. S., & Baer, J. C. (2012). Latent class analysis: A method for capturing heterogeneity. Social Work Research, 36, 336-339. doi:10.1093/swr/svs006      84 Sadeh, N., & McNiel, D. E. (2013). Facets of anger, childhood sexual victimization, and gender as predictors of suicide attempts by psychiatric patients after hospital discharge. Journal of Abnormal Psychology, 122, 879-890. doi:10.1037/a0032769 Sandler, J. C., & Freeman, N. J. (2007). Typology of female sex offenders: A test of Vandiver and Kercher. Sexual Abuse: Journal of Research and Treatment, 19, 73-89. doi:10.1177/107906320701900201 Seto, M. C., Kingston, D. A., & Stephens, S. (2015). Sexual offending. In B. L. Cutler & P. A. Zapf (Eds.), APA handbook of forensic psychology, Vol. 1: Individual and situational influences in criminal and civil contexts (pp. 351-379). Washington, DC: American Psychological Association. doi:10.1037/14461-011 Sewall, L. A., Krupp, D. B., & Lalumière, M. L. (2013). A test of two typologies of sexual homicide. Sexual Abuse: Journal of Research and Treatment, 25, 82-100. doi:10.1177/1079063212452617 Sharpe, D. (2015). Your chi-square test is statistically significant: Now what? Practical Assessment, Research & Evaluation, 20, 1-10. Retrieved from http://pareonline.net Statistics Canada. (2013). Trends in sexual offences. Retrieved from http://www.statcan.gc.ca/pub/85f0033m/2008019/findings-resultats/trends-tendances-eng.htm Statistics Canada. (2014). Police-reported crime statistics, 2013. Retrieved from http://www.statcan.gc.ca/daily-quotidien/140723/dq140723b-eng.htm Tripodi, S. J., Onifade, E., & Pettus-Davis, C. (2014). Nonfatal suicidal behavior among women prisoners: The predictive roles of childhood victimization, childhood neglect, and childhood positive support. International Journal of Offender Therapy and Comparative Criminology, 58, 394-411. doi:10.1177/0306624X12472879     85 Tripodi, S. J., & Pettus-Davis, C. (2013). Histories of childhood victimization and subsequent mental health problems, substance use, and sexual victimization for a sample of incarcerated women in the US. International Journal of Law and Psychiatry, 36, 30-40. doi:10.1016/j.ijlp.2012.11.005 Vandiver, D. M., & Kercher, G. (2004). Offender and victim characteristics of registered female sexual offenders in Texas: A proposed typology of female sexual offenders. Sexual Abuse: Journal of Research and Treatment, 16, 121-137. doi:10.1177/ 107906320401600203 Vermunt, J. K., & Magidson, J. (2002). Latent class cluster analysis. In A. J. Hagenaars & A. L. McCutcheon (Eds.), Applied latent class analysis (pp. 89-106). New York, NY: Cambridge University Press. Vermunt, J. K., & Magidson, J. (2003). Latent class models for classification. Computational Statistics & Data Analysis, 41, 531-537. doi:10.1016/S0167-9473(02)00179-2 Vrieze, S. I. (2012). Model selection and psychological theory: A discussion of the differences between the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). Psychological Methods, 17, 228-243. doi:10.1037/a0027127 Wolfgang, M. (1961). Pioneers in criminology: Cesare Lombroso. Journal of Criminal Law and Criminology, 52, 361-391. Retrieved from http://scholarlycommons.law.northwestern.edu/jclc/ Wormith, J. S., & Hanson, R. K. (1992). The treatment of sexual offenders in Canada: An update. Canadian Psychology/Psychologie Canadienne, 33, 180-198. doi:10.1037/h0078715 Wortley, R., & Smallbone, S. (2014). A criminal careers typology of child sexual abusers. Sexual Abuse: Journal of Research and Treatment, 26, 569-585. doi:10.1177/ 1079063213503689     86 Wright, R. E. (2000). Survival analysis. In L. G. Grimm & P. R. Yarnold (Eds.), Reading and understanding more multivariate statistics. Washington, DC: American Psychological Association.       87 Appendices Appendix A: Smallest Space Analysis Plot from Lundrigan and Mueller-Johnson (2013)        88 Appendix B: Latent Class Analysis Profiles from Busina (2014) 	 89 Appendix C: HLM Results Plot from Vandiver and Kercher (2004)        90 Appendix D: Coding Guide – Site Visit Section 1 – Background Characteristics  At the beginning of the report, there is a section entitled “Background”. This is where the majority of the information for the variables below will be found.  1. Presence of problematic drug/alcohol use by primary caregiver(s) No Yes [Text] Select Yes only if alcohol/drug use related to neglect, abuse, etc. of the offender before they were 18.  Provide details in provided text box. State whether drugs, alcohol, or both. 2. Presence of non- problematic drug/alcohol use by primary caregiver(s) No Yes [Text] Select Yes if drug/alcohol use described that was non-problematic (see above) before the offender was 18.  