Abstract
Existing evidence suggests that offenders tend not to specialize in sexual offending in general but that there is some specialization in particular types of sexual offending. This study examined the sexual histories and reoffending of a large, national data set of offenders convicted of a sexual offense and managed in England and Wales by the National Offender Management Service (N = 14,804). The study found that specialization in sexual offending compared to nonsexual offending was most evident for offenders with convictions for accessing indecent images. We also found considerable evidence of specialization within sexual offending, most notably for noncontact offenders, especially again indecent images offenders. Crossover between sexual offense types was very rare for those with contact adult offenses and for noncontact offenders although those with child contact offenses sometimes crossed over to indecent images reoffending. If specialization within sexual offending exists, the use of single risk assessment instruments to predict all types of sexual recidivism may be less effective than previously assumed. A comparison of different prediction models indicated that some items presently used in one-size-fits-all risk tools to predict any sexual reoffending only effectively predict certain subtypes of sexual offending. Statistically there appear to be some potential benefits to creating specialist risk predictors for different subtypes of offending, but further work is needed to justify the implementation demands that would be caused by abandoning one-size-fits-all tools.
Introduction
The question of whether offenders specialize in certain types of offending has been one of perennial interest in criminology and correctional psychology. In relation to sexual offenders, this question becomes of considerable importance. Restrictive risk management procedures are typically applied to this group, due not only to recognition of the harmfulness of this type of offending but also the popular assumption that “sexual offenders” are an identifiable group of persistent, specialist, dangerous, pathological offenders whose offending, furthermore, escalates in terms of harmfulness over time (Zimring, 2004). However, if this assumption is incorrect, then the selected risk management procedures may be unnecessary or less than ideal, perhaps failing to prevent offending as they should or even potentially increasing offending risk.
Soothill, Francis, Sanderson, and Ackerley (2000) identified two forms of specialization in relation to sexual offending. Specialization in sexual offending refers to the tendency for an offender to commit only sexual offenses throughout his (or her) criminal career. Specialization within sexual offending refers to the tendency for someone who has committed a sexual offense to reoffend with the same type of sexual offense—for example, against a new victim of the same age and gender as the original victim. As Soothill, Fitzpatrick, and Francis (2009) noted, the second question is not necessarily conditional on the first: it is plausible that some offenders may commit a wide range of offense types but only one sexual offense subtype, while other offenders may be specialist sexual offenders who commit multiple sexual offense subtypes. They also noted that specialization does not need to be seen in absolute terms: offenders may acquire and demonstrate a tendency to specialize in particular offense types or subtypes, but perhaps not to the exclusion of all other offenses.
Another important reason to understand whether sexual offenders are specialists is to inform the development and application of risk assessment tools. A number of actuarial risk assessment tools have been developed for sexual offenders. These tools code factors that research has reliably linked to reoffending outcomes to produce a score that indicates the risk of reconviction for a sexual offense over a specified period of time (Beech, Fisher, & Thornton, 2003). The most recent meta-analytic study of the accuracy of recidivism risk assessments for sexual offenders found that actuarial assessments outperformed unstructured clinical judgment and fared as well or better than structured clinical judgment (Hanson & Morton-Bourgon, 2009). For these reasons, the use of actuarial risk assessment tools has become common practice among professionals who deal with sexual offenders, and the resultant risk ratings are used to guide key decisions on the prioritizations of offenders for treatment and restrictive risk management. However, if sexual offenders specialize in certain types of sexual offending behavior, it is likely that these behaviors are driven by different risk factors, and so applying a risk tool that is scored using a single set of risk factors might not predict each sexual offense type with the same degree of accuracy. It is therefore plausible that the use of multiple risk tools that employ different weights or entirely different risk factors might improve prediction of the different types of sexual offense. It does though remain to be established whether any tendencies to commit particular types of sexual offense are strong enough to justify the development of such specialized actuarial risk assessment instruments, given the complex demands that operating multiple risk assessment tools could place on correctional organizations. Alternatively, a single tool that focuses on risk factors that predict the most harmful forms of sexual reoffending might better help these organizations meet their public protection goals.
Specialization in Sexual Offending
Several methodologies have been proposed to identify the extent of specialization within a particular type of offending, some of which are reviewed in detail by Soothill et al. (2009). These include (a) the “percentage rule” (e.g., Miethe, Olsen, & Mitchell, 2006) where the percentage occurrence of a particular offense type within the offender’s entire criminal history is identified. There is no agreed threshold for what constitutes specialization, with Miethe et al. (2006) using 50%, 75%, and 100% thresholds, while Harris, Smallbone, Dennison, and Knight (2009) did not include offenders with one sexual sanction and no other sanctions (“first-timers” in our text below) in such calculations. (b) The Forward Specialization Coefficient (FSC; Farrington, Snyder, & Finnegan, 1988) where specialization is said to occur when the number of offenses of a certain type significantly exceeds the number expected by chance alone. (c) Transition probability, which identifies the probability of the same type of offending occurring at adjacent time periods (e.g., adjacent arrests or convictions) in an offender’s criminal history. (d) The Diversity Index (Piquero, Mazzerolle, Brame, & Dean, 1999), which identifies the probability that any two offenses drawn from an offender’s record are of separate offense categories.
These methodologies do not represent the only ways of identifying specialization and there are limitations to each method. For instance, Soothill et al. (2009) noted that the FSC has been criticized for not being able to take account of the time lapse between subsequent offenses, as well as its focus on the “principal offense,” which may obscure a pattern of specialization within hidden lesser offenses, and its overdependence on the ordering of convictions. While the Diversity Index overcomes some of these problems, it is in turn limited by its failure to specify what level of diversity would be expected by chance alone. There is therefore scope for exploration of new methodologies for identifying specialization.
