Date of Award

Summer 2019

Document Type

Open Access Dissertation

Department

Educational Studies

First Advisor

Christine DiStefano

Second Advisor

Brian Habing

Abstract

This study focused on understanding how several data characteristics associated with the investigation of effect heterogeneity (i.e., mixing weights, predictor distributions, and the inclusion of covariates) affected enumeration and parameter recovery with regression mixture models. The inclusion of C on X paths, where the latent class, C, is regressed on the predictor, X, allows predictor means to vary across classes, at two points in the model building process—during and after enumeration—was of interest. This main aim was accomplished by comparing the correct enumeration rates and parameter coverage rates with and without freely estimated predictor means across classes for models with two classes, considerable separation between groups, and a total sample size of 500. Findings from this study, in accordance with previous work, indicated that C on X paths, should only be included after enumeration (e.g., Nylund-Gibson & Maysen, 2014). Inclusion of C on X paths functionally frees the estimation of associated predictor means across classes. If these paths are included in the enumeration phase, over-extraction is typical when predictor variance differences are present. Results from this study supported findings from previous research that demonstrated the necessity of including the C on X path when predictor means vary across classes (Lamont, Vermunt, & Van Horn, 2016). Therefore, once the number of classes has been determined, C on X paths should be included in models just as researchers would freely estimate residual variances across classes.

Rights

© 2019, Phillip Sherlock

Share

COinS