The Versatility and Power of PROC MIXED
Robin
High
Statistical Programmer and Consultant
robinh@uoregon.edu
In past issues of Computing News (http://cc.uoregon.edu/cnews/), I’ve
published several articles introducing linear models computed with SAS
PROC MIXED. The full text of these articles is available on my website
at http://www.uoregon.edu/~robinh/statistics.html.
If you’re analyzing continuous response data that fall under the
category of linear statistical models, PROC MIXED offers a powerful replacement
for PROC GLM. In addition to being far more flexible in handling complex
designs, PROC MIXED also computes correct standard errors in situations
where PROC GLM did not. To learn more about how PROC MIXED has essentially
made PROC GLM obsolete, see the collection of articles at http://www.uoregon.edu/~robinh/mixed_sas.html.
Two of these articles, "The Unequal Variance ANOVA Model" and "Power
Analysis for Complex ANOVA Designs," describe PROC MIXED’s
very important contributions to data analysis.
One of the inherent assumptions of an analysis of variance model is that
residuals have equal variances across groups. The first article in this
series, "The Unequal Variance ANOVA Model" (http://www.uoregon.edu/~robinh/glm10_homog_var.txt),
describes how to run an unequal variance linear model using PROC MIXED.
It first examines methods of checking this assumption, and then shows
how to run a weighted least squares analysis, which has long been a remedy
for the unequal variance situation. The article demonstrates how to apply
this technique, and then proposes a different solution offered by PROC
MIXED that is much simpler to implement and interpret if variances across
groups should not be considered equal.
The second article, "Power Analysis for Complex ANOVA Designs" (http://www.uoregon.edu/~robinh/glm14_power.html),
addresses power calculation for complicated experimental designs, which
has always been problematic, if not impossible. This article shows how
PROC MIXED serves as a very important study planning tool to compute
power for a given number of subjects for many experimental designs you
encounter, including complex repeated measures.
The key to making PROC MIXED work is understanding how to structure the
variance/covariance matrix to compare correlated group means, and how
to enter the variance component elements on a PARMS statement. The result
allows you to compute the minimum effect size you want to be able to
detect. After that, computing power to detect a significant effect for
a given number of subjects is a very simple task with a DATA step.