Lussier (2005) reviewed the specialization literature in relation to sexual offending and concluded that specialization in sexual offending is less frequently observed than specialization in other types of offending such as violence or property crime. In Lussier’s review, studies using variants of the percentage rule indicated that sexual offending has less specialization than other types of offending. For example, Lussier, LeBlanc, and Proulx (2005) reported that only 4% of rapists met the threshold of 50% of their crime being sexual although the proportion of child molesters reaching this threshold was much higher at 41%. Two studies using the Forward Specialization Coefficient (which varies between 0 and 1, with higher values indicating greater specialization) reported very low FSCs (.09 and .18, respectively) for rapists (Blumstein, Cohen, Das, & Moitra, 1988; Britt, 1996) although the FSC rose to 0.45 when the definition of sexual offending was broadened to any sexual crime (Stander, Farrington, Hill, & Altham, 1989). Studies examining reconviction for a second sexual offense in a given follow-up period (e.g., Langan & Levin, 2002; Sample & Bray, 2003) indicated again very low specialization in sexual offending, with percentages of less than 10% compared to much higher rates for other types of offending such as robbery, violence, and property offending. It can be observed, however, that sexual offenses are much rarer among all offenders than violent or property offenses. An individual offender’s behavior therefore has to differ greatly from all-offender norms to count them as a specialist sexual offender but differ more modestly to count them as a specialist violent or property offender, assuming a constant percentage threshold for specialization.
Two very large scale studies of specialization in sexual offending deserve particular attention. Soothill et al. (2000) conducted a 32-year follow-up of 6,000 offenders convicted of a sexual offense in England or Wales in 1973 and concluded that there was evidence of general versatility, as opposed to specialization, in reoffending. That is, only around one quarter of the sexual offenders in their sample were reconvicted at least once for another sexual offense. A greater proportion (41%) was reconvicted for theft/stolen goods offenses than for violence (20%) and burglary (22%). However, there was a distinctive pattern by sexual offense type in that those originally convicted of indecent assault on a male had a 41% sexual reconviction rate, substantially higher than those convicted of the three other sexual offense subtypes studied (indecent assault on a female, indecency between males, and unlawful sexual intercourse with a girl under 16).
A subsequent U.S. study by Miethe et al. (2006) examined rearrest data on a sample of about 10,000 sexual offenders released from prison in 1994. The probability of repeating the same offense over successive arrest cycles was substantially lower for sexual offenders (26%) than violent or property offenders (33% and 56%, respectively). Taking the percentage concentration approach, only 5% of sexual offenders were pure specialists (i.e., 100% of their arrests were for sexual offenses) and, applying the 50% threshold standard, 23% had sexual offenses for at least half of their offense records (compared with 16% for violent offenders, 37% for property offenders, and 49% for public order offenders). Miethe et al. concluded that their data provided strong evidence that sex offenders are not specialist offenders.
In summary, studies to date provide little reason to believe that sex offenders are a highly specialist group, although there is some indication that certain types of sexual offender (child molesters, those who sexually assault males) are more likely to specialize in sexual offending than others (especially those who sexually assault adult women).
Specialization Within Sexual Offending
Soothill et al. (2000) found evidence of specialization within sexual reoffending. Splitting their sample into groups with convictions for indecent assault on a female, indecent assault on a male, indecency between males and unlawful sexual intercourse with a girl under 16, they observed a “remarkably clear-cut” pattern that if an individual was convicted of another sexual offense, this was most likely to be of the same type of sexual offense as that for which they had previously been convicted (p. 62). Only the sexual offenders who had committed indecent assault on a male, although generally following this pattern, showed some evidence of crossover to other sexual offending types, with 12% being subsequently convicted of indecent assault on a female and 9% of gross indecency with a child. One shortcoming of this study was its focus on the most common sexual offense types in 1973. These included indecency between men, a charge used to prosecute consensual behavior that is no longer illegal.
One new type of sexual offending that has emerged since Soothill et al.’s (2000) sample was originally convicted is sexual offending involving the Internet. Eke, Seto, and Williams (2011) followed 541 registered sexual offenders convicted of offenses relating to indecent images of children, for an average of 4.1 years. The sample demonstrated some specialization in sexual offenses, with 32% being charged for any reoffense and 11% being charged with any sexual reoffense. However, they particularly demonstrated specialization within sexual offending: those with a history of contact or violent offending were four times as likely to have a subsequent conviction for a contact sexual reoffense than those whose only recorded sexual offending related to noncontact offenses. In this sample, the number of prior contact sexual offenses was significantly predictive of contact sexual reoffending, while prior indecent image offenses was negatively predictive of all forms of recidivism other than new indecent image offending.
Although not a study of specialization, Langton et al. (2007) investigated contact sexual, combined nonsexual violent and contact sexual, and noncontact (whether sexual or not) reoffending outcomes, on a sample of 468 offenders followed for 3 years. Of six risk assessment tools investigated, the Static-2002 (Hanson & Thornton, 2003) and the Sex Offender Risk Appraisal Guide (SORAG; Quinsey, Harris, Rice, & Cormier, 2006) predicted serious (violent and/or sexual) recidivism better than the Rapid Risk Assessment for Sex Offense Recidivism (RRASOR; Hanson, 1997). The Minnesota Sex Offender Screening Tool–Revised (MnSOST-R; Epperson et al., 2000), the Violence Risk Appraisal Guide (VRAG; Quinsey et al., 2006), the Static-99 (Hanson & Thornton, 2000), Static-2002 and the SORAG were all superior to the RRASOR in predicting noncontact (nonviolent, nonsexual) offending. No one assessment was superior to the others in predicting sexual recidivism. The differential performance of these tools across outcomes suggests that different risk factors were associated with different types of reoffending, which could be taken as an indicator of specialization.
In summary, the evidence to date indicates that there might be specialization within sexual offending, at least in that contact sexual offenders seem to specialize in contact sexual offending and noncontact offenders specialize in noncontact sexual offending. Furthermore, different risk factors may be associated with contact and noncontact sexual offending. This matter deserves further exploration: if this pattern is replicated in large data sets, there are implications for risk assessment and public protection. That is, risk tools for generic sexual offending may show differential performance depending on whether the outcome is contact or noncontact sexual offending. It is assumed that public protection policy is likely to be more concerned with contact sexual offending, and this could therefore mean that a specialist risk assessment tool for contact sexual offending is more valuable than a generic risk tool for any sexual offending.
Current Study
This study sets out to contribute to the specialization debate and, more specifically, aims to
Explore the criminal histories of a recent large data set of sexual offenders in England and Wales for evidence of specialization both in sexual offending and within sexual offending.
