# Multilevel Modeling

It is hard to ignore the importance of context when considering any scientific inquiry that is of social or health interest. Context is particularly critical as we move an intervention from a randomized trial to implementation phase. At a more simplistic level, we can account for contextual factors as main effects in our models, for instance, examining informal care in urban versus rural settings. This approach disaggregates group-level information to the individual level so that all predictors in the regression model are tied to the individual unit of analysis. This approach can be problematic as all of the unmodeled contextual effects are pooled into a single error term at the individual level. It is also problematic because individuals from the same context will presumably have correlated errors. Traditional techniques such as ANOVA and ANCOVA ignore the random variability associated with group-level characteristics. Newer analytic methods such as multilevel models allow for more rigorous approaches to testing contextual and structural effects taking into account the nonindependent observation. In these multilevel models, sampling errors are simultaneously appropriated at each level of analysis, which is often not possible using ordinary least squares approaches (see Kenny, Bolger, & Kashy, 2002). Multilevel models are also more flexible in handling missing data and unbalanced designs. Of importance is identifying the contextual factors vis-a-vis a particular intervention and how such factors will be measured.