Log in
Register|
earlybird ☆ 2010-04-09 09:54 (5512 d 21:08 ago) Posting: # 5059 Views: 9,652 |
|
|
Dear D_Labes, I am not a R user, but as I learned you can do meta-analysis (combined p-values analysis) of studies (hint: BE guideline: Chapter 4.1.8 / "Alternatively when relevant a combined analysis of all studies can be provided in addition to the individual study analysis." Or how do you solve that in SAS? Greetings earlybird Edit: Opened as a new thread (was linked to Survey: R use in BE/BA, category changed. [Helmut] |
|
d_labes ★★★ Berlin, Germany, 2010-04-09 10:35 (5512 d 20:27 ago) @ earlybird Posting: # 5060 Views: 8,867 |
|
|
Dear earlybird, ❝ I am not a R user, but as I learned you can do meta-analysis (combined p-values analysis) of studies (hint: BE guideline: Chapter 4.1.8 / "Alternatively when relevant a combined analysis of all studies can be provided in addition to the individual study analysis." ❝ ❝ Or how do you solve that in SAS? Although I have not done any serious meta-analysis in SAS upto now I guess it is possible. All is possible with ![[image]](img/uploaded/image2.gif)  See for instance SUGI 27: SAS® Tools for Meta-analysis — Regards, Detlew |
|
Helmut ★★★   Vienna, Austria, 2010-04-09 22:33 (5512 d 08:29 ago) @ d_labes Posting: # 5070 Views: 8,889 |
|
|
Dear D. Labes & earlybird! ❝ SUGI 27: SAS® Tools for Meta-analysis Is it really that easy? See the example at page 2: Suppose we have [...] two identical studies, following the same protocol, using the same treatments, using the same response variable, and suppose once again that we can obtain the data. We could then fit a mixed model (if we wanted to treat study as a random effect), using PROC MIXED, or an ANOVA model (if we wanted to treat study as a fixed effect), using PROC GLM. It took me while to find suitable data in my files - who repeats a study without any modifications? This was the closest I could get: AUC-data, same protocol (clinical site, sampling schedule, analytical method,...). Different batches of test and reference were used (actually the reference came from two countries). Single studies Fixed: sequence+treatment+periodNaïve pooling (study not in the model) Fixed: sequence+treatment+periodInteresting: The CV is higher than in either study despite the higher sample size. OK, let's add the study to the model... Fixed: sequence+treatment+periodAha No.1  And code all as fixed... Fixed: sequence+treatment+period+subject(sequence)+studyAha No.2 Besides some minor differences in the second decimal of the lower CL all models came up the same. Why all the fuzz with sequential designs, alpha-adjustements and the like - if I can simply throw all data into the big pot, stir it up, and voilá... Again: Is it really that easy? Both studies were already evaluated, the alpha is 'consumed' - any professional statisticians out there? — 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, 2010-04-10 00:02 (5512 d 07:00 ago) @ Helmut Posting: # 5071 Views: 8,587 |
|
|
Hi HS, I do not have the definitive answer, but I have a comment and a question. Comment: Study will be a between-subject factor. Thus it should not influence the intra/residual. Adding "study" as a factor thus might not mean much in terms of the outcome as compared to an anova and the subsequent CI without that factor. Question: Can you show a bit more? I wonder why the combined CV's (9.31%) are higher than those from the two individual studies. Were the subjects uniquely coded - i.e. "Subject 1" was uniquely assigned to either study 1 or 2 etc? Best regards EM. — Pass or fail! ElMaestro |
|
Helmut ★★★ Vienna, Austria, 2010-04-10 00:11 (5512 d 06:51 ago) @ ElMaestro Posting: # 5072 Views: 8,597 |
|
|
Dear ElMaestro! ❝ Study will be a between-subject factor. Thus it should not influence the intra/residual. Adding "study" as a factor thus might not mean much in terms of the outcome as compared to an anova and the subsequent CI without that factor. Right. ❝ Can you show a bit more? Just leaving the office. Maybe next year.  ❝ I wonder why the combined CV's (9.31%) are higher than those from the two individual studies. Surprised me too. ❝ Were the subjects uniquely coded - i.e. "Subject 1" was uniquely assigned to either study 1 or 2 etc? Yes. Codes were study (1-2) treatment (test/reference) subject (001-013 in study 1, and 101-112 in study 2) period (1-2) P.S.: Nice signature. — 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, 2010-04-10 00:20 (5512 d 06:42 ago) @ Helmut Posting: # 5073 Views: 8,683 |
