Title: Latent variable regression analysis of longitudinal multivariate data with irregular and informative observation times
Authors: Zhenke Wu - University of Michigan (United States) [presenting]
Abstract: In many mobile health studies, phone surveys such as Ecological Momentary Assessments (EMA) are increasingly adopted because they are less susceptible to recall bias and are sensitive to contextual factors. For example, they hold great potentials in smoking cessation studies to probe subjects' time-varying psychological states such as vulnerability (risk for lapse) and receptivity (ability and willingness to engage with self-regulatory activities). Inference and prediction of these states may inform just-in-time adaptive intervention development. However, the observation times of these EMAs may correlate with survey responses. For instance, some EMAs are delivered and answered with lower positive emotions when they were triggered by a recent smoking episode detected by on-body sensors. Such dependence must be accounted for to obtain valid inference. We propose a latent variable regression approach for longitudinal multivariate discrete data analysis with irregular and informative observation times. The goal is to infer the distribution of scientifically meaningful latent variables over time as a function of covariates. The observed dependence between the survey responses and observation times is assumed to be induced by unobserved random effects and observed covariates. We demonstrate the utility of the proposed model through simulation studies and an analysis of data from Break Free study among African Americans who attempt to quit smoking.