Log in
Register|
ratnakar1811 ★ India, 2008-04-02 16:07 (5866 d 09:57 ago) Posting: # 1750 Views: 17,523 |
|
|
Dear All, This is regarding the missing time points, we do write in the protocol that if the two or more time points are missing the inclusion of the subject will be decided by the statistician considering the effect of missing time points on the out-come of the PK data. But one of our sponsor says that we should have a priori criteria in percentage. For example if he had completed 90% of samples then he will be included and if not then will be excluded from the analysis. Plz comment. Ratnakar -- Edit: Category changed. [Helmut] |
|
Helmut ★★★   Vienna, Austria, 2008-04-02 16:54 (5866 d 09:11 ago) @ ratnakar1811 Posting: # 1751 Views: 15,943 |
|
|
Dear Ratnakar, I think your approach is reasonable, but consider also:
❝ But one of our sponsor says that we should have a priori criteria in percentage. For example if he had completed 90% of samples then he will be included and if not then will be excluded from the analysis. I would call this some kind of formalistic nonsense.  Imagine the situation of a delayed release formulation which shows a lot of variability in tlag due to gastric emptying. Rate of absorption and elimination are pretty nicely stable within subjects. If we have 18 sampling points, the first two after the predose sample are missing (e.g., vial broken in centrifugation) and the next two samples are still BLQ (due to tlag), your sponsor would really want to exclude the subject and decrease the power of the study?  — Dif-tor heh smusma 🖖🏼 Довге життя Україна! ![[image]](https://static.bebac.at/pics/Blue_and_yellow_ribbon_UA.png) Helmut Schütz ![[image]](https://static.bebac.at/img/CC by.png) The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
atish_azad ☆ 2008-07-29 13:03 (5748 d 13:01 ago) @ Helmut Posting: # 2090 Views: 15,577 |
|
|
Dear HS and all,
Atish |
|
ElMaestro ★★★ Denmark, 2008-07-29 13:21 (5748 d 12:43 ago) (edited by Jaime_R on 2008-07-29 15:25) @ atish_azad Posting: # 2091 Views: 15,637 |
|
|
Hi, in the EU Q&A there is a section (number 3) which deals with classification of subjects as outliers. Although this is only borderlining your situation, the section gives one general and important sentence which is relevant to this problem: "Exclusion of data can never be accepted on basis of stat analysis or for PK reasons alone." EM. -- Edit: Full quote removed. Please see this post! [Jaime] |
|
Helmut ★★★ Vienna, Austria, 2008-07-29 23:40 (5748 d 02:24 ago) @ ElMaestro Posting: # 2097 Views: 15,784 |
|
|
¡ Hola ElMaestro ! ❝ "Exclusion of data can never be accepted on basis of stat analysis or for PK reasons alone." Sure – although debatable.  But Ratnakar’s and Atish’s posts dealt with exclusion of subjects based on missing values. If three samples in the area of the expected Cmax were broken in centrifugation we probably will fall into a deep pit (especially if CVintra is low and the sample size is small – a single ‘outlier’ will kill the study). Unless we start fiddling around with cubic splines we will never be able to get a reasonable estimate of Cmax. All NCA methods will grossly underestimate Cmax – which may blow the study. Cmax is nasty in many respects. Even if just a little ‘noise’ is present, PK modeling regularly has its problems getting close to the measured data in this region. IMHO defining ‘critical regions’ of the profile in the protocol is not a bad idea. No “[…] PK reasons alone”, but common sense added! If one wants to come up with a hard-core justification of the ‘critical regions’: perform PK modeling (pilot study or literature data) and plot the partial derivatives of parameters vs. time; extremes (min/max) will indicate most influential time points. Remark: never seen this in a protocol – maybe freaks like me do it at home.  — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
ElMaestro ★★★ Denmark, 2008-07-31 12:38 (5746 d 13:26 ago) @ Helmut Posting: # 2113 Views: 15,326 |
|
|
