Download
Filename Size Access Description License

Abstract

Ecological Momentary Assessment (EMA) studies collect self-reported activities, behaviors and emotions intensively throughout the entire study span, and provide valuable information about how subjects' psychological activities evolve over time. While inherited some similarities,from the traditional longitudinal data, the abundance and intensity create its unique characteristics for the EMA data and pose new challenges to the existing statistical analysis.,Statistical methodologies investigating the associations between risk factors and mood regulation in EMA studies have not been studied thoroughly, and there is recent evidence that mood variability, together with mood assessment level, are important metrics in understanding,subjects tendency to problem behaviors. In this dissertation, a series of three studies was conducted to systematically investigate the effect of certain psychosocial factors on mood regulation in EMA studies, using both metrics of mood variability and mood assessment level. The novel statistical methods can be extended to more general frameworks where a broad spectrum of related statistical and substance problems can be solved.,The methods developed in this dissertation were motivated by an EMA adolescent mood study. First, a three level mixed effect location scale model that includes multiple random subject and wave effects in both the mean and within variance model was developed for hierarchical EMA data where subjects were measured at multiple waves, and at each wave, data were intensively collected over time on each subject. The proposed model allows heterogeneous variance at baseline as well as variance change over time, adjusting for observed covariates. Second, a shared parameter model was framed to address the non-ignorable missing responses in EMA studies, where missing indicator and longitudinal outcomes are jointly modeled by sharing common but unobserved subject level information. The missing indicator was modeled via random intercept logistic regression, and outcome by random location and scale intercept regression, with the three random effects all representing subject specific traits. The model allows subject's missing propensities to influence his/her mood level and variability. Third, a mixed location scale Hidden Markov Model was explored to classify subjects into distinct mood states at each time point, with latent mood states at sequential time points form a Markov Chain. This model allows differential effects on mood regulation for subjects with different mood states.,All models in the above studies were estimated via Bayesian sampling framework by Stan. The model estimation procedures are computational more efficient compared to the maximum likelihood based methods. Extensive simulation studies were conducted to validate the model performance and compare with the existing methods. The proposed models were applied to each of the motivating data set with interpretable results and insightful conclusions. Finally, a discussion of the advantages, limitations as well as future directions were included,in the end of each method chapter.

Details

Additional Details

Actions

from
to
Download Full History