• Print |
• E-mail  |
• Bookmark

If the TEST option is requested on the RANDOM statement and
if the reported ERROR term for a random effect test is
"almost all" MS(ERROR), then PROC GLM and PROC ANOVA will incorrectly
report the ERROR DF, ERROR SS, ERROR MS, F-Value, PR > F under the Tests
of Hypothesis for Mixed Model Analysis of Variance table.

The exact criterion for "almost all" is if

   sum_i (l(i) * MS(i))**2 / DF(i) < 1e-8

where: l(i) is the coefficient of the mean square of the non-error
            components
       MS(i) is the mean square for the ith component, and
       DF(i) is the degrees of freedom for the ith component

In the example below there is only one non-error component:

   Error: 0.0001*MS(A*B) + 0.9999*MS(Error)

with

   Source      DF    Type III SS    Mean Square   F Value   Pr > F

    A*B        11     0.01666472     0.00151497      1.21   0.3138

so

         sum_i (l[i] * MS(i))**2 / DF(i)

      ~ (0.0001 * 0.00151497)**2 / 11 < 1e-8.


There is no circumvention to this problem in PROC GLM or PROC ANOVA.

However, you can use PROC MIXED with the METHOD=TYPE3 option on the
PROC statement to obtain the correct results.  In general, all RANDOM
effects will go on the RANDOM statement (and not on the MODEL statement)
in PROC MIXED.


A Technical Support hot fix for Release 8.2 TSLEVEL TS2M0 for this
issue is available at:

http://www.sas.com/techsup/download/hotfix/82_sbcs_prod_list.html#008764

For customers running SAS with Asian Language Support (DBCS), this
hot fix should be downloaded from:

http://www.sas.com/techsup/download/hotfix/82_dbcs_prod_list.html#008764

Operating System and Release Information

A fix for Release 8.2 (TS2M0) for this issue is available at:

A fix for Release 8.2 (TS2M0) with Asian Language Support (DBCS) for this issue is available at: