Abstract:
|
Although prediction models have come under fire for potentially promoting disparities, we argue that, when properly fit and validated, they are the only way to bring about equity. Our example is the 2013 US Preventive Services Task Force (USPSTF) lung cancer screening guidelines that were based on dichotomizing age/pack-year/quit-year criteria. These guidelines engendered racial/ethnic disparities that USPSTF claimed were reduced in their new 2020 guidelines. We suggest that the disparities, if anything, may have actually increased. We propose that the only proper way to bring about equity is to equalize on an individual’s predicted risk or benefit: “Equal Management of Equal Risk/Benefit”. We demonstrate pitfalls in the prediction model approach, in particular, choice of metric to equalize, and propose a counterfactual approach to reducing the impact of unfairness in estimating benefit for disadvantaged groups who may have lower benefit (than should) because of their higher propensity of receiving worse heath care.
|