Log in
Register|
boonchai_l ☆ Thailand, 2010-03-11 09:04 (5942 d 15:06 ago) Posting: # 4893 Views: 7,528 |
|
|
Dear all, Suppose that you have to run BE study (standard design) with 80 subjects and your housing is able to contain only a half. What should you do to solve this problem? Thanks in advance! BL |
|
Helmut ★★★   Vienna, Austria, 2010-03-11 13:36 (5942 d 10:35 ago) @ boonchai_l Posting: # 4896 Views: 6,737 |
|
|
Dear Boonchai! ❝ […] BE study (standard design) with 80 subjects and your housing is able to contain only a half. ❝ ❝ What should you do to solve this problem? Perform the study in two groups of fourty subjects. Use the same protocol, mention the split, keep the time interval between groups as short as possible, and see this thread for the statistics. — 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 |
|
ElMaestro ★★★ Denmark, 2010-03-11 22:09 (5942 d 02:01 ago) @ Helmut Posting: # 4899 Views: 6,663 |
|
|
Dear Boonchai and all, I agree with HS's post, but would like to strech it a little. What if you do a two-stage approach, so that once the first batch of subjects are dun, you have the usual interim stats and then set forth towards recruiting all the rest (of course, with some luck you'd even need to include fewer than first feared).  Best regards EM. |
|
Dr_Dan ★★ Germany, 2010-03-12 10:04 (5941 d 14:06 ago) @ ElMaestro Posting: # 4901 Views: 6,611 |
|
|
Dear all Advantage of interim stats: with some luck you'd even need to include fewer subjects than first feared. Disadvantage: with some bad luck you'd even need include more than first feared due to analyses conducted at adjusted significance levels. Furthermore an interim analysis takes time (you have to perform the bioanalysis of the PK samples before starting the second group). For a generic company this is a strong argument. If you follow Helmut's suggestion ensure that bioanalysis will not start before the subjects of the second group are enrolled. Kind regards Dan — Kind regards and have a nice day Dr_Dan |
|
Helmut ★★★ Vienna, Austria, 2010-03-12 13:21 (5941 d 10:49 ago) @ Dr_Dan Posting: # 4907 Views: 6,644 |
|
|
Dear Dan! ❝ Advantage of interim stats: with some luck you'd even need to include fewer subjects than first feared. Agree with you and EM. Furthermore it's the only possible solution if you don't 'know' the CV beforehand. ❝ Disadvantage: with some bad luck you'd even need include more than first feared due to analyses conducted at adjusted significance levels. In Boonchai’s example there’s a limit in the clinical capacity. His expected sample size is 80 and he has only the option to start with 40. Or more generally: Anybody opting for a sequential design should power the first stage as a common 2×2 study - with all the side conditions in mind (point estimate, variability, drop-out rate,…). So if everything works out like you expected, you are in the left branch of Potvin’s method, and the study stops (no adjusted alpha!). The right branch might save the day. What are your options in a conventional (failed) study? BTW: the penalty for adjusted alpha is not really an issue. For the common ±5%, 80% power, CV 20% we need n=20 (alpha 0.05) and n=24 (alpha 0.0294). Most people would go for 90% power: n=26 (alpha 0.05), n=30 (alpha 0.0294). See it as a price to pay for a study which otherwise will fail anyway. Again: if you plan to be BE in the first stage, and your assumptions are fulfilled - no penalty at all! ❝ Furthermore an interim analysis takes time (you have to perform the bioanalysis of the PK samples before starting the second group). For a generic company this is a strong argument. Agree with the first part. But if you don't have really good information of the CV (and by good I don't mean literature data or a lousy pilot study) there is always a chance of failure. The only way to deal with a uncertain CV in conventional studies is to add safety margins. According to ICH E9 you have to include a sensitivity analysis in the protocol ('what-if' scenarios). From my consultancy business: A company tried to be 'on the safe side' and submitted a protocol to the IEC with a power of 95% in the worst case (maximum drop-out rate, deviation from reference 8%, CV 10% higher than ones seen in previous studies). Power without these obstacles would have been 98%. The company didn't care on money and ethics and only wanted to get the product approved 'as fast as possible and for sure'. The IEC rejected the protocol. — 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 |
|
Dr_Dan ★★ Germany, 2010-03-12 14:04 (5941 d 10:06 ago) @ Helmut Posting: # 4908 Views: 6,615 |
|
|
Dear Helmut I totally agree: Sequential design = opportunity. If I plan to be BE in the first stage, and my assumptions are fulfilled - everything is perfect. There is no right or wrong. Your decision how to plan the study will depend on the risk you want to take, the time you need for the study and the money that you have to spend. In Boonchai's example there's a limit in the clinical capacity. His expected sample size is 80 and he has only the option to start with 40. What do you think about a replicate design study? A replicate design study with 40 subjects will be better than a two way cross-over study with 80 subjects devided into two groups since you can claim the widened acceptance range for Cmax. Kind regards Dan — Kind regards and have a nice day Dr_Dan |
|
Helmut ★★★ Vienna, Austria, 2010-03-12 16:01 (5941 d 08:09 ago) @ Dr_Dan Posting: # 4909 Views: 6,627 |
|
|
Dear Dan! ❝ In Boonchai’s example there’s a limit in the clinical capacity. His expected sample size is 80 and he has only the option to start with 40. What do you think about a replicate design study? A replicate design study with 40 subjects will be better than a two way cross-over study with 80 subjects divided into two groups since you can claim the widened acceptance range for Cmax. Yes, why not? Though we are a little bit in the business of reading tea leaves until Boonchai tells us more. From the sample size a HVD/HVDP is likely and more often than not Cmax is the most critical metric. But the widened AR is not suitable is some regulations. Another problem is the drop-out rate and the higher blood-loss. It’s a pity that no method is available right now for a sequential replicate design.  — 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 |
