NIOSHTIC-2 Publications Search
Analysis of large hierarchial data with multilevel logistic modeling using PROC GLIMMIX.
Li-J; Alterman-T; Deddens-JA
Proceedings of the 31st Annual SAS Users Group International Conference, March 26-29, 2006, San Francisco, California. Cary, NC: SAS Institute Inc, Paper No. 151-31, 2006 Mar; :1-5
Studies that combine individual-level and aggregate data are common in epidemiologic research. Such studies are often subject to ecological fallacy which arises from confounding of the individual-level relationship due to heterogeneity of exposure variables and covariates within groups. One approach to address this concern is to use multilevel modeling. The advantage of using multilevel modeling is that it takes the hierarchical structure of the data into account by specifying random effects at each level of analysis, and thus results in a more conservative inference for the aggregate effect. In this study, we combined data from two databases for analysis. Data from the National Occupational Mortality Surveillance System (NOMS) containing individual-level information from death certificates was linked by occupation to the Occupational Information Network (O*NET) which contains job characteristics at the occupational level. We examined the adjusted association between job characteristics and select causes of death. A recently available generalized linear mixed models procedure, PROC GLIMMIX, was used to fit the multilevel logistic regression model to our data. Results are compared to those obtained from logistic regression modeling that ignores the hierarchical structure of the data. Results demonstrate the potential of drawing incorrect conclusions when multilevel modeling is not used. Problems encountered from use of PROC GLIMMIX with large data sets will be discussed.
Models; Epidemiology; Mortality-rates; Mortality-data; Occupational-health; Occupational-hazards; Surveillance-programs
Jia Li, Constella Group, LLC, 5555 Ridge Ave, Cincinnati, OH 45213
Work Environment and Workforce: Special Populations
Proceedings of the 31st Annual SAS Users Group International Conference, March 26-29, 2006, San Francisco, CA