Introduction:
COMPAS, or Correctional Offender Management Profiling for Alternative Sanctions, serves as a risk assessment tool for US courts that predicts the likelihood of a defendant reoffending. Our group aimed to criticize this algorithm for its lack of accuracy and root out potential bias it contains. Data was gathered from a group of 10,000 defendants in Broward County, Florida. We completed a detailed comparison of risk scores as compared to actual recidivism rates 2 years after defendants were assigned their risk score. Analysis of the tool involved finding false positives and negatives within the data set and finding which variables within could influence these incorrect findings.
Purpose:
-
Evaluate the predictive performance of the COMPAS tool.
-
Identify instances of racial bias in risk classification.
-
Use qualitative analysis to uncover cultural and political implications embedded in the algorithm’s data structure.
Methods:
-
Quantitative Analysis: We compared COMPAS risk scores to actual recidivism outcomes over a 2-year period, identifying false positives and false negatives.
-
Qualitative Analysis: Using connotative and deconstructive reading strategies, we explored how cultural assumptions, misrepresentation, and systemic bias may influence data collection and algorithmic outcomes.
Results and Analysis:
-
COMPAS demonstrated a low accuracy rate, with only 20% of its predictions for violent recidivism being correct. This raises concerns about the reliability of its assessments in dangerous situations.
-
Racial disparities in COMPAS assessments revealed that Black defendants were misclassified as high risk for violent recidivism at twice the rate of White defendants. This finding indicates a significant bias in the algorithm's predictions.
https://prezi.com/view/2O5R8kE1gk4YhWTXdj1r/
