Risk Assessment Project II Interim Report 2 Validation of a Risk Assessment Instrument by Offense Gravity Score for All Offenders

Transcription

1 Highlights Risk Assessment Project II Interim Report 2 Validation of a Risk Assessment Instrument by Offense Gravity Score for All Offenders [February 2016] The purpose of this report is to present the findings from a validation study of a risk assessment instrument that was developed to identify offenders at risk of recidivism. This phase of the risk assessment project was a follow-up to our original study that developed a risk instrument for mid-level seriousness offenders. This follow-up study developed a risk scale for all offenders and, in doing so, developed a risk scale for each Offense Gravity Score [OGS] of the sentencing guidelines. This resulted in nine separate risk assessment tools. 1 Highlights from the study are as follows: Development of risk assessment scales by Offense Gravity Scores for three reasons: o The large difference in the number of offenders in the different OGS levels o The difference in the recidivism rate by OGS, which was a non-linear relationship o The difference in the type of offenses by OGS We conducted two studies using two samples of offenders: o Offenders sentenced during [N=101,498] o Offenders sentenced during [N=131,055] The overall 3-year recidivism rate was 46% for the first sample and 44% for the second sample. While the factors predicting recidivism were similar for the two studies, there were some differences in the operationalization of the factors for some of the scales. Second, based upon our analyses, we concluded that the risk assessment scales developed with the later sample performed the best. The factors found to be predictive of recidivism, and thus used in the final risk assessment scales, are: 2 o Age (9) o Number of prior arrests (9) o Gender (8) o Current offense type (8) o Prior Record Score (5) o Prior offense type (5) o Multiple current convictions (4) o Prior juvenile adjudications (2) 1 The guidelines have 14 Offense Gravity Scores. We combined OGSs 9-14 as these categories had too few offenders to allow for the necessary analyses, and those OGSs compose Level 5 of the sentencing guidelines, which represent the most serious offenders. 2 The number in parentheses indicates how many of the nine risk scales included that factor: 1

2 Background Act 95 of 2010, mandated the Pennsylvania Commission on Sentencing to develop a risk assessment instrument to assist the court at sentencing. Specifically, legislation mandated that the Commission undertake the following: Adopt a risk assessment instrument to be used at sentencing Consider the risk of re-offense and threat to public safety Help determine if the offender is a candidate for alternative sentencing programs [RRRI, CIP,SIP,BC] Develop an empirically based worksheet using factors predicting recidivism Risk Assessment Project: Phase I [Levels 3 and 4 offenders] To address this mandate, the Commission undertook the Risk Assessment Project. The initial project, which was started during the summer of 2010, resulted in the development of an initial risk assessment tool for offenders sentenced under Levels 3 and 4 of the sentencing guidelines [mostly mid-level theft and drug offenders]. These levels were chosen for several reasons: 1) the offenses at these levels encompassed a wide variety of offense seriousness [OGS ranging from 2 to 8 depending upon prior record]; 2) the sentence recommendations provided for a variety of sentence types [prison, SIP, BC, jail, probation, IP]; and 3) we could use a sample sentenced during 2004, 2005, and 2006 [SGS Web data years], which allowed for a three year tracking period for most offenders. Phase I reports. The initial project resulted in the following nine interim reports that documented the progression of work that led to the development of the initial risk assessment tool. [These reports can be found on the Commission s website: Interim Report 1: Review of Factors used in Risk Assessment Instruments Interim Report 2: Recidivism Study: Initial Recidivism Information Interim Report 3: Factors that Predict Recidivism for Various Types of Offenders Interim Report 4: Development of Risk Assessment Scale Interim Report 5: Developing Categories of Risk Interim Report 6: Impact of Risk Assessment Tool for Low Risk Offenders Interim Report 7: Validation of Risk Scale Interim Report 8: Communicating Risk at Sentencing Special Report: The Impact of Juvenile Record on Recidivism Risk Implementation of a Risk Assessment Tool: Focus Groups. In December 2012, the Commission made the decision that the initial risk assessment tool for Level 3 and Level 4 offenders would define low risk offenders as having a risk score ranging from 0 to 4 on a 14 point scale. The next step was to determine how best to implement the risk scale. Toward that end, the Commission assembled focus groups to provide input on the implementation process. These focus groups consisted of personnel who utilized the sentencing guidelines in the four counties represented by the judicial appointments to the Commission: Allegheny, Blair, Philadelphia, and Westmoreland. The Commission held a meeting with the four focus groups to provide them with information on the background and development of the risk scale, The next step involved the development of a survey to determine the best way to convey risk information. This survey was sent to 1,000 criminal justice personnel in the four focus group counties, and follow-up meetings 2

3 were held in each of the four counties to provide the survey findings, and receive additional input into the implementation of a risk assessment tool. Implementation of a Risk Assessment Tool: Beta Testing. The Commission decided to conduct beta testing of the risk assessment tool with real cases in the four focus group counties. However, the beta testing of the risk assessment tool in the field was delayed for two reasons. First, during the Focus Group sessions that we held on the initial risk assessment tool, one issue raised was the possible utilization of Common Pleas Case Management System [CPCMS] to populate the criminal history record part of the risk assessment module of the sentencing guidelines. For our risk study, we had used the criminal history records from the State Police, and for people in the field to use those records would require manual determination of an offender s criminal history. In 2014, we worked with the Administrative Office of the PA Courts [AOPC] to receive data from their system so we could determine the feasibility of using CPCMS for criminal history information. Upon replication of our original analyses using data obtained through CPCMS, we found consistent results. We concluded that using CPCMS was a viable option for the criminal history information used in the risk assessment scale, which would eliminate the need for the county users to manually obtain and enter this information. Thus, all of the information used in the risk scale would then be pre-populated into the risk module of the SGSWeb guideline software. The second reason that beta testing was delayed was because the Commission decided to develop a risk assessment tool for all offenders [not just Level 3 and 4 offenders] prior to implementation. Thus, the Commission next undertook a follow-up risk study utilizing offenders from all 5 levels of the sentencing guidelines. Risk Assessment Project: Phase II [Levels 1-5 offenders] The original risk scale was developed for all offenders sentenced under Levels 3 and 4 of the sentencing guidelines. In the follow-up study, we developed a scale for all offenders sentenced under all five levels of the guidelines. In the original study the seriousness of the offense, as defined by the Offense Gravity Score, was one of the factors controlled for in the analysis. As it was found to be predictive of recidivism, it became one of the factors in the risk scale. In order to refine the utility of the scale, however, it was decided to develop a risk assessment scale for each of the Offense Gravity Scores, rather than use it as an independent factor. This decision is discussed in more detail later. Phase II Reports. The second phase of the risk project has two previous reports documenting the progression of work that led to the development of the risk scales by Offense Gravity Score [These reports can be found on the Commission s website: Phase II: Special Report. The Impact of Removing Age, Gender, and County from the Risk Assessment Scale Phase II: Interim Report I. The Development of a Risk Assessment Scale by Offense Gravity Score for All Offenders. Two Risk Studies. For our follow-up recidivism project, we conducted two studies using two samples of offenders 3 For the first study we included offenders sentenced during to allow for a threeyear tracking period for those sentenced to prison for a lengthy period of time. The second study included 3 DUI offenders were not included in either the original or follow-up study as these offenders were part of a separate project. 3

