It doesn't depend on the significance of the interaction--the mere inclusion of the interaction in the model will lead to differences between the solution values and the F test values.  I found this to be true even with balanced data.  Thus, I would trust the F test values, for a couple of reasons.  It considers the effect after "factoring out" all the other effects--it is meant to be an orthogonal test to all the other tests.  The theory part of my brain seems to be non-functional this morning, but I know this difference should be apparent to me in the face of including interactions.  The agreement will probably get worse with imbalance, especially with a Kenward-Rogers adjustment. Steve Denham ... View more

That seems odd.  Can you share your PROC GLIMMIX code, and the entirety of your Type3 output?  I have a hunch you have some interactions with gender that might be significant, but I'm really just guessing. Steve Denham ... View more

Do you have access to PROC CAPABILITY? The following code was generated by EG4.3.  I used anml_no as a weight, and it generated a pretty nice looking lognormal dataset this way.  The output had the goodness-of-fit tests for a lognormal distribution. PROC CAPABILITY DATA = WORK.SORTTempTableSorted CIBASIC(TYPE=TWOSIDED ALPHA=0.05) MU0=0 ; WEIGHT anml_no; VAR IgG; ; HISTOGRAM IgG / LOGNORMAL ( W=1 L=1 COLOR=LIME ZETA=EST THETA=EST SIGMA=EST) CAXIS=PURPLE CTEXT=BLACK CFRAME=WHITE CBARLINE=BLACK CFILL=GRAY ; /* ------------------------------------------------------------------- End of task code. ------------------------------------------------------------------- */ RUN; QUIT; Good luck. Steve Denham Message was edited by: Steve Denham Unfortunately, after digging a little deeper, I don't think the goodness-of-fit tests are weighted.  I get exactly the same values with and without the weight statement.  Only the basic stats and tests for location appear to actually use the weights. ... View more

Is it as simple as the default time for the x command to 'sleep'?  The behavior sounds kind of like the x command is closing before the task is completed.  I'm just guessing on this one, and I am sure that you have checked it out, but if not, maybe there is a solution. Steve Denham ... View more

Hmm.  I now wonder if there is something unusual in the particular dataset you are analyzing.  You say that you simulated the covariance structures--do you have distinctly different datasets so you could find out if this occurs for all the datasets.  And I have a truly bizarre approach, if the compound symmetry and autoregressive products are working and the unstructured by unstructured is not.  Try scrambling the dataset, so that it is NOT sorted.  I know that sounds completely counterintuitive, but we recently had a similar problem (although not with a doubly repeated measures design) that we were able to solve by doing this--we got a heterogeneous compound symmetry model to converge and have appropriate behavior with the Hessian matrix by randomizing the dataset.  I believe it all has something to do with the initial steps of the sweep operator in inverting X'VX, but I really am guessing. The reduction to 4 time points really puzzles me, unless there are not enough observations to estimate under the 6 time points, I will be watching to see if someone else has an idea. Steve Denham ... View more

You might try changing your repeated statement to the following: repeated period time / type=un@un subject=id*trt r; Another approach would be to remove period entirely as an effect, and recode time as 1 thru 12.  If you are truly interested in period means, they could then be constructed from estimate or lsmestimate statements.  The draw back to this approach is that you now must estimate 66 covariance parameters, rather than 30 with the un@un  syntax. Yet another approach might be: repeated time / type=un subject=id*trt group=period r; which should give block diagonal estimates in the covariance matrix, blocked by period. Good luck. Steve Denham ... View more

I hope Dale drops by.  He has posted a lot on NLMIXED, and I believe he has addressed the multiple random effect by a vectored approach. ... View more

This is far from a solution.  I count 15 classification variables (I don't know how many levels for each, and some of which do not appear in the model statement, e.g. s1_12), and 8 continuous covariates (age, famsize, s1_13, tot_rm, rm1_11, rm1_12, rm1_nets, and age*sex).  Some of these look like classification variables as well, but you may well wish to treat them as continuous.  So, I wonder about two things--complete collinearity (or separation) between some of the continuous covariates, and whether there is enough data to estimate everything.  Given that these are not problems, what happens if you delete the INITGLM option?  You may still get mivque(0) problems, but it is a cheap try.  I would also try dropping the ddfm=betwithin, as I don't see any within subject variability being modeled.  As a last resort, try running this in PROC MIXED and save the covariance parameters to a data set that you call as starting values with a PARMS statement. ... View more

Well.  So much for the easy way. The PROC CALIS makes a lot of references to what I would call two pass analyses.  You run it once, and get several output matrices, then use those as inputs.  One that is prominent is OUTWGT= which should give you the weightings you need to start the second pass. I am somewhat concerned by the error statement, and what I tracked down in the documentation.  You should be able to do a weighted least squares analysis without specifying INWGT=.  Since the error statement also mentions DATA=, check your syntax in the PROC CALIS statement to make sure that you are correctly specifying the input data set. Good luck. ... View more

Can you get the covariance matrix from PROC CORR? Perhaps this might work: proc corr data=indata pearson; var x1 x2 x3; ods output covar=cov; run; The dataset work.covar should have the covariances. Alternatively, create a dummy response variable (say dummy=1 for all records), and use PROC REG proc reg data=indata; model dummy=x1 x2 x3/acov; ods output acov=acovest; quit; And the dataset work.acov would have what you could use.  Good luck.  I will be curious if either method gives what you need. ... View more

keckk wrote: I reply in bold: Thank you very much for your detailed response. I have changed the lsmeans statement as you specified, and get all pairwise oddsratios with corresponding p-values in the "Differences of genotype Least Squares Means" table, but they are always in reference to the variable sweek i.e. either I define a value using the AT= suboption, of if I do not, proc glimmix takes the mean value of sweek per default. This is as it should be.  The inclusion of the continuous covariate by group interaction means that the lsmeans and ORs will be different at different values of sweek. I guess, that's why the p-values in the "Differences of genotype Least Squares Means" table are not the same as given in the "Solution for fixed effects" (at least the p-values for the three comparisons of the genotype levels with the reference level) - Am I right ? Which p-values are the ones to give for the reader, apart from the fact that I have only three out of six in the "Solution for fixed effects". I would report the comparisons at various values of sweek--something early on, something near the mean, and something at the end of observed values of sweek.  Look for consistency across time, or changes in consistency. If I drop the diff=all suboption from the parenthesis clause, I will only have confidence limits for the comparisons of the genotype levels with the reference level (which per default and in my case is the last level) in the "Odds Ratio Estimates" table. My bad.  I was hoping that by dropping it that it would leave comparisons for the covariates... For the odds ratio of the covariate, yes, I get the oddsratios comparing sweek= -2 to sweek= 6 for each of the genotypes. But how can I know if the changes over time has been significant ? How can I produce p-values for the targeted comparisons ? sweek is a continuous variable, and the overall effect is significant, but the lsmeans statement is only for class effects I am a bit confused ... Me too, and I hope this doesn't confuse things even more.  I looked for a way to add confidence limits to the odds ratio estimates in the MODEL statement, in the hope that you could look at overlap, or inclusion of 1, as a measure of significance.  The ESTIMATE statement doesn't allow for AT=, and the LSMESTIMATE wouldn't let me use multiple values.  Maybe someone has a work-around, but right now I am stumped. ... View more