APPENDIX 7.1
FUNDAMENTAL CONCEPTS IN SEX OFFENDER RISK ASSESSMENT RESEARCH
OPERATIONALIZING RECIDIVISM
Most risk assessment studies have defined recidivism as new arrests, charges, or convictions for a criminal offense and use official records to collect information about these outcomes. As discussed by Quinsey et al. (2006) and others, official records are an underestimate of the incidence of crime. Many victims do not report sexual crimes; some offenders are not apprehended; and some guilty individuals are not successfully prosecuted. Approximately one third (30%) of sexual assaults committed against minors are reported to police (Finkelhor & Ormrod, 1999; Snyder, 2000), and data from three states with relatively complete data collection suggest that less than one quarter (22%) of reports made to child welfare agencies are also known to the police (Finkelhor & Ormrod, 2001).
These sources of “noise” may obscure the identification of risk factors for recidivism to the extent that they are related to the risk factors. For example, a psychopath who has no ties to others, is highly mobile, and is deceitful may be more likely to escape detection for a new offense than someone who has many family members and relatives and has lived in the same community for many years. Psychopaths are at higher risk for recidivism than nonpsychopaths, all other things being equal. It is assumed that recidivists are more likely to be detected to the extent that they commit more serious offenses or commit offenses at a greater frequency. Although many of these offenses do not appear in official records, the logic is that the men who are the most likely to reoffend will eventually be detected (i.e., the probability of being successfully detected and convicted for a particular crime might be only 10%, but someone who commits many such crimes is highly likely to be detected and convicted for one of these crimes). Snyder (2000) found that the likelihood that a child or adolescent victim report of a sexual offense would result in an arrest was positively associated with the number of victims involved, number of offenders involved, older victims (only 19% of alleged offenses involving children under the age of 6 resulted in an arrest compared with 33% of children ages 6 to 11 or 32% of minors older than 11), involvement of a girl victim, a known offender, and offenses occurring in a residence and was negatively associated with victim injury; these correlates are listed in order of importance. It is not clear whether these correlates would increase (number of victims) or decrease (younger victim age and boy victims) the proportion of arrested sex offenders who are pedophilic.
Researchers have differed in the recidivism outcome they have focused on in their follow-up studies. Some have focused on sexual recidivism, usually defined as a sexually motivated crime involving physical contact with a victim (e.g., criminal charges of child molestation, rape, or sexual assault). In my research, I have specifically examined this type of recidivism but have also focused on violent recidivism
, defined as a new nonsexually violent offense or a new sexual offense involving physical contact with a victim. Sexual recidivism is a subset of violent recidivism in this operationalization. I and many of my colleagues have studied both recidivism outcomes because both are relevant to public safety. Moreover, some apparently nonsexually violent offenses identified in criminal records are actually sexually motivated (e.g., attempted sexual assaults that result in convictions for assault or forcible confinement and sexually motivated murders that are recorded only as homicides in official records; Rice, Harris, Lang, & Cormier, 2004). Specifically, sexual recidivism is a conservative criterion; all of the reoffenses are deemed to be sexually motivated, but some sexually motivated offenses will be missed by focusing on this criterion alone. I still discuss sexual recidivism because of the legal mandate of sex offender civil commitment laws in the United States. The distinction between violent recidivism, which includes all sexual offenses as well as nonsexually violent offenses, and sexual recidivism, which includes only sexual offenses but misses some sexually motivated crimes, should be kept in mind in reading this chapter and the rest of the book.
SHORT-TERM VERSUS LONG-TERM RISK
Some writers have suggested that the predictors of long-term risk to reoffend may differ from those that predict recidivism in the near future (e.g., Doren, 2002; Litwack, 2001). G. T. Harris, Rice, and Cormier (2002) examined this question and found that Violence Risk Appraisal Guide (VRAG; G. T. Harris, Rice, & Quinsey, 1993) scores were still significantly predictive of violent recidivism among forensic patients in the short term, such as 6 months or 1 year. Also, cross-validation studies have typically followed sex offenders for shorter time periods than those obtained by the original scale developers, yet the studies still find that the actuarial risk scales are significantly predictive of recidivism (e.g., Barbaree et al., 2001).
