Ecological inferences and multilevel studies

Abstract

Describing area-based differences in health outcomes has a long history (Kawachi and Berkman Neighborhoods and health. Oxford University Press, Oxford, 2003), and evidence of ecologic variations in health sparks interest from multiple perspectives. In particular, researchers investigate these ecologic variations for surveillance and monitoring of health disparities (Krieger et al. Am J Public Health 95:312-323, 2005) and to understand the impacts contexts have on individuals. In the latter category, causal questions motivated by evidence of area-based differences in health include the following: How much do contexts, such as neighborhoods, impact health? What is the impact of a specific contextual exposure on health? How do contexts mediate the effects of individual-level health risk factors? Ecologic factors may have tremendous importance for population health (Kawachi and Berkman Neighborhoods and health. Oxford University Press, Oxford, 2003), underscoring the value of recognizing opportunities and methodological challenges for causal inference when ecological variations in health are present. We address these issues as follows: we begin by identifying what constitutes a multilevel data analysis and present a discussion on how a range of data structures that are observed in the real world, or due to sampling design, can be accommodated within a multilevel framework. We discuss the types of research questions that typically motivate multilevel analyses and contrast the application of multilevel methods against other approaches for answering such questions with an emphasis on causal inference. After laying down the substantive motivation to utilize multilevel methods, key statistical models are specified with a description of the properties of each model. We close by presenting extensions to the basic multilevel model that allow us to incorporate realistic complexity in our analyses.