03-06-2013 09:12 AM
Dear SAS users,
We have some doubts with a longitudinal analysis using GLIMMIX and hope someone could advise us. We have daily movement data (Euclidean distances) of 45 individuals of a large bird of prey monitored during seven years which, in total, summed 18574 observations. Birds are evenly distributed in three different regions and no bird moved between regions. We are interested to know if daily movements differed among regions after controlling for bird age. The model we constructed was:
PROC glimmix DATA=COMP.PREDISP method=laplace;
CLASS birdidentity region year;
model distdiar= region age region*age /dist=XXX s;
random int / sub=birdidentity;
random int /sub=year;
LSMEANS REGION/DIFF CL;
Model results (Type III tests of fixed effects) show 15298 denominator degrees of freedom. The model is considering both within and between individual variation, so we are wondering if this procedure “artificially” inflates (in our particular case) degrees of freedom to the point that results could be wrong. Any suggestion?
Thank you very much in advance
03-07-2013 07:59 AM
PROC GLIMMIX defaults to the containment method to determine denominator degrees of freedom for models with only random effects like yours. You should consider first sorting your data by your random effects--BIRDIDENTITY and YEAR--so that all observations for a particular bird across the observation period are ordered together. Then consider using another method for determining denominator degrees of freedom (for example, specifying the MODEL statement option, DDFM=KENWARDROGER or DDFM=SATTERTHWAITE).
03-07-2013 08:26 AM
Watch out though for these two options for DDFM= when the method is LAPLACE or QUAD--they are not compatible (see Table 41.4 in the GLIMMIX documentation). Try DDFM=BW (between/within).
03-13-2013 04:28 AM
Thanks a lot. Both solutions seem to work, although the BW option with LAPLACE method needs a nested design of random effects -birdidentity(region)- to correct the d.f. As we are not especially interested in obtaining very accurate estimates or AIC values the DDFM=KR (or SAT) with residual ML or PL is also a good option.