Transcript Behavioral Prediction and the Problem of Incapacitation

Incapacitation, Recidivism and
Predicting Behavior
Easha Anand
Intro. To Data Mining
April 24, 2007
Background




Crime Control Act of 1984 and USSC
Idea in U.S. is deterrence, rather than
punishment
Tending toward formulae—USPC in D.C.
uses 14 variables
U.S. prison pop. topped 2 million,
parole/probation topped 7 million
Strategies for Incapacitation

Charge-based


Historically the case; most USSC guidelines
Selective


USPC and D.C. Code offenders—based on
individual’s characteristics
New research focuses on “criminal career”
and predicting patterns therein
(participation, frequency, seriousness,
length, patterning)
Rationale



The tendency is toward objective
decision-making processes to improve
accuracy.
More and more variables codified as we
can track offenders.
Sophistication of statistical methods
used to combine predictors seems to be
relevant to outcomes.
The Dataset





6,000 men incarcerated in the 1960s, chosen
at random
Collected life history info, official institutional
record, inmate questionnaire, psychological
tests
26 years later, followed up with Bureau of
Criminal Statistics
Offenses characterized along six dimensions:
Nuisance, physical harm, property damage,
drugs, fraud, crimes against social order
Used 4,897 records
Dataset (cont’d)
Original Offense
Final Dataset
Unusable
Died
Burglary
Other
Purged
Armed
Robbery
Forgery
Homicide
Narcotics
Other
Violent
Offenses
Usable
Problems With Data


Dichotomous dependent variable for
behavior?
Purging = potential bias



Done after age 70 OR
When 10 years arrest-free
No record of out-of-state crimes
Philosophical Problems

Metric for success



False positives: 30,000 arrests could have
been prevented!
False negatives: 1,413 people jailed
unnecessarily…
Reduced crime could have to do with
repentance, increased policing, age,
etc. and not with incapacitation at all
Data Pre-processing




Only used records where had both 1962
and 1988 data
Priors: # of previous convictions
weighted by severity of crime
PriorsP: # of previous periods of
incarceration weighted by length
Inst_(M,P,V,F,etc.): # of arrests
weighted by severity of crime in each of
six categories
# of Arrests to Desistance
(R^2 = .159)
Predictor
Priors
Age
Drugs
Serious
Free
PriorsP
Type
Alias
Regression Coeff
1.115
-.104
-2.155
-.015
-.899
-.413
-.706
.343
Standardized Reg.
Coeff
.270
-.144
-.154
-.058
-.062
-.085
-.05
.046
T
11.02
-6.39
-7.94
-2.92
-3.18
-2.37
-2.31
2.31
# of Arrests to Desistance
(Violent Crimes Only—n=1,998)
Predictor
Priors
Age
InstP
PriorsP
Regression Coeff
-.022
.134
.253
-.066
R^2 = .061; p<.05
Standardized Reg.
Coeff
-.174
.184
.076
-.077
T
-7.85
7.45
3.35
-2.91
What Next?

Multiple Linear Regression



Try using different things as class—
nuisance only, arrest rate, crime-free time
Try different predictors—have 119 variables
BUT


No reason to believe predictors are linearly
independent
No reason to believe non-linear correlation
What Next?

Better technique: Decision trees



“White Box” model mimics human
decisionmaking
Use some kind of feature-selection
algorithm?
Maybe ensemble learning, once featureselection is in place?
Acknowledgements



Trevor Gardner, UC Berkeley
Don Gottfredson, Rutgers University
Bureau of Criminal Statistics