Abstract: Bayesian Multilevel Latent Class Analysis: Exploring the Impact of Social Context (Society for Prevention Research 27th Annual Meeting)

422 Bayesian Multilevel Latent Class Analysis: Exploring the Impact of Social Context

Schedule:
Thursday, May 30, 2019
Regency B (Hyatt Regency San Francisco)
* noted as presenting author
Rashelle J. Musci, Ph.D., Assistant Professor, The Johns Hopkins University, Baltimore, MD
Amie F. Bettencourt, Ph.D., ChiPP Project Director, The Fund for Educational Excellence, Baltimore, MD
Albert Delos Farrell, PhD, Professor, Virginia Commonwealth University, Richmond, VA
Katherine Masyn, PhD, Professor, Georgia State University, Atlanta, GA
Introduction: With a growing literature related to latent classes/profiles of aggression and peer victimization during childhood, there has been a push towards increasing model complexity to incorporate school level information. Increases in complexity may not be explicitly feasible under the most common estimation framework but could be easily implemented within a Bayesian framework. Bayesian Multilevel latent class (BMLC) models can capture heterogeneity in the data at both level 1 and level 2 and include information from prior research to take advantage of the current state of science (Vidotto, Vermunt, & van Deun, 2018). Because it is becoming increasingly common for studies to focus on school contextual influences on aggression and peer victimization (e.g., Henry et al., 2011), we propose using BMLC to model the role of school climate on aggression and peer victimization among students nested within schools, across four sites.

Method: Data from the Multisite Violence Prevention Project (MVPP) on two cohorts of students (N=5,106) from 37 schools in four communities (MVPP, 2004) are used. Data were collected as part of a randomized trial where schools were randomly assigned to one of four conditions: a universal intervention, a selective intervention, a combined intervention, or no intervention. This study uses data from fall of 6th grade prior to the intervention for each cohort. Students completed measures of aggression, victimization and school climate including norms supporting aggression and nonviolence, student-teacher relationships, and awareness and reporting of peer victimization; their teachers completed measures of student-teacher relationships, staff relationships, and awareness and reporting of peer victimization. We use BMLC models to identify classes of aggression and peer victimization (level 1) and examine the influence of school climate (level 2) on classes, building on models developed by Vermunt (2003) and implemented, among others, in Henry and MutheĢn (2010)

Results: The priors used for the BMLC will be captured from a study by Bettencourt and Farrell, 2013, which found 4 latent classes: non-victimized aggressors, aggressive-victims, predominantly victimized and well-adjusted. Additional analyses are being conducted to examine the influence of school climate at level 2 on level 1 classes.

Discussion: Use of Bayesian estimation methods are not terribly commonly in prevention science but could prove to be incredibly useful with the increasing complexity of data available to researchers. The Bayesian approach is appealing because of the use of informative priors, reduced computation time, and the ability to handle a small number of clusters.