Explaining the Predictive Performance of Police in Cases of Domestic Abuse -- by Jeffrey Grogger, Andrew Jordan, Tom Kirchmaier

TEXT START: Police in England and Wales are asked to predict the likelihood of serious recidivism in domestic abuse cases, with little support beyond a flawed questionnaire.

The Dissection

The paper is an empirical autopsy of a legacy risk-assessment system. It shows that police predictions are weak, distorted by censoring, and shaped by representativeness bias, overreaction, categorization, and selective attention. Higher-skill officers partially escape the questionnaire’s limits; lower-skill officers turn extra information into additional noise.

The text is really doing two things: measuring institutional incompetence and locating that incompetence inside individual decision processes. It exposes a broken forecasting layer while leaving the broader institutional structure largely untouched.

The Core Fallacy

Relative to Discontinuity Thesis mechanics, the central error is omission: the paper treats prediction quality as a problem of officer skill, bias, and questionnaire design rather than asking whether the human forecasting function should remain human-controlled at all.

Its findings point toward cognitive automation. A system that can integrate more information consistently, correct for censoring, and identify nonlinear risk patterns would attack the exact weaknesses documented here. Training and better forms may improve the lagging institution, but they do not establish a durable human moat. The paper diagnoses poor human performance without confronting the replacement pressure its evidence creates.

Hidden Assumptions

• That police officers will remain the relevant decision-makers as predictive systems improve.
• That better questionnaires or training can materially repair a structurally noisy human judgment process.
• That predictive accuracy is the primary bottleneck, rather than the larger problems of intervention capacity, legal constraints, data quality, and institutional coordination.
• That the target being predicted remains stable even though police intervention changes the probability of the outcome.
• That human discretion has durable value merely because some high-skill officers extract signal from information absent from the questionnaire.
• That identifying bias is equivalent to possessing the organizational capability to eliminate it.

Social Function

Primary classification: partial truth and transition management.

The paper is valuable because it punctures the comforting fiction that frontline institutional judgment is reliably informed. But it remains transition management when read systemically: it converts a replacement-grade problem into a reform agenda of better training, better questionnaires, and better officer classification. That may reduce harm during the lag period. It does not preserve the human monopoly over risk assessment.

The Verdict

The paper finds a machine-shaped vacancy inside policing. Human officers are being asked to perform high-stakes prediction with weak tools, unstable targets, and inconsistent cognition. The immediate failure is institutional; the terminal implication is labor substitution. Once a superior predictive system can integrate evidence without the documented biases, lower-skill officers become redundant, while higher-skill officers become temporary servitors whose remaining value lies in implementation, verification, and legally accountable intervention—not in prediction itself.