5 replies to this topic

[BEQool]

Hello everyone!

 

I was calculating CV_WR (in replicate designs) of different compounds in Phoenix WinNonlin using method described in section "3.5 Calculation of CVWR and Scaled Acceptance Range" in these instructions (starting on page 6): 

https://bebac.at/dow...hoenix_v3.4.pdf

 

When I am calculating within reference CV (CV_WR) for full-replicate 4-period design with 2 sequences (S1: ABAB, S2: BABA) everything looks fine and I get same CV_WR regardless of the input order in Model specification Fixed effect (e.g. no difference when the input is "Sequence+subject+period", "period+subject+Sequence" ...) in Linear Mixed Effects (LME). 

 

But when I am calculating CV_WR for (semi- or full-?)replicate 3-period design with 2 sequences (S1:ABB, S2:BAA) I get 2 different CV_WR. The CV_WR I get depends on the input order in Model specification Fixed effect in Linear Mixed Effects (LME). This happened in 2 occasions (same design but 2 different studies and compounds): 
A) For example, for compound X I get CV_WR= 33,58% for three input orders "sequence+period+Subject", "period+Subject+sequence" and "period+sequence+Subject". On the other hand, for compound X I get CV_WR= 33,91% for other three remaining input orders: "sequence+Subject+period", "Subject+period+sequence" and "Subject+sequence+period". 

B)Similarly for the same design (ABB, BAA) but different study, for compound Y I get CV_WR= 34,82% for same three input orders as mentioned above for compound X ("sequence+period+Subject", "period+Subject+sequence" and "period+sequence+Subject") and CV_WR= 35,18% for same other three remaining input orders ("sequence+Subject+period", "Subject+period+sequence" and "Subject+sequence+period"). 

 

So my question would be, why is this so and which input order should be used in LME and gives the correct CV_WR?

 

Additional question (just to check if this is the correct procedure) would be, if a subject 36 (sequence ABB) has a missing period 2 but returned for period 3 - like this:

 

Subject period sequence AUCi AUCt Cmax

36             2               ABB  

36             3              ABB      495.60 479.6 48.5

 

We probably have to exclude whole subject 36 (both periods 2 and 3) when calculating CV_WR? Because we get 2 different CV_WR depending whether subject 36 is included (as above) or not?

 

Thank you and best regards

BEQool

 

 

[Simon Davis]

Can you perhaps send the project into Support for further investigation?  It maybe the unbalanced data fro Subject 36 causing issues

The CVWR (coefficient of variation of within-subject residuals) is a measure of the model's goodness of fit, indicating the amount of variability in the residuals after accounting for the fixed effects in the model.

From the information you provided so far, I'm not sure it is not possible to determine why there is a slight difference in the CVWR between the different input orders. It is possible that the order of the fixed effects may affect the model's ability to account for the variability in the data, but it could also be due to chance or other factors that are not apparent from the information given.

In general, the input order of fixed effects should not affect the results of the analysis significantly, as long as the correct model assumptions are met, and the model includes all relevant fixed and random effects. Therefore, you should be able to use any of the six input orders that you tested.

Statistically I have read that it is recommended to choose an input order that is meaningful in the context of the study,  For example, if the sequence is a crucial factor in the study, it may be useful to include it as the first fixed effect in the model. Similarly, if the period is a more critical factor, it may be included as the first fixed effect in the model.  In BE we hope to demonstrate that there is not a sequence or period effect.

In summary, the choice of input order should be based on the study's goals and the relevance of the fixed effects in the model, rather than the slight difference in CVWR values obtained with different input orders.

[Helmut Schütz]

Hi BEQool,

 

here you are as well!

 

Check out the recordings of a recent webinar. Of course, his personal opinion, but in the Q&A-session Donald Schuirmann pointed out that in a partial replicate design only subjects completing all periods should (read: must?) be included in the calculation of s_wR / CV_wR. Why? No idea.
 

[BEQool]

Sorry for my late response. Thank you both for your answers!

 

Now that I re-checked data it seems logical to exclude subject if he/she has a missing period with the treatment we want to get CV_WR for (for example excluding subject 36 as described above) because then there is no treatment replication for this subject. 

 

Can you perhaps send the project into Support for further investigation?  It maybe the unbalanced data fro Subject 36 causing issues

The CVWR (coefficient of variation of within-subject residuals) is a measure of the model's goodness of fit, indicating the amount of variability in the residuals after accounting for the fixed effects in the model.

From the information you provided so far, I'm not sure it is not possible to determine why there is a slight difference in the CVWR between the different input orders. It is possible that the order of the fixed effects may affect the model's ability to account for the variability in the data, but it could also be due to chance or other factors that are not apparent from the information given.

In general, the input order of fixed effects should not affect the results of the analysis significantly, as long as the correct model assumptions are met, and the model includes all relevant fixed and random effects. Therefore, you should be able to use any of the six input orders that you tested.

Statistically I have read that it is recommended to choose an input order that is meaningful in the context of the study,  For example, if the sequence is a crucial factor in the study, it may be useful to include it as the first fixed effect in the model. Similarly, if the period is a more critical factor, it may be included as the first fixed effect in the model.  In BE we hope to demonstrate that there is not a sequence or period effect.

In summary, the choice of input order should be based on the study's goals and the relevance of the fixed effects in the model, rather than the slight difference in CVWR values obtained with different input orders.

 

I think I have figured it out. I think that none of the input orders I stated above are correct. 
Since we have sequences ABB and BAA, for each treatment (A and B ) there is just one sequence where the treatment is replicated (sequence ABB for treatment B and sequence BAA for treatment A). Therefore in this case, sequence shouldnt affect the CV_WR and sequence should not be put in the LME model. In LME in Model specification Fixed effect there should only be "subject+period" regardless of the order (i.e. "period+subject" should give and gives the same result). So I think that the CV_WR I get by putting just "subject+period" in the LME model in this case is the correct one (and again, it is different than the other two CV_WR mentioned above). 

 

I still dont know why the input order affects the (wrong?) calculation of CV_WR but in this case when we have a design with sequences ABB/BAA, we should probably omit "sequence" from the LME model (just as we omit "treatment" as we have just one) and just put "subject+period". 

 

Hi BEQool,

 

here you are as well!

 

Check out the recordings of a recent webinar. Of course, his personal opinion, but in the Q&A-session Donald Schuirmann pointed out that in a partial replicate design only subjects completing all periods should (read: must?) be included in the calculation of s_wR / CV_wR. Why? No idea.
 

 

Hey, yes, I have a lot of questions 

Interesting. So we should (must?) exclude the whole subject when calculating CV_WR even if for example his sequence is TRTR and he has a first missing period (xRTR). In this case he still has replicated reference treatment ® to calculate CV_WR but we should exclude him anyway?

 

Best regards

BEQool

Edited by BEQool, 18 April 2023 - 09:28 AM.

[Helmut Schütz]

Hi BEQool,

 

…we should probably omit "sequence" from the LME model (just as we omit "treatment" as we have just one) and just put "subject+period".

 

Bingo – not only in replicate designs. See there for the simple 2×2×2 crossover.

 

 

So we should (must?) exclude the whole subject when calculating CV_WR even if for example his sequence is TRTR and he has a first missing period (xRTR). In this case he still has replicated reference treatment ® to calculate CV_WR but we should exclude him anyway?

 

Personally I think that excluding such a subject is crap. The more data we keep, the more accurate the estimate. That’s not rocket science.

 

Agencies differ in their approaches to estimate CV_wR anyway. See the very end of this post for an example.

[BEQool]

I agree regarding the inclusion of a subject. Thank you for your answers!