Bioequivalence and Bioavailability Forum

Main page Policy/Terms of Use Abbreviations Latest Posts

 Log-in |  Register |  Search

Back to the forum  Query: 2017-09-23 13:07 CEST (UTC+2h)
 
Ohlbe
Hero

France,
2010-01-28 18:31

Posting: # 4655
Views: 15,671
 

 EMA BE guideline - final version [Regulatives / Guidelines]

Dear all,

According to the report of January's CHMP meeting the revised version of EMA's BE guideline was adopted last week. We can expect to see it posted on EMA's web site in the next few days.

Regards
Ohlbe

Regards
Ohlbe
Helmut
Hero
Homepage
Vienna, Austria,
2010-01-29 14:29

@ Ohlbe
Posting: # 4659
Views: 13,595
 

 Final version published

Dear all!

The eagle has landed. The final BE-Guideline was published today.
The Guideline will be effective with 1 August 2010.

Fascinating!
  • […] the 90% confidence interval for the ratio of the test and reference products should be contained within the acceptance interval of 80.00-125.00%. To be inside the acceptance interval the lower bound should be ≥80.00% when rounded to two decimal places and the upper bound should be ≤125.00% when rounded to two decimal places.
    Funny. 79.995%-125.004% is bioequivalent, but your software will print out “Bioequivalence not shown.” Time for a revision of Phoenix/WinNonlin. ;-)
  • No statistical test for any effect (carry-over, period, treatment), as well as no confidence intervals.
  • Two-stage design: 94.12% confidence interval given as an example for the combined analysis. That’s the one of Potvin et al. (based on Pocock’s).
  • Interesting section on exclusion of subjects.
  • Scaling for Cmax made it, but the widening of the acceptance range is limited to a maximum of 69.84–143.19% (CVref≤50%).The suggested method is (following the terminology of Tóthfalusi et al.) ABEL (Average Bioquivalence with Expanding Limits). Like in the U.S. a constraint of 80.00%-125.00% is set to the GMR.
  • tmax is back (instead of partial AUC truncated at tmax,ref) – but no statistical analysis (“no apparent difference in median tmax and its variability”).
  • Css,min has vanished.


Edit: Link corrected for new EMA’s website. 2010-07-16 [Helmut]

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-01-29 17:02

@ Helmut
Posting: # 4660
Views: 13,370
 

 Final version published

Bong Sure,

4.1.8 (Stats):

"The assessment of bioequivalence is based upon 90% confidence intervals for the ratio of the population geometric means (test/reference) for the parameters under consideration."
Population geometric means.. I am sure they will continue to mean LS Means although this is not what that sentence means. Business as usual.

"The terms to be used in the ANOVA model are usually sequence, subject within sequence, period and formulation. Fixed effects, rather than random effects, should be used for all terms."
Implies no more mixed effect modeling? Now, if we do a normal linear model on data from a replicated study then we are certain to violate some basic requirements (correlated errors; those that the Mixed Model intends to solve for us). That also implies that we will not be able to give unbiased or good estimates of the intra-subject variability. In 4.1.10 (HVD's):
"The applicant should justify that the calculated intra-subject variability is a reliable estimate and that it is not the result of outliers."
The wording leaves anybody doing a replicated study in a dilemma. Either use a mixed model and violate 4.1.8 but get healthy withins, or using a fixed effects model as required and get toxic withins.

Next issue of course, is that if we comply with "Fixed effects, rather than random effects, should be used for all terms." then there will NOT be individual withins for Ref (or Test). The recommendation of using swr for scaling purposes does not make sense as long as no mixed effecs model is allowed.

I will stop here, but let me just quote John McEnroe:
"You cannot be serious!"

EM.
Helmut
Hero
Homepage
Vienna, Austria,
2010-01-29 17:30

@ ElMaestro
Posting: # 4661
Views: 13,206
 

 Statistix

Hi ElMaestro!

» "The ratio of the population geometric means"
» Business as usual.

Well, if you perform a study in the entire population of patients:
population means = sample means. Isn't it? :lol3:

» "The terms to be used in the ANOVA model are usually sequence, subject within sequence, period and formulation. Fixed effects, rather than random effects, should be used for all terms."

Oh goodness, I missed that one. :angry:
Terrible.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-01-29 18:11

@ Helmut
Posting: # 4662
Views: 13,162
 

 Statistix

Hi HS,

I hope your life is good.

» Well, if you perform a study in the entire population of patients: population means = sample means. Isn't it? :lol3:

It is a very economical use of words here. As long as population means is the same as "means across populations defined by their sequences" people know what to do although I am not sure people - including myself - know what they are doing, but that is a discussion about the relevance of LSMeans I will not dare to start.

EM.
Helmut
Hero
Homepage
Vienna, Austria,
2010-01-29 19:22

@ ElMaestro
Posting: # 4663
Views: 13,256
 

 Calories ;-)

Hi ElMaestro!

» […] Business as usual.

Found another gem. I know that units are a very difficult topic. FDA and our Canadian friends failed as well in distinguishing calories and kilocalories.

4.1.4 Study conduct, Fasting or fed conditions

[...] the meal should be a high-fat (approximately 50 percent of total caloric content of the meal) and high-calorie (approximately 800 to 1000 calories) meal. This test meal should derive approximately 150, 250, and 500-600 calories from protein, carbohydrate, and fat, respectively. The composition of the meal should be described with regard to protein, carbohydrate and fat content (specified in grams, calories and relative caloric content (%)).


OK, here comes my suggestion:
                   caloric breakdown (cal)
--------------------------------------------------
                        carbo-
              protein   hydrate    fat     total
--------------------------------------------------
123 mg toast    46.1     203.0     33.5    282.6
 50 mg butter    0.0       1.7    452.0    453.7
350 µL milk     47.3      58.3    124.2    229.8
--------------------------------------------------
total calories  93.5     263.0    609.6    966.1
target cal.    150       250     500-600  800-1000
% of total       9.7      27.2     63.1

Not perfect, but a first try. Protein is difficult to match, maybe I add a quarter soybean to the meal and serve only 100 mg of toast instead.

