Dear Simon and all,
Lack of IntersubjectCV for untransformed data in BE analysis
#1
Posted 04 September 2017  09:46 AM
#2
Posted 04 September 2017  10:20 AM
Dear Zhang, Please can you expand a little on your study design, to whom you would like to submit and which BE model you are using? Even better please post your project, Phoenix/WinNonlin’s default in BE is:
fixed: Sequence+Formulation+Period
Random: Subject(Sequence)
In which case the intersubject CV% is found in Final Variance Parameters when logtransforming the data using the option on the Fixed Effects tab which is what the guidances expect. Maybe I am missing something in your question ?
Simon.
#3
Posted 04 September 2017  11:41 AM
Hi Simon,
Maybe I am missing something in your question ?
You do. Not you fault. ;)
@Zhang Yong. If you crosspost, please refer to the ongoing thread at the BEBAForum.
Best regards,
Helmut
#4
Posted 04 September 2017  11:55 AM
Thanks  that background helps  if I summarise the thread correctly we don't currently have a published reference to calculate CV% in untransformed data? If you do have one please let us know.
ALso we may need to consider reviewing this formula when using untransformed data.
» intrasubject CV = sqrt(Var(Residual)) / RefLSM
Simon.
#5
Posted 04 September 2017  12:21 PM
Hi Simon,
… we don't currently have a published reference to calculate CV% in untransformed data? If you do have one please let us know.
ALso we may need to consider reviewing this formula when using untransformed data.
» intrasubject CV = sqrt(Var(Residual)) / RefLSM
As Michael wrote, the formula is given as (34) on p. 85 of the PHX/WNL 7.0 User’s Guide. He assumed that it is referring to the book by Chow & Liu (which is given on p. 415). Can’t find the formula in there… I guess the correct reference is Phillips, which is also on p. 415. However, I think that it is wrong.
I would drop the “intrasubject CV” of untransformed data from the output in PHX.
Best regards,
Helmut
#6
Posted 04 September 2017  01:11 PM
Hi Simon,
You do. Not you fault. ;)
@Zhang Yong. If you crosspost, please refer to the ongoing thread at the BEBAForum.
Dear Helmut,
Thanks for your citing giving our Simon the calrification of the background.
#7
Posted 04 September 2017  01:35 PM
Thanks  that background helps  if I summarise the thread correctly we don't currently have a published reference to calculate CV% in untransformed data? If you do have one please let us know.
ALso we may need to consider reviewing this formula when using untransformed data.
» intrasubject CV = sqrt(Var(Residual)) / RefLSM
Simon.
Dear Simon,
First question:

Let look back to the currently used formula by PHX 7.0 (for untransfrormed data BE analysis)
intrasubject CV = sqrt(Var(Residual)) / RefLSM
Yes, I guess the above formula (for untransformed data) may be cited from Page 153 of Chow and Liu's book:
Second question:

IntersubjectCV for BE analysis with untransformed data. If no reference can be cited, Will PHX calculate it in the next version?
I guess the answer is NO.
Attached Files
#8
Posted 04 September 2017  01:42 PM
Hi Simon,
As Michael wrote, the formula is given as (34) on p. 85 of the PHX/WNL 7.0 User’s Guide. He assumed that it is referring to the book by Chow & Liu (which is given on p. 415). Can’t find the formula in there… I guess the correct reference is Phillips, which is also on p. 415. However, I think that it is wrong.
I would drop the “intrasubject CV” of untransformed data from the output in PHX.
Attached is the paper form Phillips, these is no formula for IntrasubjectCV or IntersubjectCV in this paper.
If the PHX cann't provide more references for "currently used formula" for IntrasubjectCV, maybe I will drop the “intrasubject CV” of untransformed data from the output in PHX too.
At this situation, I think PHX will not provide IntersubjectCV in the next version.
Attached Files
#9
Posted 04 September 2017  02:08 PM
Dear Zhang,
1.
you are completely right regarding your results, they are the same as in the book.
