FORUM on Corrections Research

Sex offender recidivism prediction

Recent research⁽²⁾ has established an actuarial model for predicting sexual recidivism. The Sex Offender Risk Appraisal Guide has been shown to be relatively effective in predicting both sexual and non-sexual recidivism.

This is very encouraging, as dangerous behaviour has historically been difficult to predict. However, this prediction instrument has been somewhat criticized. Given that its test sample was made up of sex offenders assessed or treated at a psychiatric facility, it is not clear whether the Sex Offender Risk Appraisal Guide would be useful when applied to sex offenders in the criminal justice system.

This article, therefore, summarizes an effort to preliminarily test the predictive validity of this instrument with convicted offenders through a follow-up study of sex offenders released from Correctional Service of Canada institutions. Methodology The Sex Offender Risk Appraisal Guide was developed based on a follow-up study of 178 sex offenders released from a maximum-security psychiatric facility. After an average follow-up of 59 months, 27.5% of the sex offenders sexually recidivated and 40.4% of the sex offenders were arrested, convicted or returned to the psychiatric facility for a violent offence. Regression analyses revealed correlations between recidivism and a variety of predictor variables including the Hare Psychopathy Checklist scores, criminal history (both sexual and non- sexual), and a physiological measure of sexual arousal.

These predictor variables were combined to construct the Sex Offender Risk Appraisal Guide.⁽³⁾ The results obtained with this guide have correlated highly with both sexual and non-sexual recidivism (0.45 and 0.46, respectively). The guide was also demonstrated to have correctly identified 72% of violent offenders and 77% of sex offenders.

To preliminarily test the predictive validity of the guide with sex offenders in the criminal justice system, a sample of 57 male sex offenders was studied. These offenders were released from minimum-, medium- and maximum-security Correctional Service of Canada Quebec Region institutions between January 1, 1991 and January 31, 1993. All of the offenders were released to a half-way house with the specific condition that they participate in a community-based sex offender treatment program. None of the offenders had received institutional treatment.

The average age of this sample at the time of their incarceration was 34.3 (with an age range from 21 to 55). As for offence type, 21 of the offenders had victimized adult women, while 36 had victimized children under the age of 14.

Data for the Sex Offender Risk Appraisal Guide was obtained from institutional files and clinical interviews. The results of a physiological assessment of sexual arousal were also obtained for each offender.

Of the 57 files, 23 were randomly selected and coded by two raters to establish an estimate of inter-rater reliability. Raters agreed in 94.7% of the cases and the disagreements were resolved by a third rater.

Follow-up was conducted from the moment an inmate was released and continued until their sentence expired. Recidivism was defined as being charged with a new sex offence, being charged with a new non-sex offence, or a conditional release violation serious enough to return the inmate to a federal institution. Arrest, conviction and conditional release violation information was obtained from police reports, the case management team at the half-way house, and therapists in the community-based treatment program.

Given the small sample size, readers should be cautious about generalizing any results beyond the scope of this study. However, the positive results seemingly suggest that further study of the guide is warranted. Results Of the 57 sex offenders, 43.9% were returned to a federal institution within 40 weeks of release (see Table 1). Approximately 29.8% of the offenders were arrested for either a sex offence or a non-sex offence, while 14% were incarcerated for a conditional release violation. A point biserial correlation revealed a significant linear relationship between scores on the Sex Offender Risk Appraisal Guide and recidivism (p < .01).

Table 1

Sex Offender Recidivism and Offence Category
Recidivism offence	Number of offenders
Sex offence	19.3%
Non-sex offence	10.5%
Conditional release violation	14.0%
Total	43.9%

To further test the validity of the instrument, a computer simulation was used to identify correct and incorrect release decisions for each Sex Offender Risk Appraisal Guide category score. The Sex Offender Risk Appraisal Guide has nine prediction categories ranging from 0.00 to 1.00 probability of violent recidivism. For each prediction category, the simulation calculated the percentage of the sample correctly identified by the instrument, the percentage of false negatives (offenders predicted to succeed who recidivated), and the percentage of false positives (offenders successful in the community who had been predicted to fail).

For example, if it had been predicted that all inmates would fail in the community, 43.9% of the total sample (those who recidivated) would have been correctly identified but 56.1% would have remained incarcerated unnecessarily.

The accuracy of prediction and types of errors produced by the model varied over the nine prediction categories (see Table 2). The prediction categories between 0.00 and 1.00 yielded varying levels of overall accuracy, with a maximum of 75.4% of the total sample being correctly identified (at the 0.55 category).

Table 2

Computer Simulation Evaluation of Sex Offender Risk Appraisal Guide Release Decisions
Predictian category	Offenders correctly Identified	False negatives (success predicted but offender recidivated)	False positives (recidivism predicted but offender succeeded)
0.00	43.9%	0.0%	56.1%
0.08	43.9%	0.0%	56.1%
0.12	43.9%	0.0%	56.1%
0.17	54.4%	0.0%	45.6%
0.35	70.2%	7.0%	22.8%
0.44	71.9%	14.0%	14.0%
0.55	75.4%	19.3%	5.3%
0.76	63.2%	35.1%	1.7%
1.00	57.9%	42.1%	0.0%

The overall accuracy of the model at each prediction category is of considerable theoretical and empirical interest, but the results of the computer simulation suggest that maximizing overall accuracy does not account for the relative costs of prediction errors.

For example, specifying a cutoff at the model's maximum prediction capacity (75.4% correct) would result in misclassifying approximately 24.6% of the sample, and most of the errors would be false negatives (releasing offenders who should have been kept in custody). A lower cutoff score of 0.35 would lower the proportion of correct predictions (70.2%), but would also decrease the false negatives from 19.3% to 7.0%. Clearly, risk prediction must balance the costs of recidivism against the costs of continued incarceration.

Of course, the best way to reduce the costs of both recidivism and continued incarceration is to provide effective sex offender treatment. The Sex Offender Risk Appraisal Guide can be useful in determining both treatment need and preferred treatment form. For example, low-risk offenders (those scoring 0.17 and below on the instrument) should receive either no treatment or low-intensity programming. High-risk offenders (those scoring 0.35 and above) should be referred to intensive pre-release treatment programs and supervised closely after release.

---

⁽¹⁾C.P. 6128, succursale Centre-ville, Montréal, Québec H3C 3J7. The authors would like to thank V. L. Quinsey and M. Lalumière for their comments on an earlier version of this article.
⁽²⁾V. L. Quinsey, M. E. Rice and G. T. Harris, "Actuarial prediction of sexual recidivism," The Journal of Interpersonal Violence, 10 (1995): 85-105.
⁽³⁾G. T. Harris, personal communication. For a description of the closely related Risk Appraisal Guide, see C. D. Webster et al., The Violence Prediction Scheme: Assessing Dangerousness in High Risk Men (Toronto: Centre of Criminology, 1994).