4 offenders sentenced during to provide for a more recent time period as well as improvements in data quality. Measuring Recidivism. As our measure of recidivism, we used re-arrest and, for offenders sentenced to state prison, re-incarceration on a technical violation. We obtained arrest information from the criminal history records maintained by the State Police. In determining exposure time, we used date of sentence for probation cases, expiration of minimum sentence for county jail sentences, 4 and date of release for state prison sentences. The Department of Corrections provided the date of release, as well as information on technical violations that resulted in return to prison, which we took into account for the recidivism of offenders sentenced to prison. Description. Table 3 shows the sample description for the two samples [ and ]. The first sample consisted of 101,498 offenders and the second sample consisted of 131,055 offenders. The percentage of offenders who recidivated during the three year follow-up period was 46% for the earlier sample and 44% for the later sample. Overall, for both samples, most of the offenders were white [55%, 60%], male [82%, 80%], from an urban county [78%, 77%], and under age 30 [52%, 53%]. 5 Offenders were most likely to be convicted of a property [32%, 32%], personal [27%, 24%], or drug [26%, 30%] offense, with the most frequent Offense Gravity Score being 3. About 38% of the offenders in both samples had multiple convictions for which they were being sentenced. The majority of offenders [75%, 76%] had a prior arrest with the mean number of prior arrests being 3.4 for the earlier sample and 3.7 for the later sample. As with their current offense, the most prevalent type of prior arrest was for property [53%. 54%], personal [39%, 41%], and drug [33%, 39%] offenses. The majority of offenders did not have a prior conviction, as indicated by the Prior Record Score [56%, 55%], and a small percentage had a prior juvenile adjudication [5%; 6%]. 6 [See Appendix A for sample descriptions by Offense Gravity Score.] 4 For county jail sentences, date of release was unavailable. Since the judge has paroling authority in these cases, offenders can be released prior to the expiration of the minimum. Findings from another study the Commission had conducted indicated that about a third of the offenders are released prior to their minimum, about a third at the expiration of their minimum, and about a third post minimum sentence. Thus, we decided that expiration of the minimum sentence was the best estimate for date of release for county jail sentences and we used that to determine exposure time. 5 In the sample description, the first number in the parenthesis represents offenders in the sample and the second number represents offenders in the sample 6 Under the sentencing guidelines, an offender can have a prior conviction for one misdemeanor offense and have a Prior Record Score of 0. Prior juvenile adjudications do not count for M2, M3, unclassified misdemeanors, and some M1 misdemeanors. 4

5 Table 1. Description for [N=101,498] and [N=131,055] samples N % N % Race *** Multiple Convictions*** White 55,795 78,761 55% 60% Yes 38,206 50,214 38% 38% Black 37,398 41,500 37% 32% No 63,292 80,841 62% 62% Hispanic 7,364 9,452 7% 7% Juvenile Adjudications * Other 941 1,342 1% 1% Yes 5,421 7,275 5% 6% Sex*** No 96, ,780 95% 94% Male 82, ,236 82% 80% Female 18,560 26,819 18% 20% Prior Convictions*** County*** Yes 44,273 59,519 44% 45% Philadelphia 15,202 15,194 15% 12% No 57,225 71,536 56% 55% Allegheny 12,094 13,748 12% 10% Prior Offense Type Other Urban 51,442 71,542 51% 55% Prior Property * Rural 22,760 30,571 22% 23% Yes 54,035 70,381 53% 54% Age*** No 47,463 60,674 47% 46% < 21 17,816 18,363 18% 14% Prior Personal/Sex*** ,477 32,005 22% 24% Yes 39,896 53,501 39% 41% ,193 17,243 13% 13% No 61,602 77,554 61% 59% ,903 31,744 28% 24% Prior Drug *** ,666 23,238 14% 18% Yes 33,929 51,402 33% 39% 50+ 4,443 8,462 4% 6% No 67,569 79,653 67% 61% Mean Prior Weapon Number of Prior Arrests*** Yes 16,333 21,169 16% 16% 0 25,799 30,853 25% 24% No 85, ,886 84% 84% 1 19,564 24,138 19% 18% Prior Traffic*** ,636 30,304 23% 23% Yes 20,688 22,958 20% 18% ,595 17,106 12% 13% No 80, ,097 80% 82% 6-7 6,985 10,236 7% 8% Prior Public Adm*** 8+ 12,919 18,418 13% 14% Yes 16,443 26,811 16% 20% mean No 85, ,244 84% 80% Prior Public Order*** Current Offense*** Yes 26,786 36,742 26% 28% burglary 3,125 3,670 3% 3% No 74,712 94,313 74% 72% property - misd 16,854 20,991 17% 16% Type of Sentence*** property - felony 12,445 17,102 12% 13% Prison 10,419 16,024 10% 12% personal - misd. 17,671 20,362 17% 16% Jail 34,608 42,906 34% 33% personal - felony 7,633 8,283 8% 6% County IP 5,245 6,570 5% 5% sex offense - felony 1,216 1,767 1% 1% Probation 48,714 62,488 48% 48% sex offense - misd. 1,372 1,643 1% 1% Fines/restitution 2,512 3,067 2% 2% drug - felony 14,429 20,334 14% 16% Recidivism [within 3 years]*** drug - misd. 12,065 18,188 12% 14% Yes 47,052 57,311 46% 44% other traffic 2,207 2,746 2% 2% No 54,446 73,744 54% 56% firearms 2,602 3,937 3% 3% other weapons 959 1,256 1% 1% public order 4,460 3,794 4% 3% public adm 4,460 6,982 4% 5% Offense Gravity Score*** 1 13,055 17,309 13% 13% 2 11,521 12,851 11% 10% 3 36,110 46,534 36% 36% 4 4,651 6,275 5% 5% 5 13,885 18,214 14% 14% 6 8,554 10,722 8% 8% 7 4,947 7,407 5% 6% 8 2,297 4,026 2% 3% 9 2,783 3,759 0% 0% 10 2,183 2,608 0% 0% % 0% % 0% % 0% % 0% mean * sig. at.05 level ** sig. at.01 level *** sig. at.001 level 5