|
|
Ahoy! ❝ Yes. Codes were ❝ study (1-2) ❝ treatment (test/reference) ❝ subject (001-013 in study 1, and 101-112 in study 2) ❝ period (1-2) There's a period 1 and 2 in both studies, so one could code it as the period-by-study interaction in analyses where study is included. I have not thought this fully through and if I tried my brain would prolly just idle because I am not a statistician. ❝ P.S.: Nice signature. Dank je wel. Best regards EM. — Pass or fail! ElMaestro |
|
Helmut ★★★ Vienna, Austria, 2010-04-10 02:49 (5512 d 04:13 ago) @ ElMaestro Posting: # 5074 Views: 8,844 |
|
|
Ahoi, ElMaestro! ❝ […] I have not thought this fully through […] because I am not a statistician. Me not either. Maybe codes for period should be simply 1-4. Pooled: n=25, 95.19% (91.11% - 99.46%), CVintra 9.03%Results identical for all models (yeah, treatments still within). Well, how to set up your interaction? Goede wind! — 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, 2010-04-10 10:13 (5511 d 20:49 ago) (edited on 2010-04-10 12:34) @ Helmut Posting: # 5075 Views: 8,591 |
|
|
Goede morgen, ❝ Well, how to set up your interaction? I only know R (Study:Per)1. But it could also be faked by calling the periods of the second study "3" and "4" etc. Then no need to specify the interaction, I think. EM. Addendum: Just re-read your post - you are suggesting exactly the same if I interpret your post right? 1: Sorry wrote Trt:Per at first - typo. |
|
Helmut ★★★ Vienna, Austria, 2010-04-10 16:46 (5511 d 14:16 ago) @ ElMaestro Posting: # 5076 Views: 8,531 |
|
|
Ahoi! ❝ Just re-read your post - you are suggesting exactly the same if I interpret your post right? Yes. After RTFM and setting up the model with a study × period interaction term, I'm getting exactly the same results as with faked four periods. There are only minor differences between the 'faked 4 period' and interaction set-ups, probably due to numerical issues. More significant digits below: Faked Periods: 95.1926% (91.1103% - 99.4578%), CVintra 9.02706%Apart from the technical issues, my main concern remains alpha in the combined analysis.  Though a sailor and a scuba-instructor supposedly are familiar with murcky waters, I guess we would need some statistical consultancy here. Wist het anker!  — 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, 2010-04-11 13:44 (5510 d 17:17 ago) @ Helmut Posting: # 5077 Views: 8,480 |
|
|
Hi HS, ❝ I guess we would need some statistical consultancy here. Completely agree. Wouldn't it be wonderful if for example sdavis joined in with a comment here in this forum, in this thread, dealing with this topic? Best regards EM. |
|
Helmut ★★★ Vienna, Austria, 2010-04-11 16:27 (5510 d 14:35 ago) @ ElMaestro Posting: # 5078 Views: 8,583 |
|
|
Dear ElMaestro! Yes, Simon is a nice and clever lad - not only here, but also in the real world. But this problem is not specific for WinNonlin. I'm not sure whether Simon is monitoring all posts here. If you think it would be nice joining us - contact him. — 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, 2010-04-12 15:32 (5509 d 15:30 ago) @ Helmut Posting: # 5087 Views: 8,484 |
|
|
Dear Helmut! ❝ Is it really that easy? ❝ ... Why all the fuzz with sequential designs, alpha-adjustements and the like - if I can simply throw all data into the big pot, stir it up, and voilá... ❝ ❝ Again: Is it really that easy? ❝ Both studies were already evaluated, the alpha is 'consumed' - any professional statisticians out there? Since I'm only a bloody raw recruit in statistics and since I haven't done serious meta-analysis up to now I can't answer these questions really professional. But what I have read scarcely about that topic there is actually no such thing in meta-analysis like alpha-adjustment or alpha-spending. How