Equal Opportunity in the Presence of Risk Distribution Differences

Oregon Convention Center 

The proliferation of risk prediction models and algorithm-assisted decision making has spurred demands for methods to identify, quantify and reduce algorithm bias. One popular notion of fairness, equal opportunity, requires parity in true positive rate (TPR) across subgroups. While intuitively appealing, models constrained to satisfy the equal opportunity condition suffer from a loss in overall accuracy. Moreover, even models with perfect prediction can fail to satisfy equal opportunity if the risk distribution differs between subgroups. In the healthcare setting, the true risk distribution can be expected to differ between subgroups; therefore, fairness measures that account for these differences are needed. We investigate how TPR and related error-rate based metrics are expected to differ between subgroups in the presence of differences in underlying risk distribution and/or subgroup-dependent calibration bias through a combination of theoretical and numerical work. We further propose a modified formulation of the equal opportunity criterion and apply it to a risk prediction model implemented in a large, urban health system.

algorithm bias

fairness

equal opportunity

true positive rate

calibration bias 

Health Policy Statistics Section