Closed case files can seed race-neutral predictive models for prosecutorial decisions

Evidence

Fortunately, prosecutors’ offices are in possession of information that can be used to form clear and unbiased baselines: closed case files. Employing advanced methods of statistics and computer science, they could use closed case files as a corpus from which to build a model that can produce raceneutral case outcomes. Specifically, they could build a computational model that enables introduction, early in the development of a new criminal case, of an anchor to prosecutorial decision-making. That is, the model would receive the police report and would produce a prediction as to how the case would resolve (based on tens of thousands of previous cases) if the defendant were treated race-neutrally (i.e., all else held equal, the defendant will not receive a different outcome if he is, say, White rather than Black). This prediction could be used by the prosecutor as an anchor; if he or she deviates from the model’s prediction, the reason for the deviation can be described by the prosecutor and incorporated into the model moving forward. — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1215

Categorization

Category: Inspiration and Speculation

• QUESTION: Would prosecutors actually engage with an anchoring prediction they can override, and does the override documentation become another bias channel? Worth chasing: any pilots of anchored sentencing/charging systems.
• SEED: A piece on ‘algorithmic self-rating’: how software that predicts the race-neutral outcome could shift responsibility in prosecutorial discretion, and whether override logs discipline or merely launder bias.
• XREF: Connects to literature on anchoring effects in judgment and on COMPAS-style risk scoring criticized for encoding bias in features.

Related

• Prosecutors can benchmark race-neutral decisions using models built from closed case files. (similarity 0.93)
• Predictive models can suggest case outcomes under counterfactual race-neutral processing. (similarity 0.81)
• A suggestive model removes race to predict race-neutral case outcomes (similarity 0.78)