Thresholddiscrimination

Thresholddiscrimination refers to discriminatory outcomes that arise when decisions are made by applying fixed or adaptive thresholds to an underlying scoring system. In many predictive processes, a continuous score (such as a risk, credit, or likelihood score) is converted into a binary decision by comparing it to a cutoff. If score distributions differ across groups, the same threshold can yield different false positive and false negative rates, producing disparate impact even when the model itself is not overtly biased.

Causes include differences in base rates between groups, calibration gaps where scores do not map to outcomes uniformly across populations, and thresholds tuned for overall accuracy without regard to fairness constraints. Thresholds may be static or updated over time, potentially reinforcing disparities if monitoring and adjustment are neglected.

Examples commonly discussed include lending decisions, where a single cutoff can approve applicants from one group at higher rates than another; hiring or admissions scenarios where test scores or assessments translate into cutoffs; and criminal justice risk assessments where fixed thresholds can produce unequal error rates across demographics.

Measurement and detection rely on comparing error rates and outcome rates for different groups at the chosen threshold, and on evaluating metrics such as equalized odds, calibration within groups, and disparities in positive predictive value. Visual tools like score distributions and threshold-Outcome analyses can aid assessment.

Mitigation approaches include group-aware or calibrated thresholding to equalize error rates, fairness-constrained optimization that jointly tunes scores and thresholds, and exploring alternative decision rules that reduce reliance on single cutoffs. Ongoing monitoring and context-specific fairness goals are essential, given trade-offs with overall accuracy and utility. See also fairness in machine learning, disparate impact, calibration, and post-processing techniques.