Iterative learning algorithm for exposure assessment used in generalized linear epidemiological models.

Studying exposure effects on health is a key area in environmental epidemiology. As direct measurement is impossible, the estimation requires exposure assessments. Spatio-temporal data are combined with ground-true measurements to build an \emph{exposure model}. The exposure model's predictions at the residences of the subjects of the epidemiological study are then used as covariates in an \emph{epidemiological model}, where parameter estimation is the goal. This process is sometimes called a two-stage study. We call upon ideas from the Domain Adaptation and Experimental Design literature to suggest an iterative importance-weighted algorithm that may replace the two stage process, to achieve more accurate estimates in an epidemiological generalized linear model (GLM). We show the optimality of our algorithm In simulation analysis.​ ​