05-21-2014 01:24 PM
Hi, I am planning to run my data using Glimmix a set of variable has a non-normal distribution.
The statements I used was :
proc glimmix data=steps;
class blk diet house;
model steps=diet house diet*house;
lsmeans diet house diet*house;
ods output diffs=ppp lsmeans=mmm;
ods listing exclude diffs lsmeans;
The Glimmix output however produces P-value, SE and estimates almost similar to Mixed procedure. Is it supposed to be like this ?
The estimation technique used was the restricted maximum likelihood (Glimmix default is RSPL if random effect is used) and I found out that to use other techniques, I will have to use the LINK option.
Also, because my experimental design is a 2x3 RCBD, I am using blk as my random effect. However, using the random statement as above will show the variance matrix as Not Blocked.
If I use the statement (as below), I will get the variance matrix as Blocked. All this however still using the restricted maximum likelihood as the estiamtion technique ( as opposed to RSPL which is the Glimmix default)
Any comments on this ?
05-22-2014 08:39 AM
I may be missing the point here, but if you want to exclude some tables from the listing from a procedure, I think the ODS listing exclude statement must precede the call to the procedure.
The statement: "The estimation technique used was the restricted maximum likelihood (Glimmix default is RSPL if random effect is used) and I found out that to use other techniques, I will have to use the LINK option" is not really correct. Check out the METHOD= option to the PROC GLIMMIX statement.
The LINK option is designed to set the linearizing function appropriate for the distribution of interest.
Lastly, the two random statements are exactly equivalent under the default method, but there are powerful implications regarding which to use. If you should choose METHOD=LAPLACE or METHOD+QUAD, the random statement MUST be of the latter form, where subject is specified. Additionally, the latter form, with subject specified, is almost always faster to converge.