B1210
Title: Granularity in deriving environmental exposures \& community characteristics: Impacts on predictive accuracy
Authors: Anushka Palipana - Cincinnati Children's Hospital Medical Center (United States) [presenting]
Emrah Gecili - Cincinnati Children's Hospital Medical Center (United States)
Rhonda Szczesniak - Cincinnati Children Hospital Medical Center (United States)
Erika Rasnick - Cincinnati Children's Hospital Medical Center (United States)
Andrew Vancil - Cincinnati Children (United States)
Daniel Ehrlich - Cincinnati Childrens Hospital Medical Center (United States)
Teresa Pestian - Cincinnati Childrens Hospital Medical Center (United States)
Eleni-Rosalina Andrinopoulou - Erasmus Medical Center (Netherlands)
Pedro Miranda Afonso - Erasmus University Medical Center (Netherlands)
Ruth Keogh - London School of Hygiene and Tropical Medicine (United Kingdom)
Yizhao Ni - Cincinnati Childrens Hospital Medical Center (United States)
John Clancy - Cystic Fibrosis Foundation (United States)
Patrick Ryan - University of Cincinnati (United States)
Cole Brokamp - University of Cincinnati (United States)
Abstract: Nearly 50\% of the variability in lung function measurements from individuals with cystic fibrosis (CF) is attributable to environmental influences. Little is known about how the resolution with which these environmental exposures and community characteristics (geomarkers) are measured leads to biased and/or imprecise estimates and predictions in longitudinal modeling. Although differing resolution has been shown to yield biased associations with health outcomes, research has focused on a limited number and type of geomarkers. In this empirical study, we evaluate geomarker measurements derived for a local CF center cohort ($n = 148$, aged 6 to 20 years, followed from 2012 to 2017) to determine the extent to which geomarker granularity, coded based on US postal zip code or residential address, impacts the accuracy of dynamic prediction modeling of lung function decline. We employ a stochastic linear mixed effects model with target functions tailored to prediction of clinically relevant thresholds of rapid lung function decline in CF. A novel Bayesian selection approach for clinical and geomarker covariates is presented. Findings from the two different derivation types (zip code and residential address) are compared for the real-world CF clinical data and simulation settings. Implications and trade-offs of each derivation are discussed.