Title: Statistical properties of sub-cohort selection when testing interactions in biomarker studies
Authors: Leann Long - University of Alabama at Birmingham (United States)
Stephanie Tison - University of Alabama at Birmingham (United States)
Suzanne Judd - University of Alabama at Birmingham (United States)
George Howard - University of Alabama at Birmingham (United States)
Mary Cushman - University of Vermont (United States)
Leslie McClure - Drexel University (United States) [presenting]
Abstract: The Reasons for Geographic And Racial Differences in Stroke (REGARDS) study is a cohort of over 30,000 participants concerned with understanding racial and regional disparities in cardiovascular disease risk factors and stroke in the US. When examining the impact of biomarkers on cardiovascular disease risk factors, and particularly the differential effect of biomarkers by race, it is not financially feasible to assay these biomarkers in all participants. One strategy is to measure the biomarkers in a sub-sample of the cohort, but it isn't clear how to best choose the sample. We assessed different approaches to selecting the sub-cohort, in order to maximize power to detect interactions, while minimizing bias and maximizing the coverage of the 95\% confidence interval. We simulated sub-samples of $n=4000$, with characteristics of the participants based on the REGARDS cohort, to estimate these operating characteristics. We considered 3 assumptions: simple random sampling, sampling with equal allocation across race, and sampling with equal allocation across sex-race groups, and compared the statistical properties observed to those when using the full cohort. We will discuss the results of our simulations and practical implications for implementing the sampling and subsequent analyses.