Multilevel Modeling

Multilevel modeling is the simultaneous use of more than one source of data in a hierarchical structure of units [Page 695]and is useful for analysis of clustered or longitudinal data. Single-level models are unable to accommodate the characteristics of hierarchical data structures, such as unit of analysis, aggregation bias, state dependency, and within-group and between-group heterogeneity, resulting in misestimations. By accommodating these characteristics, multilevel modeling allows researchers to use data more fully and efficiently and to assess the direction and magnitude of relationships within and between contextual and individual factors.

In the past few years, multilevel modeling, also known as hierarchical linear models, mixed-effects models, random-effects models, and random-coefficient models, has become increasingly common in public health research, partly due to growing interest in social determinants of health and partly due to the growing availability of multilevel statistical methods and software.

Multilevel modeling is a powerful research method that promises to complicate and explicate the field of epidemiology. The inclusion of multiple levels of data in simultaneous analysis models allows for more efficient use of data and provides greater information about within-group and between-group effects and relationships. However, attention must be paid to group-level measurement issues and complex data structures, and better social epidemiology theoretical models are needed to guide empirical research.

With the development of research into social inequalities in health in the late 20th century came an interest in social epidemiology that sought to draw focus away from decontextualized individual characteristics and toward their social or ecological context. Such research can be thought of as encompassing multiple layers surrounding and intersecting with the individual, as illustrated by the multiple layers or levels of the ecological model, including intrapersonal, interpersonal, institutional, community, and policy. At the same time, significant contributions to statistical theory for multilevel methods were being made by Dennis V. Lindley and Adrian F. M. Smith, Arthur P. Dempster, Nan M. Laird, and Donald B. Rubin, and others.