6 Developing a risk scale for each OGS. For this study, we focused on refining the risk assessment tool and determining whether a risk tool could be developed for each individual Offense Gravity Score [OGS] of the guidelines, rather than developing one tool that would be applied to all OGS levels. This decision was made for three reasons: 1) the large difference in the number of offenders in the different OGS levels; 2) the difference in the recidivism rate by OGS, which was a non-linear relationship; and 3) the difference in the type of offenses by OGS The sentencing guidelines have 14 OGS levels. As the number of cases in the six highest OGS levels was too small to give us confidence in the results of the analyses, we collapsed OGS levels 9 through 14, which represented the most serious offenders [Level 5 of the guidelines]. Thus, we essentially conducted nine separate recidivism studies. Number of offenders by OGS. The first reason for developing a risk score by OGS was the wide range in the number of offenders depending upon the OGS level. Table 4 shows that less than 1% of the offenders had an OGS of 12, 13, or 14, while 36% had an OGS of 3. With over a third of the offenders at OGS 3, those offenders would be overrepresented in the analysis, even controlling for OGS level. Table 2. Number of Offenders by OGS for and s. Number Of Offenders Percent OGS 1 13,055 17,309 13% 13% 2 11,521 12,851 11% 10% 3 36,110 46,534 36% 36% 4 4,651 6,275 5% 5% 5 13,885 18,214 14% 14% 6 8,554 10,722 8% 8% 7 4,947 7,407 5% 6% 8 2,297 4,026 2% 3% 9 2,783 3,759 3% 3% 10 2,183 2,608 2% 2% % 1% % 0% % 0% % 0% Total 101, , % 100% Number of Offenders 50,000 45,000 40,000 35,000 30,000 25,000 20,000 15,000 10,000 5,000 0 Number of offenders by OGS sample sample Offense Gravity Score [OGS] Recidivism by OGS. The second reason to develop a risk score by OGS was to better address the varying levels of seriousness represented by the various OGS levels. Table 5 shows variation in the recidivism rate of offenders varies by OGS. While the overall recidivism rate was 46% [ sample] and 44% [ sample], there was not only a wide range in the percent recidivating depending upon the OGS, the recidivism rate was also not linearly related to OGS. For example, for the sample, the highest recidivism rates were for OGS 6, 4, 7, and 9 [50%, 50%, 52%, and 56%], while the lowest recidivism rates were for OGS 13, 12, and 1 [31%, 35%, and 37%, respectively.] 6

7 Table 3. Percent of Offenders Recidivating by OGS Total Number of Offenders Percent Recidivating OGS 1 13,055 17,309 43% 37% 2 11,521 12,851 49% 42% 3 36,110 46,534 45% 43% 4 4,651 6,275 41% 50% 5 13,885 18,214 49% 47% 6 8,554 10,722 52% 50% 7 4,947 7,407 52% 52% 8 2,297 4,026 42% 41% 9 2,783 3,759 53% 56% 10 2,183 2,608 49% 46% % 41% % 35% % 31% % 42% Total 101, ,055 46% 44% Percent of Offenderw 60% 50% 40% 30% 20% 10% 0% Percent of offenders recidivating by OGS Overall Average sample sample Offense Gravity Score [OGS} Thus, some offenders committing less serious offenses have higher recidivism rates than those who commit the more serious offenses. Development of a risk tool for all offenders as a group would potentially result in the less serious offenders having higher risk scores while more serious offenders would have lower risk scores. Application of a risk tool within the context of the offense seriousness [as reflected by the OGS] was viewed as being both consistent with the guideline structure and potentially more useful to judges at sentencing. Offense type by OGS. The third reason to develop a risk scale by OGS was the difference in the type of offense by OGS. Table 4 shows that the lower OGSs [1-5] had the highest rate of property offenses, the higher OGSs [10-14] had the highest rate of personal offenses, and the middle OGSs [6-8] had the highest rate of drug offenses. 7

8 Table 4. Type of Offense by OGS. Offense Property Personal Drug Other Gravity Percent Number Percent Number Percent Number Percent Number Score 1 20% 17% 2,592 2,875 5% 6% 588 1,098 36% 43% 4,733 7,413 39% 34% 5,142 5, % 63% 7,487 8,113 2% 2% % 0% 0-33% 34% 3,823 4, % 33% 11,133 15,555 39% 34% 14,237 15,716 25% 28% 8,910 12,896 5% 5% 1,830 2, % 43% 1,900 2,698 41% 35% 1,886 2,218 0% 0% 0-19% 22% 865 1, % 52% 7,485 9,442 14% 14% 2,009 2,634 26% 27% 3,582 4,944 6% 7% 809 1, % 5% % 27% 2,091 2,844 69% 66% 5,918 7,035 2% 3% % 27% ,966 30% 24% 1,476 1,771 36% 46% 1,788 3,375 6% 4% % 0% % 28% 1,083 1,119 47% 49% 1,070 1,990 6% 22% % 0% % 31% 1,040 1,179 1% 0% % 68% 1,709 2, % 0% % 80% 1,990 2,079 9% 16% % % 0% % 59% % 41% % 0% % 100% % 0% % 0% % 33% % 67% % 0% % 0% % 100% % 0% - - 0% 0% - - Total 32% 31% 32,361 41,215 27% 24% 27,878 31,917 26% 29% 26,473 38,483 14% 15% 14,786 19,440 Charts a-d. Percentage of offenders by type of conviction offense and OGS. Percent of Offenders 100% 80% 60% 40% 20% 0% a. Property Offenses Offense Gravity Score sample sample Percent of Offenders 100% 80% 60% 40% 20% 0% b. Personal Offenses Offense Gravity Score sample sample 100% c. Drug Offenses 100% d. Other Offenses 80% 80% Percent of Offenders 60% 40% 20% sample sample Percent of Offenders 60% 40% 20% sample sample 0% % Offense Gravity Score Offense Gravity Score 8

9 Study 1: Development of risk scales with the sample Bivariate analysis: factors related to recidivism. For the first study, we randomly split the sample into two samples: development sample [N= 50,743] and validation sample [N=50,758]. The initial analysis involved using the development sample to determine which factors were significantly related to recidivism. Table 5 shows the recidivism rates for the factors considered in the analysis for the full development sample. Overall, offenders were more likely to recidivate if they were: Black, male, young, from an urban county, convicted of burglary or firearms offenses, had multiple current convictions, had a greater number of prior arrests, had a Prior Record Score [which measures number and seriousness of prior convictions and serious juvenile adjudications], and had a prior juvenile adjudication. Table 5. Recidivism rates for variables used in bivariate analysis: Development [N=50,743]. Percent Number Success Failure Success Failure sig Overall Race *** White/Other Race ,071 11,300 Black ,217 10,478 Hispanic ,800 1,877 Gender *** Male ,326 20,190 Female ,762 3,465 Age *** < ,728 5, ,225 8, ,031 6, ,404 2, , County *** Philadelphia ,281 4,291 Allegheny ,561 3,447 Other urban ,054 11,676 Rural ,192 4,241 Offense type *** Burglary Property_M ,446 3,932 Property_F ,980 3,220 Personal_M ,385 3,419 Personal_F ,864 1,939 Sex Offense_F Sex Offense_M Drug F ,535 3,655 Drug M ,284 2,854 Other Traffice Firearms Other Weapons Public Order , Public Administration ,129 1,056 Multiple charges *** Yes , No , Prior Arrests *** ,136 3, ,894 3, ,608 3, ,967 4, ,416 3, , ,318 2,966 Type of Prior Offense [yes indicated] Personal/Sex ,527 11,324 *** Property ,876 15,034 *** Drug ,773 10,105 *** Firearms/Weapons ,082 5,042 *** Traffic ,086 5,274 *** Public Order ,823 7,582 *** Public Administration ,105 5,072 *** Prior Record Score *** Yes ,860 12,186 No ,228 11,469 Juvenile Adjudication Yes ,785 *** No ,156 21,870 * Significant at.05 level ** Significant at.01 level *** Significant at.001 level 9