❝ ❝ "Exclusion of data can never be accepted on basis of stat analysis or ❝ for PK reasons alone." ❝ IMHO defining ‘critical regions’ of the profile in the protocol is not a bad idea. No “[…] PK reasons alone”, but common sense added! ❝ If one wants to come up with a hard-core justification of the ‘critical regions’: perform PK modeling (pilot study or literature data) and plot the partial derivatives of parameters vs. time; extremes (min/max) will indicate most influential time points. Remark: never seen this in a protocol – maybe freaks like me do it at home. Hi, Dat's a good point to make, HS. But I think the current trend is towards more 'cook book' and less science or judgment. It might be difficult to make a recipe for this situation which is not based on some degree of subjectivity (subjectivity- sometimes disguised as 'common sense' - is something that can be tricky to handle for regulators as well as applicants, I think). Let me emphasize, I like the idea nevertheless. However, I could also imagine regulators thinking "Plan the study well, make sure it is robust towards 'eventualities'. If the loss of two samples (or whatever) is remotely/reasonably possible then of course a study should be planned so that such an event does not scr¤w it all up. If a study is planned well it will not be necessary to fiddle around with data." or something along those lines. N'est-ce pas? EM. -- Edit: Full quote removed. Please see this post! [Jaime] |
|
Helmut ★★★ Vienna, Austria, 2008-07-31 16:36 (5746 d 09:28 ago) @ ElMaestro Posting: # 2122 Views: 15,408 |
|
|
¡ Hola ElMaestro ! ❝ But I think the current trend is towards more 'cook book' and less science or judgment. I came across the term ‘cook book’ for the first time in Lisbon last year.  To quote Les Benet: Even though it’s applied science ❝ It might be difficult to make a recipe for this situation which is not based on some degree of subjectivity (subjectivity- sometimes disguised as 'common sense'- is something that can be tricky to handle for regulators as well as applicants, I think ). […] When I started learning my first programming language in 1976 (haha, that’s a sidenote for youngsters) my instructor told us the story of the ‘apple pie algorithm’ (well known amongst IT people): Target is to write an algorithm (the ‘apple pie recipe’) which will enable anybody (not only cooks) to come up with an apple pie tasting as the one’s your granny always baked. First you think it’s simple, but digging deeper into the problem, you realize that you assume too much (temperature, duration – OK, but which kind of stove, etc.). Morale of the story: cook books are well sold, having nice pictures in it, but:
or consider becoming a cook.  ❝ However, I could also imagine regulators thinking "Plan the study well, make sure it is robust towards 'eventualities'. If the loss of two samples (or whatever) is remotely/reasonably possible then of course a study should be planned so that such an event does not screw it all up. If a study is planned well it will not be necessary to fiddle around with data." or something along those lines. My yesterday’s example is a realistic one. Consider an IR formulation of a drug with a fast half life and very low variability (CVintra 7%). If the study is planned for the minimum sample size (12 in many regulations) and two reserves allowing for drop-outs (which actually did not happen) you are in the nasty situation of an overpowered study anyhow (the point-estimate-not-within-100%-business is likely). Even if you have a lot of sampling points around tmax, missing data will kill you (I’m referring to an example where due to only two lost samples in one subject CV increased to 22%). How to make such a study ‘robust’? Running it in 24 subjects - which will raise ethical questions – or, if ‘nothing happens’ will result in a CI like 93–97% (haha). Or follow Martin’s suggestion and exclude a subject if a single value is missing… The second example was the never-ending story of omeprazole single dose fed state. Recently I saw one of these 100+ subject studies where 40 (!) blood samples (OK just 1 mL) where drawn within 24 hours. The sponsor tried to make the study ‘robust’ (to be protected against high variability in Cmax caused by variability in gastric emptying mainly), unfortunately not robust enough. Maybe next time they will go for more samples… ❝ N'est-ce pas?
— Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
![[image]](img/uploaded/Magritte.jpg)
|
Helmut ★★★ Vienna, Austria, 2008-07-29 23:18 (5748 d 02:46 ago) @ atish_azad Posting: # 2096 Views: 15,708 |
|
|
Dear Atish! ❝ 1. What are the reasons that can be given for exclusion of subject from statistical analysis, having more than two missing time points. Where does this number two come from? ❝ 2. On what basis we can decide the uncritical region. Depending on the drug, formulation, sampling scheme,… Examples of critical regions are the absorption phase in the case of an IR formulation, around tmax and the (late) elimination phase. ❝ 3. By interpolation (linear in the rising part, log/linear after tmax) do you mean that we have to show un-transformed or raw data to be linear in rising part and after tmax ln-transformed data to be linear. No, you cannot show that (because we don’t now the ‘true’ time course). Lin-up/Log-down is a crude – but useful – approximation in NCA. If you start a poll amongst NCA experts – that’s the one probably the majority will opt for. ❝ Kindly clarify and give a breif idea about what do you mean by interpolation. OK, have a look at one of my lectures (slides 51-53). The profile comes from the ‘real world’; afterwards (knowing the results already) I cracked the 48 hours’ sample vial on my PC. The measured value was 3.805 ng/mL – but since I cracked the vial, suddenly I didn’t know the value any more: If we draw a straight line (e.g. don’t attempt to deal with the situation) – that’s the linear trapezoidal rule within 36 and 72 hours. Unexpressed we treat the missing value as 4.996 ng/mL; AUC will be over-estimated. If we perform a log-linear interpolation we would get an estimated value of 3.850 ng/mL which is much closer to the ‘true’ one. If you don’t know about this game and just treat the timepoint as missing you will end up with an overestimated AUC. Unfortunately that’s exactly the default in WinNonlin. State the method for missing values in the protocol – and change the option in this nice piece of software. What’s actually happening then is a calculation by the linear trapezoidal method and a switch to log/linear between 36 and 72 hours. If you don’t like this idea, you may opt for lin-up/log-down in all cases (not just the one with missing data). — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
martin ★★ Austria, 2008-07-29 15:56 (5748 d 10:08 ago) @ ratnakar1811 Posting: # 2092 Views: 15,418 |
|
|
Dear Ratnakar! according to my point of view you have to define of a full analysis data set (FADS) and a per-protocol analysis data set (PADS) according to ICH E9. for FADS you have to analyze the data as they are whereas for the PADS you can specify exclusion/inclusion criteria (e.g. a minimum of x time points prior Tmax and a minimum of y time points after Tmax) where Tmax can be calculated from data of a pilot study. best regards martin |
|
Helmut ★★★ Vienna, Austria, 2008-07-30 00:18 (5748 d 01:46 ago) @ martin Posting: # 2098 Views: 15,583 |
|
|
Dear Martin! ❝ according to my point of view you have to define of a full analysis data set (FADS) and a per-protocol analysis data set (PADS) according to ICH E9. ❝ for FADS you have to analyze the data as they are whereas for the PADS you can specify exclusion/inclusion criteria […] ![[image]](img/uploaded/image3.jpg) To quote ICH E9 (Section 5.3):“If no procedure for dealing with outliers was foreseen in the trial protocol, one analysis with the actual values and at least one other analysis eliminating or reducing the outlier effect should be performed and differences between their results discussed.” (my emphasis in blue) Reading the Q&A document quoted by ElMaestro in his post again I’m getting the impression members of the CHMP/PK group don’t (want to?) know this guideline… If I anticipated events and stated a decision-tree in the protocol, there’s no need to come up with two bloody evaluations. To quote from the movíe ‘Highlander’: There can be only one! What's the full data set if a subject comes up with a zero response after the reference formulation?Is T/R of this subject ‘infinity’ or only ‘undefined’? Ask a mathematician. My kind of software simply cannot deal with such a value… — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
