Dear all we plan to calculate the 90% CI for partial replicate study by using average bioequivalence method in phoenix winnonlin 6.3. we are getting error message for only Cmax . for AUCt and AUCinf we got the average bioequivalence values. pls find the error message and give your suggestion. error message There was an error while executing Workflow.Bioequivalence Core Warning Messages: 1 Warning 11091: Newton’s algorithm converged with modified Hessian. Output is suspect. Core Error Messages: 1 ERROR 11070: Error in Satterthwaite DF. Try using Residual DF option. [file name=Winnonlin_error.docx size=92589]Certara | Drug Development Solutions (90.4 KB)
The selected model isn’t fitting the data well. You could try changing the variance structure on the Random tab from FA0 to CSH (Heterogeneous Compound Symmetry) for Cmax, to see if that converges without problems. As indicated in the last error message, you could try changing the DF option on the General Options tab to Residual, although that often does not help. If neither of these works, you could send the project into Support to look at. Regards, Linda
Dear Karthik, it is not helpful posting a blurred screenshot pasted into of Word-document.As Linda suggested please post a project-file in the future. The partial replicate (TRR, RTR, RRT) is a – very1 – stupid designs since T is not repeated. No idea why FDA seems to love it. Depending on the code and software you will get different results – if the algo converges at all. See this current thread. A better three-period design is a two-sequence fully replicated: TRT, RTR.
Hi Karthik, You can also try to remove the subject-by-formulation effect from the model. In the [Random 1] tab select Type ‘Compound Symmetry’. It worked in some of my data sets.
Hi Linda, do you have any idea on how to ‘fix’ csDiag_11 (the variance of the subject-by-formulation interaction) to zero? Or – maybe even better – set up a model without this effect? For an example data set see this post and the following discussion.
Hi Helmut, So you just want csBlock for every element in the Variance Structure. I looked through all the variance structure options we have to try to find a way of doing this, with no luck. Sorry. I read the discussion on your forum, and saw the question about whether SAS will estimate a negative variance. I know at least in the case of the Variance Components structure that, if SAS MIXED ends with a negative variance component, it will output zero. This partial replicate design has certainly opened a can of worms, as the expression goes. I’ll watch the discussion on your forum to see what develops. Regards, Linda
Hi Linda! [quote]I looked through all the variance structure options we have to try to find a way of doing this, with no luck.[/quote] So did I. ;-). [quote]I know at least in the case of the Variance Components structure that, if SAS MIXED ends with a negative variance component, it will output zero.[/quote] Yep. But even if not (i.e., getting a non-zero estimate) I duno whether it would be possible in SAS to force it to zero. [quote]This partial replicate design has certainly opened a can of worms,…[/quote] X-actly! Since FDA’s progesterone guidance was published in 2010 more and more people opt for RSABE (~30% of drugs are highly variable). The partial replicate design is really wacky. Quite often scaling is allowed for Cmax (higher variability) but not for AUC. If we follow FDA’s ABE-code (which is the default setting in PHX – and was in classical WNL as well) we might run into convergence problems. See Kaushik’s post above and John’s two data sets in my forum. The can is opened and the worms look ugly. It’s crazy that people designed studies according to FDA’s requirements only to find out that they cannot evaluate them.