Accounting for Errors from Predicting Exposures in Environmental Epidemiology and Environmental Statistics

摘要: In environmental epidemiology and related problems in statistics, it is typically not practical to directly measure the exposure for each subject. Environmental monitoring employed with a statistical model assign exposures individuals. The result form of misspecification that can complicated errors health effect estimates if naively treated as known. error neither “classical” nor “Berkson”, so standard regression calibration methods do apply. We decompose estimation into three components. First, are too small field correlated, independent variability estimating parameters. Second, because they account Third, there bias from using approximate parameters place unobserved true ones. outline three-stage correction procedure separately these errors. A key insight we second part (sampling exposure) by averaging over simulations posterior surface informative outcome. This amounts samples parameters, call “parameter simulation”. One implication preferable use parametric correlation (e.g., kriging) rather than semi-parametric approximation. While latter approach has been found be effective mean fields, does provide needed decomposition non-informative illustrate properties our corrected estimators simulation study present an example statistics. focus this paper on linear models uncorrelated outcomes, but extensions generalized correlated outcomes possible.