Chapter 4: Causal Inference in Multilevel Settings

Not many statistical phrases make it into the common vernacular. "Correlation is not causation" is one of them. This phrase is no less true in a multilevel context. Multilevel regression models can be quite complex relative to their standard regression counterparts. Nonetheless, just like standard regression models, the building blocks of multilevel regression models are correlations and partial correlations. Therefore, causal claims from data cannot be justified solely on the basis of a multilevel regression analysis of a data set. Instead, such claims can be justified only by some combination of features of the research design and untestable assumptions. This chapter explores the ways in which multilevel models can (and cannot) help to facilitate appropriate causal inferences from data. In particular, multilevel models may allow a researcher to account for aspects of the research design that would not arise in non-causal settings (e.g., random assignment). Additionally, the assumptions necessary to obtain unbiased (or close to unbiased) estimates of average causal effects may be more credible when multilevel models are utilized than assumptions associated with other analytic approaches. Finally, using multilevel models in settings where causal inference is desired confers many of the same advantages that accrue when using these models in purely descriptive settings.