Comparing Surface and Linear Estimates of the Black White Hazard Gap in Mortality Modeling

Black–White disparities in U.S. mortality persist. Conventional regression-based hazard models used to estimate group risk ratios may inadvertently distort
the representation of excess risk among Black Americans when mortality is rare, samples are imbalanced, and risk factors are inter-connected. To examine
this representational possibility—what we term inequitable error—we compare a standard discrete-time hazard model with a nonlinear, machine-learned
hazard representation using survival records from the Health and Retirement Study (154,639 person-intervals from 15,259 respondents). A neural-network
architecture is used diagnostically to evaluate how alternative hazard representations affect both the detection and description of mortality disparities.
Whereas the regression model yields a nonsignificant adjusted Black–White coefficient, the nonlinear representation reduces false negatives among Black
women and men and reveals localized peaks in excess mortality that contrast with the more age-normative pattern observed among Whites. Counterfactual
suppression of bio-physical and psychometric inputs does not eliminate this structure. Together, these findings indicate that late-life disparities differ not
only in level but also in age-structured form, and that average hazard representations can render this structure less empirically visible under rare-event and
subgroup-imbalanced conditions.