LMIXED Procedure

PROC LMIXED is designed to fit a subclass of large (in number of observations and number of levels in a subject) linear mixed models. The procedure takes advantage of distributed and multithreading capabilities of your system by working on subsets of the data in parallel, and it aggregates results in the estimation process, which can lead to substantial speed improvements. This subclass of models requires that the matrix have a block-diagonal structure, meaning that the data can be grouped according to a subject effect. This subject effect is referred to as a "common subject effect."

The common subject effect should be the only subject effect that is used in a program, and it must be specified as the same SUBJECT= effect in all RANDOM statements.

For example, a model that has a single RANDOM statement with a SUBJECT= specification has a common subject effect by definition:

random intercept age / subject=student;

You can have multiple RANDOM statements. The same SUBJECT= effect is the common subject:

random time          / subject=student*school;
random intercept age / subject=student*school;

In contrast, a single RANDOM statement without SUBJECT= specification, as shown in the following statements, implies that the model does not have a common subject effect:

random time;

More than one subject effect appearing in a program, as shown in the following statements, implies that the model does not have a common subject effect:

random time          / subject=student*school;
random intercept age / subject=student;

If you have a variance component model (TYPE=VC), in some situations, you can use a random effect as a subject effect, which leads to the same model specification but with a common subject effect. For example, the following two RANDOM statements are equivalent (the second specification has a common subject effect in time):

random time;

random int / subject=time;

Another equivalency is that

random time          / subject=student*school;
random intercept age / subject=student;

can be rewritten as

random time*school     / subject=student;
random intercept   age / subject=student;

where the student becomes the common subject effect.