10 Table 6 shows a summary of the factors found to be significant at the bivariate level for the each of the OGSs. Age, number of prior arrests, current offense, PRS, prior offense type, and juvenile adjudications were found to be significantly related to recidivism for all OGSs. Gender was found to be significantly related to recidivism for all but one OGS [OGS 1], while having multiple current convictions was found to be significant for four of the OGSs [OGSs 3, 4, 5, and 7]. [See Appendix B for specific bivariate results by OGS.] Table 6. Summary of Bivariate Findings by OGS for Development [ ]. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 n=6,673 n=5,687 18,021 n=2,328 n=6,946 n=4,126 n=2,599 n=1,140 n=3,221 Gender *** *** ** *** *** *** *** *** Age *** *** *** *** *** *** *** *** *** Multiple charges *** *** * ** Total prior arrests *** *** *** *** *** *** *** *** *** Current offense type * ** *** *** *** *** *** *** *** PRS *** *** *** *** *** *** *** *** *** Type of prior arrest Personal *** *** *** *** *** *** *** *** *** Property *** *** *** *** *** *** *** *** *** Drug *** *** *** *** *** *** *** *** *** Firearms/weapons *** *** *** *** *** *** *** *** *** Traffic *** *** *** *** *** * Public Order *** *** *** *** *** ** *** ** *** Public Administratio *** *** *** *** *** *** *** *** *** Juvenile Adjudications *** ** *** ** *** *** *** *** *** * p <.05 ** p <.01 *** p <.001 Multivariate analysis: factors predicting recidivism. Multivariate analysis procedure. The next step was to conduct the multivariate analysis to determine which factors predicted recidivism controlling for all factors simultaneously. In conducting the logistic analysis, we rotated all of the categories for factors that were multi-categorical, not dichotomous, categories. That is, we used one category as the comparison category in the first analysis, a second category as the comparison category in the second analysis and so on for all of the categories. This procedure was necessary to ensure that reported differences were real and not simply due to the particular comparison category. We took a conservative approach and included a factor only if the category was significantly different from all of the other categories for that factor. Tables 7 and 8 show two examples of this process, and provide excerpts from the logistic model rotations. 7 Table 7 shows the rotation done for age at OGS Offenders who were age were 7 Subsequent to this analysis we discovered that some offenses listed in the criminal history records were not actual criminal activity. For example, registration for Megan s Law offenses was indicated in the offense field and date of registration was listed in the arrest field, even though this was not an offense, but rather a requirement that sex offenders register with the state police under the Megan s Law. We removed these noncriminal offenses, which resulted in some differences in the findings for some OGSs. Thus the final logistic models included in Appendix D vary somewhat from the one used for these analyses. 10

11 significantly less likely to recidivate than offenders who were less than 21, and significantly more likely to recidivate than offenders in the 30-39, 40-49, and over 49 age group categories. However, there was no difference among offenders who were in the and age groups categories [shaded grey in Table 7], so those two categories were collapsed into one. Table 7. Example of Variable Rotation:Age categrories for OGS 9-14 [excerpt from logistic regression model outcome] < 21 years years years years 50+ years Odds Ratio Odds Ratio Odds Ratio Odds Ratio Odds Ratio < 21 Reference 1.901*** 2.801*** 3.680*** 9.947*** years 0.526*** Reference 1.474*** 1.936*** 5.233*** years 0.357*** 0.679*** Reference *** years 0.272*** 0.516*** Reference 2.703*** *** 0.191*** 0.282*** 0.370*** Reference *** sig. at.001 level Note: grey shaded area shows where the categories were not significantly different from each other. Table 8 shows the rotation done for current offense type at OGS 6. Offenders convicted of personal misdemeanor offenses had the lowest recidivism rate at this OGS. If this offense is used as the reference category in the analysis, we find that offenders convicted of personal felony, drug, and escape offense are significantly more likely to recidivate than offenders convicted of personal misdemeanor offenses. Further, there is no difference in the recidivism of offenders convicted of burglary, property, sex related offense, or other offenses when compared to offenders convicted of personal misdemeanors. Thus, this pattern of significant differences would result in offenders convicted of personal felony, drug, and escape offenses receiving one point on the risk assessment scale, while all other offenders would receive no points for their current offense. However, this possibly would result in some offenders receiving a point in the risk scale while being no more likely to recidivate than some offenders receiving no points. For example, if personal felony offenses are used as the reference category, we find that they are not more likely to recidivate than offenders convicted of burglary and property offenses. Yet, using the original comparison [personal misdemeanor offenses as the reference offense] offenders convicted of personal felony offenses would have received one point while offenders convicted of burglary and property offenses would have received 0 points in the risk scale. Table 8. Example of Variable Rotation: Offense caegrories for OGS 6 Burglary Property Personal (M) Personal (F) Drug Offenses Sex Related Escape Other Offenses Odds Ratio Odds Ratio Odds Ratio Odds Ratio Odds Ratio Odds Ratio Odds Ratio Odds Ratio Burglary Reference * Property Reference * Personal (M) Reference 0.454* 0.377** ** Personal (F) * Reference *** Drug Offenses ** Reference 2.488*** Sex Related Offenses 0.567* 0.602* *** 0.402*** Reference 0.365*** Escape ** *** Reference Other Offenses Reference * Sig. at.05 level ** Sig. at.01 level *** Sig. at.001 level 11

12 Multivariate analysis findings. Table 9 shows a summary of the factors found to be predictive of recidivism for each of the OGSs. Age and the number of prior arrests were predictors for all of the OGSs, while gender was predictive for all but OGS 1. Prior offense type was predictive for five of the OGSs that represented less serious offenders [OGSs 1-5]. Having multiple current convictions also was a significant predictor for five OGSs [OGS 3-7]. Having a prior juvenile adjudication was predictive for three OGSs [3, 5, and 6], while having Prior Record Score greater than zero was predictive for two OGSs [7 and 8]. Current offense type was predictive for only OGS 3. [See Appendix C for the final logistic models by OGS.] Table 9. Summary of Predictors of Recidivism by Offense Gravity Score for Development s [ ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 n=6,673 n=5,687 18,021 n=2,328 n=6,946 n=4,126 n=2,599 n=1,140 n=3,221 R Gender ns Yes Yes Yes Yes Yes Yes Yes Yes Age Yes Yes Yes Yes Yes Yes Yes Yes Yes Current offense ns ns Yes ns ns ns ns ns ns Num. of Prior Arrests Yes Yes Yes Yes Yes Yes Yes Yes Yes Prior Offense Type Yes Yes Yes Yes Yes ns ns ns ns Multiple charges ns ns Yes Yes Yes Yes Yes ns ns PRS ns ns ns ns ns ns Yes Yes ns Prior juv. Adjud ns ns Yes ns Yes Yes ns ns ns The significant predictors of recidivism were used to develop the risk scale for each OGS. 8 Table 10 shows how these factors counted in the risk assessment scales for each OGS. The risk scale ranged from 0-8 [OGS 8] to 0-19 [OGS 3]. Before finalizing the risk scales, and conducting the analyses on the validation sample, we addressed two additional issues: 1) the predictive accuracy of the risk scales with and without each of the predictors, and 2) the predictive accuracy of the OGS specific scale compared to the full sample scale. 8 While race and county were found to be significant predictors of recidivism, they are not included in the risk scale. They are, however, statistically controlled for in the analyses, which means that the effects of the other factors are included only after eliminating the effects of race and county.. 12