Examine whether evidence of both types of specialization is present in the reoffending patterns of these offenders by determining whether different patterns of nonsexual and sexual reoffending are found for those with different types of sexual offending history. Based on the evidence presented above, we hypothesized that we would find limited evidence of specialization in sexual offending, although based on Soothill et al.’s (2000) findings we expected some variation in specialization in sexual offending for different sexual offense types. However, we expected that specialization will be observed within specific subtypes of sexual offending.
Fit statistical models to predict different types of sexual reoffending, to see if there is likely to be genuine value in developing alternative risk assessment tools that take specialization into account. This study does not, however, develop or validate these tools.
In addition, the study will explore some new methodologies for assessing specialization.
Method
Participants
To identify offenders with any history of sexual offending, the criminal records of offenders assessed using NOMS’ Offender Assessment System (OASys; Home Office, 2006) by March 31, 2008 were retrieved from the Ministry of Justice’s Police National Computer (PNC) database. In all, 222,279 cases were successfully matched while applying rigorous consistency checks on OASys and the PNC’s records of offenders’ index offense conviction or sentence date, date of birth, and gender. Of these matched offenders, 14,804 were retained in the OASys male sex offender sample. These either had
incurred at least one sanction before their follow-up start date (see below), for statutory sexual offending, and/or
an index offense marked as having a sexual element (OASys question 2.2F) or motivation (OASys question 2.9).
Of these 14,804 cases, the mean follow-up length was 37.3 months, with a standard deviation of 20.3 months, and 13,295 (90%) could be followed up for at least 24 months. The mean age of the offenders was 38.1, with a standard deviation of 13.6. Approximately two-fifths (39%) were on license from a custodial sentence. Principal index offenses were statutory sexual for approximately half (48%) of the cases, not statutory sexual but with a sexual element or motivation (“element/motivation”) for 9%, and nonsexual for 42%. This latter group was therefore included in the sample on the grounds of previous convictions for sexual offending.
Measures
The Police National Computer (PNC) research database
The Police National Computer (PNC) is the operational system used by all 42 police forces in England and Wales to record details of suspected and proven offenders as well as details of crimes solved and under investigation. The Ministry of Justice’s PNC research database contains extracts of PNC criminal records data on cautioned and convicted offenders. It is available to researchers through the Ministry of Justice’s Analysis and Statistics group. It is the source of data on previous sanctions and proven reoffending.
Offender Assessment System (OASys)
The Offender Assessment System (OASys; Home Office, 2006) is a structured clinical risk/needs assessment and management tool. It is used to inform court reports on offenders convicted awaiting sentence and manage those serving custodial sentences of at least 12 months (which are usually partially served in the community) or community sentences involving supervision. Community sentences most frequently last 1 or 2 years (Bewley, 2012). Assessments are reviewed periodically over the course of the sentence. In 2010/2011, approximately 860,000 assessments were completed on 360,000 offenders by 18,500 staff. OASys has strongly influenced the design of the offender assessment systems of several other European countries (van Kalmthout & Durnescu, 2008). Data from completed assessments are copied to a research and statistics office within the National Offender Management Service (NOMS) headquarters, to form the OASys research database. Data completeness and integrity checks on this database are undertaken within this office before producing subsets for analysis.
Procedure
Development of offense groups
Four sexual offense type groups were first defined. Contact child offenses involve physical contact offending where the victim is known to be a child (i.e., aged under 16, or aged 16 or 17 in the specific statutes concerning contact between 16/17-year-olds and those aged 18 and above) either from the statutory offense code or from OASys victim information. Contact adult offenses involve physical contact offending where the victim is known to not be a child or, in a small number of cases, of unknown age. (We avoid the term “rapist” to describe those in our sample who have committed contact adult offenses, as English and Welsh statutory offenses include “rape of a girl aged under 16” and similar offenses.) Indecent images offenses involve the making, distribution, showing, and advertisement of indecent images of children. Indecent images usually involve the Internet, but this is not inherent in the offense, and indeed statutorily defined grooming offenses, which typically involve the Internet, are included in the contact child group, as the motivation of this offense is to make sexual contact with a child. The remaining group, paraphilia, includes offenses resulting from a range of sexual interests that are usually most easily gratified through criminal behavior; victims may be of any age. Most paraphilia offenses that result in criminal sanction are prosecuted as indecent exposure (i.e., exhibitionism) and are therefore noncontact offenses, while some of the other offenses in the group are related to voyeurism and zoophilia.
Four aggregated sexual offense groups were also defined. Two groups cover contact offenses, and two almost exclusively cover noncontact offenses; therefore, they are aggregated as contact sexual and noncontact sexual, respectively. Similarly, the two groups that exclusively have child victims are aggregated as child sexual and the other two groups as nonchild sexual.
A distinction was also made between statutory and element/motivation sexual offenses. A statutory sexual offense is one that is defined as sexual on the basis of the legal charge for which the offender has been cautioned or convicted. An element/motivation offense is one that is not a statutory sexual offense but has been identified as sexual on the basis of professional judgment that the offending behavior included a sexual element or was sexually motivated. For example, offenses such as rape, sexual assault, incest, or indecent exposure are all statutory, whereas convictions for theft or actual bodily harm could only be classified as sexual on the basis of clinical observation of an element or motivation. Where an offender’s only known sexual offending history was a current element/motivation offense, they are referred to as an “element/motivation only” offender/case. Element/motivation offenses can only be identified at the index conviction as this information is not available in relation to past offending. This results in underestimates of past sexual offending by some individuals; however, our understanding from clinical experience and user consultation is that such underestimates are also made routinely in mainstream correctional practice due to the limited offending history information typically available.
Among nonsexual offenses, a distinction was made between violent and nonviolent offenses. This uses the broad violent offense categorization developed for NOMS’ OASys Violence Predictor (OVP; Howard & Dixon, 2011), encompassing violence against the person, robbery, aggravated burglary, public order, criminal damage, weapon possession, threats, and harassment offenses.
