Log in
Register|
Helmut ★★★   Vienna, Austria, 2012-06-02 20:54 (4344 d 18:39 ago) Posting: # 8656 Views: 11,277 |
|
|
Gossip: One regulator (I leave it to your imagination which one of EMA’s working parties he belongs to) told me that Our [i.e., his agency’s] statisticians don’t believe* that in Potvin’s Method C the patient’s risk is preserved at ~5%. They suggest to use only Method B because application of Pocock’s α 0.0294 is well known. I suggested to tell them to have a closer look at the empiric alphas (the endless tables in the paper). BTW, specifically Method C (not B!) is recommended by the FDA and in the recent Canadian GL.
— 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 |
|
d_labes ★★★ Berlin, Germany, 2012-06-04 12:34 (4343 d 02:59 ago) @ Helmut Posting: # 8657 Views: 9,039 |
|
|
Dear Helmut, ❝ Quoting my late father: “If you don’t want to know, but to believe – go to church.” Nothing to add to your late father. But unfortunately the regulatory body concerned seems to alter into some sort of church we are obliged to go to. Many Articles of Faith and Commandments in their guidances and advices. — Regards, Detlew |
|
ElMaestro ★★★ Denmark, 2012-06-05 13:45 (4342 d 01:48 ago) @ Helmut Posting: # 8662 Views: 8,969 |
|
|
Hi HS, that's a complex area. I can understand why regulators prefer to stay with things they know, like Pocock's 0.0294. But as you say, in the context of equivalence there is no proof that this value provides superior protection against inflated type I error. In my humble experience method C and B perform equally well by and large, possibly with a little advantage in terms of type I error to method B on the whole cf. table I in their first publication. Or perhaps the logic is that if you use alphas of 0.05 first and 0.0294 next (= method C) then the overall type I error cannot be less than 0.05? This latter argument is flawed as the tables show. We have an analogous situation with endpoints in pharmacodynamic asthma equivalence studies. The classical endpoint in asthma is FEV1 (a measure of lung function), but it turns out in practice that this endpoint is rather useless once we enter the field of equivalence. It took EU regulators more than 10 years to realise this after these facts were widely published (and trust me, in the name of self-criticism, it took me very long time to realise it myself). On one hand we/they/someone need to be very careful when we/they/someone just blindly assume that what works with superiority will also work for equivalence. On the other hand, we should appreciate signals from regulators. They give us something we can base our actual plans on. And besides, I could be completely wrong as I usual am, in practice I think it is just as easy to plan a mehod B study as a method C study and their sample sizes appear equal also. I am happy with either. I have a feeling the two-stage approaches are going to become more and more prominent for generic developers so there is certainly a need to have these debates. — Pass or fail! ElMaestro |
|
Helmut ★★★ Vienna, Austria, 2012-06-07 17:45 (4339 d 21:49 ago) @ ElMaestro Posting: # 8673 Views: 8,961 |
|
|
Hi ElMaestro! ❝ I can understand why regulators prefer to stay with things they know, like Pocock's 0.0294. Do you think that they know/understand the context of Pocock’s original 0.0294 (superiority testing, parallel groups, fixed total sample size k, normal distribution – Z, known and equal σ², interim look at k/2)? ❝ On one hand we/they/someone need to be very careful when we/they/someone just blindly assume that what works with superiority will also work for equivalence. Well said. ❝ On the other hand, we should appreciate signals from regulators. They give us something we can base our actual plans on. And besides, I could be completely wrong as I usual am, in practice I think it is just as easy to plan a method B study as a method C study and their sample sizes appear equal also. Total sample sizes are equal (see Table II), but with very few exceptions Method B more often goes into Stage 2 than Method C. That’s a consequence of B’s penalty in Stage 1. ❝ I am happy with either. I’m happy with B and very happy with C.  ❝ I have a feeling the two-stage approaches are going to become more and more prominent for generic developers so there is certainly a need to have these debates. Yep. Was the topic in one thirds of the chats I had in the coffee-breaks of this week’s workshop in Budapest. — 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, 2012-06-07 19:01 (4339 d 20:32 ago) @ Helmut Posting: # 8676 Views: 8,955 |
|
|
Hi HS, ❝ Do you think that they know/understand the context of Pocock’s original 0.0294 (superiority testing, parallel groups, fixed total sample size k, normal distribution – Z, known and equal σ², interim look at k/2)? Very good point and I like what you are hinting at. Defaulting to 0.0294 whenever the word "two-stage" pops up is an understandable thing to do today although I'd like to see that change tomorrow. In the absence of more papers discussing the two-stage designs I can't imagine the change will happen quickly. I guess I will store it mentally along with the likes of 2.9979×108 m/s and 6.626×10-34 Js, and I will call it "Pt" for 'Pocock's universial two-stage constant'. Communication and education is necessary. Let N1 be the number of subjects in stage 1, let a1, a2 be the two alphas and E the overall type I error rate, let CV be the coefficient of variation and T:R the test:ref ratio. Then:
Furthermore, the crackpot paper demonstrated that given for any set of (a1, N1, CV, T:R) we can identify an a2 which keeps E under 0.05 or whatever level is deemed useful. Hey, next time they call from IQPC or Informa I will suggest such a talk in an upcoming conference. ❝ (...) the topic in one thirds of the chats I had in the coffee-breaks of this week’s workshop in Budapest. I can imagine. Coffee breaks don't get better than that. — Pass or fail! ElMaestro |
|
Helmut ★★★ Vienna, Austria, 2012-06-07 20:28 (4339 d 19:05 ago) @ ElMaestro Posting: # 8678 Views: 9,019 |
|
|
Dear ElMaestro! ❝ Defaulting to 0.0294 […] I guess I will store it mentally along with the likes of 2.9979x108 m/s and 6.626x10-34 Js, and I will call it "Pt" for 'Pocock's universial two-stage constant'. Communication and education is necessary. Yep. Nice term. ❝ […] can we produce proof that […] ❝ Elementary my dear Watson, the answers are given by Potvin herself. Not quite. Don’t mix up (mathematical/statistical) proof with (pseudoempiric) evidence obtained by simulations. Since the latter contains endless tables it can easily be discerned from the former, which looks like: From the statement __________ I invited a group from the universities of Vienna and Lancaster to work on the problem a good while ago. They said that ‘in principle’ a full adaptive design should be possible for any cross-over, including replicates. Seems to be more challenging that they thought in the first place… ❝ Furthermore, the crackpot paper demonstrated … 
— 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 |
Ing. Helmut Schütz