Multilevel Modeling in Obesity Intervention Research

Purpose: The prevalence of childhood obesity is increasing in the United States and globally. Associated with numerous comorbid conditions, childhood obesity is recognized as a risk factor for multiple chronic conditions and premature mortality in adult life. Healthcare providers in school-based health care centers are at a crucial juncture at which to have measureable impact on childhood obesity. The inherent developmental characteristics of healthcare providers’ practices overtime requires longitudinal, multilevel research designs in order to account for different individual initial practices and their changes. We describe the design decisions and analytic techniques of a longitudinal, multi-site study using the exemplar of a prospective, cluster-randomized controlled trial of web-based training with and without technological decision support for introducing evidence-based, family-centered, culturally sensitive, guidelines for obesity prevention into practice in school-based health centers. 

Methods:  Designing a study to test the efficacy of HeartSmartKids on healthcare providers’ practices requires examination of variability at multiple levels: random samples of children that are nested in school-based health centers, which are nested in states. 

Results: Based on the design and analysis of baseline data from 24 healthcare providers in 24 school-based health centers in 6 states, the results of multilevel multiple regression and multilevel logistic regression are presented with a particular focus on statistical methodological decisions such as centering (i.e., not centering/using raw scores, group mean centering, and grand mean centering), estimating intraclass correlations and design effects, alternate parameterizations of time, and data imputation.  Additionally, we will address estimating sample sizes for similar studies. 

Conclusion: This presentation will provide researchers who use an ecological framework (necessitating a multilevel approach) with a better understanding of rigorous methods to design and subsequently model and interpret their findings.