Identifying and Predicting Unobserved Pathways in Change of States: Introduction to Latent Class Progression Analysis with Covariates

Add to Calendar
Date and Time: 
Monday, November 14, 2016
2:00 pm – 3:30 pm
Ying Liu

Longitudinal data are frequently used to extract and identify the different ways in which people’s belief and behavior change over time. Often times, such trajectories are postulated to be continuous or have some type of functional form (e.g., linear, quadratic). However, this assumption may become inappropriate when the latent state is discrete or discontinuous. For example, a student may navigate the immense space of skills and find her unique way of learning; a smoker’s psychological state in a long-term cessation program may move from being anxious to apprehensive and then to confident as he quits smoking; a credit market may be segmented by people’s varying change in borrowing patterns when they cope with a lengthy period of recession. These different types of pathways, which move from one discrete state to another, are often of primary interest by researchers who want to understand when to best time an intervention or what factors affect the type of trajectory an individual may take.

This talk proposes an exploratory longitudinal methodology, Latent Class Progression Analysis (LCPA). Developed based on latent class analysis, this new technique identifies an individual’s level on a discrete latent variable at a specific time, referred to as his status, and extracts the most likely trajectories of statuses over multiple time points in a population, referred to as progressions. In addition, covariates may be introduced to predict the progressions. The technique is illustrated by a data example in which common pathways of adolescents’ change in risk behavior (e.g., smoking, drinking, and drug using) are identified and further predicted by the person’s mental health as well as the risk behavior among the peers.