Accounting for heterogeneity in large-scale observational studies

Abstract

Large-scale observational studies that collect data from wearable devices provide a unique opportunity to analyze population level data on an individual scale. There is significant heterogeneity both within individuals (intra-individual) and between individuals (inter-individual). Intra-individual heterogeneity refers to variability due to personal day-to-day changes. Inter-individual heterogeneity focuses on the differences between people, possibly due to sociodemographic or behavioral factors. It is essential to use statistical models that address these complexities. We focus on the sleep-wake cycle and physical activity, which have been shown to be important determinants of mortality. Using data from the National Health and Nutrition Examination Survey (NHANES), we leverage quantitative physical activity and light data from wearable physical activity monitors to accurately quantify the sleep-wake cycle and physical activity. We propose a novel joint latent class model (JLCM) that combines a mixture of hidden Markov models for the longitudinal physical activity and light data with a Cox proportional hazards model for survival outcomes. From our models, we identify distinct behavioral profiles associated with mortality risk. Reduced wake physical activity and increased sleep physical activity are both linked to increased mortality. Furthermore, we demonstrate that compared to our JLCM, conventional two-stage modeling approaches underestimate these effects. Our JLCM provides a robust framework for uncovering latent behavioral classes that contribute to mortality risk.