Conditional Mediation in Latent Class Analysis

• In Event: 61.063 - Mixture Modeling in Practice

Abstract

The incorporation of auxiliary variables within a Latent Class Analysis (LCA) can be extended to explore the estimation of conditional mediation with the latent class multinominal variable as mediator. An applied example is used to illustrate how conditional mediation can be implemented in mixture modeling context. Mclarnon & O’neill (2018) provide detailed procedures on how to implement mediation and moderation in mixture models. Expanding on these modeling ideas, this presentation provides a pedagogical implementation for specifying a moderated mediation model in which class-specific indirect effects are estimated allowing for inferences to be made about the mediating relationship specific to each latent class. This important contribution allows for the testing of whether sub-groups (varying by type) mediate the relationship between predictor and outcome auxiliary variables. In mixture modeling, with complex profiles of response patterns, a mediation effect summed across classes may miss valuable information about the class-specific direction and magnitude of indirect effects.
Mplus, 8th Edition (Muthén, & Muthén, 1998-2017) was used to conduct all analyses and annotated Mplus syntax will be provided for pedagogical purposes. This study utilized the Educational Longitudinal Survey (2002-2012) dataset where the latent class construct of interest, ecological expectations. This variable was measured by academic expectations of the students’ influential people in their life (e.g., teachers, parents, peers), measured in 10th-grade, for students who have been perceived to have a disability. The unconditional LCA measurement model was composed of six binary indicators, reported by a cross-section of respondents (i.e., Peer, Parent, Math teacher, English teacher). We focus on students with perception of disability (POD) as reported by one or more influential ecological actors (School administrator (IEP), Parent, Math Teacher, English Teacher, or Student reported participation in special education programming) which was N= 3,515 students (see Hornstra et al., 2010). A binary predictor variable, student expectations of college, and a polytomous distal outcome variable, academic attainment, measured 10 years post-baseline survey were used. To demonstrate the conditional mediation model, the latent class variable ecological expectations (M) served as the mediator between student expectations (X) and student outcomes (Y).
The results of this quasi-experimental design study utilized a causal affects decomposition approach (Mclarnon & O’neill, 2018; Hayes & Preacher, 2014; Muthen & Asparouhov, 2015). The 3-step manual procedure was used to incorporate auxiliary variables after specification of the measurement model (see Nylund-Gibson et al., 2014). In this study, significant class-specific (conditional) indirect effects were found. Direct effects between student expectations and academic attainment were no longer significant after partitioning out the variance of the meditator. Steps will be provided for pedagogical purposes with emphasis on the interpretation of Mplus output files and translation to meaningful results accessible to a diverse educational audience. Reporting standards in mediation models (Hayes & Preacher 2014; Muthen, Muthen & Asparouhov, 2016), in-particular the reporting of bootstrapped asymptotic confidence intervals will be discussed. This paper provides an example of conditional mediation within a mixture modeling context, including an applied demonstration, as well as how it is implemented in the Mplus software program.

Authors

• Adam Garber, University of California - Santa Barbara

• Mian Wang, University of California - Santa Barbara

• Karen L. Nylund-Gibson, University of California - Santa Barbara