A Framework for Considering the Impact of Denominator Uncertainty on Small-Area Spatial Regression Studies

Lauren Hund, University of New Mexico

Area-level spatial regression modelling is an important tool for characterizing spatiotemporal trends in health effects analyses. Studies within socioeconomically homogeneous units often outperform aggregation at other levels of spatial resolution. For such studies, small-area population counts stratified by sociodemographics are used as population denominators. We propose a framework for incorporating denominator uncertainty into the spatial regression model. We conceptualize errors in denominators as model misspecification and use model ensembling to incorporate model uncertainty using data-driven model weights. This framework relies on a rich set of projection models; and a validation metric to evaluate the models. We propose a novel set of candidate projection models, along with an ex post facto validation metric. We apply the proposed framework to assess the impact of denominator uncertainty on quantifying temporal socioeconomic trends in breast cancer incidence in Los Angeles county from 1980-2000.