|
Abstract
|
A motivating example from criminal recidivism
Optimal decisions about bail or early release rely on predictions for how likely the person seeking these outcomes is to engage in future criminal behavior. As in suicide risk prediction, this process increasingly involves standardized models. ProPublica recently analyzed over 10,000 of the actual predictions from a popular recidivism prediction model (COMPAS), finding clear evidence of racial bias (6). Black defendants were twice as likely as white defendants to receive a false positive...
Uneven base rates cause unavoidable predictive disparities
The paradoxical disparity in the COMPAS model is unlikely due to biased data, questions, or even the model. Instead, the predictive disparity is likely caused by uneven base rates on the outcome being predicted.(5) To explain, there are many important accuracy metrics for predictive models to maximize. The most popular are (a) sensitivity, (b) specificity, and (c) probability of false alarms (i.e., proportion of false positives over all test positives, equal to 1 - positive predictive value...
The cost of predictive disparities is non-ignorable
All predictive models make errors and all errors have costs. For even a false alarm on a suicide screener, the patient whose future was mis-predicted will often (a) pay a monetary fee for unnecessary follow-up, (b) and potentially a cost of lost trust in the provider - an issue especially salient for historically marginalized communities health systems already struggle to serve...
The new frontier of predictive fairness in suicide prediction
The growing field of algorithmic fairness (2,3) proposes that a model's fairness can be measured by how far it deviates from equally accurate prediction across any metric of interest.2 As an example, in Table 1, we would say that fairness is high, when the false alarm disparity is low. Fortunately, the trade-off between...
Where do we go from here?
A new commitment is needed to large scale and prospectively designed studies that investigate the full scope of this problem and the optimal alternatives, beyond just what is currently possible with existing secondary data sources. These new studies must address not only traditional measures of cost (e.g., money, time to patients), but also new and more inclusive measures of screening mistakes (e.g., eroded community trust in healthcare systems), as well as community stakeholders' preferences...
Language: en |