Time-varying coefficient cumulative gap time models for intensive longitudinal ecological momentary assessment data with missingness

• Published in: Journal of applied statistics Vol. 49; no. 2; pp. 498 - 521
• Main Authors: Li, Xiaoxue, Anderson, Stewart J., Shiffman, Saul, Zhang, Bo
• Format: Journal Article
• Language: English
• Published: England Taylor & Francis 25.01.2022; Taylor & Francis Ltd
• Subjects: 62-00; Application Notes; Coefficients; Data analysis; Ecological momentary assessment data; gap time models; Health behavior; intensive longitudinal data analysis; inverse probability weighting; Missing data; Other; Patterning; recurrent event analysis; Smoking; Statistical methods; Ecological momentary assessment data; 62-00; gap time models; recurrent event analysis; intensive longitudinal data analysis; inverse probability weighting; missing data
• ISSN: 0266-4763; 1360-0532
• DOI: 10.1080/02664763.2020.1815676

Ecological momentary assessment (EMA) studies investigate intensive repeated observations of the current behavior and experiences of subjects in real time. In particular, such studies aim to minimize recall bias and maximize ecological validity, thereby strengthening the investigation and inference of microprocesses that influence behavior in real-world contexts by gathering intensive information on the temporal patterning of behavior of study subjects. Throughout this paper, we focus on the data analysis of an EMA study that examined behavior of intermittent smokers (ITS). Specifically, we sought to explore the pattern of clustered smoking behavior of ITS, or smoking 'bouts', as well as the covariates that predict such smoking behavior. To do this, in this paper we introduce a framework for characterizing the temporal behavior of ITS via the functions of event gap time to distinguish the smoking bouts. We used the time-varying coefficient models for the cumulative log gap time and to characterize the temporal patterns of smoking behavior, while simultaneously adjusting for behavioral covariates, and incorporated the inverse probability weighting into the models to accommodate missing data. Simulation studies showed that irrespective of whether missing by design or missing at random, the model was able to reliably determine prespecified time-varying functional forms of a given covariate coefficient, provided the the within-subject level was small.