Multilevel models in health outcomes research: Part II: Statistical and analytic issues

Abstract

Health outcomes research often involves analysis of data that are nested within hierarchical groups, such as patients or nurses in hospitals. An example is study of the outcomes for nurses and patients with HIV/AIDS of hospital care on dedicated AIDS units or conventional units (Aiken, Lake, Sochalski, & Sloane, 1997). The sample comprised 1,304 patients and 687 nurses from two nursing units in each of 20 hospitals. This hierarchical data structure poses serious methodological issues that challenge the assumptions of traditional data analysis methods. A fundamental problem is that clustered observations are not independent. Recall that an assumption of parametric statistics is that observations are independent. The dependency of observations in the same context can lead to misestimated standard errors. A second problem is that traditional regression methods are limited in their ability to model response and effect variables: Only one level can be modeled in one regression. It is well known, however, that the analysis of variables on any of these levels separately can be misleading. These two methodological issues and statistical approaches that address them are described below. Several excellent resources are available to guide the researcher in designs and analyses to address multilevel research questions (Cho, 2003; Park & Lake, 2005; Wu, 1995).