Transcript California Static Risk Assessment

California Static Risk
Assessment (CSRA)
Susan Turner, UCI
October 16, 2008
CDCR Has Embraced Risk
Based Decision-Making
• Recommended by reviewers of CDCR practices
– Expert panel recommendations
• Risk assessment part of the California Logic Model
– Strike Team
• Realigning parole resources with offender risk
• Actuarial risk tools are used in criminal justice for a
variety of purposes
– Pretrial release, sentencing guidelines, parole release and
revocation decisions, probation caseload assignment, priority for
programming
How Actuarial Risk Prediction
Works—Auto Insurance Example
• Insurers want to know the likelihood that a driver
will be in an accident
• They use their extensive records data to
determine what factors are related to drivers
experiencing an accident
• The model:
Risk (accident)=age + gender + zip code
+ driving experience +…etc.
States are Currently Using Risk
Assessment Tools in Parole
• Pennsylvania uses LSI and Static 99 along
with violence indicator, institutional
programming and behavior
• Maryland uses crime and 6-item risk
assessment for parole eligibility
Actuarial Risk Prediction Has
Benefits and Drawbacks
• Strengths
– Promotes efficiency, consistency and objectivity in
decision-making
– Has an empirical basis
– Is more accurate than clinical judgment
• Limitations
– Decisions are based on aggregate, or group,
performance
• “Good” people with “bad” characteristics penalized
• “Bad” people with “good” characteristics get a break
A Number of Outcomes Can Be
Predicted in Corrections
• Most common is recidivism over a certain time
period (e.g. three years)
• Other outcomes may be important for program
evaluation
– Drug use, employment, mental health status, etc.
• Availability/ease can drive this choice
– Return to Custody (CDCR data)
– Arrest (DoJ data)
– Conviction (DoJ data)
Each Recidivism Outcome Offers
Something Different
• Arrest
– Captures the most criminal behavior
– Most likely to “over-capture”
• Conviction
– Highest standard of proof
– Many instances of criminal behavior do not result in conviction
for a new offense
• Return to Custody
– Most direct impact on institutions population
CDCR has elected to use arrest as the outcome
• Most conservative outcome for public safety protection
UCI Asked to Assist with Risk
Prediction for the CDCR Population
• Develop an actuarial risk prediction
instrument using available data
• Validate the instrument to determine
predictive power for the CDCR population
• Tool will operate as a “plug-in” to the
existing COMPAS system
UCI Drew on Washington State
Work to Develop CSRA
• Washington State Institute for Public Policy
(WSIPP) started by testing the items on the LSI-R
– One of most widely used risk/needs assessments
– Includes static and dynamic factors
• WSIPP removed items that did not have predictive
usefulness
• WSIPP added items that improved predictive
accuracy
– Detailed juvenile and adult criminal history items
• The resulting tool uses static factors only
• WSIPP tool predicts reconviction (arrest data not
available)
CSRA Uses Multiple Data Sources
• CDCR OBIS
– Demographics
– Return to custody outcomes
• DOJ Automated Criminal History (“Rap Sheets”)
– Arrests
– Convictions
– Parole/probation violations
• Juvenile criminal history data not available
(reliably) in California
Test Development Followed Standard
Procedure
• Large sample of 103,000 individuals
released from CDCR institution in FY
‘02/’03
• Sample divided randomly into construction
and validation groups
• Developed items and weights on the
construction group
• Validated instrument on the validation
group
UCI Refined the Model to Fit the
California Population
• Test the predictive power of the Washington tool’s
items and scales using available CA data
– No juvenile criminal history record data
• Examine CDCR data to see if they had items that
added predictive power to the instrument
• Experiment with different cut points within items and
counting rules within the prediction model
• Weight items based on the strength of their
relationships to recidivism
• Develop predictive models for arrests, reconvictions,
and returns to custody
Resulting CSRA Uses 22 Items to
Predict Recidivism
• Demographics
– Age at release, gender
• Number of felony sentences
• Felony sentences for murder/ manslaughter, sex, violent,
weapons, property, drug and escape offenses
• Misdemeanor sentences for assault, sex, weapons,
property, drug, alcohol and escape offenses
• Revocations of probation or parole supervision
• The model:
Risk (felony arrest)= age + gender +…# of violent
felony convictions +…# of misdemeanor property
convictions +…# of probation/parole violations
CSRA Scores Offenders on
Three “Nested” Sub-Scales
1. Violent Sub-Scale
2. Property & Violent Sub-Scale
3. Any Felony Sub-Scale
This allows CDCR to differentiate risk by type
of recidivism
CSRA Risk Group Is
Determined Hierarchically
Violent Score
103 or higher?
