Academic journal article
By Gatsonis, Constantine; Daniels, Michael J.
Journal of the American Statistical Association , Vol. 94, No. 445
The presence of substantial variations in the delivery and outcome of health care has been documented extensively in recent years (Dartmouth Medical School, Center for Evaluative Clinical Sciences 1996; Paul-Shaheen, Clark, and Williams 1987; Wennberg and Gittelsohn 1982). Researchers in this area report often large differences in the use of medical care and its outcomes across geographic regions, hospitals, and individual health care providers. The major aim of the analysis is to measure the magnitude of the variations and to assess the role of contributing factors, including patient, regional, and provider characteristics (Diehr 1984; Diehr, Cain, Connell, and Volinn 1990; Gatsonis, Epstein, Newhouse, Normand, and McNeil 1995). In subsequent steps, researchers study the effects of variations in health care use on patient outcomes; they examine the relationship between measures of process, which define regional or hospital practice patterns, and measures of outcome, such as patient mortality, morbidity, functioning, and satisfaction with care. In this formulation, the analysis of variations is focused on comparing the performance of health care providers, such as hospitals, practice groups, or individual physicians, and is commonly referred to as provider profiling (Goldstein and Spiegelhalter 1996; Normand, Glickman, and Gatsonis 1997a; Normand, Glickman, and Ryan 1997b; McNeil, Pedersen, and Gatsonis 1992).
Data on health care use and outcomes have a multilevel structure, usually with patients at the first level and physicians, hospitals, and geographic regions forming the upper-level clusters. Cluster size varies substantially at each level of the hierarchy, and covariates are often available measuring, for example, disease severity and comorbidity for individual patients, and location, size, and organizational characteristics for hospitals. A key analytic goal is to provide cluster-specific estimates of a particular response, such as the rate of utilization of a procedure by hospital or geographic region, adjusted by patient characteristics. Another key goal is to derive estimates of covariate effects, such as differences in health care utilization between patients of different gender or race and practice differences between urban and rural hospitals.
Hierarchical regression modeling provides a general analytic approach that can accomplish these goals. The main purpose of our article is to discuss the application of a broad class of such models, called hierarchical generalized linear models (HGLMs). The response variable in these models is distributed according to a one-parameter exponential family, such as binomial, Poisson, exponential, and Gaussian. Covariates can be discrete or continuous. Models of this type have previously been widely used in the analysis of longitudinal data (Gilks, Wang, Yvonnet, and Coursaget 1993; Laird and Ware 1982; Lindstrom and Bates 1988; Longford 1987). They have also been developed and used in the analysis of effects of contextual factors in the social sciences (Wong and Mason 1985, 1991), in education (Bryk and Raudenbush 1992; Goldstein 1995), in quality control (Natarajan, Ghosh, and Maiti 1998), and in the analysis of spatial data and small area estimation (Breslow and Clayton 1993; Ghosh, Natarajan, Stroud, and Carlin 1998). Applications to the analysis of health care data are relatively recent and have concentrated primarily on binary or multinomial response variables (Calvin and Sedransk 1991; Daniels and Gatsonis 1997; Gatsonis et al. 1995; Gatsonis, Normand, Liu, and Morris 1993; Kahn and Raftery 1996; Malec, Sedransk, Moriarity, and LeClere 1997; Normand et al. 1997a; Normand et al. 1997b). Models for other types of response variables were recently used by Goldstein and Spiegelhalter (1996).
We illustrate the analysis with two examples drawn from a study of variations in the use of cardiac procedures for elderly Medicare beneficiaries who sustained a heart attack. …