By the way the question is: could we blindly trust what Chow does say?
As I wrote before: try to substitute Test and Reference treatments in the Bioequivalence object interface and execute it. Your results will be different. But that's a nonsense! CV cannot change just because you reverse the order of the drugs!
2.
I think that's correct
No justification = no implementation
So as far as you (or someone else) can justify which formula is appropriate for interCV using untransformed data (using math or scientific links), Simon can ping some devs responsible for this.
BTW there are many things which cannot be done in current PHX release and they do not have so simple solution as just dividing one variable by another one
Bests,
Mittyright
#10
Posted 04 September 2017  02:23 PM
Dear Mittyri, Happy to meet you here.Dear Zhang,
1.
you are completely right regarding your results, they are the same as in the book.
By the way the question is: could we blindly trust what Chow does say?
As I wrote before: try to substitute Test and Reference treatments in the Bioequivalence object interface and execute it. Your results will be different. But that's a nonsense! CV cannot change just because you reverse the order of the drugs!
2.
I think that's correct
No justification = no implementation
So as far as you (or someone else) can justify which formula is appropriate for interCV using untransformed data (using math or scientific links), Simon can ping some devs responsible for this.
BTW there are many things which cannot be done in current PHX release and they do not have so simple solution as just dividing one variable by another one
Bests,
Mittyright
Yes. I agree with you. PHX is a tool, not a statistician or a biostatistician. It's clever for PHX not involved into an ambiguous issue.
Edited by ZhangYong, 04 September 2017  02:30 PM.
#11
Posted 04 September 2017  02:24 PM
Dear Mittyri, Happy to meet you here.Dear Zhang,
1.
you are completely right regarding your results, they are the same as in the book.
By the way the question is: could we blindly trust what Chow does say?
As I wrote before: try to substitute Test and Reference treatments in the Bioequivalence object interface and execute it. Your results will be different. But that's a nonsense! CV cannot change just because you reverse the order of the drugs!
2.
I think that's correct
No justification = no implementation
So as far as you (or someone else) can justify which formula is appropriate for interCV using untransformed data (using math or scientific links), Simon can ping some devs responsible for this.
BTW there are many things which cannot be done in current PHX release and they do not have so simple solution as just dividing one variable by another one
Bests,
Mittyright
Yes. I agree with you. PHX is a tool, not a statistician or a biostatistician. It's clever for PHX not involved into an ambiguous issue.
Dear Simon, sorry for replicate reply. Thank you for deleting one.
Edited by ZhangYong, 04 September 2017  02:30 PM.
#12
Posted 04 September 2017  03:05 PM
Hi Yong,
Let look back to the currently used formula by PHX 7.0 (for untransfrormed data BE analysis)
intrasubject CV = sqrt(Var(Residual)) / RefLSM
Yes, I guess the above formula (for untransformed data) may be cited from Page 153 of Chow and Liu's book:
Design and Analysis of Bioavailability and Bioequivalence Studies, Third Edition
I missed this in Chow & Liu’s book. Not their only crime. Mr Chow joined the FDA recently. ;)
See the attached project. Since the study was balanced, I could simply use the global mean.
CVintra (based on R): 0.1566
CVintra (based on T): 0.1611
CVintra (based on global mean): 0.1588
After logtransformation: 0.1947
Now look at the second workflow. I divided all Tdata by 10. Now:
CVintra (based on R): 0.1688
CVintra (based on T): 1.7357
CVintra (based on global mean): 0.1711
After logtransformation: 0.1947
Since we are playing around with mean (which we shouldn’t) the CVs change. No surprise. On the other hand, the CV in logscale is exactly the same.
Simply: If you want the CV for sample size estimation (what else?) the estimate will depend on the mean(s). That’s crap – unless you expect in the next study exactly the same T–R. Do you?
