Exploration of Confounding Due to Poor Health in an Accelerometer–Mortality Study

Supplemental digital content is available in the text. ABSTRACT Purpose Confounding due to poor health is a concern in accelerometer-based studies of physical activity and health, but detailed investigations of this source of bias are lacking. Methods US adults (n = 4840) from the National Health and Nutrition Examination Survey (2003 to 2006) wore an accelerometer for 1 to 7 d (mean = 5.7 d) and were followed for mortality through 2015. Logistic regression was used to examine odds ratios between poor health (chronic conditions, self-reported health, mobility limitations, frailty) and low physical activity levels; Cox models were used to estimate adjusted hazard ratios (HR) and 95% CI for mortality associations for a 1 h·d−1 increase in moderate-to-vigorous–intensity physical activity (MVPA) using two commonly used cut-points (MVPA760, MVPA2020). Modeling scenarios with shorter and longer follow-up time, increasing adjustment for poor health, by age group, and after excluding early years of follow-up were used to assess bias. Results Over a mean of 10.1 yr of follow-up, 1165 deaths occurred. Poor health was associated with low MVPA760 levels and increased mortality risk. In fully adjusted MVPA760 models, HR was 26% stronger comparing 0 to 4 yr (HR = 0.46) with 0 to 12 yr of follow-up (HR = 0.62), particularly in older adults (65 yr and older). Increasing statistical adjustment for poor health attenuated MVPA760 associations by 13% to 15%, and exclusion of the first 2 yr of follow-up had limited effects. Comparable results were obtained with the MVPA2020 cut-point. Conclusions We did not find evidence that confounding by health status resulted in entirely spurious MVPA–mortality associations; however, potential bias was appreciable in modeling scenarios involving shorter follow-up (<6 yr), older adults, and more limited statistical adjustment for poor health. The strength of MVPA–mortality associations in studies reflecting these scenarios should be interpreted cautiously.