During the past decade, there has been a dramatic increase in the use of multilevel modeling in epidemiologic analysis. There were 10 times as many papers in Index Medicus identified by the terms "multilevel' ' and "epidemiology" in 2005 than there were in 1995. In recent years, the American Journal of Epidemiology has published papers that use multilevel analyses to consider the associations between group-level characteristics and a range of health indicators (1-7). Accompanying these

more »

... pecific contributions have been several discussions of the strengths and limitations of multilevel modeling techniques (8, 9) and the rationale for applying these methods to public health (10, 11). In this issue of the Journal, Wight et al. (12) examine whether neighborhood context influences the cognitive function of older adults. This paper contributes to the literature that considers the role of group-level education as a determinant of individual health (13, 14) . The authors show that older adults living in areas characterized by low aggregate education have lower cognitive function than do older adults living in areas characterized by high aggregate education, independent of the participants' own education. In many ways, this analysis is representative of the growing multilevel analysis literature. This study makes use of a large, nationally representative cross-sectional sample in the United States. Participants were geocoded to their census tract of residence, and US Census data were used to characterize census tracts (equated here with "neighborhoods") and to test for an association between neighborhood educational level and individual-level cognition. However, this paper is especially successful on several fronts and provides us with an opportunity to reflect on directions that might fruitfully move the field forward. It is the purpose of this commentary to offer thoughts about 1) specificity of association, 2) multilevel causal pathways, and 3) heterogeneity of associations, motivated by the work of Wight et al. and building on the extant literature on multilevel epidemiologic analyses. SPECIFICITY OF ASSOCIATIONS One key challenge facing studies of the macrolevel context and health is the relative paucity of theoretical development compared with the rapid progress of methodological development (9, 15, 16) . This limitation places epidemiology in sharp contrast with other disciplines, such as sociology, where substantial theoretical work has developed strong frameworks that provide the basis for posing and testing hypotheses on the relations between macrolevel exposures and individual outcomes (17-19) . The principal danger of having underdeveloped theoretical frameworks is that multilevel analyses run the risk of becoming another form of "black box epidemiology" (20), where the associations between "exposures" and "outcomes" are explored, and statistically significant associations are reported, without a theoretical basis for why such associations should exist and why specific factors would cause particular diseases. Although atheoretical epidemiologic explorations can reveal associations that then generate mechanistic hypotheses that can guide future research (21), they risk unearthing statistically significant relations that represent only the play of chance. Multilevel analysis in all aspects of health would do well to focus on evaluating specific theoretically driven hypotheses, transcending black box approaches in order to fruitfully identify macrolevel domains that may influence particular health outcomes. Where no such theoretical or mechanistic work has been conducted for a particular macrolevel determinant, it would be productive for the field if we focused our energy on developing it. Wight et al. (12) present an analysis predicated on a specific neighborhood