P.S.: I’m quite happy with my body mass of 72 kilograms. 72 grams with my size would give my enough bouyancy to reach the stratosphere. On the other hand keeping my specific weight and BMI I would only measure 5.7 centimetres (slightly smaller than a mouse). Luckily I don’t have a cat at home.


Edit: Was corrected in March 2010. See here. [Helmut]

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-01-29 20:43

@ Helmut
Posting: # 4664
Views: 13,177
 

 Calories ;-)

Hi HS,

» 123 mg toast
» 50 mg butter
» 350 µL milk

What a feast that is, especially if you add a quarter soy bean. To preserve good table manners, do you suggest to eat all that with knife and forceps?

EM.

I could be wrong, but…


Best regards,
ElMaestro

- since June 2017 having an affair with the bootstrap.
Helmut
Hero
Homepage
Vienna, Austria,
2010-01-29 21:08

@ ElMaestro
Posting: # 4665
Views: 13,148
 

 Calories ;-)

Dear ElMaestro!

» What a feast that is, especially if you add a quarter soy bean. To preserve good table manners, do you suggest to eat all that with knife and forceps?

I would go with a micro-applicator, preferably teflon-capped to prevent problems with the notoriously sticky butter. A calibrated pipette with a disposable tip would do for the 350 µl of milk.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
d_labes
Hero

Berlin, Germany,
2010-02-01 12:05

@ Helmut
Posting: # 4673
Views: 13,097
 

 Variable calories

Dear Helmut!

» Found another gem. I know that units are a very difficult topic. FDA and our Canadian friends failed as well in distinguishing calories and kilocalories.
(emphasis in bold by me)

This would make me very pensively! Great Nations, famous regulators(!), native speakers and wrong terms?
Have a look at Large and small calories. :-P
Ok! Not very scientific, you are right!

I am glad Your are not vanished beyond the stratosphere only because of units. So stay on earth. Who shall else admin this forum? :-D

Regards,

Detlew
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-10 23:41

@ ElMaestro
Posting: # 4748
Views: 12,611
 

 ANOVA only - no doubts

Hi ElMaestro!

» "The terms to be used in the ANOVA model are usually sequence, subject within sequence, period and formulation. Fixed effects, rather than random effects, should be used for all terms."
» Implies no more mixed effect modeling?

See the comment-document (linked here, page 138):

One objective of the new guidance was to completely standardise the method of analysis. While mixed models are generally useful, for bioequivalence ANOVA is considered adequate. […] a mixed linear models approach would not be acceptable, and subjects with valid data for only one of the two treatments should be excluded. No change. The phrase "or equivalent parametric method" removed to make clear that we are insisting on ANOVA.

(my emphasis) I lost also in my comment on lines 500-501/574-576 (page 140, top) – and of course on nonparametrics (page 141, top). ANOVA only. Full stop.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-02-12 21:09

@ Helmut
Posting: # 4757
Views: 12,359
 

 ANOVA only - no doubts

Dear HS,

» ANOVA only. Full stop.

This is becoming more and more mysterious. Anova and mixed effects modeling are not mutually exclusive, and it sounds to me like they have not realised the difference between a model and an anova.
They will sooner or later realise that if the guideline in its present form is to be followed there is no way to come up with correctly estimated intra-subject variabilities in a replicated study.
Furthermore, the entire section 4.1.10 has no relevance unless mixed modeling is used.

This is regrettable.
EM.

I could be wrong, but…


Best regards,
ElMaestro

- since June 2017 having an affair with the bootstrap.
d_labes
Hero

Berlin, Germany,
2010-02-01 11:53

@ Helmut
Posting: # 4672
Views: 13,085
 

 Final EMA oracle

Dear Helmut, dear all!

» The eagle has landed. The final BE-Guideline was published today.
»
» Fascinating!

Really!
All in all we can live with that I mean (of course we must, no chance to do not). Some goodies, some 'badies', some half cooked. Some clear like the delphian Pythia.

But I have to read it again and again to have more than one question :-D.

This one is actual for me:
Section 4.1.8 Evaluation
Subject accountability
It is stated: "... In studies with more than two treatment arms (e.g. a three period study including two references, one from EU and another from USA, or a four period study including test and reference in fed and fasted states), the analysis for each comparison should be conducted excluding the data from the treatments that are not relevant for the comparison in question." (emphasis by me)

How shall I interpret this :confused:.

Do they claim that such studies should evaluated as if they consist of a series of classical 2×2 cross-over's? ("Extraction of 2×2 comparisons" in Helmut's famous lectures)

Anybody (hopefully an EMAist!) out there to enlighten me?

Regards,

Detlew
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-01 12:28

@ d_labes
Posting: # 4674
Views: 13,031
 

 Final EMA oracle

Dear D Labes!

» Section 4.1.8 Evaluation
» Subject accountability
» "In studies with more than two treatment arms ..."

Yes, ‘discovered’ this piece in the meantime also…

» Do they claim that such studies should evaluated as if they consist of a series of classical 2×2 cross-over's? ("Extraction of 2×2 comparisons" in Helmut's famous lectures)

I can’t believe that. I would guess that one should only omit information on the 'foreign' reference as far as possible (descriptive statistics, plots, etc.).

The section mentioning the 4-way fed/fasted study is a mystery to me.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-02-01 12:50

@ d_labes
Posting: # 4675
Views: 13,031
 

 Final EMA oracle

Dear d_labes,

» Do they claim that such studies should evaluated as if they consist of a series of classical 2×2 cross-over's? ("Extraction of 2×2 comparisons" in Helmut's famous lectures)

Contrary to HS' opinion, I think this is exactly how that sentence should be interpreted for now. I would not be surprised if a slight revision is due soon which addresses this issue and/or the fixed factor problem.


