Bayesian Nonhomogeneous hidden Markov models to leverage routine in physical activity monitoring with informative wear time

Abstract

Missing data is among the most prominent challenges in the analysis of physical activity (PA) data collected from wearable devices, with the threat of nonignorabile missingness arising when patterns of device wear relate to underlying activity patterns. We offer a rigorous consideration of assumptions about missing data mechanisms in the context of the common modeling paradigm of state space models with a finite, meaningful, set of underlying PA states. Focusing in particular on hidden Markov models, we identify inherent limitations in the presence of missing data when covariates are required to satisfy common missing data assumptions. In response to this limitation, we propose a Bayesian non-homogeneous state space model that can accommodate covariate dependence in the transitions between latent activity states, which in this case relates to whether patients' routine behavior can inform how they transition between PA states and thus support imputation of missing PA data. We show the benefits of the proposed model for missing data imputation and inference for relevant PA summaries. Our development advances analytic capacity to confront the ubiquitous challenge of missing data when analyzing PA studies using wearables. We illustrate with the analysis of a cohort of adolescent and young adult (AYA) cancer patients who wore commercial Fitbit devices for varying durations during the course of treatment.