Defining previous sanctions and proven reoffending
Previous sanctions for an offense group are the number of formal criminal sanctions (convictions, cautions, reprimands, and final warnings) the offender had received for that offense group up to and including the index offense. Where an offender was sanctioned for more than one offense within an offense group on the same occasion, only one sanction was counted. Each previous sexual offending sanction was counted as a sanction for exactly one of the four sexual offense subgroups, by classifying the primary offense, which is flagged on the PNC database (see below) on the basis of sentence severity. Element/motivation offenses were classified using OASys data, as paraphilia if there was no victim contact, as contact child if there was a child victim, and contact adult or child depending on the age of any victim. (Element/motivation offenses were never classified as indecent images; neither OASys data nor clinical practice indicates that indecent image offenses are ever charged as nonsexual offenses.) For simplicity, the Results do not reiterate these counting rules, but all findings should be interpreted accordingly.
The extent of multiple-subgroup sanctions within the sample was examined. In the sample of 14,904 offenders, 931 had a current element/motivation sanction and no other known sanction for sexual offending. Of the remaining 13,973 offenders, there were 1,250 sanctions involving offenses in more than one of the sexual offense subgroups (multiple-subgroup sanctions). There was little indication that some offenders were particularly likely to have multiple-subgroup sanctions: the distribution of 12,771 with no multiple-subgroup sanctions, 1,163 with 1, 32 with 2, 5 with 3, and 2 with 4 compares with an expected distribution of 12,777 / 1,143 / 51 / 1.52 / 0.03 from a Poisson distribution with a parameter of 0.08946 (i.e., 1,250 / 13,973). Of all 18,424 statutory sexual sanctions in the sample’s criminal histories, 17,174 (93.2%) sanctions were for offenses in one of the four statutory sexual offense groups only, 1,194 (6.5%) were for offenses in two groups, and 56 (0.3%) were for offenses in three groups. It was concluded that multiple-subgroup sanctions were an issue to be aware of, but did not pose a strong threat to the validity of the study.
Proven reoffending comprises offenses committed after the date of community sentence or release from custody (the follow-up start date) and by December 3, 2009. Prior to matching with OASys data, the PNC research database had been last updated on December 3, 2010, so a “buffer period” of 12 months after the follow-up period (which, as noted above, varied in length with a mean duration of 37.3 months) allowed conviction to occur and data to be entered onto the PNC. A survival analytic follow-up method was preferred to allow the use of all available reoffending data for each offender. Therefore, days from follow-up start to first reoffending for each of the four groups, four aggregate groups, and any sexual offense were coded as outcome events, and days to the earliest date of imprisonment for any reoffense and December 3, 2009 were coded as censoring events.
A potential problem was noted when considering both previous sanctions and proven reoffending for nonsexual offenses, which were subdivided into violent and nonviolent offenses. It is arguable whether nonsexual offenses sanctions when occurring together with sexual offenses should be counted separately, as they will sometimes have been a part or consequence of the sexual offense (e.g., theft from a rape victim; aggression toward an arresting police officer). Alternative analyses excluding such offenses were calculated, but the effects were too minimal to substantially affect any result. As such, to maintain consistency with the existing literature, all reported results include nonsexual offenses that led to caution/conviction on the same date as sexual offenses.
Coding of risk factors
A range of static risk factors were coded from OASys and PNC data, as shown in Table 4 below. Some were familiar from existing assessment tools, while others were coded from clinical insight or experience modeling other types of reoffending. Offenders who had not offended sexually as adults were included in the sample—although some tools (e.g., Risk Matrix 2000/Sexual offending scale; RM2000/S; Thornton et al., 2003) recommend that no prediction can be made for such offenders, they were sufficiently numerous in our sample to be modeled with some confidence, with their lack of adult offending modeled as a risk factor. Among other risk factors, note that the “stranger sexual victim” and “family child sexual victim” items were limited to data that were recorded in OASys, for the index offense only, as is the case in routine correctional practice. It was also considered that the stranger item would be more meaningful if it were limited to contact offenses. Offenders were coded as having a male victim if they had ever been convicted of a statutory offense of this type (in English law, most contact sexual offenses exist in separate male and female victim versions) or if the current offense was sexual and victim data from OASys indicated a male victim. Years since last sexual sanction is measured from the index conviction date rather than follow-up start date, as the item is intended to reflect the predictive value of offenders without an index sexual offense having passed some time at risk of committing a sexual offense without doing so. Its utility may be compromised both by time in custody after the last sexual sanction and before the index conviction (which is not evident from PNC data, as English and Welsh prisoners are discharged from custody into community supervision with variable proportions of the total sentence remaining) and, as with all criminal history items, by offending that did not result in criminal sanction.
Results
Sexual and nonsexual offending histories
Of the 14,804 offenders, 11,934 had one previous sanction for sexual offenses, while 1,988 had two, 445 had three, and 437 had at least four such sanctions. Sets of offenders sharing particular combinations of sexual offense group counts were created, with all counts capped at three to limit the number of offenders dispersed between many low-n sets. Table 1 presents set frequencies, together with four measures of specialization in sexual offending: the proportions of set members who had any nonsexual nonviolent, any nonsexual violent, and all nonsexual sanctions, and the proportions for whom various percentages of all previous sanctions featured a sexual offense, with first-time offenders separated from repeat offenders whose entire criminal histories featured sexual offenses. This allows the use of several different thresholds for specialization (Miethe et al., 2006) and the choice of whether first-time offenders should be included in specialization calculations at all (Harris, Smallbone, et al., 2009). All sets of offenders of one or two previous sanctions are shown, together with other sets with n of at least 50, with the final row ensuring that all 14,804 offenders are represented.
Nonsexual Criminal Histories, by Combination of Sexual Offense Histories.
Note. A “sexual specialist” offender was one who offended sexually in at least 50% of all their sanction occasions. Offense group counts were capped at three.
Table 1 indicates that the extent of specialization in sexual offending differed between sexual offender types. Overall, about two thirds of offenders had some history of nonviolent offending, two-thirds also had some history of nonsexual violent reoffending, and three-quarters had some history of nonsexual offending. In all, 24% had only ever been sanctioned for sexual offending, 12% had been sanctioned at least as often for sexual as nonsexual offending, and 64% had been sanctioned less often for sexual than nonsexual offending. The major exceptions to this overall pattern were any sets including indecent images offending. The proportion with some history of nonsexual offending was 81% among those with no history of indecent images offending (n = 13,334), but just 31% among those with any such history (n = 1,470). Only 31% of those with no history of indecent images offending were sexual specialists (on a definition where at least 50% of their past and index sanctions included sexual offenses, allowing first-time offenders to be counted as specialists), compared with 85% of those who had a history of indecent image offending. Table 1’s various sets combining indecent image and other sexual offense groups suggest that indecent image offending was dominant in this sense: that is, such offenders were similar to indecent image specialists in that their criminal history included little nonsexual offending, and indeed the majority of indecent image offenders with multiple sanctions for sexual offenses had never been sanctioned for a nonsexual offense. Differences between sets not involving indecent image offending were relatively modest, with contact adult offenders having somewhat more nonsexual violent offending and somewhat less sexual specialization, and the opposite being true of contact child offenders.