13 Table 10. Risk Scales by OGS [ Development ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 n=6,673 n=5,834 n=23,206 n=3,156 n=4,126 n=2,348 n=1,959 n=3,849 n=3,221 Scale Gender Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Age <21=3 <21= =2 <21=3 <21=3 <21=3 <21=3 <21=3 <21=2 <21= = = =2 to29= = = =2 39= = =1 >49= =1 44= = = = =1 >49=0 >49=0 >44=0 >39=0 >49=0 >49=0 >49=0 Current offense Number of Prior Arrests Prop Fel=1 All other=0 none=0 none=0 none=0 none=0 none=0 none, 1=0 none=0 none =0 0=0 1=1 1=1 1=1 1 to 2 = 1 1=1 2=1 1=1 1 to 4 =1 1=1 2 to 4 = 2 2=2 2=2 3 to 8 = 2 2 to 4 = 2 3 to 6 = 2 =2 =2 2 to 4 = 2 5 to 9 = 3 3 to 6 = 3 3 to 4 = 3 =3 5 to 7 = 3 over 6=3 =3 5 to 7 =3 over 9 =4 4 5 to 7 =4 over 7 = 4 =4 over 7 =5 Prior Offense Drug=1 Drug=1 Drug=1 Drug=1 Drug=1 Type Public order=1 Property=1 Public adm=1 Public adm=1 Multiple charges Yes =1 Yes =1 Yes =1 Yes =1 Yes =1 No =0 No =0 No =0 No =0 No =0 PRS Yes =1 Yes =1 No =0 No =0 Prior juv. Adjud Yes=1 Yes=1 Yes=1 No/ unknown=0 No/ unknown=0 No/ unknown=0 Additional analysis on the development sample. We conducted two additional series of analyses with the development sample; 1) the contribution of specific factors to the prediction of risk, and 2) whether the OGS specific scales performed better than an overall risk model. These two issues are discussed in more detail in the previous report, Development of a Risk Assessment Scale by Offense Gravity Score for All Offenders. Thus, only the overall findings and conclusions of these analyses are presented here. Contribution of specific factors. In order to determine how well the scale performed with and without each factor, we conducted a series of block testing analyses, adding one variable at a time. We started with the two variables that were consistently the best predictors of recidivism, age and number of prior arrests, and added each variable found to be a significant in the order of their importance [based upon the odds ratio]. 9 For most of the OGSs, most of the factors in the risk scale provided additional 9 For another report, we conducted additional analysis to determine the contribution that three demographic factors made toward the prediction of arrest. In that report, we found that, overall, age contributed about 25% toward the prediction of arrest, and number of prior arrests contributed about 50% toward the prediction of arrest. Removing age significantly reduced the accuracy of the scale s prediction of arrest. Since our focus was on demographic factors, we had not tested number of prior arrests, but since they contribute most toward the prediction of arrest, we think it is safe to assume the removal of that factor would also 13

14 improvement. We decided to keep all of the significant predictors of recidivism in the scale for each OGS as we concluded that any improvement in the scales accuracy, even if modest, [and not statistically significant] was useful. 10 OGS specific scale vs. full scale. In addition to examining the extent to which the individual factors impacted the prediction accuracy of the scale, we also examined how well the risk assessment scale developed for the specific OGS performed compared to a scale that would be based upon findings for the entire sample. In other words, are we doing better in developing OGS specific scales than using one scale for all offenders, and controlling for OGS? In order to determine how well the OGS specific risk scale performed compared to an overall risk scale, we again conducted a series of ROC [receiver operating characteristic] analyses. The OGS specific risk model predicted better than the overall risk scale for all OGSs. This finding was significant, however, for only 3 of the 9 OGSs, and approached significance for another OGS. We concluded that the development of an OGS specific risk scale was the best direction to take, even if it resulted in only modest improvement in the accuracy of the risk prediction for some OGSs. Validation of the risk scales: Bivariate analyses: were the same factors significantly related to recidivism? We used the validation sample to determine whether the same factors were related to recidivism, predicted recidivism, and were operationalized the same way in the risk scales. Table 11 shows the bivariate analyses findings, which indicate that the factors found to be significant in the development sample were also found in the validation sample. The exception to this, however, was for multiple current convictions, which was significant in the development sample for OGS 5 and OGS 7, but not in the validation sample. Further, that factor was not found to be significant in the development sample, but was significant in the validation sample for OGS 2. Additionally, prior traffic offenses were significantly related to recidivism in the validation sample for OGS 1, OGS 7, and OGSs 9-14 but not in the development sample. [See Appendix B for the specific bivariate findings by OGS.] significantly reduce the scale s accuracy. See Special Report The Impact of Removing Age, Gender, and County from the Risk Assessment Scale on Commission s website [ 10 When this analysis was conducted, current offense was found to be a significant predictor of recidivism for OGS 6 and OGS Subsequent to this analysis, however, we discovered that some offenses listed in the criminal history records were not criminal activity. For example, registration for Megan s Law offenses was indicated in the offense field and date of registration was listed in the arrest field. We removed the noncriminal offenses, which resulted in some differences in the findings for some OGSs. 14

15 Table 11. Summary of bivariate findings by ogs for development and validation sample. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. n=6,673 n=6,382 n=5,687 n=5,834 18,021 n=18,089 n=2,328 n=2,323 n=6,946 n=6,939 n=4,126 n=4,428 n=2,599 n=2,348 n=1,140 n=1,157 n=3,221 n=3,257 Gender *** *** *** *** ** *** *** *** *** *** *** *** *** *** *** *** Age *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Multiple charges ** *** *** *** ** * ** Total prior arrests *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Current offense type * *** ** ** *** *** *** *** *** *** *** *** *** *** *** * *** *** PRS *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Type of prior arrest Personal *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Property *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Drug *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Firearms/weapons *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Traffic ** *** *** *** *** *** *** *** *** *** *** ** * ** ** Public Order *** *** *** *** *** *** *** *** *** *** ** *** *** *** ** ** *** *** Public Administratio *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Juvenile Adjudications *** *** ** *** *** *** ** ** *** *** *** *** *** *** *** *** *** *** * p <.05 ** p <.01 *** p <.001 (Chi-square/ T-test for variable) Multivariate analyses: were same factors predictive of recidivism? Table 12 shows a summary of the factors found to be significant predictors of recidivism for each of the OGSs. For four of the OGSs, the same factors were found to be predictive in the validation sample as with the development sample {OGSs 1, 3, 5, 9-14]. For five OGSs, however, we found that there were some differences in the factors predicting recidivism. The factor for which the findings changed the most was current multiple convictions; for three OGSs this factor was not significant in the validation sample {OGSs 4, 6, 7] ; for one OGS [OGS 2] it was significant in the validation sample, but not in the development sample. For two OGSs [OGSs 7 and 8] the PRS factor [measuring prior convictions] was no longer significant in the validation sample. In the validation sample, prior juvenile adjudications became a significant factor for OGS 2, current offense became a significant factor for OGS 7, and prior offense type became a significant factor for OGS 8. [See Appendix C for the final logistic models by OGS.] Table 12. Summary of Predictors of Recidivism by Offense Gravity Score for Development and Validation s [ ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. n=6,673 n=6,382 n=5,687 n=5,834 18,021 n=18,089 n=2,328 n=2,323 n=6,946 n=6,939 n=4,126 n=4,428 n=2,599 n=2,348 n=1,140 n=1,157 n=3,221 n=3,257 R Gender ns ns Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Age Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Current offense ns ns ns ns Yes Yes ns ns ns ns ns ns ns Yes ns ns ns ns Num. of Prior Arrests Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Prior Offense Type Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ns ns ns ns ns Yes ns ns Multiple charges ns ns ns Yes Yes Yes Yes ns Yes Yes Yes ns Yes ns ns ns ns ns PRS ns ns ns ns ns ns ns ns ns ns ns ns Yes ns Yes ns ns ns Prior juv. Adjud ns ns ns Yes Yes Yes ns ns Yes Yes Yes Yes ns ns ns ns ns ns Scale accuracy in predicting arrest. While none of these factors mentioned above were the major contributors to the prediction of arrest, we wanted to ensure that their inclusion [or exclusion] would not 15

