Hello all, I am working on a 3 period, 4 treatment, crossover PK study in which the randomization has a partially balanced design of 4 sequences. There are 20 subj with 5 subj/sequence (ABC, BAC, ABD, BAD). If I use the bioequivalence wizard to determine BE with say, B as the reference, the LSMs are “not estimable” and as a result, no BE estimates can be obtained. The error message in the log says: “ERROR 11016: Numerical problem when computing satterthwaite df.” The BE setup is standard (form, seq, per = fixed, sub(seq) = random) . Can someone shed light on how Winnonlin handles LSM for unbalanced designs of >2 periods? Thanks in advance, Mike
Mike, I’m not sure if this is a general problem or data-specific. It’s possible that you’ve got problems with the default model, and we could try another model. Please can you attach some exemplar data or better yet the project? (then we can check settings etc.) Simon
Simon, Thanks for your prompt response. Attached is the project file; I removed some parameters to simplify the analysis. For some reason, WNL doesn’t flag the error anymore, but still get the same results (ie: non-estimable LSMs etc..). Mike [file name=Test.phxproj size=196903]Certara | Drug Development Solutions
Test.phxproj (192 KB)
Since the only sequences in the data are:
BAD
ABC
ABD
BAC
it is really only a crossover for A vs. B. If I exclude the C and D data, then the model will successfully complete for the A vs. B data. When I get a chance, I’ll run this in SAS PROC MIXED to see how it handles the case where the C and D data are included.
Another model that can be tried is the default model for a Parallel design - this model will complete successfully. I’ve attached an updated project with these two additional model runs. [file name=Test-20120813.phxproj size=488496]Certara | Drug Development Solutions (477 KB)
Thanks..If you run it using Proc Mixed with the C and D data, you run into the same problems. I don’t think the parallel model is applicable here.. Sorry, I can’t seem to open your file?
Hi Linda and Mike! That’s a balanced incomplete block design (BIBD). Rare, but not exotic. 
For details see Chow/Liu, 3rd. ed. pp326 - and please forget carry-over. [quote]I don’t think the parallel model is applicable here.[/quote]Sure. But you can misuse it for other models as well. That’s why it says ‘Parallel/Other’. But here I’m afraid we have to go with LinMix. Map SUBJ, SEQUENCE, TRTA, and PERIOD to Classification. Set up the usual model. Fixed: PERIOD+TRTA+SEQUENCE, Random: SUBJ(SEQUENCE). Define 3 Estimates. Drag TRTA to Effect. Give a nice title. For the comparison of A with B set the Coefficients to A=1, B=-1, C=0, D=0. Similar for the others (C or D to 1, B always to -1). Set CL to 90. Execute and send the Estimates to the Data Wizard to back-transform. For AUClast I got A/B 100.52% (90% CI: 89.63–112.74%), C/B 124.55% (111.05–136.68%), and D/B 79.87% (69.59–91.67%). CV[sub]intra[/sub] 21.7%. Wider limits than after Linda’s data exclusion because we have a common variance in the model and cannot ignore its structure. Does this make sense? [quote]Sorry, I can’t seem to open your file?[/quote]I could (though with a warning) in 6.3. Which version are you using?
Thanks Helmut! Don’t we also need to assign coefficients for PERIOD? For the D/B comparison, for instance, D never appears in Period 1 and 2, but since B does and never appears in Period 3, do we not have to consider the unequal period weighting? Mike
By the way, I’m using version 6.1 which probably explains why I can’t open Linda’s file..
Hi Mike, I wouldn’t say so. Period is already a fixed effect in the model. BTW, is strongly suggest to update to v6.3. The new Data Wizard is a great feature and makes life much easier. A nasty bug in calculating the CI by the BE Wizard was also corrected.
6.3 is the current version and can be downloaded freely by all current licensees from; ftp://support_ftp:pt9JTmD9GdXQLtCgPkvL@ftp.certara.com/support/desktop/Phoenix/Phoenix_1.3.zip I would suggest you speak with your internal IT contact to make sure he passes on information from Pharsight about these updates. Simon.