P1: FQPMarch 16, 2000 10:47 Annual Reviews CHAP-09?Annu. Rev. Public Health. 2000. 21:171–92Copyrightc 2000 by Annual Reviews. All rights reservedMULTILEVEL ANALYSIS IN PUBLIC HEALTHRESEARCHAna V. Diez-RouxDivision of General Medicine, Columbia College of Physicians and Surgeons, andDivision of Epidemiology, Joseph T. Mailman School of Public Health, ColumbiaUniversity, New York, New York; e-mail: [email protected] Words methods, contextual effects, random effects, social epidemiology,ecologic■ Abstract Over the past few years there has been growing interest in consider-ing factors defined at multiple levels in public health research. Multilevel analysishas emerged as one analytical strategy that may partly address this need, by allowingthe simultaneous examination of group-level and individual-level factors. This paperreviews the rationale for using multilevel analysis in public health research, summa-rizes the statistical methodology, and highlights some of the research questions thathave been addressed using these methods. The advantages and disadvantages of multi-level analysis compared with standard methods are reviewed. The use of multilevelanalysis raises theoretical and methodological issues related to the theoretical modelbeing tested, the conceptual distinction between group- and individual-level variables,the ability to differentiate “independent” effects, the reciprocal relationships betweenfactors at different levels, and the increased complexity that these models imply. Thepotentialities and limitations of multilevel analysis, within the broader context of un-derstanding the role of factors defined at multiple levels in shaping health outcomes,are discussed.INTRODUCTIONThe term multilevel analysis (or hierarchical modeling) has been used in the fieldsof education (4), demography (44, 71), and sociology (19) to describe an analyticalapproachthatallowsthesimultaneous examinationoftheeffectsofgroup-levelandindividual-level variables on individual-level outcomes. Over the past few years,interest in the use of multilevel analysis to investigate public health problems(14, 23,96) has grown. This growth has been stimulated in part by a resurgence ofinterest in the potential ecological-, macro-, or group-level determinants of healthand the notion that variables referring to groups or to how individuals are relatedto each other within groups may be relevant to understanding the distribution ofhealthoutcomes(14, 22, 84, 93,96). A seconddrivingforcein the useof multilevel0163-7527/00/0510-0171$14.00 171Annu. Rev. Public. Health. 2000.21:171-192. Downloaded from arjournals.annualreviews.orgby JOHNS HOPKINS UNIVERSITY on 07/04/07. For personal use only.P1: FQPMarch 16, 2000 10:47 Annual Reviews CHAP-09?172 DIEZ-ROUXmethods has been the accelerated development of the statistical methods them-selves (as well as the accompanying software) and the recognition that they haveapplications in a broad range of circumstances involving nested data structures.The availability of these complex statistical methods challenges public healthresearchers to articulate theories of the causes of disease that bring together factorsdefined atdifferent levels. This willensurethatthe method does notbecomean endin itself, but rather serves as a tool to investigate more sophisticated and hopefullymore realistic models of disease causation. This paper (a) reviews the rationalefor using multilevel analysis in public health research; (b) describes the funda-mentals of the methods involved and how they compare with traditional methods;(c) highlights selected areas in which these methodologies havebeenapplied in theliterature; and (d) summarizes the potential and limitations of multilevel analysisin achieving a more comprehensive understanding of the determinants of healthand disease. Although the focus of this review is on the use of multilevel analysisto investigate research questions involving groups and individuals nested withinthem, other applications are also briefly mentioned.RATIONALE FOR THE USE OF MULTILEVEL ANALYSISTheideathat individualsmay be influenced bytheirsocial contextisa keynotion inthe social sciences, and has led to much debate and empirical research on the inter-actionsbetweenattributes of groupsandattributes ofindividuals (2,19,46,47, 95).Incontrast,despitethe factthat healthanddisease occurinsocial contexts,researchinto the determinants of health has often been characterized by individualization,that is, explaining individual-level outcomes exclusively in terms of individual-level independent variables. The underlying assumption is that alldiseasedetermi-nants are best conceptualized (and consequently best measured) at the individuallevel. Group-level variablesare used only asproxiesfor individual-leveldatawhenthe latter are unavailable. Populations (or groups) are thought of as collections ofindependent individuals, rather than entities with properties that may affect indi-viduals within them. Consequently, there is generally little interest in examininggroup-to-group variation per se. Although there has been abundant discussion inthe epidemiologicliteratureofthe fallacy inherentinusingdata at one level to drawinferences at another level (specifically of the ecological fallacy), until recentlythere has been relatively little discussion of the substantive problem of ignoringpotentially important variables that are best conceptualized and measured at thegroup level. Just as studies examining differences between groups may need totake into account possible differences in group composition (i.e. characteristics ofthe individuals within them), studies of individuals may need to take into accountdifferences in the properties of the groups to which individuals belong (14).In explaining the occurrence of a given phenomenon, researchers can appeal todifferent types of theories, which may be more or less relevant depending on theAnnu. Rev. Public. Health. 2000.21:171-192. Downloaded from arjournals.annualreviews.orgby JOHNS HOPKINS UNIVERSITY on 07/04/07. For personal use only.P1: FQPMarch 16, 2000 10:47 Annual Reviews CHAP-09?MULTILEVEL ANALYSIS 173particular question being investigated (9,19). In the simplest case, the outcome atone level is explained by independent variables that apply to the same level. This istheapproachcommonlytakeninepidemiologywhenindividual-leveloutcomesareexplained in terms of individual-level variables (as in traditional cohort