EM.
d_labes
Hero

Berlin, Germany,
2010-02-01 13:30

@ ElMaestro
Posting: # 4677
Views: 13,023
 

 More then two EMA oracles

Großer Meister,

» Contrary to HS' opinion, I think this is exactly how that sentence should be interpreted for now. I would not be surprised if a slight revision is due soon which addresses this issue and/or the fixed factor problem.

I can't believe that :no:! We have only one study! And shall treat them as if we had some independent 2×2 studies?

What if we have period effects? In the '2×2 extracted evaluation' we are then mixing periods (for instance in T1T2R for the pair T1/R T1 comes from period 1, in sequence T2T1R from period 2).

And what about sequences? For the above example T1T2R and T2T1R are the same in evaluating T1/R?

And what about the error (MSE). In the 3×3 cross-over we have then two (distinct) MSE's and also CV's?

Questions over questions. With the simple answer: This is statistically not so very sound, to be political correct (and not saying nonsense).

Hopefully this will be cleared before the magic 10-Aug-2010.

Regarding the fixed effects issue you questioned above in your post:
I can't imagine, what this means at all :ponder:, especially if I think about the sequence test. Shall we use the error term as the denominator as is in a fully fixed effects model?
We all know that this is the wrong (already Grizzle in his basic paper noticed the right denominator is the subject(sequence) MS).

Hopefully this will be clarified before the magic 10-Aug-2010.


JE Grizzle
The two-period change over design and its use in clinical trials
Biometrics 21, 467-480 (1965)
[Helmut]

Regards,

Detlew
ElMaestro
Hero

Denmark,
2010-02-01 15:29

@ d_labes
Posting: # 4684
Views: 13,021
 

 Less of an issue

Goede middag d_labes,

» Regarding the fixed effects issue you questioned above in your post:
» I can't imagine, what this means at all :ponder:, especially if I think about the sequence test. Shall we use the error term as the denominator as is in a fully fixed effects model?
» We all know that this is the wrong (already Grizzle in his basic paper noticed the right denominator is the subject(sequence) MS).

This I don't think is an issue at all. Even in spite of the bogus statement, Proc Glm fits the entire model as a fixed effects model, which is good enough for 2-seq 2-per 2-trt data. So even though subject in such a design is treated as fixed, it does not imply that the seq effect cannot be tested using the between subject variabilty in the denominator. This is to the best of my knowledge what the bogus statement allows for and does post-hoc, and this is compliant with the guideline's wording.

EM.
d_labes
Hero

Berlin, Germany,
2010-02-01 16:53

@ ElMaestro
Posting: # 4686
Views: 13,004
 

 Less or More of an issue

Dear ElMaestro,

» This I don't think is an issue at all. Even in spite of the bogus statement, Proc Glm fits the entire model as a fixed effects model, which is good enough for 2-seq 2-per 2-trt data. ...

I don't know if you are right here.
As far as I had understand it to now (but my brain is here some smaller than a walnut ;-)) the natural way in an ANOVA model with all effects fixed is to use the error MS as denominator in the F-test of all of the effects.
This is why all statistical software, which I know off, come out with exactly these tests if no option is chosen.

I think you are mixing up estimation with test methods based on the ANOVA model. You are right in stating that Proc GLM fits the model as fixed, i.e. the the model parameters are estimated via least squares, which is identical to the likelihood estimates in that case.
But then after that it modifies the tests (called in a post hoc manner) according the model with subject as random by the RANDOM statement. Call it bogus if you like.
Or you make the appropriate test "by hand" as in the bear R-Code using lm() or in the former SAS code for cross-over studies we had here in the forum. Again you are estimating by least squares and adjust the tests appropriately.

I strongly assume (but had questioned because not sure) that the text of the guidance is aimed with respect to the estimation method by least squares. My reasoning is along your reasoning in respect to the 3x3 or 4x4 crossover study above: straight forward simple method, known and at hand by regulators, against 'complicated' iterative method in case of 'real' mixed model software, giving not always the same result, lacking also the well known ANOVA table of the least square solution.

Regards,

Detlew
ElMaestro
Hero

Denmark,
2010-02-01 17:22

@ d_labes
Posting: # 4687
Views: 12,995
 

 Less or More of an issue

Dear d_labes,

I think much of the confusion comes from the use of word "random".
It may lead to the impression that a model is fit with one or more random terms in addition to the error term, in other words, a mixed model. This is not the case.
When Proc GLM is executed it means in our context: "Fit a fixed effects model and do a standard anova". As pointed out this gives rise to a test of seq with the wrong error term.
When Proc GLM is executed with the proper bogus statement it means in our context "Fit a fixed effects model, and do an anova but when doing that make sure to do a proper calculation of P for the sequence effect".

I call it a bogus statement simply for linguistic reasons. It certainly makes sense to do what the statement does, but syntactically it is unfortunate that it is called "random" because it can lead the user to think (s)he is doing a mixed model fit.
If in any doubt, just leave out the bogus statement, get the normal anova and then do a posthoc correction for sequence effect with manual coding.

Best regards
EM.
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-02 00:55

@ d_labes
Posting: # 4690
Views: 12,989
 

 Effective with 1 Aug 2010

Dear D Labes!

» Hopefully this will be cleared before the magic 10-Aug-2010.

Why should it? And by whom?
After so many internal drafts, hundreds of pages of comments (will they ever be published?)*
I would guess the “EWP-PK drafting group’ is happy to survived and is right now running out of steam.