State whether drugs, alcohol, or both in provided text box. 3. Marital status Married Co-habiting Non co-habiting Divorced/Separated Widowed Single Select marital status at time of the pre-sentencing report. 4. Number of children 0 1 2-3 4-5 6+ Includes biological, step, adoptive, or foster children. 5. Relationship to children Biological parent Step parent Adoptive parent Foster parent [Text] If more than one, please identify the additional relationships in the provided text box. 6. Self-Reported Ethnicity Aboriginal Asian Black or African American Hispanic or Latino Other Not applicable If other is selected, please provide additional information in the provided text box.  If no information is provided, please select not applicable. 7. Education level Did not complete high school High school diploma Some college      91 2-year degree 4-year degree Graduate studies/degree  Section 2 – Offender Characteristics  Within the report there are sections entitled “Past Medical History,” “Past Psychiatric History,” “Drug and Alcohol History,” “Past Forensic History,” “Mental State Examination,” and “Diagnosis.” This is where the majority of the information for the variables below will be found.  8. Presence of drug arrests No Yes Usually noted in Past Forensic History. Indicate Yes if there is any reference to drug arrest; did not have to result in conviction. 9. Presence of intellectual disability No Yes [Text] Usually noted in Mental State Examination. Does not have to be officially diagnosed. Identify the disability in the provided text box, as well as who provided the information.  10. Previous mental health hospitalizations No Yes [Text] Usually noted in Past Forensic History or Past Psychiatric History. Provide details in the provided text box. Identify who provided the information. 11. Current/past DSM diagnoses No Yes [Text] Usually noted in Past Psychiatric History, Mental State Examination, or Diagnosis. Identify the diagnoses in the provided text box. 12. Presence of psychopathy No Yes No PCL-R Score Usually noted in Past Psychiatric History, Mental State Examination, or Diagnosis. PCL-R score of 30 or above equates to a “Yes” on Psychopathy. 13. Self-reported sexual victimization history No Yes Indicate Yes if there is any reference to sexual victimization of the offender before the age of 18. 14. Self-reported pornography use. No Yes [Text] Indicate Yes if there is any reference to use of pornography. Note in available text box if there were abnormal qualities (violent, child pornography, addiction).  Section 3 – Previous Index/Offence Information  Within the report there are sections entitled “Index Offence” and “Collateral Sources of Information.” This is where the information for the variables below will be found.  15. Victim gender Male Female Multiple [Text] Identify whether victim of index offence was male or female. If there were multiple victims, please give details in the provided text box.     92 16. Victim age 0-5 6-11 12-15 16-18 19+ If there were multiple victims, please select the age of the youngest victim. 17. Victim relationship Stranger Acquaintance Relative [Text] If the victim was not a stranger, indicate the relationship in the provided text box. 18. Use of aggression No Yes Indicate yes if aggression was used at any point during the offence. Examples: verbal abuse, physically struck, or restrained. 19. Use of weapon No Yes Indicate yes if weapon was used at any point during the offence. Items used as weapons should be included (e.g., using a stick to beat someone).  20. Sexual/non-sexual Sexual Non-sexual Whether the index offence was classified as a sexual or non-sexual offence. 21. Type of sexual offence [Text] Indicate the index offence charge.  Section 4 – Static-99R  The Static-99 Revised assessment instrument is provided in each file, so the answers will be coded directly from the instrument. The data in the files looks similar to the below table, with the answer for each question bolded.  22. Age Aged 18 to 34.9 Aged 35 to 39.9 Aged 40 to 59.9 Aged 60 or older 1 0 -1 -3 23. Ever lived with lover for at least two years? Yes No 0 1 24. Index non-sexual violent – any convictions No Yes 0 1 25. Prior non-sexual violence – any convictions No Yes 0 1 26. Prior sex offences Charges None 1-2 3-5 6+ Convictions None 1-2 3-5 6+  0 1 2 3 27. Prior sentencing dates (excluding index) 3 or less 4 or more 0 1 28. Any convictions for non-contact offences No Yes 0 1     93 29. Any unrelated victims No Yes 0 1 30. Any stranger victims No Yes 0 1 31. Any male victims No Yes 0 1 32. Total score/Risk category  -3 through 1 2, 3 4, 5 6+ Low Moderate-low Moderate-high High   Section 5 – Risk for Sexual Violence Protocol  The Risk for Sexual Violence Protocol assessment information is provided in each file. Each section goes over a brief description of the assessment category, and which items are partially/fully endorsed by the individual of interest. The coders will code for those items that are confirmed as endorsed, and select whether they are partially or fully endorsed.  