martin ★★ Austria, 2008-07-30 12:28 (5747 d 13:36 ago) @ Helmut Posting: # 2099 Views: 15,354 |
|
|
dear hs ! thank you for comment, I think I mixed up definition of data sets and handling of missing values, here is my point of view in more detail. based on my experience with these kind of things usually two data sets are used (artificial example for definitions of a cross-over trial): Full analysis data set (intent-to-treat): subjects that were randomized and received at least one study treatment. Per-protocol analyses data set: subjects that were randomized, met all inclusion criteria, were dosed according to the protocol, received both treatments with measurements at all assessment time points with both study drugs, etc. based on these example definitions, handling of missing values is applicable for the full analyses dataset only. handling of missing values for calculation specific PK parameters (e.g. primary endpoint) and circumstances when these approach will be applied should be stated in the protocol (decision tree) or can be documented after a blinded review of the data. for PK analyses, both data sets are analyzed and the results for the full analyses data set is the more important as pointed out by the ICH E9 guideline (i.e. primary analysis). for safety analyses, only the full analysis data set is used. best regards martin |
|
Helmut ★★★ Vienna, Austria, 2008-07-30 15:30 (5747 d 10:34 ago) @ martin Posting: # 2102 Views: 15,507 |
|
|
Dear Martin! ❝ Full analysis data set (intent-to-treat): subjects that were randomized and received at least one study treatment. Full ACK. ❝ Per-protocol analyses data set: subjects that were randomized, met all inclusion criteria, were dosed according to the protocol, received both treatments with measurements at all assessment time points with both study drugs, etc. IMHO only almost.  I disagree with: ❝ […] met all inclusion criteria, […] In the ‘real world’ sometimes a subject was included although an exclusion criterion should have prevented this. A ‘classical example’ is a violation of limits set for the Body Mass Index. Due to rounding sometimes a subject finished the study and the violation becomes only evident during QAU data clearing. In such a case I would handle it according to ICH E6 (Section 10.2 Protocol Deviations), and keep the subject in the evaluation. ❝ […] measurements at all assessment time points with both study drugs […] You’re suggesting to exclude a subject if one sample vial is broken? ❝ based on these example definitions, handling of missing values is applicable for the full analyses dataset only. Why? ❝ for PK analyses, both data sets are analyzed and the results for the full analyses data set is the more important as pointed out by the ICH E9 guideline (i.e. primary analysis). Disagree. See NfG on BA/BE (Section 3.8 Reporting of results): ‘All results should be clearly presented and should include data from subjects who eventually dropped-out. Drop-out and withdrawal of subjects should be fully documented and accounted for.’ At least by means of an ANOVA it’s simply not possible to run a BE analysis of the full data set in a conventional 2×2 cross-over if the second period is missing. Of course PK parameters have to be calculated as far as possible and presented. A nice example is the case where a subject drops out let’s say in the middle of the elimination phase of the second period. In such a case it would be possible to base the assessment of Cmax on the data set of n subjects and of AUC of n-1 subjects in order to maintain maximum power.❝ for safety analyses, only the full analysis data set is used. 100% ACK. — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
d_labes ★★★ Berlin, Germany, 2008-07-30 15:39 (5747 d 10:25 ago) @ martin Posting: # 2103 Views: 15,315 |
|
|
Dear Martin ❝ ... ❝ for PK analyses, both data sets are analyzed and the results for the full analyses data set is the more important as pointed out by the ICH E9 guideline ... ^^^^^^ Just to quote ICH E9 paragraph 3.3.2 Trials to Show Equivalence or Non-inferiority: "... The equivalence (or non-inferiority) trial is not conservative in nature, so that many flaws in the design or conduct of the trial will tend to bias the results towards a conclusion of equivalence. For these reasons, the design features of such trials should receive special attention and their conduct needs special care ..." IMHO this calls for the per protocol analysis as the primary one. Best would be if you have only per protocol volunteers/subjects. — Regards, Detlew |