Evidence of past specialization within sexual offending was found when studying the 1,988 offenders with two sexual sanctions. Figure 1 illustrates how the combinations of these sanctions would be distributed if there were no specialization and compares this with their actual frequencies. Expected frequencies for each combination under the no-specialization assumption were generated based on the actual frequencies of each sanction; there were 1,270 (32%) contact adult, 1,719 (43%) contact child, 728 (18%) paraphilia, and 259 (7%) indecent images sanctions among the 3,976 total sexual sanctions. For example, each of the 1,988 offenders had an expected probability of ((1,270 / 3,976) * (1,270 / 3,976)) of being in the set with two contact adult sanctions, and an expected probability of (((1,270 / 3,976) * (1,719 / 3,976)) + ((1,270 / 3,976) * (1,719 / 3,976))) of being in the set with one contact adult and one contact child sanction. As the frequencies of the combinations were interdependent, conventional techniques could not be used to test statistical significance. Instead, ranges of expected frequencies were generated by random simulation under the assumption of statistical independence and thus no specialization: the above probabilities of each combination were applied to 1,000 data sets of 1,988 offenders each, and the highest and lowest frequencies of each combination noted. For emphasis, the vertical lines of Figure 1 show the minimum and maximum of these 1,000 expected values for each combination.

Combinations of offense types for those with two sexual sanctions (N = 1,988). Expected frequencies in the absence of specialization in sexual offending, frequency ranges in simulated samples without specialization, and actual frequencies.
For all four offense groups, the actual number of offenders specializing by having two identical offenses was at least 1.5 times the expected number. The extent of specialization was greatest for the noncontact groups, with ratios of 2.7 (179 actual, 67 expected) for paraphilia and 6.4 (54/8.4) for indecent images. Combinations found less frequently than expected included contact adult and paraphilia (19 actual, 83 expected; ratio 0.23), and contact child and paraphilia (133/315; ratio 0.4), whereas indecent images and paraphilia occurred more frequently than expected (61/47; ratio 1.3). Ratios between 0.6 and 0.8 were found for the other three combinations. For 9 of the 10 combinations, the actual value lay outside the range of 1,000 expected frequencies. For one paraphilia and one indecent images sanction, 95% of expected values were no higher than 59, compared with the actual value of 61. This demonstrated that the differences between actual frequencies and those expected under statistical independence were far greater than could have arisen by chance and therefore that some degree of past specialization did occur among those with two sexual sanctions.
Rates of Proven Nonsexual Reoffending and All Reoffending
Table 2 presents rates of proven reoffending for nonsexual violent, nonsexual nonviolent, all nonsexual, and all offenses. The strongest pattern evident involved the far lower rates of all four types of reoffending associated with offenders with a history of indecent images offenses. Among those with one sexual sanction, offenders with an indecent images sanction had an overall reoffending rate below 10%, whereas the other three sets had overall reoffending rates above 40%. Among those with two sexual sanctions, the four sets with indecent images sanctions had rates ranging from 7% to 20%, while other sets ranged from 32% to 50%.
History of Each Sexual Offense and Proven Reoffending Rates for Nonsexual and All Offenses.
A secondary pattern suggested lower rates of nonsexual reoffending among those with contact child offenses, especially those with clear evidence of such specialization within sexual offending: the numerous two contact child sanction only set (n = 554) had easily the lowest nonsexual violent reoffending rate (14.3%) of the six sets without indecent images sanctions, while all rates for the three contact child sanction set were well below the mean rates for all offenders with three or more sexual sanctions. Nonsexual reoffending rates were similar between contact adult and paraphilia offenders, while those of indecent images offenders were much lower. Across all offender sets, there were also very limited differences between the patterns of nonsexual nonviolent and nonsexual violent reoffending.
Rates of Proven Sexual Reoffending
Table 3 presents rates of reoffending in the four sexual offense groups, four aggregate groups, and overall sexual offending, again by sexual offense history combination. Comparison of the four sets of offenders with a single sexual sanction shows the extent of specialization within sexual offending: in each set, the most frequent reoffense type was the previous offense type, and for three of the four sets (i.e., all sets except contact offenses against a child) it was more frequent than the other three combined. Overall sexual reoffending rates were about twice as high among those whose previous offense was noncontact, but this difference was wholly due to noncontact reoffending, with contact reoffending rates being higher among the contact adult and contact child offenders.
History of Each Sexual Offense and Proven Reoffending Rates for Four Sexual Offense Groups and Five Aggregate Sexual Offense Groups.
Note. CA = contact adult; CC = contact child; PP = paraphilia; II = indecent image.
Among offenders with two previous sanctions, specialization among those with exclusively noncontact histories (i.e., paraphilia and/or indecent images offending) was apparent, with high levels of noncontact reoffending and little contact reoffending. In Risk Matrix 2000/s, ever having committed noncontact sexual offenses is an aggravating factor for all sexual reoffending: this appears to be true in the sense that the overall level of sexual reoffending is high, but it is restricted to noncontact offenses. Among those with exclusively noncontact histories, whether one or two previous sanctions, reoffending patterns also show that offenders seldom crossed over to the other noncontact group. This pattern of considerable specialization within sexual offending is complicated, as noted above, by a set of offenders with offending histories and high reoffending rates in both noncontact groups (n = 61, with one sanction in each group).
Those with only contact adult histories also showed considerable specialization, while those with only contact child histories sometimes crossed over to indecent image reoffending.
