Introduction to Mixed Modeling Procedures


Comparing the MIXED and GLM Procedures

Random- and mixed-effects models can also be fitted with the GLM procedure, but the philosophy is different from that of PROC MIXED and other dedicated mixed modeling procedures. The following lists important differences between the GLM and MIXED procedures in fitting random and mixed models:

  • The default estimation method for covariance parameters in the MIXED procedure is restricted maximum likelihood. Covariance parameters are estimated by the method of moments by solving expressions for expected mean squares.

  • In the GLM procedure, fixed and random effects are listed in the MODEL statement. Only fixed effects are listed in the MODEL statement of the MIXED procedure. In the GLM procedure, random effects must be repeated in the RANDOM statement.

  • You can request tests for model effects by adding the TEST option in the RANDOM statement of the GLM procedure. PROC GLM then constructs exact tests for random effects if possible and constructs approximate tests if exact tests are not possible. For details on how the GLM procedure constructs tests for random effects, see the section Computation of Expected Mean Squares for Random Effects, in ChapterĀ 45: The GLM Procedure. Tests for fixed effects are constructed by the MIXED procedure as Wald-type F tests, and the degrees of freedom for these tests can be determined by a variety of methods.

  • Some of the output of the GLM procedure applies only to the fixed effects part of the model, whether a RANDOM statement is specified or not.

  • Variance components are independent in the GLM procedure and covariance matrices are generally unstructured. The default covariance structure for variance components in the MIXED procedure is also a variance component structure, but the procedure offers a large number of parametric structures to model covariation among random effects and observations.