V.T. Notes
on pre-trial criminal risk assessment...
Direct Link to File. 791 words, 4 minute read, 2 paperback pages
communities drive change, funding
arnold ventures
Crime & Justince INstitute…
pretrial justice conference
classification: depends on policy and jurisdiction,
Can read The Public Safety Assessment (PSA), transparent. can choose scores & a flag…
one score per outcome: 6 point scale
- failure to appear (FTA)
- new criminal arrests (NCA)
- flag: new violent criminal arrests (flag, yes or no, 1-3 no, 4-6 yes)
release conditions matrix, no longer recommend detention, better for supervision
FTA vs NCA, matrix , scores get bucketed release conditions, https://www.researchgate.net/publication/338840197_The_impact_of_overbooking_on_a_pre-trial_risk_assessment_tool/figures
aggregate, statistically find the breaks between the likelihoods of failure
maybe score predictive but not the levels
“failure” term getting turned around. even people in high risk is likely to succeed, “success rate” instead
influence of yet another of the same offense in the same day…
judges more lenient on certain times / days
help as a bulwark against how they judge is feeling
stats + professional judgment is better than prof judgment alone
Propublica paper controversial within pretrial risk assessment specialist community, = debunked? Used a different ==definition of fairness==, TP vs FP rates. equally well for everyone?
The people making decisions are not technies, for techies to educate them
in the 1920s they were using actuarial tools for risk assessment
Externalities:
Costs to incarcerating someone. 3 days of detainment tends to be a threshold because people lose their jobs. (model feeding back in on itself, no mechanism to catch the feedback loop)
correlation/regression, similar strength between black & white doesn’t mean its equal.
or just as accurate on AUC-ROC, are you really
how to you extract the bias
proxies:
best practice, use as few relevant factors as possible, Data Scientists love ot throw everything into their model, COMPAS has questions like family ties, job turnover (proxies for race), transparency… PSA has 8 or 9 questions… don’t include race or gender, …[commuity ties, attitudes toward law inforcement not predictive],
what is a proxy for race
often they try to break down by race…
don’t do interview-based tools, PSA can be automated,
try to take out as much proxy for race…
is race correlated with risk score?
Chi squared vs R…
control for age, race, gender,…
multivariate logistic reg
redlining, governments denying resources, running an interstate, lack of the same resources,…, associations indicative of systemic problems.
why are young men more reckless drivers
man on man…
recidivism vs treatment
booking vs filing, e.g. resisting arrest (=violent)
the way that you get categories can impact you in massive ways
labeled as potential gang members
stuck in a category, it follows you
algorithms are labeling you whether you want it or not, and it’s hard to get un-labeled, because often the person is not involved…
adolescent brain science, time in connecting crime with consequence
Cyntoia Brown
high risk, more supervision, more supervision = more likely to fail, poor outcomes, domino effect
judicial discretion - judges may not understand the tool, or it wasn’t implemented well so it’s not helpful, judges are very much about their discretion, “another tool in the toolbox”,
more common for defense attorneys asking for copies of risk assessment, Loomis vs. Wisconsin
pretrail risk assessment is for bail
there are also RA’s used for sentencing. “entered as evidence” in sentencing.
how do we classify RA tools.
vs actuarial methods: more data…Random Forests…?
bail bondsman very slick, manipulating people, lots of money involved
RA is not for setting bail.
RA is supposed to be a solution to over incarceration, to release more people
40% of people in jail are pretrial
is bail linked to outcomes? typically not
“day fine” is % of your income, bail is not linked to income except for maybe flight risk
wealthy person gets released the next day, poor person sits in jail until trial
3,000 different criminal justice systems – municipal (city), county (sherrifs), justice of the peace, – geographically specific, across state lines,… different counties have different laws,… (Dukes of Hazard)
750,000 defendants in PSA… locally validated
real obstacle is having to get everyone to agree to something
PSA can be taken for free and then they can tweak it, just rename it
magistrates do the late night in booking,…
supervision officer is kind of like a probation officer (bringing probation to the pretrial process)
arraignment is often done via video in jail already (not transporting people)
FTA - childcare, transport, babysitter, scared, phone, mail, internet,
Release Conditions Matrix in PSA…
Chan Zuckerberg “clean slate” initiative for expungement, Code for America
Code for Nashville / Memphis…
Scraping & Python,
School of Public Policy & Urban Affairs… EDF
The Jim Crow, the (New Jim Code too)
Cf. Kristian Lum et al’s Jan 2020 paper on PSA: https://www.researchgate.net/publication/338840197_The_impact_of_overbooking_on_a_pre-trial_risk_assessment_tool