Combining the two and three previous contact child sanction only sets (n = 712), rates were 2.0% for contact child and 2.1% for indecent images reoffending. Offenders with histories of both types of contact reoffending reoffended in both of those groups, with very low crossover to noncontact offending. Among the four mixed contact and noncontact sets, all of which included small numbers of offenders, no clear reoffending pattern was apparent among the two sets with contact adult history and the reoffending of the two contact child history sets was dominated by noncontact offenses. Strong specialization was apparent among those with three previous contact adult or paraphilia sanctions. While, as noted above, few offenders had more than three sexual sanctions of any sort, some offenders had a very large number of paraphilia sanctions. For example, while only 243 offenders (1.6% of the whole sample) had five or more total sexual sanctions, 77 (3.6% of those with any paraphilia sanctions) had five or more sanctions for paraphilia alone. This is indicative of frequent repetition of this offense, and therefore extremely high reoffending rates, among a small subset of offenders.
Examination of the data underlying Table 3 also reveals that the 36% of the sample with some history of contact adult offending provided 66% of contact adult reoffenders, while the respective pairs of proportions for the other three sexual offense groups were 49% and 67% for contact child, 15% and 72% for paraphilia, and 10% and 53% for indecent images. Those with a history of any contact offending comprised 80% of the sample and 88% of the contact reoffenders (an odds ratio of 1.79), whereas the respective figures for noncontact offending and reoffending were 25% and 70% (OR 7.46). There was less difference in specialization levels between victim types: the respective figures for child victim offending were 58% and 84% (OR 3.99), and for nonchild victim offending were 50% and 82% (OR 4.70).
Predictive Models for Proven Sexual Reoffending
Tables 4 and 5 display Cox regression models predicting each of the nine sexual offense outcomes (four simple sexual offense types and five aggregated sexual offense types). All risk factors were retained in each model, as variable selection methods are unstable in the presence of correlated variables (Steyerberg, 2009) and make comparison of model parameters difficult (i.e., because each model is likely to include a different set of variables). Predictive validity was assessed using the Concordance Index (C Index; Harrell, Lee, & Mark, 1996). This measures whether a reoffender has a higher score or risk category than a nonreoffender, provided that the nonreoffender was at risk for at least as long as the reoffender. This qualification is important when studying survival analytic data sets, for which Area Under Curve (AUC) measurements would include comparisons between nonreoffenders at risk for shorter periods than reoffenders where no valid predictive validity inferences can be drawn. As with AUCs, a C Index of 0.5 indicates that the predictor is no better than a chance method such as tossing a coin; a C Index of 1 indicates perfect prediction. While no guidelines exist for what might be considered an acceptable C Index in criminal risk prediction, the similarity of this metric to AUC suggests that the advice of Rice and Harris (2005) can be copied over. That is, a C Index above .639 represents a medium effect size, and a C Index above .714 represents a large effect size. We describe models with these C Indices as predicting acceptably and well, respectively, and models with C Indices above .8 as predicting very well.
Cox Regression Models of Four Simple Sexual Reoffending Groups.
Note. N = 14,804. Model n (%) reoffending and change in −2 log likelihood from null model (14 df) by outcome: contact adult, 170 (1.2%) reoffending, 190.9 change; contact child, 111 (0.8%) reoffending, 53.5 change; paraphilia, 164 (1.1%) reoffending, 373.7 change; indecent images, 125 (0.8%) reoffending, 188.5 change.
Cox Regression Models of Five Aggregate Sexual Reoffending Groups.
Note. N = 14,804. Model n (%) reoffending and change in −2 log likelihood from null model (14 df) by outcome: contact, 269 (1.8%) reoffending, 165.7 change; noncontact, 275 (1.9%) reoffending, 400.1 change; child, 221 (1.5%) reoffending, 160.2 change; nonchild, 329 (2.2%) reoffending, 411.5 change; all, 520 (3.5%) reoffending, 375.3 change.
Examining first the models for the four simple offense groups, in Table 4, contact adult and paraphilia offenses appeared to be more associated with general criminality than contact child and indecent image offenses, as the first-time entrant item was strongly associated with the former but not the latter. (First-time entrants are those with no sanctions prior to the index offense. As well as the nonsignificant nonsexual sanction counts shown in Tables 4 and 5, alternative models Trialled separate nonsexual violent and nonsexual nonviolent terms with similar lack of success.) Older age indicated lower recidivism probability for all four outcomes though this was not significant for indecent images offenses. Risk may persist more over time for the two contact groups: the juvenile offending terms were less pronounced for these groups, while the “years since last sexual offense” term indicated that desistance effects were stronger for noncontact than contact offenses. Past sanctions for each offense group tended to be significantly associated with reoffending in that offense group only, with the exception that past contact child sanctions were significantly associated with indecent images reoffending. Having ever had a male sexual offense victim was predictive of the two noncontact reoffence types but not the two contact reoffence types. The stranger victim item, restricted as stated above to contact index offenses, was associated with raised risk of contact adult reoffending. (In RM2000/s, both of these items are aggravating factors for all sexual reoffending.) Having a current child victim who was a family member was not significantly predictive of any type of reoffending. As the associations of the male and stranger victim items with the reoffense types appeared to vary considerably, these regression coefficients were formally compared, with the model with the highest coefficient acting as the reference group. The male victim coefficient in the paraphilia model was significantly higher than the equivalent coefficient in the contact child model (z = 2.19, p = .03) but not in the contact adult model (z = 1.19, p = .23) or indecent images model (z = 0.17, p = .87). The stranger victim coefficient in the contact adult model was significantly higher than the equivalent coefficients in neither the contact child (z = 1.06, p = .29), paraphilia (z = .82, p = .41) nor indecent images (z = 1.54, p = .12) models.
Table 5 fits similar models for the aggregate and overall sexual offense groups. Comparing contact and noncontact models, both outcomes were less likely for first-time and older offenders. Juvenile-only offenders were less likely to commit offenses in any group, but especially noncontact offenses. Noncontact reoffenses—but not contact reoffenses—were also less likely for those with a long period since their last sexual offense and more likely for those with their first sexual offenses at an older age. Past contact sanctions were predictive of contact reoffending; contact adult sanctions were more predictive than contact child sanctions, perhaps because contact adult reoffending was more frequent. Paraphilia sanctions were also somewhat predictive of contact reoffending. Both types of noncontact sanction were highly predictive of noncontact reoffending. Offending against male victims was predictive of noncontact reoffending, and contact offending against stranger victims was predictive of contact reoffending. Table 5 also shows that in decreasing order of parameter estimate size, indecent images, contact child, and paraphilia sanctions were significantly predictive of reoffending with child victims, while paraphilia and contact adult sanctions were significantly predictive of reoffending with nonchild victims. The male and stranger victim items were also predictive of nonchild but not child reoffending.