should it be? Meta analysis was invented for pooling studies (mostly from literature) which generally are not under control of the meta-analyst and had generally already been evaluated with some alpha. Therefore there is little or no room for alpha-spending at all. Thats why meta-analyses are not generally accepted as confirmatory proof. But there are regulatory concerns, f.i. in CPMP/EWP/2330/99 "Points to consider with 1. meta-analysis 2. one pivotal study". On page 2 is stated: "When a meta-analysis is included in an application it should be performed in accordance with a protocol specifying ... -Statistical methods (including methods for investigation of heterogeneity). In case where the meta-analysis provides the pivotal evidence for an indication, a p-value more extreme then the conventional significance level of 0.05 would generally required, and narrow confidence intervals for the treatment effect would be expected. The required degree of significance will be judged on a case by case basis considering factors such as amount of supportive data, plausibility of hypothesis tested and whether the analysis is pre-specified or not." Emphasis by me. BTW: I would incorporate a treatment by study interaction in the model as test of heterogeneity. BTW2: What big differences did you expect with such a very low intra-subject variability? BTW3: See Chow/Liu "Design and Analysis ...", Chapter 16 "Meta-analysis for bioequivalence review" to be more confused .— Regards, Detlew |
|
Helmut ★★★ Vienna, Austria, 2010-04-12 16:25 (5509 d 14:37 ago) @ d_labes Posting: # 5088 Views: 8,531 |
|
|
Dear D. Labes! Agree with your points (e.g., not confirmatory). I did not know this EMA-document, THX. Interesting Section II.1.3: A retrospective meta-analysis of only two studies originally intended to stand on their own is not expected to add any useful information. In particular, a meta-analysis cannot be used to reconcile the conflicting results of one positive and one inconclusive study. ❝ BTW: I would incorporate a treatment by study interaction in the model as test of heterogeneity. OK. ❝ BTW2: What big differences did you expect with such a very low intra-subject variability? I didn't expect anything. This was just the only dataset I had coming closest to two identical studies. As said in my previous post it is unlikely to repeat a study with any modifications. ❝ BTW3: See Chow/Liu "Design and Analysis ...", Chapter 16 "Meta-analysis for bioequivalence review" to be more confused Oh.  — 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, 2010-04-12 17:03 (5509 d 13:59 ago) @ Helmut Posting: # 5091 Views: 8,396 |
|
|
Dear Helmut! ❝ Interesting Section II.1.3: ❝ A retrospective meta-analysis of only two studies originally intended to stand on their own is not expected to add any useful information. In particular, a meta-analysis cannot be used to reconcile the conflicting results of one positive and one inconclusive study. THX for highlighting this. IMHO its not so very sound, not to say ... urrgh (cervix blocked to to keep a civil tongue ), especially if the study sizes differ to a greater extent.Fortunately this is weaker formulated in the new BE guidance, text pointed to by earlybird, page 16: "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." Hopefully this overrules. — Regards, Detlew |
|
Helmut ★★★ Vienna, Austria, 2010-04-12 17:18 (5509 d 13:44 ago) @ d_labes Posting: # 5093 Views: 8,437 |
|
|
Dear D. Labes! ❝ […] cervix blocked to to keep a civil tongue Cervix?! Not larynx? — 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, 2010-04-12 17:33 (5509 d 13:29 ago) @ Helmut Posting: # 5094 Views: 8,626 |
|
|
Dear Helmut! ❝ Cervix?! Not larynx? Oh, what a shame. False page in my anatomy atlas  .(Leo "Hals") — Regards, Detlew |
|
Helmut ★★★ Vienna, Austria, 2010-04-12 17:40 (5509 d 13:22 ago) @ d_labes Posting: # 5095 Views: 8,451 |
|
|
Dear D. Labes! ❝ Oh, what a shame. False page in my anatomy atlas Maybe correct image of the body, but upside down. Hint: if you experience difficulties in reading the text - turn around. — 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