A Comparison of Two Procedures to Fit Multi-Level Data: PROC GLM versus PROC MIXED

This paper presents the advantages of using PROC MIXED versus PROC GLM as a solution for hierarchical data. Analyzing multi-level, non-independent data requires a different methodology from the standard general linear model that is implemented in PROC GLM. A random coefficient (RC) regression model utilizing the SAS ® procedure PROC MIXED can be used when the assumption of independence is not satisfied due to group structure in the data. Data from a health care survey administered to beneficiaries in several different health plans were examined. In this case, beneficiaries are grouped by health plan, and there is a distinct possibility that the beneficiaries ' survey responses are more similar within plans than across plans. In the preliminary analysis to determine the most appropriate model-to-data agreement, the intraclass correlation coefficient (ICC) was investigated. The ICC assesses the degree of clustering or dependency among subsets of cases in nested data. The ICC revealed that clustering was present in the data, and that PROC MIXED was the most statistically appropriate procedure to implement. The PROC MIXED analysis presented here focuses on the detection of clustering and on determining both individual level (level 1) and group level (level 2) predictor variables of physical health. The data are also analyzed using PROC GLM, and the results between the two procedures are compared