Given that there were both similarities and differences between the four simple offense group models, their overall similarity was tested formally by comparing the four models with the overall sexual offense model using an omnibus chi-square test (Singer & Willett, 2003). Using the four models rather than the overall model reduced −2 log likelihood by a highly significant amount, χ2(42) = 431.3, p < .001, indicating that the use of separate models did improve prediction across the four offense types.
Table 6 presents Concordance Indices summarizing the validity of each model in predicting reoffending in those offense groups and subgroups overlapping with the model’s predicted outcome. Contact adult reoffending was predicted well by models of either contact adult or all contact reoffending, slightly less well by a model of nonchild reoffending and only acceptably by a model of all reoffending. Contact child reoffending was predicted less well (i.e., with a lower C Index) generally; it was best predicted by a contact child model, followed by a child model, with models of contact and all reoffending failing to provide acceptable predictive validity. Paraphilia was predicted very well by all four relevant models. Indecent image offending was predicted very well by indecent image and child models, and well by a noncontact model, but only acceptably by a model of all sexual reoffending.
Concordance Indices for Nine Predictive Sexual Reoffending Models, Applied to Similar Sexual Reoffending Outcomes.
Note. The models tested were those listed in Tables 4 and 5. For example, Table 4’s model of contact adult sexual reoffending predicted contact adult reoffending with a Concordance Index of 0.765 and predicted all contact reoffending with a Concordance Index of 0.688. Concordance Indices were not estimated when the modeled outcome and reoffense were mutually exclusive.
Each of the four aggregate groups was predicted best by a model designed to predict that specific outcome. Contact and noncontact reoffending both suffered losses of several points (hundredths) of C Index if predicted by a model of any sexual reoffending. The prediction of offenses against children was most badly compromised by the use of a model of any sexual reoffending, with a loss of 10 points of C Index. The outcome of all sexual reoffending was predicted well only by a model of all sexual reoffending, though losses of predictive validity were least when models of paraphilia, all noncontact, or all nonchild reoffending were used.
Among outcomes not represented in the table, it may be noted that a model of contact sexual reoffending would have C Indices of 0.688 for paraphilia and 0.449 for indecent images reoffending. However, very basic models have some adequacy for these outcomes: the count of previous paraphilia sanctions has a C Index of 0.803 for paraphilia reoffending, and the equivalent model for indecent images has a C Index of 0.715.
Discussion
This study had three main aims. First, we aimed to explore the extent to which sexual offenders specialize in sexual offending (as compared to other types of offending) and within sexual offending (i.e., is there specialization within different subtypes of sexual offending?). The previous offending data for the sample suggest that specialization in sexual offending is relatively limited, with only one quarter of the sample having no other type of conviction. However the main notable evidence of specialization related to indecent image offenders. Of this group, 85% could be defined as sexual specialists using the threshold that 50% or more of their offending was sexual, and 69% had only sexual offending in their histories of whom most were first-time offenders. Looking at the reoffending data for the sample, the same pattern was observed. While there were lower rates of nonsexual reoffending among contact child offenders (especially those who were classed as sexual specialists), indecent images offenders had far lower rates of nonsexual offending than other groups.
Supporting previous findings that rapists are more versatile in their offending than child molesters (e.g., Harris, Mazerolle, & Knight, 2009; Harris, Smallbone, et al., 2009), we found that among those who had committed contact sexual offenses against adults, having a history of nonsexual violent offending was more frequent than being classed as a specialist sexual offender (even on the most inclusive definition, classing as specialists all those for whom at least 50% of previous proven offending was sexual), while the reverse was true for contact child sexual offenders. In addition, higher rates of nonsexual reoffending were observed in the contact adult group than in the contact child group, corroborating the results of Harris, Knight, Smallbone, and Dennison (2011). This supports the notion that a general antisocial orientation is more likely to explain sexual offending among rapists (or, in our classification, contact adult offenders) than those who offend sexually against children (Hanson, 2002).
The second aim was to explore the extent to which sexual offenders specialize within sexual offending. Examination of past offending data, as shown in Table 1, indicates this to be very much the case. This form of specialization—committing the same kind of sexual offense as was previously committed—was observed for all types of sexual offender. Most notably, noncontact offenders were very often specialist within sexual offending, while contact offenders had a more moderate tendency to be specialist within sexual offending. The noncontact offenders also tended to specialize in the type of noncontact offending for which they were previously convicted; that is, indecent image offenders, if reconvicted, tended to be convicted for further indecent image offenses, while most of the paraphilia offenders’ sexual reoffenses were paraphilic. The former finding is in line with the results of a recent meta-analysis of online offenders’ contact offending, which led to the conclusion that there is a distinct group of Internet offenders whose only sexual offending involves indecent images of children, and that online offenders rarely go on to have proven contact sexual reoffenses (Seto, Hanson, & Babchishin, 2011). The only real exception to the pattern of specialization within sexual offending was that those whose sexual history was all contact child offending (but not mixed-contact offenders) could cross over to indecent images; however, this crossover did not happen in the opposite direction.
It is possible that specialization in sexual offending may partly appear greater among Internet offenders because such criminal acts are always very likely to be prosecuted as statutory sexual offenses if they are prosecuted at all, given that indecent image offending typically involves little incidental property or nonsexual violent criminal activity. In contrast, it seems likely from the large number of element/motivation offenses that many noncurrent past contact offenses and also many contact reoffenses will have been prosecuted as nonsexual offenses, resulting in an underestimation of the extent of sexual specialization by contact sexual offenders. This weakness in our study could only be overcome with detailed examination of full criminal records, which would only be feasible in a much smaller sample. A related weakness—the use of official criminal records—is discussed below.