SPEED AND SEVERITY OF RECIDIVISM
In addition to whether a sex offender with child victims is likely to reoffend during a specified time period, some assessors may be interested in how quickly any new offenses will occur and how severe any new offenses will be. Quinsey et al. (2006) reported that VRAG scores were negatively correlated (r
= −.34) with latency to new offenses in a sample of 343 sex offenders. In other words, sex offenders with higher VRAG scores were more likely to violently reoffend than those with lower VRAG scores, and among those who reoffended, sex offenders with higher VRAG scores were quicker to reoffend than those with lower VRAG scores.
Quinsey et al. (2006) reported that both VRAG and SORAG scores are positively correlated (r
= .35) with the severity of any new offenses among sex offenders. In other words, sex offenders with higher scores are more likely to reoffend, and those who do reoffend cause more victim injury than those with lower scores. Similar analyses have not been reported for the Rapid Risk Assessment of Sexual Offense Recidivism (Hanson, 1997) or the Static-99 (Hanson & Thornton, 2000).
TIME AT RISK
Knowing the base rate of recidivism is not enough when comparing the results of different follow-up studies. Longer follow-up periods are associated with higher base rates of recidivism. Thus, one study may report a sexual recidivism rate of 7%, and another study may report a sexual recidivism rate of 47%, but this is not very informative until one also knows that the first study followed sex offenders for 2 years and the second study followed them for 25 years. Given the relatively low rate of sexual reoffending observed after 4 to 5 years of follow-up (an average of 13% in the meta-analyses by Hanson & Bussière, 1998, and Hanson & Morton-Bourgon, 2004), relatively long follow-up periods are required to detect statistically significant associations between putative risk factors and recidivism and to detect statistically significant differences between groups, such as offenders who have participated in treatment and those who have not (see Barbaree, 1997). G. T. Harris and Rice (2003) have demonstrated that clearer prediction results are obtained when the offenders in a sample are followed for the same fixed follow-up time.
PREDICTIVE ACCURACY STATISTICS
Receiver operating characteristic (ROC) analysis is often used to examine the predictive accuracy of binary decisions such as “release” versus “do not release” (Swets, 1988). The primary statistic of interest is the area under the ROC curve, which is bounded by 0 and 1. An area under the curve (AUC) value of .5 indicates prediction at the chance level, and an AUC value of 1 indicates perfect prediction in the expected direction. In the prediction of recidivism, the AUC can be interpreted as the probability that a randomly selected recidivist will have a higher score on a risk variable than a randomly selected nonrecidivist.
The AUC is an attractive measure of predictive accuracy because it is less sensitive to base rate of recidivism and selection ratio (the proportion of offenders who are predicted to reoffend) than a correlation coefficient, percentage of offenders correctly classified, or the values for sensitivity and specificity at a particular cut score. These properties have led to a call for its use as a standard measure of diagnostic and predictive accuracy (Rice & Harris, 1995; Swets, Dawes, & Monahan, 2000).
The AUC has been criticized as a global index of accuracy because it is possible to get a high AUC value despite ordinary performance at particular cut scores, for example, at a high specificity for recommending civil commitment of a sex offender in the United States or at a high sensitivity when considering whether a sex offender should be released on parole. Each jurisdiction’s population of offenders has a particular base rate of recidivism, and thus each jurisdiction is expected to have its own optimal cutoff score. The optimal cutoff score will also vary according to the relative costs of false positive errors (incorrectly identifying a nonrecidivist as someone who is likely to reoffend) versus false negative errors (incorrectly identifying a recidivist as someone who is unlikely to reoffend).
Fortunately, the AUC permits the comparison of different measures developed on samples with different base rates of recidivism, and practitioners or researchers can determine for themselves the optimum cut scores for their particular purpose by examining each ROC curve. Rice and Harris (1995) have discussed how to determine cutoff scores for different base rates and situations in which the relative costs of false positive or false negative errors are not equal. A more detailed discussion of how to measure predictive accuracy is provided in Quinsey et al. (2006,
chap. 3
).