The Guideline was adopted by the CPMP on 20 Jan 2010 and will be effective with 1 Aug 2010.
Same story as it was with the current one (adopted Jul 2001, effective Jan 2002).

The sheep have 180 days left to follow their sheperds.

Not so sure
about that one any more:

[image]

On the other hand the MR-NfG was in Draft 15 state from July 1999 to Mai 2009. Then the "Corr."-Version appeared, where the word "Draft" was removed from pages 1 & 2. Do you expect that a magic hand will silently remove some words from the BE Guideline too? After five years of discussions (I've heard the "cookbook"-phrase back in Nov 2004)…


  • Yes! 248 pages – on EMA's website since 10 February 2010. See msg=4746]this post[/msg].

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-02-04 11:27

@ d_labes
Posting: # 4707
Views: 12,824
 

 Bias?

Dear d_labes,
(and HS who made similar arguments)

» What if we have period effects? In the '2x2 extracted evaluation' we are then mixing periods (for instance in T1T2R for the pair T1/R T1 comes from period 1, in sequence T2T1R from period 2).
»
» And what about sequences? For the above example T1T2R and T2T1R are the same in evaluating T1/R?
»
» And what about the error (MSE). In the 3x3 cross-over we have then two (distinct) MSE's and also CV's?
»
» Questions over questions. With the simple answer: This is statistically not so very sound, to be political correct (and not saying nonsense).

These are relevant and important points to make. I do not have a good reply and I am not a regulator. To make matters worse, I have not read any literature.
I could easily be 120% wrong, but ... All in all, I would think that whether something or not is wrong in this game comes down to whether or not the parameter on which a decision is based, is biased (sorry about this terrible sentence, I could not formulate myself better). Wrong is not the same as sub-optimal in the sense that 'minimum variance unbiased' and just plain 'unbiased' can be both correct although one may be more practical (some would say optimal) than the other. If it can be proven that some approach like the one discussed here biases the estimate of T/R then I think that could weigh in heavily. I am incompetent to tell if this is the case here.

Best regards,
EM.
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-05 02:03

@ ElMaestro
Posting: # 4712
Views: 13,072
 

 Bias?

Dear Firefighters!

» […] To make matters worse, I have not read any literature.

That’s a perfect starter!

» I could easily be 120% wrong,

Not 125.00%?

» I am incompetent to tell…

Like me. So I fired up a little simulation. I didn’t care about lognormal errors (too late in the evening).
T1 (mean 95, CV 15%), R1 (mean 90, CV 10%), R2 (mean 110, CV 15%). No identical CVs to add a little spice. Six sequence Williams's design; two setups: no period effect (response 1) and an additive period effect of 10 in P3 (response 2). My target ratios therefore are
T/R1 105.6%
T/R2  86.47%

R came up with:
subject sequence period treatment response 1 response 2
  1        1       1       T        120.6      120.6
  1        1       2       R1        81.56      81.56
  1        1       3       R2       136.7      146.7
  2        2       1       R1        96.86      96.86
  2        2       2       R2       134.0      134.0
  2        2       3       T         95.09     105.09
  3        3       1       R2        96.79      96.79
  3        3       2       T         82.46      82.46
  3        3       3       R1       101.3      111.3
  4        4       1       T        114.4      114.4
  4        4       2       R2        88.93      88.93
  4        4       3       R1        89.56      99.56
  5        5       1       R1        99.64      99.64
  5        5       2       T         97.84      97.84
  5        5       3       R2       117.8      127.8
  6        6       1       R2       101.5      101.5
  6        6       2       R1        72.80      72.80
  6        6       3       T         74.43      84.43
  7        1       1       T         76.64      76.64
  7        1       2       R1        83.70      83.70
  7        1       3       R2       116.8      126.8
  8        2       1       R1        86.61      86.61
  8        2       2       R2       116.2      116.2
  8        2       3       T         84.64      94.64
  9        3       1       R2       122.4      122.4
  9        3       2       T        104.8      104.8
  9        3       3       R1        87.69      97.69
 10        4       1       T         96.65      96.65
 10        4       2       R2       120.4      120.4
 10        4       3       R1        96.99     106.99
 11        5       1       R1        96.53      96.53
 11        5       2       T         89.58      89.58
 11        5       3       R2        77.61      87.61
 12        6       1       R2       115.7      115.7
 12        6       2       R1        94.51      94.51
 12        6       3       T         98.80     108.8

Raw mean ratios and pooled CV%s are
T/R1 105.6%  (11.9%)
T/R2  86.36% (15.3%)


WinNonlin 5.2.1 gave:
Full 3-period model:
        PE      90% CI      CV%
T/R1 103.81  94.38  114.2   13.59%  no period effect
T/R2  84.65  76.96   93.11  13.59%
T/R1 104.1   94.93  114.1   13.09% +P3 effect
T/R2  85.20  77.73   93.39  13.09%


Next I tortured the software and set R1 or R2 to ‘Missing’ (ignoring the warning messages).
Full 3-period model with missing values:
        PE      90% CI      CV%
T/R1 103.81  95.46  112.89  11.24% no period effect
T/R2  84.65  75.93   94.36  14.59%
T/R1 104.1   95.83  112.89  11.03% +P3 effect
T/R2  85.20  76.78   94.55  13.98%


Last the infamous 2×2 extractions:
        PE      90% CI      CV%
T/R1 103.81  95.43  112.93  11.41% no period effect
T/R2  84.65  76.75   93.35  13.29%
T/R1 104.1   94.71  114.31  12.76% +P3 effect
T/R2  85.20  77.20   94.04  13.39%


As expected PEs are unbiased in either method. In the full (correct, as I would dare to call it) model there is a common CV. Like D. Labes already suspected, CVs in the ‘missing data model’ and in the extracted tables are different. The outcome in some cases is anticonservative. Regulators, do you read this?