33. Sexual Violence History Chronicity of sexual violence Diversity of sexual violence Escalation of sexual violence Physical coercion in sexual violence Psychological coercion in sexual violence Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full  Did not endorse/Partial/Full 34. Psychological Adjustment Extreme minimization or denial of sexual violence Attitudes that support or condone sexual violent Problems with self-awareness Problems with stress or coping Problems resulting from child abuse Did not endorse/Partial/Full  Did not endorse/Partial/Full  Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full 35. Mental Disorder Sexual deviance Psychopathic personality disorder Major mental illness Substance abuse Violent or suicidal ideation  Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full 36. Social Adjustment Problems with intimate relationships Problems with non-intimate relationships Problems with employment Problems with non-sexual criminality Did not endorse/Partial/Full  Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full     94 37. Manageability Problems with planning Problems with treatment Problems with supervision Did not endorse/Partial/Full Did not endorse/Partial/Full Did not endorse/Partial/Full 38. Risk Scenarios Scenario most likely Scenario 1 Scenario 2 Scenario 3 Scenario 1 is a repeat of the current offence, while Scenario 2 and 3 are scenarios that include escalation in terms of nature, severity, imminence, frequency or duration, and likelihood. 39. Risk Rating Low Low to moderate Moderate Moderate to high High This is located in the section called “Opinion” in the RSVP assessment. 40. Management Suggestions [Text] Provide a brief outline of the suggestions for management. Example: possible to safely supervise in the community.      95 Appendix E: Inter-rater Reliability of Select Coded Files  File Krippendorff’s α File Krippendorff’s α File Krippendorff’s α 2 0.77 20 0.88 41 0.92 3 0.92 21 0.87 43 0.86 4 0.81 22 1.00 44 1.00 5 0.92 23 0.79 45 0.85 6 0.88 24 0.78 46 0.86 7 0.87 25 0.92 47 0.85 8 1.00 26 0.88 49 0.86 9 0.87 27 0.76 50 1.00 10 0.93 29 0.88 51 0.80 11 1.00 30 0.70 53 0.94 12 0.88 31 0.87   13 0.88 33 0.82   14 0.93 34 0.93   15 0.78 35 0.93   16 0.73 36 0.87   17 0.86 37 0.87   18 0.93 38 0.93   19 0.88 39 0.91         96 Appendix F: Demographic Characteristics  Characteristic n % Age        18 – 34.9  47 44.3      35 – 39.9 10 9.4      40 – 59.9 40 37.7      60 or older 9 8.5      Missing 3 2.8 Drug Arrests        No 76 71.7      Yes 27 25.5      Missing 3 2.8 Education level        Did not complete high school 65 61.3      High school diploma 19 17.9      Some college 11 10.4      2-year degree 2 1.9      4-year degree 1 0.9      Missing 8 7.5 Ethnicity        Aboriginal 55 51.9      Caucasian 27 0.9      Other 1 21.7      Missing 23 25.5 Intellectual disability        No 69 65.1      Yes 29 27.4      Missing 8 7.5 Marital Status        Married 20 18.9      Co-habiting 14 13.2      Non co-habiting 5 4.7     97      Divorced/separated 21 19.8      Single 43 40.6      Missing 3 2.8 Non-problematic drug/alcohol use        No 41 38.7      Yes 58 54.7      Missing 7 6.6 Number of children        0 33 31.1      1 18 17.0      2-3 34 32.1      4-5 10 9.4      6+ 8 7.5      Missing 3 2.8 Previous mental health hospitalizations        No 81 76.4      Yes 18 17.0      Missing 7 6.6 Problematic drug/alcohol use        No 89 84.0      Yes 10 9.4      Missing 7 6.6 Pornography exposure        No 26 24.5      Yes 55 51.9      Missing 25 23.6 Psychopathy        No 1 0.9      Yes 2 1.9      No PCL-R Score 103 97.2 Sexual victimization history       98      No 48 45.3      Yes 43 40.6      Missing 15 14.2            99 Appendix G: Replicated HLM Results – K-Way and Higher-Order Effects   Likelihood Ratio Pearson  K df c2 p c2 p K-way and Higher Order Effects 1 2 3 4 5 239 228 183 98 24 371.75 184.32 57.11 3.72 0.14 <.001 1.00 1.00 1.00 1.00 836.00 333.10 60.06 2.42 0.08 <.001 <.001 1.00 1.00 1.00 K-way Effects 1 2 3 4 5 11 45 85 74 24 187.43 127.20 53.40 3.58 0.14 <.001 <.001 1.00 1.00 1.00 502.90 273.04 57.64 2.35 0.08 <.001 <.001 1.00 1.00 1.00  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0305065/manifest

Comment

Related Items