Workshop

Predictive Inconsistency in Algorithmic Risk Assessment Tools

Instructor: Prof. Galit Shmueli, Chair Professor, National Tsing Hua University

Summary

This workshop examines the use of Algorithmic Risk Assessment Instruments (ARAIs) in decision-making and introduces the concept of predictive inconsistency. It uses criminal justice as the core case, but the issues are applicable across many domains where predictive tools are used.

The session explores how conceptually justified but technically different choices made by ARAI designers can lead to disparate predicted risk scores for the exact same individual. It traces the garden of forking paths across the ARAI development process, highlighting data science challenges such as operationalizing legal constructs, handling missing values, and selecting performance metrics. Finally, it introduces an adaptation of multiverse analysis for visualizing and evaluating predictive inconsistency, with an emphasis on more reproducible and democratically accountable ARAI development.

Instructor Bio

Prof. Galit Shmueli is Chair Professor at National Tsing Hua University.

Learning Outcomes

  • Understand what Algorithmic Risk Assessment Instruments (ARAIs) are and their role in criminal justice decision-making contexts.

  • Define predictive inconsistency and identify how it stems from varying data science choices made during the model development process.

  • Recognize specific data science challenges that introduce uncertainty at different stages of model building, including non-sampling errors, the reference class problem, and predictive generalization.

  • Consider adaptations of multiverse analysis for evaluating and visualizing predictive inconsistency.

Session Structure

  • Introduction: what are ARAIs.

  • Context: the role of prediction in sentencing and the concept of legal consistency.

  • Core concept: defining and exploring predictive inconsistency throughout the steps of ARAI development.

  • Evaluation: methods for visualizing and evaluating predictive inconsistency using adapted multiverse analysis.

  • Conclusions.

Pre-Reading

  • Familiarity with the data science modelling pipeline, including data collection, preprocessing, variable selection, and model testing or deployment.

  • A basic understanding of predictive modelling and familiarity with machine learning algorithms such as classification and regression trees.

Suggested Reading

  • Greene, T., Shmueli, G., Fell, J., Lin, C. F., & Liu, H. W. (2022). Forks over knives: Predictive inconsistency in criminal justice algorithmic risk assessment tools. Journal of the Royal Statistical Society Series A: Statistics in Society, 185(S2), S692-S723.