|
Dr_Dan ★★ Germany, 2010-03-15 12:21 (5938 d 11:49 ago) @ Helmut Posting: # 4914 Views: 6,547 |
|
|
Dear Helmut (or whoever can help) You wrote that ❝ It’s a pity that no method is available right now for a sequential replicate design. I am not a statistician or mathematician and therefore have to rely on expert statements. We are currently planning a sequential replicate design BE study in cancer patients. Two different statisticians had no objections against the design and now you are telling me that there is no method available. So I am confused. Please advice. I am looking forward to your reply Kind regards Dan Edit: Please stay within the thread. Linked to previous post, standard quotes restored. [Helmut] — Kind regards and have a nice day Dr_Dan |
|
Helmut ★★★ Vienna, Austria, 2010-03-15 16:37 (5938 d 07:33 ago) @ Dr_Dan Posting: # 4915 Views: 6,843 |
|
|
Dear Dan! ❝ I am not a statistician or mathematician and therefore have to rely on expert statements. You know: two experts = three opinions? ❝ We are currently planning a sequential replicate design BE study in cancer patients. Since your patients are in steady state anyway, the replicate design is easy to perform; just an additional profile in each period – if you don't run into ethical problems (blood loss). ❝ Two different statisticians had no objections against the design … See me first sentence. ❝ … and now you are telling me that there is no method available. OK. It try to sort things out. There is an abundance of methods for designing and evaluating sequential and adaptive designs. These methods have their origin in testing for a significant difference, i.e., the Null-Hypothesis H0 is ‘no difference’ and the Alternative Hypothesis Ha is a ‘significant difference’ between treatments. In bioequivalence the problem is formulated differently: H0 is ‘bioinequivalence’, and Ha is ‘bioequivalence’. The latter is accepted by inclusion of the confidence interval in the acceptance range. This distinction in the formulation of hypotheses is not trivial. Example (AR 80–125%, 90% CI):
As you see these examples lead to contradictory results if the wrong set of hypotheses are employed. The problem with the extensive literature on sequential/adaptive designs is that to my knowledge (almost) all references deal with ‘classical’ significance testing. See the introductory section of Potvin et al. (2008)¹ and a general overview in Gould (1995).² If a method based on the wrong hypotheses is used, patient's risk may be inflated (i.e., to an unknown degree higher than 5 %). None of the ‘classical’ approaches are validated for cross-over studies and the formulation of the test problem in bioequivalence (inclusion rule or two one-sided t-tests). Don’t get me wrong, of course it’s possible to push the button to run a macro in SAS – but IMHO up to now no method is published showing no inflation of the alpha-risk in an sequential/adaptive replicate design study. BTW, such a simulation study would have to be extremely large: whereas the conventional 2×2 crossover assumes a common variance, setting up a simulation for different variances of test and reference would be challenging at least.
— 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 |
|
boonchai_l ☆ Thailand, 2010-03-16 13:25 (5937 d 10:46 ago) (edited on 2010-03-17 03:51) @ Helmut Posting: # 4916 Views: 6,591 |
|
|
Dear HS, At first I think it is a design issue because I just think of replicate design or sequential design for solving. Thank you for that thread and many involving links, It makes me more understand. Now I already knew a suitable way for me to solve the problem. As you state in that thread ❝ I haven't seen a single study in the EU where a group effect was included in the statistical model... I would like to know in this situation which one you will select between including group effect and single analysis, because I would like to include group effect in the model. Neither, I haven't seen a single study in Thailand where a group effect was included in the model. So the way to do the study smoothly I ready to do, time wasting to argue with Thai FDA. Thanks in advance. BL Edit: Standard quotes restored. [Helmut] |
|
ElMaestro ★★★ Denmark, 2010-03-18 19:21 (5935 d 04:50 ago) @ boonchai_l Posting: # 4934 Views: 6,525 |
|
|
Hi all, ❝ I would like to know in this situation which one you will select between including group effect and single analysis, because I would like to include group effect in the model. Neither, I haven't seen a single study in Thailand where a group effect was included in the model. So the way to do the study smoothly I ready to do, time wasting to argue with Thai FDA. Not exactly sure I know the meaning of the question, but I am sure I need to improve my understanding of the need for the group effect when groups are separated. What do we test for specifically, and what do we conclude if the resulting p-value is low? I have not read any papers on this issue. I would think that group is a between-factor, so the anova residual will not depend on whether or not group is added as a factor in the lm. In fact I would also be inclined to think that group*treatment would be a potentially relevant test. What would you experts think of group or group*treatment coming out significant? Nothing? Something? We could in theory get LSMeans for Treatment T and R in both groups, and if group*treatment is significant it implies that T/R may differ between the two groups (which could suggest they may differ in time). This is a cosmic mindf&#ker to me. You experts out there got a qualified view on this? Many thanks and best regards, EM. |
Ing. Helmut Schütz