• Print |
• E-mail  |
• Bookmark

PROC GLM and PROC MIXED do not estimate the model using the same methods. For the estimation of random effects, PROC GLM uses method of moments, and PROC MIXED uses the maximum likelihood (ML) or restricted/residual maximum likelihood (REML) method. These two estimation methods can lead to different results in many situations. PROC GLM and PROC MIXED also differ in the approach they take to setting up the random effects in the model. PROC MIXED defines random effects as truly random, while PROC GLM defines all effects in the model as fixed and then adjusts for the random effects after they have been estimated.

In general, PROC MIXED and PROC GLM will agree if the models contain only fixed effects. If you do not have a RANDOM or REPEATED statement in your model, then the Type III F-tests and the solution vectors should be the same for the two procedures.

You might not get the same results if your model includes a RANDOM statement, especially if you have unbalanced data. PROC MIXED by default includes a lower boundary of 0 on all random effects in the model. You can remove this lower boundary with the NOBOUND option in the PARMS statement. With the NOBOUND option, you might be able to reproduce your results from PROC GLM. Remember also that by default PROC GLM's Type I and Type III F-tests for the fixed effects in the model are tested against the residual error. PROC MIXED tests the fixed effects against the appropriate error term as determined by the model specification. You might have to use the TEST option in the RANDOM statement or the TEST statement to duplicate the tests that are given by PROC MIXED. You will also need the DDFM=SATTERTHWAITE option in the MODEL statement in PROC MIXED because PROC GLM by default uses this approximation in unbalanced data situations.

PROC GLM with a REPEATED statement has different data requirements from PROC MIXED. PROC GLM requires that you have complete data for a subject. If you have just one missing time point for a particular subject, then that entire subject will be deleted from the PROC GLM analysis. PROC MIXED can take subjects with incomplete data into its analysis. You should be able to diagnose this problem by looking at the degrees of freedom for your tests. If the degrees of freedom are significantly different between PROC GLM and PROC MIXED, then you probably have missing data for at least one of your subjects. You might be able to replicate the PROC MIXED results in PROC GLM by switching to a split-plot (or univariate) approach. For details, see "Repeated Measures Analysis of Variance" in the GLM procedure chapter of the SAS/STAT User's Guide. PROC GLM also makes some assumptions about which between-subject/within-subject interactions to include in the model. You will need to include these interactions explicitly in your MODEL statement in PROC MIXED to get the results from the two procedures to match.

See this reference comparing PROC GLM and PROC MIXED for repeated measures models.

Operating System and Release Information