16 impact the performance of the scales in predicting arrest. Thus, the next step in the analysis was to conduct tests of scale accuracy. Table 13 shows the results of the AUC [area under curve] analysis, which was conducted to see whether the scales that were developed using the development sample performed as well with the validation sample. The results indicate that the scales were validated with the validation sample for all of the OGSs except OGS 2, which did not perform as well on the validation sample. As a result, we conducted additional analyses for OGS 2 to determine whether a better model could be developed. These analyses primarily entailed establishing different categories for age and prior arrests to see whether a revised scale would perform equally well on the development and validation sample. However, these changes did not alter the findings. Thus, we decided that our first step with the second study would be to test these scales with the development sample being used for the second study, which is discussed below. [See Appendix D, Table 1, for specific AUC results by OGS.] Table 13. Testing the scale prediction on the development and validation samples. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 performed the same on both samples better on dev sample performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples Study 2: Development of risk scales with the sample Testing scale performance with sample. As with the first study, before we began analysis for the second study, we split the sample into two: the development sample [n=65,5321 and the validation sample [n=65,524]. Before using the development sample to develop the risk assessment scales for this sample, we first decided to use the development sample as a second validation sample for the first study. Thus, we conducted AUC analyses [area under the curve] to determine whether the risk scales developed with the sample performed as well with the development sample. Table 14 shows the results of the analyses. For the majority of the OGSs [seven of the nine OGSs], the scales were validated. However, the scale for OGS 2 and OGS 8 performed better on the sample, indicating that the scales did not validate. For OGS 2, this meant that the original scale developed did not validate on either the sample or the sample. We next developed a new set of scales using the sample and tested whether they would validate for all OGSs. [See Appendix D, table 2, for the specific AUC results testing the scale with the sample.] Table 14. Comparing the prediction accuracy of the Development Scales between the Development and Development. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 performed the same on both samples better on data performed the same on both samples performed the same on both samples performed the same on both samples better on data performed the same on both samples better on data performed the same on both samples * blue indicates that the scale was validated 16

17 Bivariate analysis: factors related to recidivism. We used the second development sample to once again determine which factors were significantly related to recidivism. Table 15 shows the recidivism rates for the factors considered in the analysis for the full development sample. As with the first sample, offenders were more likely to recidivate if they were: Black, male, young, from an urban county, convicted of burglary or firearms offenses, had multiple current convictions, had a greater number of prior arrests, had a Prior Record Score [which measures number and seriousness of prior convictions and serious juvenile adjudications], and a prior juvenile adjudication. Table 15. Recidivism rates for variables used in bivariate analysis: Development (N=65,532). Percent Number Success Failure Success Failure sig Overall Race *** White/Other Race ,548 15,527 Black ,682 11,006 Hispanic ,780 1,989 Gender *** Male ,330 23,843 Female ,680 4,679 Age *** < ,258 4, ,036 11, ,189 6, ,349 4, ,178 1,026 County *** Philadelphia ,357 4,188 Allegheny ,188 3,685 Other urban ,851 15,004 Rural ,614 5,645 Offense type *** Burglary ,018 Property_M ,134 4,354 Property_F ,478 4,075 Personal_M ,474 3,730 Personal_F ,094 2,032 Sex Offense_F Sex Offense_M Drug F ,238 4,995 Drug M ,209 3,838 Other Traffice Firearms ,019 Other Weapons Public Order , Public Administration ,905 1,607 Multiple charges *** Yes ,609 11,466 No ,401 17,056 Prior Arrests *** ,692 3, ,749 4, ,919 3, ,681 5, ,789 5, ,184 1, ,996 4,093 Type of Prior Offense [yes indicated] Personal/Sex ,104 14,538 *** Property ,453 18,739 *** Drug ,249 14,404 *** Firearms/Weapons ,242 6,421 *** Traffic ,339 5,178 ** Public Order ,443 9,900 *** Public Administration ,287 8,106 *** Prior Record Score *** Yes ,216 15,613 No ,794 12,909 Juvenile Adjudication Yes ,273 2,407 *** No ,737 26,115 * Significant at.05 level ** Significant at.01 level *** Significant at.001 level 17

