B0232
Title: Causal identification of dynamic effects in non-stationary time series from N-of-1 observational mobile health data
Authors: Xiaoxuan Cai - Columbia University (United States) [presenting]
Abstract: Mobile technology (e.g., mobile phones and wearable devices) enables unprecedented monitoring of social interactions, symptoms, and other health and behavioral conditions among individuals. Continuous monitoring of personal data generates multivariate time series data of outcomes, exposures, and confounding variables in an N-or-1 study. Popular methods for univariate time series or longitudinal data assume stationary time series or time-invariant treatment effects, which fail to capture the dynamic treatment effect in both short- and long-term non-stationary time series. We propose a set of causal estimands for non-stationary multivariate time series in N-of-1 studies, in order to systematically summarize time-varying treatment in the short and long term, and demonstrate their identification via the g-formula in the presence of exposure- and outcome-covariate feedbacks. The g-formula employs an innovative state space model to account for the time-varying treatment effects in non-stationary time series in an N-of-1 setting. We demonstrate the estimation of proposed estimands using a smartphone observational study of bipolar patients, the Bipolar Longitudinal Study, in which we examine both the short- and long-term effects of digital social interaction on psychiatric symptoms, as well as how these effects change over time. A novel positivity validation plot for testing the positivity assumption in time series studies is proposed.
