--- U Eson < XXXX@XXXXX.COM > wrote:

> Dear all,

>

> Is there an easy explanation why the REPEATED statement has

> disappeared

> when the macro GLIMMIX was transferred into a genuine procedure?

>

> Ulf

>

I can't say definitively, but my guess is that the decision to

do away with the REPEATED statement in the procedure was to

discourage use of R-side random effects for the nonlinear models

fitted by GLIMMIX. That might be a little bit strong, since

R-side random effects can still be obtained with the GLIMMIX

procedure. Perhaps it would be better to say that the authors

wanted to encourage thoughtful use of R-side random effects.

Specifying R-side random effects yields a marginal rather than

a conditional effects model when you have a nonlinear model.

For Gaussian response with identity link function, a random

intercept model can also be fit employing a compound symmetric

residual covariance structure. Thus, for Gaussian response,

the following two models produce the same point estimates and

standard errors:

/* Random intercept model */

proc mixed;

class ID;

model y = <fixed effects> / s;

random intercept / subject=id;

run;

/* Compound symmetric residual covariance structure */

proc mixed;

class ID;

model y = <fixed effects> / s;

repeated / subject=id type=cs;

run;

This equivalence does not hold when you have a nonlinear model.

Thus, the two models presented below produce different results.

/* Random intercept model */

proc glimmix;

class ID;

model y = <fixed effects> / s dist=bin;

random intercept / subject=id;

run;

/* Compound symmetric residual covariance structure */

proc glimmix;

class ID;

model y = <fixed effects> / s dist=bin;

random _residual_ / subject=id type=cs;

run;

The latter model bears more similarity to a GEE model than to

a random effects model. The folks at SI make available to you

the capacity to estimate these marginal random effect models.

However, they have attempted to refocus your attention so that

you are aware of the different interpretations of these models.

I would note, too, that when you have a nonlinear model, the

residuals represent a multiplicative overdispersion effect. The

GLIMMIX documentation does state that "If overdispersion arises

from correlations among the observations, then you should

investigate more complex random effects structures." Again, this

conforms with discouraging use of the REPEATED statement for

adding a correlation structure to the responses.

At least, that is my impression of the reasoning behind elimination

of the REPEATED statement and replacing the REPEATED statement

with a RANDOM _RESIDUAL_ statement. It could be something more

mundane, such as that it is easier to support the procedure

with only the RANDOM statement. However, I really don't believe

that to be the case. In fact, given that residual effects

enter a different part of the model, I believe that it is more

burdensome to write the code with just the RANDOM statement.

That plus the effort expended in the GLIMMIX documentation to

instruct on the implications of use of R-side random effects

leads me to my interpretation of encouraging thoughtful use of

R-side residuals.

Dale

---------------------------------------

Dale McLerran

Fred Hutchinson Cancer Research Center

mailto: XXXX@XXXXX.COM

Ph: (206) 667-2926

Fax: (206) 667-5977

---------------------------------------

__________________________________________________

Do You Yahoo!?

Tired of spam? Yahoo! Mail has the best spam protection around

http://mail.yahoo.com