Reports and Publications
Beginning in February 2012, the Minnesota Department of Corrections (MnDOC) began using a modified version of the Minnesota Sex Offender Screening Tool-3 (MnSOST-3) to assess sex offender risk. The original version of the MnSOST-3 contains nine main effects and two interaction effects. In the modified version (hereafter referred to as the MnSOST-3.1), the two interaction effects have been removed, resulting in an instrument with nine main effects. In addition, individual item caps have been instituted to limit inflated risk estimates.
Below are the considerations involved in the decision to begin using the MnSOST-3.1.
First, in the MnSOST-3 interaction model (hereafter referred to as the MnSOST-3), both the effects of violations of orders for protection (VOFP) and recent disorderly conduct convictions on sexual recidivism risk vary according to the age of the offender at the time of release. Whereas VOFPs increase the risk for younger offenders, they decrease the risk for older offenders. Conversely, recent disorderly conduct convictions increase the risk for older offenders, while they decrease the risk for younger offenders. The interaction effects appear counter-intuitive because VOFP and disorderly conduct convictions decrease an offender’s risk in some instances. Moreover, given that the VOFP and disorderly conduct convictions are relatively new findings in the sex offender literature, prior research offers little guidance.
Second, the MnDOC is currently analyzing additional data related to the ability to assess risk for different types of recidivism, including non-sexual violent, non-sexual, felony, and first-time sexual offending. Preliminary findings suggest that VOFP’s and disorderly conduct convictions increase the risk of non-sexual violent recidivism. Therefore, although VOFP’s and disorderly conduct convictions will reduce the MnSOST-3 score for certain offenders, it does not necessarily mean that these offenders would pose less of a risk for violent recidivism. On the contrary, this finding implies that the risk for sexual recidivism may drop because the risk for other types of recidivism (non-sexual violence) increases. For example, the increased likelihood for non-sexual violent recidivism may shorten the at-risk period for sexual recidivism, especially if that recidivism leads to reincarceration (resulting in a more limited opportunity to sexually reoffend).
Third, the MnDOC’s Risk Assessment and Community Notification (RACN) Unit has, by the end of January 2012, scored more than 200 cases on the MnSOST-3. Approximately 20 percent of the offenders scored thus far have had a prior VOFP, which is more than three times the percentage observed (6 percent) in the MnSOST-3 development sample. Given the uncertainty in explaining the VOFP-age interaction, the higher percentage of offenders with VOFP convictions exacerbates concerns regarding this item’s interaction with age at release.
Fourth, as reported by Duwe and Freske (2012), the apparent area under the curve (AUC) statistic for the MnSOST-3 was .821. After adjusting for optimism (.025), the AUC was .796. To determine whether removing the two interaction effects would compromise the predictive accuracy of the MnSOST-3, analyses were performed on the nine-item main effects model. The results showed an apparent AUC of .818. After correcting for optimism (.022), however, the AUC value for the MnSOST-3.1 (.796) was the same as it was for the MnSOST-3. Thus, the results suggest the MnSOST-3.1 will likely perform no worse, but also no better than, the MnSOST-3 in predicting sexual recidivism.
The MnDOC has begun using the MnSOST-3.1 because it is simpler, easier to interpret, and does not have an adverse impact on predictive accuracy. Yet, due to the same optimism-corrected AUC values for both instruments, the MnDOC will continue to make both versions of the MnSOST-3 available to the public. Because the scoring process is similar for both versions of the MnSOST-3, recertification will not be required for those who have been trained on the MnSOST-3.
Use of the MnSOST-3.1 will result in modest changes in cut scores for presumptive risk levels. Under the MnSOST-3, the cut scores for presumptive risk levels are:
Under the MnSOST-3.1, the cut scores for presumptive risk levels are:
For offenders without VOFP and/or disorderly conduct convictions, scores from the two versions of the MnSOST-3 will vary only slightly. The major differences in scores for the two instruments will be observed among offenders who have VOFP and/or disorderly conduct convictions. Under the MnSOST-3.1, VOFP and disorderly conduct convictions will increase risk regardless of the offender’s age at the time of release.
In addition to the removal of the interaction effects, there is one other difference between the MnSOST-3 and MnSOST-3.1 worth noting. To avoid inflated risk estimates resulting from extreme values among the five continuous items on the instrument, the values for these items are capped based on the highest value observed in the MnSOST-3 development sample. For example, the largest value for male victims in the MnSOST-3 sample was four. As a result, the highest value an offender can receive for the male victim item on the MnSOST-3.1 is four. Similarly, predatory offense sentences are capped at 25, felony offense sentences at 20, VOFP convictions at 5, and recent disorderly conduct convictions at 2.
Even though the MnDOC will be using the MnSOST-3.1, efforts will be made to better understand the interaction findings from the MnSOST-3 and further assess the impact these interactions have on assessments of risk. In particular, by estimating the impact of items such as VOFP and disorderly conduct convictions on other types of recidivism besides sexual reoffending, continuing work on the global risk assessment tool may help decrease the extent to which the interaction findings appear to be counter-intuitive.
The MnDOC will also be collecting data on offenders scored on the MnSOST-3.1 to determine what their scores would have been on the MnSOST-3. It is anticipated that analyzing data on a larger number of cases, especially those scored recently, will enable the MnDOC to more fully comprehend the behavior of the interaction terms in the MnSOST-3. Efforts will be made to later share the results of these analyses with the broader corrections and forensic communities.
MnSOST-3 application (Excel document)
MnSOST-3.1 application (Excel document)
Duwe, Grant and Pamela Freske (2012). Using logistic regression modeling to predict sex offense recidivism: The Minnesota Sex Offender Screening Tool-3 (MnSOST-3). Sexual Abuse: A Journal of Research and Treatment DOI: 10.1177/1079063211429470.