Yes
High Violent
No
Property/Viol. Score 119
or higher?
Yes
High Property
No
Felony Score 127 or
higher?
Yes
High Drug
No
Property/Viol. Score
or Felony Score 96
or higher?
No
Low
Yes
Moderate
CSRA Divides the Population
into Distinct Risk Groups
100
90
82
82
82
80
69
70
Percent
60
50
Any Felony
48
Drug Felony
48
Property Felony
40
40
28
30
20
38
34
26
21 22
23
Moderate (33%)
High Drug (9%)
31 31
26
17
10 11
10
0
Low (22%)
High Property (19%)
"Any Arrest" Rates by Risk Group
High Violent (17%)
Violent Felony
CSRA Divides the Population
into Distinct Risk Groups
100
90
82
82
82
80
69
70
Percent
60
50
Any Felony
Drug Felony
48
Property Felony
37
40
21
17
20
12
10
29
28
30
8
15
25
22
18
15
10
17
13
0
Low (22%)
Moderate (33%)
High Drug (9%)
High Property (19%)
"Most Serious Arrest" Rates by Risk Group
High Violent (17%)
Violent Felony
Appendices
CSRA Reconviction Prediction
100
90
80
70
Percent
60
Any Felony
Drug Felony
46
50
44
Property Felony
41
40
Violent Felony
31
30
30
20
10
24
18
18
16
10
9
5
4
16 15 15
13
7
8
8
0
Low (22%)
Moderate (33%)
High Drug (9%)
High Property
(19%)
"Any Reconviction" Rates by Risk Group
High Violent (17%)
CSRA Reconviction Prediction
100
90
80
70
Percent
60
Any Felony
Drug Felony
46
50
44
Property Felony
41
40
Violent Felony
31
30
20
10
25
23
18
13
8
4
4
10
12
7
13
8
8
11 13
15
0
Low (22%)
Moderate (33%)
High Drug (9%)
High Property
(19%)
"Most Serious Reconviction" Rates by Risk Group
High Violent (17%)
CSRA Performs within Usual
Range for Risk Assessments
Instrument
AUC
Sample
Recidivism
Measure
Source
CSRA
0.70
103,000
releasees
Felony arrest
Current
COMPAS
0.67
515 California
parolees
Return to prison
Farabee and
Zhang (2007)
Criminal History
Computation
0.68
28,519 Federal
offenders
Re-conviction, re-arrest
w/out dispo. available,
supervision revocation
US Sentencing
Commission
(2004)
LSI-R
0.67
22,533 Wash.
offenders
Any conviction
WSIPP (2003)
Washington
Static Risk
Assessment
0.74
51,648 Wash.
offenders
Felony conviction
WSIPP (2007
100
80
Sample Item Weights—
Felony Scale
Age 20-29 (vs. 60+)
60
5 Probation/Parole
Violations (vs. None)
40
20
Male (vs. Female)
5 Prior Felony Property
Convictions (vs. None) 3 Prior Felony Drug
Convictions (vs. None)
0
-20
-40
Felony Homicide
Conviction (vs. None)
100
Sample Item Weights—
Violent Scale
80
Age 20-29 (vs. 60+)
60
40
20
5 Probation/Parole
Violations (vs. None)
Male (vs. Female)
Felony Homicide
5 Prior Felony Property
Conviction (vs. None) Convictions (vs. None)
0
-20
-40
3 Prior Felony Drug
Convictions (vs. None)