Two-Part Modeling of Semicontinuous Longitudinal Variables A Comparison of Approaches

This paper presents specification of two approaches for analyzing a longitudinally observed semicontinuous variable, in which a large proportion of observations equal zero while remaining observations follow a continuous distribution. Both approaches utilize two-part models, where Part 1 models the zero values, such that hypotheses can be examined regarding the likelihood that the outcome equals zero at a particular time point (using the Olsen & Schafer model) or regarding the likelihood of observing the initial onset of a nonzero value (using the "launch model") and Part 2 for the remaining continuous portion of the outcome variable is a standard latent growth curve model. However, interpretation of Part 2 depends on the approach used in Part 1. Parts 1 and 2 are jointly estimated, allowing them to be correlated, and covariates may have differing relationships across the two parts. The approaches are illustrated using a longitudinal study of adolescent alcohol use.