Novel Statistical Models for Intensive Longitudinal Analyses of Cancer Control Behaviors

PROJECT SUMMARY This project will develop, test, apply, and disseminate multilevel statistical models and software for estimating effects of intraindividual means, variances, slopes generated from multi-burst and continuous ILD designs to predict cancer control behaviors and outcomes. Cancer remains a leading cause of mortality. Approximately

42% of new cancer cases in the U.S. are viewed as potentially avoidable including 19% caused by smoking and 18% caused by excess body weight, physical inactivity, excess alcohol consumption, and poor nutrition. Intensive Longitudinal Data (ILD) methods, which collect many assessments captured at high density on a micro-

timescale (e.g., seconds, minutes, hours) using real-time data capture methodologies (e.g., Ecological Momentary Assessment [EMA] and accelerometry), offer enormous opportunities for insight into the dynamic nature of cancer control behaviors and outcomes. In ILD studies, it is common to have hundreds to thousands

of observations per subject, and this allows us to model intraindividual parameters comprised of time-varying variables such as means (e.g., how unhappy is a subject, on average, across occasions?), variances (e.g., how erratic is a subject’s mood across occasions?), and slopes (e.g., is a subject’s mood related to feelings of energy

across occasions?). In our prior work, we developed a software, called MixWILD, consisting of a series of two- stage multilevel statistical models testing the effects of intraindividual means, variances, and slopes on time- varying and subject-level outcomes. The next generation of ILD studies has begun to use multi-burst (e.g.,

multiple day EMA periods interspersed with days with no assessment) and continuous (e.g., 24-hour/days per week smartwatch accelerometry) measurement designs, allowing the entire study to extend across months or years. However, available data analysis techniques cannot address common substantive questions that arise

with multi-burst and continuous ILD designs such as does momentary mood variability increase across a year? Also, do month-to-month increases in momentary mood variability predict declines in sleep duration over a year? To address these gaps, we will develop multilevel models capable of (Aim 1) jointly estimating how within-burst

means, variability, and slopes differ between bursts and/or subjects, (Aim 2) testing predictors (either occasion, burst-, person-level) of how within-burst means, variability, and slopes differ between bursts and/or subjects, and (Aim 3) testing whether random effects from Aim 1 predict subject- and burst-level cancer control outcomes. We

will test and apply these statistical features by conducting secondary analyses of data from a multi-burst ILD study of cancer control behaviors and outcome, which conducted mobile sensing, EMA, and accelerometry from 246 emerging adults (ages 18-29) across 12 months. We will also develop, test, and disseminate a stand-alone

software with GUI capable of running these statistical models to be used by applied behavioral and social science researchers. The methods to be developed can easily generalize to a variety of other disease areas such asthma, disordered eating, suicide prevention, HIV risk, medication adherence, and environmental exposures.