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-02-05 09:33

@ Helmut
Posting: # 4713
Views: 12,786
 

 Bias?

Dear HS,

firefighters??

» So I fired up a little simulation.

I LUUUV simulations! In this case, though, I do not understand the details. Did you simulate one or many datasets?
Did you investigate the T/R as n approaches inf?

» As expected PEs are unbiased in either method. In the full (correct, as I would dare to call it) model there is a common CV. Like D Labes already suspected, CVs in the 'missing data model' and in the extracted tables are different.

(G)LM's discard 2x2x2 data when missing values are involved. Mixed models fit via REML that optise a likelihood given a model; that can readily be done regardless of a missing value here or there.

» The outcome in some cases is anticonservative. Regulators, do you read this?

Or: Companies, let the music play! But actually, when you say in some cases, can that be quantified, as in, how often versus how often not?

EM.
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-05 10:21

@ ElMaestro
Posting: # 4714
Views: 12,768
 

 Bias?

Dear ElMaestro!

» Did you simulate one or many datasets? Did you investigate the T/R as n approaches inf?

No. After 11 hours of workshop I was too tired to set up a proper simulation. So only one dataset - the one given above.

» (G)LM's discard 2x2x2 data when missing values are involved. Mixed models fit via REML that optise a likelihood given a model; that can readily be done regardless of a missing value here or there.

ACK.

» » The outcome in some cases is anticonservative. Regulators, do you read this?
»
» But actually, when you say in some cases, can that be quantified, as in, how often versus how often not?

Just observed that CIs marked in red were tighter than the ones of the original dataset.

P.S.: EMA implicitly favours a Williams’ design. GL page 22 concerning ODTs:
[…] the following recommendations regarding study design apply:
  • if the reference medicinal product is taken only in one way (e.g. only with water), and the test product is intended for additional ways of administration (e.g. without water), the conventional and the new method should be compared with the reference in the conventional way of administration (3 treatment, 3 period, 6 sequence design).
(my emphasis)

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
d_labes
Hero

Berlin, Germany,
2010-02-05 13:59

@ Helmut
Posting: # 4715
Views: 12,791
 

 Bias?

Dear Helmut!

» As expected PEs are unbiased in either method.

Maybe this is due to the balanced situation here. I would expect some contamination in unbalanced situations.

» The outcome in some cases is anticonservative. Regulators, do you read this?

Bingo! :clap:
Helmut, continue this way.

Regards,

Detlew
ElMaestro
Hero

Denmark,
2010-02-05 16:00

@ Helmut
Posting: # 4716
Views: 12,726
 

 Bias?

Dear HS,

» Next I tortured the software and set R1 or R2 to ‘Missing’ (ignoring the warning messages).
» Full 3-period model with missing values:
»
»         PE      90% CI      CV%
» T/R1 103.81  95.46  112.89  11.24% no period effect
» T/R2  84.65  75.93   94.36  14.59%
» T/R1 104.1   95.83  112.89  11.03% +P3 effect
» T/R2  85.20  76.78   94.55  13.98%

Something is wrong here, or I am not getting it right. If all R1's are missing then an R1 effect cannot be estimated (for example, a column with all zero's in the model matrix for that effect, or, a beta vector shortened by one because that effect is left out). Are you sure you deleted all objects prior to re-running the analysis?

Since this goes off thread, I would be happy enough if this could be colported to its own thread.

EM.


Edit: See the follow-up thread. [Helmut]
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-05 19:14

@ ElMaestro
Posting: # 4719
Views: 12,724
 

 Bias?

Dear ElMaestro!

» » Next I tortured the software and set R1 or R2 to ‘Missing’…
»
» Something is wrong here, or I am not getting it right. If all R1's are missing then an R1 effect cannot be estimated (for example, a column with all zero's in the model matrix for that effect, or, a beta vector shortened by one because that effect is left out). Are you sure you deleted all objects prior to re-running the analysis?

OK, to be more precise: from the original dataset I generated two new ones: one with T and R1 and another one with T and R2, but the allocation to sequences and periods was kept 'as is'.

In WinNonlin-speak:
subject sequence period treatment response 1 response 2
  1        1       1       T        120.6      120.6
  1        1       2       R1        81.56      81.56
  1        1       3       R2      Missing    Missing
  …        …       …       …         …          …

and
subject sequence period treatment response 1 response 2
  1        1       1       T        120.6      120.6
  1        1       2       R1      Missing    Missing
  1        1       3       R2       136.7      146.7
  …        …       …       …         …         ...

I guess missing values in SAS-speak are denoted by a period ‘.’ instead.

» Since this goes off thread, I would be happy enough if this could be colported to its own thread.

Yes, maybe later on. There’s no way to do that directly by the forum’s software – manipulating the MySQL-backend calls for a fresher brain I’m having right now.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-01 13:43

@ ElMaestro
Posting: # 4678
Views: 13,062
 

 Final EMA oracle

Dear ElMaestro!

» Contrary to HS' opinion, I think this is exactly how that sentence should be interpreted for now.

Do I get you right - really like this one?

In Bonn last year we had a heated debate about the proposed 4-way fed/fasted study.
+-----+-------+-------+-------+-------+
| S/P |   1   |   2   |   3   |   4   |
+-----+-------+-------+-------+-------+
|  1  | R-fed | R-fst | T-fst | T-fed |
|  2  | T-fst | R-fed | T-fed | R-fst |
|  3  | T-fed | T-fst | R-fst | R-fed |
|  4  | R-fst | T-fed | R-fed | T-fst |
+-----+-------+-------+-------+-------+

Do you believe, that EMA wants us to come up with 4 separate pseudo-2×2 cross-over studies (T-fst/R-fst, T-fed/R-fed, T-fed/T-fst, R-fed/R-fst)?

