4 Factorial-Split Plot Proc mixed

[SteveDenham]

Hi Caroline,

I would never code it with proc glm.  See Walter Stoup (2013) Generalized Linear Mixed Models. 

PROC GLM just does not do a good job of analyzing repeated measures or split plot models.

The fact that the crop is not the same across all dates does present a problem in terms of estimability of means.  There will be "missing" cells by design.  To get around this, I would fit a cell means model, and use LSMESTIMATE statements to derive lower order means and comparisons.

proc glimmix;

class nem trt crop blk date;

model  lgsprCysts100=nem*trt*crop*date/noint; /* This fits a cell means model */

random intercept nem|trt|crop/subject=blk;

random date/ residual subject=nem*trt*crop*blk type=arh(1); /* This may be problematical to fit, due to the small number of levels for date, and incomplete subjects.  If so, try type=chol */

run;

The LSMESTIMATE statements will need to be added later. I also have a hunch that the dependent variable is a count variable, and rather than a log transform prior to analysis, a log link should be used so that the dependence of the variance on the mean can also be captured.

Steve Denham

Message was edited by: Steve Denham

[palolix]

Thanks a lot Steve!

I used to run my models (also split plot models) as a first run with proc glm in order to pool, so I can get a clean model when using proc mixed.    In this case, I think I do not have a repeated measures, because not only the crops are different every year, but also the planting dates, so the circunstances differ every year.

Caroline

[SteveDenham]

I could give it a shot, but details of the design, including the nature of all variables, are absolutely a minimum requirement, as per .

Steve Denham