Detecting Bias in Jury Selection
Investigating the presence of human biases in decision making with interpretable methods
Was there bias in jury selection in Mississippi?
To support the case, APM Reports collected and published court records of jury strikes in the Fifth Circuit Court District and conducted analysis to assess if there was a systematic racial bias in jury selection in this district. As part of their analyses, they used a logistic regression model and concluded that there was significant racial disparity in jury strike rates by the State, even after accounting for other factors in the dataset.
Race consistently identified as important with optimal models
With Optimal Feature Selection, we still find that race is consistently selected as the most predictive variable of juror strike outcome at all levels of sparsity. This strengthens the conclusion in the original study, as we knowing that the variable selection is exact.
Furthermore, we observe a significant decrease in model performance when the race variable is removed, indicating this feature is providing unique signal in the data.
Variable importance across all sparsity levels from Optimal Feature Selection
Some subgroups have more prominent racial disparity
The remaining groups show significant disparity. For example, among jurors that have not been accused of crimes but know the defendant, there is an over 60% difference in struck rate between black and non-black.
These subgroups suggest systemic patterns of racial bias in the strike process, and provide direct characterizations of the situations in which black jurors are likely to have experienced discrimination.
Optimal Classification Trees showing the racial disparity in different subgroups
Why is the Interpretable AI solution unique?
Exact variable selection strengthens the conclusion that race is highly improtant
Automatic Subgroup Identification
Optimal Trees automatically identifies subgroups where the racial disparity is the strongest