Forget about the tricky things (imbalance, sequence/carry-over), but what about the period effect, which does not influence the treatment effect in a 2×2 cross-over? But here; probably apples-and-oranges.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-02-01 14:54

@ Helmut
Posting: # 4682
Views: 13,024
 

 Final EMA oracle

Dear d_labes, dear HS,

it was only an attempt from my side to guess how regulators would think in relation to that sentence; it was not an attempt to justify the science behind it.

What makes me think like I do has to do with
  1. Vocabulary. I think the sentence leaves little room for interpretation in its present form. The way that sentence is written I would definitely say that pruning a dataset (however bad science one may think it is) is compliant.

  2. Regulatory checking of results. In several countries across Europe, I think assessor teams do not routinely involve statisticians and the capability of an assessor to check results will in such cases be limited to simpler data like those from 2-seq 2-per 2-trt studies. If the applicant can prune the data they become widely checkable.
    Bear also in mind that the reason why only noncomportmental modeling is accepted has in large part to do with exactly the same.

Best regards
EM.
d_labes
Hero

Berlin, Germany,
2010-02-01 15:20

@ ElMaestro
Posting: # 4683
Views: 13,022
 

 Final EMA oracle

Dear ElMaestro,

Sorry, I had not the intention to put you in the shoes of defending this part of the Guidance. It was my outcry :crying: because I feared the worst you are definitely right in your interpretation.

That would mean for me a redesign of all the parts of my [image] programs dealing with crossover studies with more than two formulations! And a revision of all my internal effort to teach my folks the appropriate evaluation.

Regards,

Detlew
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-01 16:07

@ d_labes
Posting: # 4685
Views: 13,026
 

 Teaching

Dear D Labes!

» […] And a revision of all my internal effort to teach my folks the appropriate evaluation.

Yeah! I’ll be giving a workshop this Thursday in Slovenia and since last Friday I’m working (almost) day and night to update my presentations…

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
d_labes
Hero

Berlin, Germany,
2010-02-01 14:02

@ Helmut
Posting: # 4680
Views: 13,113
 

 Cmin really gone? tmax reappeared?

Dear Helmut,

» tmax is back (instead of partial AUC truncated at tmax,ref) - but no statistical analysis ("no apparent difference in median tmax and its variability").

What ever this means :surprised:! tmax reappeared as toothless tiger. Report the values, give some descriptive statistics and forget.

I find it very strange that the abhorrence for non-parametric methods leads to such a featureless evaluation guidance, even in the case tmax is clinically relevant.
Implicitly this accepts that tmax can only be analyzed non-parametrically, but "... Non-parametric analysis is not acceptable. ...". Thus we have no statistical method left.

» Css,min has vanished.

I liked this, because as we know this is a nasty parameter with sometimes high variability. Likewise it seemed fluctuation (PTF) had also gone.

But I noticed yet this is only true for immediate release products.

The CPMP/EWP/280/96 Corr "Note for Guidance on Modified Release Oral and Transdermal Dosage Forms: SECTION II (Pharmacokinetic and Clinical Evaluation)" mentions explicitly Fluctuation and Cmin in section 4.1.1. Rate and extent of absorption, fluctuation as parameters of interest in multiple dose studies.

Thus we have the situation that we need distinct sets of PK parameters for multiple dose studies depending on the product characteristics immediate or modified release. The question is: Why?

What is now really gone I think is MRT. Ok, this was only an parameter nice to have. Only reported and handled descriptively in BE studies.

Regards,

Detlew
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-01 23:30

@ d_labes
Posting: # 4688
Views: 13,033
 

 Cmin really gone. tmax reappeared - but how?

Dear D Labes!

» » "no apparent difference in median tmax and its variability").
»
» What ever this means. Report the values, give some descriptive statistics and forget.
»
» I find it very strange that the abhorrence for non-parametric methods leads to such a featureless evaluation guidance, even in the case tmax is clinically relevant.
» Implicitly this accepts that tmax can only be analyzed non-parametrically, but "... Non-parametric analysis is not acceptable. ...". Thus we have no statistical method left.

Exactly. OK, the median is mentioned. What is “no apparent difference”? Since non-parametrics are not acceptable - how far should we go? Damn, why is it not acceptable to use the right method? Are box plots acceptable? If yes, which ones? Or should we go with the weird plots suggested by

Sauter R, Steinijans VW, Diletti E, Böhm E and H-U Schulz
Presentation of results from bioequivalence studies
Int J Clin Pharm Ther Toxicol 30/Suppl1, S7-S30 (1992)


» » Css,min has vanished.
»
» But I noticed yet this is only true for immediate release products.
»
» Thus we have the situation that we need distinct sets of PK parameters for multiple dose studies depending on the product characteristics immediate or modified release. The question is: Why?

Oh, this doesn’t worry me. MR products are by their design more complicated to assess. For DR products I would look at tlag, for some products (double pulse) I would use partial AUCs (first part, second part), for CR with flip-flop-PK or transdermals my primary metric is AUCinf (not AUCt!), etc. IMHO, the one-size-fits-all concept is not applicable to MR products.

It would be of interest whether scaling for Cmax is acceptable for MR products also - and by analogy for Cmin. Time to update the MR guideline (was once announced for 2010). BTW, I’m planning for a replicate steady-state study just to see.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
d_labes
Hero

Berlin, Germany,
2010-02-01 14:52

@ Helmut
Posting: # 4681
Views: 12,981
 

 Interpol or not

Dear all!

In section 4.1.4 Study conduct under Sampling times it is stated: "... In multiple-dose studies, the pre-dose sample should be taken immediately before (within 5 minutes) dosing and the last sample is recommended to be taken within 10 minutes of the nominal time for the dosage interval to ensure an accurate determination of AUC(0-tau)...".