18 Table 16 shows a summary of the factors found to be significant at the bivariate level for the each of the OGSs. The findings were similar to those found with the earlier sample. Gender, age, number of prior arrests, PRS, prior offense type, and juvenile adjudications were found to be significantly related to recidivism for all OGSs. Current offense type was found to be significantly related to recidivism for all but one OGS [OGS 8], while having multiple current convictions was found to be significant for three of the OGSs [OGSs 1, 3, and 4]. [See Appendix B for specific bivariate results by OGS]. Table 16. Summary of bivariate findings by ogs for Development [ ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 n=8,580 n=6,432 n=23,206 n=3,119 n=9,198 n=5,421 n=3,748 n=1,959 n=3,868 Gender ** *** *** *** *** *** *** *** *** Age *** *** *** *** *** *** *** *** *** Multiple charges * *** *** Total prior arrests *** *** *** *** *** *** *** *** *** Current offense type (most serious) *** ** *** *** *** *** *** *** PRS *** *** *** *** *** *** *** *** *** Type of prior arrest Personal *** *** *** *** *** *** *** *** *** Property *** *** *** *** *** *** *** *** *** Drug *** *** *** *** *** *** *** *** *** Firearms/weapons *** *** *** *** *** *** *** *** *** Traffic *** * * Public Order *** *** *** *** *** *** *** *** *** Public Administration *** *** *** *** *** *** *** *** *** Juvenile Adjudications *** *** *** *** *** *** *** *** *** * p <.05 ** p <.01 *** p <.001 = Multivariate findings. The next step was to conduct the multivariate analysis to determine which factors predicted recidivism controlling for all factors simultaneously. 11 Table 17 shows a summary of the factors found to be predictive of recidivism for each of the OGSs. As with the first study, age and the number of prior arrests were predictors for all of the OGSs, while gender was predictive for all but one OGS. Current offense and PRS were significant predictors for five OGSs [OGSs 3, 4, 5, 7, and 9-14]. Having multiple charges was a significant predictors for four OGSs [OGSs 1, 3, 4, and 6], and having a prior juvenile adjudication was predictive for two OGSs [OGSs 1 and 3]. One of the biggest differences between the findings in this sample and the previous sample was that while current offense was significant for only one OGS in the previous study, it was significant for five OGSs in this study. This was due at least partly to the change in our methodological approach for this factor [see footnote 10.] [See Appendix C for the final logistic models by OGS.] 11 As with the first study, we rotated all of the categories for factors that were multi-categorical, not dichotomous, categories. That is, we used one category as the comparison category in the first analysis, a second category as the comparison category in the second analysis and so on for all of the categories. This procedure was necessary to ensure that reported differences were real and not simply due to the particular comparison category. In the first study, we took a very conservative approach and included a factor only if the category was significantly different from all of the other categories for that factor. We decided that this was missing some important findings, particularly if only one category was the exception to the rule and/or the number of offenders in a specific category was small. Thus, we relaxed that standard somewhat for this study, and included a factor if there was only one category that was not significantly different from the other categories. For example, if there were seven offense categories and one offense was statistically different from five of the other six offense categories, we would count offense as a factor in the risk scale. 18

19 Table 17. Summary of Predictors of Recidivism by Offense Gravity Score for Development [ ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 n=8,580 n=6,432 n=23,206 n=3,119 n=9,198 n=5,421 n=3,748 n=1,959 n=3,868 R , Gender Yes Yes Yes ns Yes Yes Yes Yes Yes Age Yes Yes Yes Yes Yes Yes Yes Yes Yes Current offense ns ns Yes Yes Yes ns Yes ns Yes Num. of Prior Arrests Yes Yes Yes Yes Yes Yes Yes Yes Yes Prior Offense Type Yes Yes Yes Yes Yes Yes Yes ns Yes Multiple charges Yes ns Yes Yes ns Yes ns ns ns PRS ns ns Yes Yes Yes ns Yes ns Yes Prior juv. Adjud Yes ns Yes ns ns ns ns ns ns Table 18 shows how these factors counted in the risk assessment scales for each OGS. The risk scale ranged from 0-8 [OGS 8] to 0-19 [OGS 3]. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 Scale Gender Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Male=1 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Female=0 Age <21=5 <21=4 <21= =4 25= =4 <21=4 <21=4 <21=4 <21=3 <21=3 <21= = = = = = = =2 25= = = = = = = = = = = =1 >49= = = = =1 >39=0 >39= =1 >49=0 >49=0 >49=0 >49=0 >49=0 >49=0 Current offense Prop Fel=1 All other =1 All other =1 All other =1 All other =1 Number of Prior Arrests All other =0 Sex =0 Personal Misd.; Sex =0 Sex =0 Burglary =2 Sex,murder=0 none=0 none=0 none=0 none=0 none=0 none=0 none=0 none=0 none=0 1=1 1=1 1=1 1=1 1=1 1=1 1-3=1 1=1 1-5=1 2-3=2 2-7=2 2-3=2 2-5=2 2-3=2 2-3=2 4-7=2 2-3=2 >5=2 4-7=3 >7=3 4-5=3 6-7=3 4-7=3 4-7=3 >7=3 4-7=3 >7=4 6-7=4 >7=4 >7=4 >7=4 >7=4 >7=5 NS Prior Offense Drug=1 Drug=1 Drug=1 Public adm=1 Drug=1 Public adm=1 Public order=1 Property=1 Type Property=1 Public adm=1 personal=1 Public adm=1 Personal=1 Multiple charges Yes =1 Yes =1 Yes =1 Yes =1 No =0 No =0 No =0 No =0 PRS Yes =1 Yes =1 Yes =1 Yes =1 Yes =1 No =0 No =0 No =0 No =0 No =0 Prior juv. Adjud Yes=1 Yes=1 No/ unknown=0 Table 18. Risk Scales by OGS [ Development ] No/ unknown=0 OGS

20 Validation of the risk scales: sample Bivariate analyses: were same factors significantly related to recidivism? We used the validation sample to determine whether the same factors were related to recidivism, predicted recidivism, and were operationalized the same way in the risk scales. Table 19 shows that, overall, the factors found to be significant in the development sample were also found to be significant in the validation sample. However, there were minor exceptions. Multiple current convictions, which was significant in the development sample for OGS 1, was not significant in the validation sample. Further, as was found in the sample, for OGS 2, this factor was significant only in the validation sample. For OGS 1, prior traffic offenses were found significant only in the validation sample, while for OGS 4 they were significant only in the development sample. Additionally, gender was not significant for OGS 1 in the validation sample. [See Appendix B for the specific bivariate results by OGS.] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. n=8,580 n=8,729 n=6,432 n=6,419 n=23,206 n=23,328 n=3,119 n=3,156 n=9,198 n=9,016 n=5,421 n=5301 n=3,748 n=3,659 n=1,959 n=2,067 n=3,868 n=3,849 Gender ** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Age *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** County *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Multiple charges * * *** *** *** *** Total prior arrests *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Current offense type *** *** ** *** *** *** *** *** *** *** *** *** *** *** *** *** PRS *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Type of prior arrest Personal *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Property *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Drug *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Firearms/weapons *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Traffic * *** *** * * ** Public Order *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Public Administrat *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** Juvenile Adjudicatio *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** * p <.05 ** p <.01 *** p <.001 Table 19. Summary of bivariate findings by ogs for development and validation samples. Multivariate analyses: were the same factors predictive of recidivism? Table 20 shows a summary of the factors found to be significant predictors of recidivism for each of the OGSs for the validation sample, as well as the development sample. While most of the factors found to be predictive of recidivism for the development sample were also found to be significant for the validation sample, for all but OGS 3, there were one or two factors for which the findings differed. Age and number of prior of arrests were found to be predictive of recidivism for all OGSs in both the development and validation sample. Prior offense type was a predictive factor in eight of the nine OGS in both samples. The findings were the same in the development and validation samples for gender and current offense for all OGSs but one [gender was not significant for OGS 1 in the validation sample, and current offense was not significant for OGS 2 in the development sample]. Multiple charges was a significant factor for OGS 1 in the development sample but not the validation sample, while the reverse held true for OGS 2. The two factors with the most variation in the findings were PRS and prior juvenile adjudications. There were four OGSs in the development sample for which PRS was a significant predictor [OGS 4, 5, 7, 9-14], but not significant in the validation sample; for one OGS the reverse was found [OGS 8]. There were also four OGSs in the development sample that did not find prior juvenile adjudications to be significant OGS 2, 4, 7, 9-14], but were found to be significant in the validation sample. [See Appendix C for the final logistic models.] 20