The third aim was to investigate whether one-size-fits-all risk assessment may not predict all types of sexual offending equally well, and to see if other models could outperform a general risk assessment model in relation to different sex offender subtypes. These analyses created some surprising results. The models in Tables 4 and 5 replicate some well-known facts about the prediction of sexual recidivism, such as showing a simple reduction in risk as age rises. But they also reveal some patterns in relation to different types of sexual offending, such as the “ever had male victim” item predicting noncontact offending only, and the “stranger victim” item predicting contact offending only. At present, these items are used in some general risk tools (e.g., Risk Matrix 2000) to predict any sexual recidivism. It should be noted that the models presented in Tables 4 and 5 are not intended as final models. In a study currently underway, we take a chosen (all contact) outcome, and additionally model it with a mixture of static and dynamic risk factors.
Table 6 in particular raises some challenging questions for those involved in assessing the risk of known sexual offenders. It appears that a one-size-fits-all risk assessment scheme only shows acceptable performance in relation to certain sexual offense types. The model of all sexual reoffending predicted contact adult offending acceptably but not contact child offending, and for each of the four main aggregate offending subtypes (adult contact, child contact, all noncontact, and noncontact) a specialist model was more effective than the one-size-fits-all model. However, the increases in validity provided by a specialist model, which were not substantial in all cases, must be weighted against the additional administrative complexity of implementing a greater number of different schemes. That is, as the number of tools increases, so does the likelihood of administrator error, which may derail risk assessment accuracy just as much.
An obvious limitation of this study is that the reoffending observations examined related to a follow-up period of less than 5 years. It can take time for sexual offenses to come to the attention of the authorities and more time for cases to be tried. A recent study of reoffending hazards among English and Welsh offenders (Howard, 2011) focuses on offense rather than conviction dates and shows that sexual offending risk is highest in the first year after release/sentencing. The decline in reoffending hazards over time is slower than for some nonsexual offense types and, as many sexual reoffenses will occur beyond the mean 3-year follow-up period of the present study and the 4-year follow-up period of the hazards study, it remains possible that longer follow-up periods may yield different patterns. However, it is not clear from the existing literature what differences might be hypothesized to exist between shorter and longer term reoffending patterns or risk factors. A desirable future research goal would be to extend the follow-up of this study’s sample beyond 5 years, which would provide greater statistical power and checking whether reoffending patterns persist in the longer term. While the differentiations in sexual reoffending rates demonstrated in Table 3 are probably ample to demonstrate that some sexual offenders pose greater public protection risk than others—to an extent that is practically as well as statistically significant—any additional information provided by longer follow-up periods can only be of public benefit.
Another caveat concerns the use of official criminal records as the sole data source. This may lead to overestimation of specialization within sexual offending. Several self-report studies (Abel, Becker, Cunningham-Rathner, Mittelmann, & Rouleau, 1988; Ahlmeyer, Heil, McKee, & English, 2000; Heil, Ahlmeyer, & Simons, 2003) have found greater crossover than the current study between sexual offense subtypes, and in particular Seto et al. (2011) found that almost one half of indecent image offenders self-reported a contact child offense despite—similarly to the current study—below one-tenth having an official criminal record of such offending. It may be that offenders who have committed both contact child and indecent image offenses, but only been convicted of indecent image offenses, tend to commit indecent image offenses more frequently than they commit contact child offenses, influencing their probabilities of conviction for each offense type. If this is so, then the specialization evidence that is observable from criminal records may, at least to some extent, tell us about which types of offense are committed more often than others, as opposed to whether each offense type is committed at all. Future self-report studies might investigate the frequency of each type of offending as well as providing opportunities to validate or disvalidate this study’s results. The fullness of these opportunities will be dependent on the combination of the reported rate and frequency of each type of sexual offending and the number of participants. Self-report studies will measure higher offending rates and frequencies than studies that use official records only, but official records can be collected for large numbers of participants, including entire national offender cohorts.
Despite these limitations identified above, the findings do provide support for the notion that noncontact and child contact offenders specialize in sexual offending and provide evidence of specialization within sexual offending. That is, particularly for noncontact offenders, if there is another sexual reoffense this is likely to be for a similar type of sexual offense for which they had previously been convicted. Further examination of those factors that predict different types of sexual recidivism also suggest that specialist models are superior to “all sex offending” models. We suggest that this finding warrants further examination in future research, in two respects. First, while our results (Table 6) indicate that models specific to certain sexual reoffending outcomes improved prediction of those outcomes compared with one-size-fits-all models, we applied the outcome-specific models to the cases on which they were developed. Any attempt to formally develop outcome-specific risk predictors must involve separate validation samples or use other procedures (e.g., bootstrapping; Steyerberg, 2009) to estimate shrinkage in predictive performance, which typically occurs when predictive models are applied to new cases. Second, the two sets of comparisons we made between the model coefficients for specific risk factors (male and stranger victims) across the four separate reoffending outcomes mostly failed to demonstrate significant differences. We did not conduct a more extensive range of comparisons due to “data fishing” concerns, and any future study that focuses on the predictive effects of specific risk factors across different outcomes should specify relevant hypotheses in advance to avoid similar multiple comparison problems. Given the large standard errors of many model coefficient estimates, as shown in Tables 4 and 5, it is evident that increased sample sizes and/or longer follow-up periods (generating more reoffenders and thus reducing standard errors) will be required to confirm the statistical significance of moderate effect size differentials.
Specialization in and within sexual offending is a sizeable research topic, with many issues remaining unresolved. As Harris et al. (2011) identified, the timing of specialization in sexual offending is unclear: Do offenders not classed as specialist sexual offenders commit their sexual offenses earlier or later in their overall criminal careers, if any such tendency exists? A related question can be framed in terms of persistence and desistance: Do most individuals desist from sexual offending before or after they desist from nonsexual offending? Studies with long follow-up periods, when there can be relative confidence that total desistance has occurred for some participants, are well placed to answer this question. Finally, this study’s predictive modeling has focused on sexual reoffending by known sexual offenders. A fuller approach would also examine nonsexual violent and nonsexual nonviolent reoffending by these offenders and sexual reoffending by those with no known history of sexual offending. From the perspective of a correctional organization, a full accounting of public protection risks requires understanding of when nonsexual offenders crossover to sexual offending as well as all the potential future harms attributable to known sexual offenders.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