My questions:
  • Do I interpret this correctly that this prohibits any interpolation in case of time deviations at the measurement of time point t=tau?
  • Is extrapolation (as we discusses at times here in case of missing samples) also prohibited?
  • Would that mean, we are not able to calculate AUC(0-tau) in case of missing samples at t=tau?
  • Does this all also apply to AUC(0-72), truncated area?

Regards,

Detlew
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-02 00:02

@ d_labes
Posting: # 4689
Views: 13,122
 

 Interpol!

Dear D Labes!

» In section 4.1.4 Study conduct under Sampling times it is stated: "... In multiple-dose studies, the pre-dose sample should be taken immediately before (within 5 minutes) dosing and the last sample is recommended to be taken within 10 minutes of the nominal time for the dosage interval to ensure an accurate determination of AUC(0-tau)...".
»
» My questions:
  • Do I interpret this correctly that this prohibits any interpolation in case of time deviations at the measurement of time point t=tau?
  • Is extrapolation (as we discusses at times here in case of missing samples) also prohibited?
  • Would that mean, we are not able to calculate AUC(0-tau) in case
    of missing samples at t=tau?
  • Does this all also apply to AUC(0-72), truncated area?

Positive thinking! :-D
I guess this section is a reminder to sample accurately at critical time points. I will continue to do exactly the same I did for many years:
  • Imputation of missing values by interpolation; linear at t<tmax, loglinear at t≥tmax
  • If a value is measured not exactly at tau, using the inter-/extrapolated one
  • Detailed description in the protocol
  • Never any questions
Following our discussion I changed my procedure. If the 72 h value is missing in a particular subject after only one formulation, I use an estimate. If both values are missing (theoretically - never seen that), I would use AUCt instead of AUC72 - for that subject only.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ray_be
Junior

India,
2010-02-03 18:31

@ Helmut
Posting: # 4704
Views: 12,945
 

 Interpol!

Dear all,

As per the new guidance
"Actual time of sampling should be used in the estimation of the pharmacokinetic parameters and In multiple-dose studies, the pre-dose sample should be taken immediately before (within 5 minutes) dosing."
Now, while inserting time in winnonlin for 2nd, 3rd.... presdose concentration, whether we have to insert the actual time point/ the protocol specified time? For example if dosing interval is 24 hours, in that case wheather to use 23.92/ 24.00 for calculation of pharmacokinetic parameters?


Regards

Ray
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-03 20:26

@ ray_be
Posting: # 4705
Views: 13,069
 

 WinNonlin/Phoenix extrapolation to t=tau

Dear Ray!

» "Actual time of sampling should be used in the estimation of the pharmacokinetic parameters […] In multiple-dose studies, the pre-dose sample should be taken immediately before (within 5 minutes) dosing." Now, while inserting time in winnonlin for 2nd, 3rd… predose concentration, whether we have to insert the actual time point/ the protocol specified time? For example if dosing interval is 24 hours, in that case wheather to use 23.92/ 24.00 for calculation of pharmacokinetic parameters?

Which version of WinNonlin? Can you come up with a numerical example?
Are you talking about something like this?
 0     0    0     0    0     0
23.92  1   24.00  1   24.08  1
36.00  2   36.00  2   36.00  2
48.00  1   48.00  1   48.00  1

A dosing interval of 24 will give an AUCτ of 36 (linear trapezoidal rule) for the first and second data set, and 35.96 for the third. WinNonlin 5.2.1 will interpolate to t=24 (the τ you have specified). In the text output the time point 24.00 will be marked (@) and the line
@) Note - the concentration at dose time was added for extrapolation purposes.
will clarify the method.
You get the same results in Phoenix 6.1; The note about extrapolation to t=τ is given in the core-output.
I don’t see a problem here - or what would you suggest?
Phoenix is more flexible in setting up individual dosings. You may even give delayed administrations and define partial areas based on actual time points, etc. But do we really need that?

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-07 17:41

@ Helmut
Posting: # 4727
Views: 12,655
 

 Meta-analysis

Dear all!

Meta-analysis was not acceptable in the draft (see this post), but made it to the final GL (Section 4.1.8 Evaluation / Presentation of data):

If for a particular formulation at a particular strength multiple studies have been performed some of which demonstrate bioequivalence and some of which do not, the body of evidence must be considered as a whole. Only relevant studies […] need be considered. The existence of a study which demonstrates bioequivalence does not mean that those which do not can be ignored. The applicant should thoroughly discuss the results and justify the claim that bioequivalence has been demonstrated. Alternatively, when relevant, a combined analysis of all studies can be provided in addition to the individual study analyses. It is not acceptable to pool together studies which fail to demonstrate bioequivalence in the absence of a study that does.

(my emphasis)

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Panks.79
Junior

2010-03-08 07:22

@ Helmut
Posting: # 4875
Views: 12,225
 

 Final version published

Hi,

This is my first post on this forum, so kindly forgive me for any mistakes.

These guidelines are for IR products. Can this be applicable to MR product also. Can we assume that the CminSS is not required for MR product also.

Also would like to know if by anyways we can justify a steady state study that has failed at steady state for Cmin.

Thanks.

Dr. Pankaj Patil
Helmut
Hero
Homepage
Vienna, Austria,
2010-03-08 18:16

@ Panks.79
Posting: # 4877
Views: 12,188
 

 Cmin for MR-formulations?

Dear Pankaj!

» This is my first post on this forum,…

Welcome to the club!

» These guidelines are for IR products. Can this be applicable to MR product also.

1. Yes. 2. Very good question. The MR guidance is a little bit outdated (1999) - and you may expect that regulators are quite aware about recent developments. The current workprogramme of EMA's EWP states that a revision of CPMP/EWP/280/96 is to be considered in 2010. Reading the comments to the drafted BE-guideline, I got the impression that EMA removed the requirement due to the high variability reported by respondents. For controlled release formulations variability might be lower than variability of an IR formulation in steady state - but for a delayed release product it might well be even worse.

