Session: Challenges in Multilevel Analysis: Implications in School Based Prevention Science (Society for Prevention Research 27th Annual Meeting)

3-057 Challenges in Multilevel Analysis: Implications in School Based Prevention Science

Schedule:
Thursday, May 30, 2019: 3:00 PM-4:30 PM
Regency B (Hyatt Regency San Francisco)
Theme: Innovative Methods and Statistics
Symposium Organizer:
Rashelle J. Musci
Discussant:
Jessaca Spybrook
Schools and other institutions are prime locations for prevention programming. This multi-level context raises a number of methodological challenges in terms of addressing power, exploring causal inferences, and modeling complex associations. The goal of this symposium is to discuss challenges associated with multi-level analysis and offer innovative modeling solutions which are flexible enough to be applied in many contexts. This symposium represents the conference theme of innovative methods and statistics

A primary goal of prevention research is to unpack the processes whereby an intervention impacts an individual- or cluster-level mediator variable and how changes in the mediator are then translated into changes in an outcome. However, prior work provides limited guidance on the best way to model cluster-level mediators. The first paper, “Experimental Design for Studies Probing Multilevel Mediation,” presents a flexible framework to direct the effective and efficient design of multilevel mediation studies under a wide range of considerations. This knowledge has important implications in the implementation of prevention programs.

Unbalanced cluster sizes are another common problem encountered in school-based prevention research. The second paper, “Unbalanced cluster sizes and variation in sampling in cluster randomized controlled trials: What’s the impact on power?,” uses simulated data to examine the effects of unbalanced cluster sizes and assumptions behind that variability on power and parameter estimation. Understanding the impact these factors have on study power is essential as we move toward a research framework emphasizing rigor and reproducibility. Results are discussed with an emphasis on incorporating level 1 sample sizes as well as the sampling coverage/proportion of level 1 units into power calculations.

Appropriately modeling complex contextual influences in studies with small numbers of clusters is another common challenge in school-based research. The third paper, “Bayesian Multilevel Latent Class Analysis: Exploring the impact of Social Context,” uses data from two cohorts of over 5,000 students from four locations in the United States who participated in the Multisite Violence Prevention Project to characterize student level behavior and model the impact of school-level school climate on student-level behaviors. A Bayesian multilevel Latent Class model is proposed as a potential solution, offering reduced computational time and use of informative priors.

Finally, a discussant will highlight commonalities among the papers, discuss implications for prevention, and moderate a discussion between the presenters and the audience.


* noted as presenting author
420
Experimental Design for Studies Probing Multilevel Mediation
Ben Kelcey, PhD, University of Cincinnati; Kyle Cox, BS, University of Cincinnati; Yanli Xie, BS, University of Cincinnati
421
Unbalanced Cluster Sizes and Variation in Sampling in Cluster Randomized Controlled Trials: What’s the Impact on Power?
Joe Kush, PhD, University of Virginia; Catherine Bradshaw, PhD, University of Virginia
422
Bayesian Multilevel Latent Class Analysis: Exploring the Impact of Social Context
Rashelle J. Musci, Ph.D., The Johns Hopkins University; Amie F. Bettencourt, Ph.D., The Fund for Educational Excellence; Albert Delos Farrell, PhD, Virginia Commonwealth University; Katherine Masyn, PhD, Georgia State University