Why spatial residuals matter as much as accuracy

When a model reaches a strong AUC, it is tempting to stop there. But an aggregate accuracy score can hide geographic blind spots: a model can perform well on average while systematically over- or under-predicting in specific regions. Mapping out-of-fold residuals as a spatial layer surfaces these blind spots directly, and in DHS-based work these clusters often align with real differences in health system access, provider availability, or terrain. Treating spatial residual mapping as a standard diagnostic step, not an optional add-on, has become part of how I evaluate every model I build.