» Can we assume that the CminSS is not required for MR product also.

Even better question. I don’t think you should assume that. The MR NfG is still applicable; in the light of the new IR BE-GL I would suggest to go for a scientific advice (probably in a difficult country).

» Also would like to know if by anyways we can justify a steady state study that has failed at steady state for Cmin.

Hhm, post-hoc decisions are never a good idea. You can’t justify anything afterwards - you may only try to start an argument.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Marcel
Regular

2010-04-23 10:09

@ Helmut
Posting: # 5196
Views: 11,785
 

 Cmin for MR-formulations?

» The MR guidance is a little bit outdated (1999) - and you may expect that regulators are quite aware about recent developments. The current workprogramme of EMA's EWP states that a revision of CPMP/EWP/280/96 is to be considered in 2010.

Is this still the case? I'm looking through the current work programme for 2010 and can't find anything on this in it. If this is still the case, in your opinion can we expect a draft guidance out in 2010?
Helmut
Hero
Homepage
Vienna, Austria,
2010-04-23 12:34

@ Marcel
Posting: # 5197
Views: 11,751
 

 MR-Guideline?

Dear Marcel!

You did not dig deep enough. At the end of the table on page 60 it is stated:
CPMP/EWP/280/96 – Note for guidance on Modified Release Oral and Transdermal Dosage Forms: Section II – Revision to be considered

» If this is still the case, in your opinion can we expect a draft guidance out in 2010?

Don’t know. I’m not optimistic for a draft that early. As an example: regulators talked at conferences about a revision of the (IR)BE-GL in mid-2004, the Q&A was published as some kind of ‘intermediate’ in 2006, the ‘Recommendation on the Need for Revision of NfG on BA/BE’ in 2007, and the draft in 2008…

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Helmut
Hero
Homepage
Vienna, Austria,
2010-02-10 18:33

@ Ohlbe
Posting: # 4746
Views: 12,827
 

 Overview of comments published

Dear all,

the “Overview of Comments Received on Draft Guideline on the Investigation of Bioequivalence” (EMA/CHMP/EWP/26817/2010) was published today.
As expected the document is quite heavy (248 pages, 1.50 MB).
Enjoy, but take your time!

An interesting one (page 157) for people following the MR-guideline (Whenever multiple dose studies are performed it should be demonstrated that steady state has been reached):

Achievement of steady state can be evaluated by collecting pre-dose samples on the day before the PK assessment day and on the PK assessment day. A specific statistical method to assure that steady state has been reached is not considered necessary in bioequivalence studies. Descriptive data is sufficient.

No need for a statistical test (at least for EMA). I have to revise my previous recommendations (like in this post).

A comment on replicate designs (page 185) As there is still some controversy around replicate design studies and the statistical analysis, it would be very helpful if more detailed recommendations could be given here. What statistical analyses to be used? How to evaluate a replicate design in an average BE approach?
Was answered by

It is out of the scope of the guideline to give details on how to analyse the data of a replicate design, since it is standard statistical analysis.


[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
d_labes
Hero

Berlin, Germany,
2010-02-11 08:19

@ Helmut
Posting: # 4749
Views: 12,541
 

 Comments commented

Dear Helmut,

» A comment on replicate designs (page 185) As there is still some controversy ...
» ... What statistical analyses to be used? How to evaluate a replicate design in an average BE approach?

» Was answered by: It is out of the scope of the guideline to give details on how to analyse the data of a replicate design, since it is standard statistical analysis.
(emphases by me)

How wonderful! Congratulation! :clap:
In German: Da bleibt einem ja die Spucke weg.
(no translation of this idiom. roughly: leaves one speechless)


Edit: See the follow-up thread. [Helmut]

Regards,

Detlew
Ravi
Regular

India,
2010-03-13 12:05

@ Helmut
Posting: # 4910
Views: 13,416
 

 Definition of Cmin by EMEA

Dear All,

According to Overview of comments received on draft Guideline on the investigation of bioequivalence by EMEA Cmin,ss is defined as "By Cmin,ss we mean the concentration at the end of the dosage interval, i.e. Ctrough".
But according to WinNonlin User guide Cmin at steady state is defined as "Minimum concentration between dose time and dose time + Tau".

Now my question is Which definition is to be followed. What is the exact definition of Cmin according to other regulatory bodies (e.g. FDA, Health Canada, ANVISA). If definition by EMEA is to be followed then we can not use Cmin value calculated by using WinNonlin for submission. Please correct me if I am wrong.

Waiting for your reply.

Thanks & Regards
Ravi Pandey
Helmut
Hero
Homepage
Vienna, Austria,
2010-03-13 12:52

@ Ravi
Posting: # 4912
Views: 12,331
 

 EMA's Cmin & WinNonlin

Dear Ravi!

» Now my question is Which definition is to be followed.

The one you have stated in your approved protocol. ;-)
See this little survey.

» If definition by EMEA is to be followed then we can not use Cmin value calculated by using WinNonlin for submission.

Since Cmin in WinNonlin is always the true (=global) minimum within τ, you have to improvise. If all your concentrations at t=τ are ≥LLOQ, use Clast instead. If not, use the same set of rules you would use for values <LLOQ. Describe the procedure in an SOP and – even better – in the protocol.

[image]All the best,
Helmut Schütz 
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Back to the forum Activity
 Thread view
Bioequivalence and Bioavailability Forum | Admin contact
17,323 Posts in 3,705 Threads, 1,068 registered users;
27 users online (0 registered, 27 guests).

It’s easy to lie with statistics;
it is easier to lie without them.    Frederick Mosteller

The BIOEQUIVALENCE / BIOAVAILABILITY FORUM is hosted by
BEBAC Ing. Helmut Schütz
XHTML/CSS RSS Feed