|
Helmut ★★★ Vienna, Austria, 2008-07-30 16:34 (5747 d 09:31 ago) @ d_labes Posting: # 2106 Views: 15,418 |
|
|
Dear D. Labes! ❝ Just to quote ICH E9 paragraph 3.3.2 Trials to Show Equivalence or Non-inferiority: […] ❝ IMHO this calls for the per protocol analysis as the primary one. Full ACK. BTW haven’t seen anything else. ❝ Best would be if you have only per protocol volunteers/subjects. Yeeees!  — Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
martin ★★ Austria, 2008-07-30 16:50 (5747 d 09:14 ago) @ d_labes Posting: # 2107 Views: 15,474 |
|
|
dear hs and dlabes ! thanks for your comments; the example criteria "met all inclusion criteria" can of course be dropped and/or limited to specific criteria (i.e. explicitly defined major protocol violations) important for a specific study and yes I would exclude a subject from the PADS even when one vial is broken (ICH E9: the availability of measurements of the primary variable). however, you can also specify a specific pharmacokinetic analysis set (PKAS) using the following example criteria for a cross-over trial: pharmacokinetic analysis set: subjects received both study treatments and have suitable data for PK analysis (a) samples are available at pre-dosing and 15 min post each study drug administration (b) for at least x out of y time points specified in the protocol, etc. etc. using this approach you don't have to run PK analyses for the FADS and PADS. you have to define your method how missing values are handled for this data set only. however, I have seen both approaches in practice (a) analyses of FADS and PADS separately and (b) analysis of PKAS only. regarding the importance of FADS and PADS: the FADS should reflect the clinical situation in practice better than the rather artificial PADS and is according to my point of view more important. I personally think that the importance of a data set should not depend on the hypothesis and/or parameters investigated. I have to admit that I am not fully convinced by the ICH E9 in this respect but I may have to think about this topic in more detail. best regards & thanks for the very interesting discussion martin |
|
Helmut ★★★ Vienna, Austria, 2008-07-30 19:14 (5747 d 06:50 ago) @ martin Posting: # 2109 Views: 15,446 |
|
|
Dear Martin, I’m with you in the definition of data sets (especially PKAS(b) etc. etc.)… The important point is an a-priori strategy, which has to be adapted according to the PK characteristics of the drug, etc. ❝ regarding the importance of FADS and PADS: the FADS should reflect the clinical situation in practice better than the rather artificial PADS and is according to my point of view more important. Yes, but phase I trials never reflect the clinical situation. IMPs in a BE study are not labeled for sole use in healthy male Caucasian* subjects, preferably non-smokers of a BMI [19-27] fasting four hours after administration together with 150 mL non-carbonated water of ambient temperature avoiding methyl-xanthine containing food and beverages as well as grapefruit juice, charcoal broiled meat and broccoli during the study (did I forget something?). A properly designed BE study should be powered to be able to detect differences between formulations in the smallest possible number of subjects – which we hope to attain by a high degree of standardization. This has nothing in common with the clinical situation. ❝ […] thanks for the very interesting discussion Same to you!
— Dif-tor heh smusma 🖖🏼 Довге життя Україна! Helmut Schütz The quality of responses received is directly proportional to the quality of the question asked. 🚮 Science Quotes |
|
d_labes ★★★ Berlin, Germany, 2008-07-31 11:18 (5746 d 14:46 ago) @ Helmut Posting: # 2112 Views: 15,269 |
|
|
Dear HS, ❝ [...] ❝ IMPs in a BE study [...] http://en.wikipedia.org/wiki/IMP # Imp, a fantasy creature. # Imp, a type of fictional lower level devil in the Dungeons & Dragons role-playing game. ... — Regards, Detlew |
Ing. Helmut Schütz