Attached Files
Best regards,
Helmut
#13
Posted 05 September 2017  12:38 AM
Hi Yong,
I missed this in Chow & Liu’s book. Not their only crime. Mr Chow joined the FDA recently. ;)
See the attached project. Since the study was balanced, I could simply use the global mean.
CVintra (based on R): 0.1566
CVintra (based on T): 0.1611
CVintra (based on global mean): 0.1588
After logtransformation: 0.1947
Now look at the second workflow. I divided all Tdata by 10. Now:
CVintra (based on R): 0.1688
CVintra (based on T): 1.7357
CVintra (based on global mean): 0.1711
After logtransformation: 0.1947
Since we are playing around with mean (which we shouldn’t) the CVs change. No surprise. On the other hand, the CV in logscale is exactly the same.
Simply: If you want the CV for sample size estimation (what else?) the estimate will depend on the mean(s). That’s crap – unless you expect in the next study exactly the same T–R. Do you?
Dear Helmut,
Thanks for your kind calculation and demostration.
But my results are:
See the attached project. Since the study was balanced, I could simply use the global mean.
CVintra (based on R): 0.1566
CVintra (based on T): 0.1611
CVintra (based on global mean): 0.1588
After logtransformation: 0.1947
Now look at the second workflow. I divided all Tdata by 10. Now:
CVintra (based on R): 0.1688
CVintra (based on T): 1.7357
CVintra (based on global mean): 0.3076 tihis is different from yours, would you like to confirm this?
After logtransformation: 0.1947
I like to calculate the parameters by hand (in MS Excel via Excel's VBA) to verify the formula used in PHX and other software. Just for my hobby.
#14
Posted 06 September 2017  09:56 PM
Dear Zhang,
just want to confirm your findings, there's a little mistake in data mapping in the project:
CVintra using global mean = sqrt(194.12374)/45.29 = 0.3076
I like to calculate the parameters by hand (in MS Excel via Excel's VBA) to verify the formula used in PHX and other software.
please look into R world. More fun. I promise.
Mittyright
#15
Posted 06 September 2017  11:29 PM
Dear both,
just want to confirm your findings, there's a little mistake in data mapping in the project:
CVintra using global mean = sqrt(194.12374)/45.29 = 0.3076
I stand corrected.
please look into R world. More fun. I promise.
Yessir! Absolutely.
Best regards,
Helmut
#16
Posted 07 September 2017  04:42 AM
Dear Zhang,
just want to confirm your findings, there's a little mistake in data mapping in the project:
CVintra using global mean = sqrt(194.12374)/45.29 = 0.3076please look into R world. More fun. I promise.
Mittyright
Thank you for your confirmation and recmmendation.
Edited by ZhangYong, 07 September 2017  05:06 AM.
#17
Posted 08 September 2017  05:31 AM
Dear Helmut,
Thanks for your kind calculation and demostration.
But my results are:
See the attached project. Since the study was balanced, I could simply use the global mean.
CVintra (based on R): 0.1566
CVintra (based on T): 0.1611
CVintra (based on global mean): 0.1588
After logtransformation: 0.1947
Now look at the second workflow. I divided all Tdata by 10. Now:
CVintra (based on R): 0.1688
CVintra (based on T): 1.7357
CVintra (based on global mean): 0.3076 tihis is different from yours, would you like to confirm this?
After logtransformation: 0.1947
I like to calculate the parameters by hand (in MS Excel via Excel's VBA) to verify the formula used in PHX and other software. Just for my hobby.
Also tagged with one or more of these keywords: IntersubjectCV, BE, untransformed
Certara Forums →
Phoenix WNL basics →
ABE using GLMStarted by sirikalyan , 25 Jun 2020 BE, ABE, Crossover 



Certara Forums →
Phoenix WNL basics →
SAS codes for BE assessment in PHX WinnonlinStarted by Mastan , 07 Dec 2018 SAS Winnonlin, BE, Connect 


0 user(s) are reading this topic
0 members, 0 guests, 0 anonymous users