Attribute-contrained performance estimates
raamana opened this issue · comments
Often, performance is estimated in aggregate on the entire test-set (in a split), regardless of their covariate- or attribute characteristics.. hence looking into performance estimates segregated by certain characteristics of covariates would reveal additional insights - such as difference in AUC in men vs women, or across different age bins, or between sites etc.
This is closely related to calculations of a model bias towards men, women and other sub-groups etc
check this for more: https://twitter.com/melanieganzben1/status/1270732563297026048?s=20