2005 HSR&D National Meeting Abstract
3079 — The Use of Multilevel Modeling in Implementation Research: An Example from SCI QUERI
To describe a statistical approach to address two common challenges that commonly occur in implementation research: the clustering of subjects within treatment facilities, and repeated measures of outcomes.
Using two years of survey data from 2,250 subjects at 23 SCI centers, the relationship between influenza vaccination receipt and site interventions, patient characteristics and patient attitudes was examined using multilevel logistic regression models. Chi-square likelihood-ratio tests were used to compare the fit of random-effects models that accounted for clustering within sites and individuals, within sites only, within individuals only, and traditional logistic regression models that do not control for potential correlations within sites or within individuals. Models were estimated using STATA and the GLLAMM package.
In this study, the majority of unexplained variation in vaccination receipt appears to result from differences between and within individuals. While the models that accounted for site clustering did not result in a statistically significant improvement in fit over the model that accounted for repeated measures or the traditional logistic regression model, there were observable changes in standard errors for the site level variables. The model that accounted for clustering within individuals that occurred as a result of the repeated survey of some of the subjects over the two years did have a significantly better fit than the models that did include a random effect for individuals (p =0.000).
Implementation research often involves intervention at patient, provider, and/or institutional levels over time. The advantages of multilevel modeling include the ability to control for site characteristics that may impact outcomes and at the same controlling time for correlations within subjects (in a repeated measures design) or within practitioners. Failure to account for clustering may lead to misleading results. A number of statistical programs estimate multilevel models; the advantages of the GLLAMM package in STATA include the ability to measure up to 3 levels (e.g. time, patient, hospital) simultaneously and the flexibility to estimate models with binary and count outcomes.