21 Table 20. Summary of Predictors of Recidivism by Offense Gravity Score for Development and Validation s [ ] OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. Valid. n=8,580 n=8,729 n=6,432 n=6,419 n=23,206n=23,328 n=3,119 n=3,156 n=9,198 n=9,016 n=5,421 n=5301 n=3,748 n=3,659 n=1,959 n=2,067 n=3,868 n=3,849 R , Gender Yes ns Yes Yes Yes Yes ns ns Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Age Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Current offense ns ns ns Yes Yes Yes Yes Yes Yes ns ns Yes Yes Yes ns ns Yes Yes Num. of Prior Arrests Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Prior Offense Type Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ns ns Yes Yes Multiple charges Yes ns ns Yes Yes Yes Yes Yes ns ns Yes Yes ns ns ns ns ns ns PRS ns ns ns ns Yes Yes Yes ns Yes ns ns ns Yes ns ns Yes Yes ns Prior juv. Adjud Yes Yes ns Yes Yes Yes ns Yes ns ns ns ns ns Yes ns ns ns Yes Scale accuracy in predicting arrest. While the findings predicting recidivism were very similar for the development and validation samples, there were some differences. Even though the differences involved factors that were not the major contributors to the prediction of arrest, we wanted to ensure that these differences did not impact the accuracy of the scales Thus, the next step in the analysis was to determine whether there were differences in the performance of the scales prediction accuracy between the development and validation samples. Table 21 shows that the results of the AUC [area under curve] analysis indicate that there were no differences in the prediction accuracy of the scales, and thus the scales were validated. [See Appendix D. Table 3, for the specific AUC results.] 12 Table 21.Comparing the prediction accuracy of the Development Scales between the Development and Validation. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples performed the same on both samples As a final step in determining the risk assessment scales, we conducted an additional test that involved comparing the performance of the two sets of risk assessment scales developed from each of the two studies [ and ]. If the scales developed using the sample did better or the same with the sample, we would conclude that those scales would be best to use. Alternatively, if the scales developed using the sample did better or the same with the sample, then we would conclude that the scales developed using the sample would be better. Table 22 shows the results of this testing [using the AUC analysis]. The first row shows the results of comparing the two scales using the development sample. For most OGSs there was no difference in the prediction accuracy. The exception was OGS 2 for which the scale did better 12 To provide further assurance that the scales as developed performed essentially the same with both samples [and thus validated], we conducted some additional analyses for factors that were found to be significant in the development sample, but the validation sample. This involved calculating the factor s Z scores to test whether there were significant differences in the coefficients between the development and validation samples. This analysis confirmed that there were no significant differences in the factors tested, and that they performed essentially the same in the logistic regression model for both the development and validation samples. 21

22 and OGS 3 for which the did better. The second row shows the results of comparing the two scales using the development sample. For most OGSs the scale did better. The exception to this was OGS 2 for which both scales performed the same. Thus, we concluded that the risk assessment scales that were developed with the scales were the best to use. Table 22 Comparing the prediction accuracy of the two scales [ and 04-04] with the two Development s [ and ]. OGS 1 OGS 2 OGS 3 OGS 4 OGS 5 OGS 6 OGS 7 OGS 8 OGS 9-14 sample no difference scale does better scale does better no difference no difference no difference no difference no difference no difference sample scale does better no difference scale does better scale does better scale does better scale does better scale does better scale does better scale does better Recidivism rates by risk score. Once the risk scales had been developed, we looked at the recidivism rates associated with each of the risk scores. Table 23 shows the recidivism rate by risk score for each of the OGSs. On the graph, the red line indicates the mean risk score for that OGS, and the grey box shows where the majority of offenders lie along the risk scale range [68% or one standard deviation from the mean as indicated by grey box]. 22

23 Table 23. Recidivism Rates by Risk Assessment Score by OGS. 13 OGS 1 Risk Total Number Percent Score NumberClean Arrest Clean Arrest ,508 1, ,078 1, ,925 1, ,579 2,300 1, ,778 1,560 1, , Total 17,309 10,970 6, Mean 5.65 Median 6 SD 2.07 Mean + 1 SD Mean - 1 SD % in Mean Group 65.6% Percent 100% 80% 60% 40% 20% 0% OGS 1 32% 28% 21% 22% 7% 13% 62% 58% 52% 44% 36% 74% 65% Risk Score OGS 2 Risk Total Number Percent Score NumberClean Arrest Clean Arrest , ,840 1, ,866 1,812 1, ,869 1,514 1, , , , Total 12,851 7,471 5, Mean 4.42 Median 4.00 SD 1.66 Mean + 1 SD 6.08 Mean - 1 SD 2.76 % in Mean Group 75.9% 13 The red line indicates the mean risk score for that OGS; the grey box shows where the majority of offenders lie along the risk scale range [one standard deviation from the mean]. 23

24 Table 23 [cont.] Recidivism Rates by Risk Assessment Score by OGS. OGS 3 Risk Total Number Percent Score NumberClean Arrest Clean Arrest ,393 1, ,244 1, ,068 2, ,220 3,055 1, ,257 3,520 1, ,770 2,973 1, ,445 2,585 1, ,298 2,312 1, ,945 1,916 2, ,881 1,661 2, ,448 1,245 2, , , , , Total 46,534 26,558 19, Mean 8.02 Median 8.00 SD 3.39 Mean + 1 SD Mean - 1 SD 4.63 % in Mean Group 66.2% Percent 100% 80% 60% 40% 20% 0% 14%12% 16%20%24% 28% 33%38% OGS 3 42% 46%51% 64% 69%74%76% 57% 80%81% Risk Score 92% OGS 4 Risk Total Number Percent Score NumberClean Arrest Clean Arrest , , Total 6,275 3,738 2, Mean 5.8 Median 6.0 SD 2.2 Mean + 1 SD 7.9 Mean - 1 SD 3.6 % in Mean Group 62.9% Percent 100% 80% 60% 40% 20% 0% OGS 4 83% 75% 62% 66% 53% 40% 33% 23% 15% 5% 7% 9% Risk Score 24

25 Table 23 [cont.] Recidivism Rates by Risk Assessment Score by OGS. OGS 5 Risk Total Number Percent Score NumberClean Arrest Clean Arrest , ,477 1, ,837 1, ,245 1, ,223 1,184 1, ,384 1,147 1, ,611 1,008 1, , , , Total 18,214 9,689 8, Mean 7.08 Median 7.00 SD 2.57 Mean + 1 SD 9.66 Mean - 1 SD 4.51 % in Mean Group 62.0% Percent 100% 80% 60% 40% 20% 0% OGS 5 85% 85% 79% 72% 61% 52% 47% 38% 30% 21% 15% 8% 11% Risk Score OGS 6 Risk Total Number Percent Score NumberClean Arrest Clean Arrest , , , Total 10,722 5,478 5, Mean 6.32 Median 6.00 SD 2.31 Mean + 1 SD 8.62 Mean - 1 SD 4.01 % in Mean Group 67.8% Percent 100% 80% 60% 40% 20% 0% 0% 9% 13% 21% OGS 6 47% 36% 27% 88% 84% 78% 73% 65% 55% Risk Score 25