Comparing demographic and neural-network-based representations of the black–white mortality hazard gap

Abstract

Black–White disparities in U.S. mortality persist, but existing analyses may inherit a representational problem that I term inequitable error —socially patterned false-negative error that can obscure observed deaths among Black Americans even in the presence of significant average mortality differences. This study proposes an analytic probe into inequitable error by rendering alternative representations of mortality risk via a neural network to recover risk that conventional benchmarks may leave obscured. Using Health and Retirement Study survival records from 154,639 person-intervals nested within 15,259 respondents, regression benchmarks affirm age-graded Black–White mortality patterns but correctly classify few deaths under conventional and prevalence-calibrated thresholds, with lower sensitivity among Black Americans. The neural-network-based predictions contain substantially fewer false negatives across race–sex groups while locating potential sources of distortion among Black Americans in relation to their biophysical and psychometric inputs. The NN hazard representation of Black and White mortality risk can therefore complement regression-based demographic benchmarks by providing insight into when, where, and for whom observed deaths are most vulnerable to false-negative displacement in scalable applications across clinical, academic, and intervention settings and varied data environments.