Log in
Register|
jag009 ★★★ NJ, 2013-01-28 17:01 (4900 d 19:57 ago) Posting: # 9916 Views: 9,135 |
|
|
Hi everyone, If I want to run a 5-way crossover study, would 5-sequence be sufficient or I need to go 10? Crazy stuff... Thanks John |
|
d_labes ★★★ Berlin, Germany, 2013-01-28 17:37 (4900 d 19:21 ago) @ jag009 Posting: # 9917 Views: 8,289 |
|
|
Dear John, ❝ If I want to run a 5-way crossover study, would 5-sequence be sufficient or I need to go 10? Crazy stuff... what do you mean with five-way  .Assume you mean a cross-over study with 5 treatments in 5 periods. If you want to go with a Williams design you need 10 sequences. If you don't need the extra-balance of a Williams design, and IMHO you don't need it if you have no carry-over term in your evaluation model (specifically "first order" carry-over which is totally flawed, see 1)) you can go with 5 sequences (5x5 Latin square). Search for "Williams design" to find a lot of discussions we had here in the forum. (1)S. Senn Cross-over Trials in Clinical Research John Wiley & Sons, Chichester, pp162-163 (2nd ed. 2002) Chapter 10 — Regards, Detlew |
|
jag009 ★★★ NJ, 2013-01-28 18:18 (4900 d 18:40 ago) @ d_labes Posting: # 9918 Views: 8,122 |
|
|
Thanks d_labes, Yes a 5-period, 5-treatment, crossover study. There is always a first(in my case). I have done some article research this morning and most studies went with a 5-treatment, 5-sequence crossover instead of 10-sequence. I don't anticipate carryover from washout... Thanks John |
|
jag009 ★★★ NJ, 2013-01-29 16:12 (4899 d 20:46 ago) @ d_labes Posting: # 9923 Views: 7,985 |
|
|
Hi d_labes, I re-thought about what I was trying to accompolish in that experiment. What I really wanted to do was to simulate a set of data with a specific intrasubject CV. As I mentioned in previous post, I have a set of test and reference data already so I know the intrasubject CV. How do I generate different simulations (increase/decrease the intrasubject CV) with these test and reference data? Maybe a better question is how do I vary the intrasubject CV of a given set of test and reference data? Thanks John |
|
d_labes ★★★ Berlin, Germany, 2013-01-29 16:28 (4899 d 20:30 ago) @ jag009 Posting: # 9924 Views: 7,998 |
|
|
Hi John, can you be a bit more elaborate and specific? Don't get your points. What are you attempting to do? BTW: Don't mind to call me Detlew. — Regards, Detlew |
|
jag009 ★★★ NJ, 2013-01-29 17:04 (4899 d 19:54 ago) @ d_labes Posting: # 9925 Views: 8,025 |
|
|
Hi Detlew, I think I wrote the question in the wrong thread but I hope that you can help me as well. In a nutshell here is what I am trying to do. I have a set of data from a parital 3-way replicate study (1 Test, reference given twice). I want to find out what happens if I simulate this to a full replicate study. Therefore I need to 1) Generating a 2nd set of data for the test product 2) I also want to introduce some intrasubject CVs on the test data Question 1 can be done with bootstrap on the existing data(Correct?). But how should I approach question 2 so that I can see BE/SABE scenarios with different intrasubject CV on the test? thanks John |
|
d_labes ★★★ Berlin, Germany, 2013-01-29 17:52 (4899 d 19:06 ago) @ jag009 Posting: # 9927 Views: 7,955 |
|
|
Hi John, ❝ In a nutshell here is what I am trying to do. I have a set of data from a parital 3-way replicate study (1 Test, reference given twice). I want to find out what happens if I simulate this to a full replicate study. Therefore I need to ❝ 1) Generating a 2nd set of data for the test product ❝ 2) I also want to introduce some intrasubject CVs on the test data ❝ ❝ Question 1 can be done with bootstrap on the existing data(Correct?). May be. I'm not quite sure. Depends on the goal of your simulations I think. ❝ But how should I approach question 2 so that I can see BE/SABE scenarios with different intrasubject CV on the test? Simulating replicate studies is beyond the capacity of my brain. — Regards, Detlew |
|
jag009 ★★★ NJ, 2013-01-29 18:00 (4899 d 18:58 ago) @ d_labes Posting: # 9928 Views: 8,001 |
|
|
Thanks Detlew, ❝ ❝ Question 1 can be done with bootstrap on the existing data(Correct?). ❝ ❝ May be. I'm not quite sure. Depends on the goal of your simulations I think. The goal is to use existing data and just simulate. John |
|
ElMaestro ★★★ Denmark, 2013-01-30 11:10 (4899 d 01:48 ago) @ jag009 Posting: # 9932 Views: 7,976 |
|
|
Hi John, ❝ In a nutshell here is what I am trying to do. I have a set of data from a parital 3-way replicate study (1 Test, reference given twice). I want to find out what happens if I simulate this to a full replicate study. Therefore I need to ❝ 1) Generating a 2nd set of data for the test product ❝ 2) I also want to introduce some intrasubject CVs on the test data ❝ ❝ Question 1 can be done with bootstrap on the existing data(Correct?). But how should I approach question 2 so that I can see BE/SABE scenarios with different intrasubject CV on the test? I assume at this point you have estimates of intrasubj. sigma for Ref and you have an estimated T/R ratio and a between-subject variance estimate. Usually, when we talk replicated studies it implies a mixed model with compound-symmetric covariance matrix (FDA) or something similarly evaluated through a linear model (EMA). The model is Y = Treatment + Sequence + Period + (fixed interactions, who cares?) + Subj + error. It is the model that determines how you simulate. The error is the intra-subject term with mean zero and some variance. The variance is depending on whether the observation comes from Test or Reference. The Subj term is fixed as preferred by EMA or you can use random (mean zero and some variance) for FDA. Thus when you simulate the i'th subject who is given the j'th treatment (pardon me if I use i and j in other meanings than Chow and Liu, I don't remember what their i and j and k's were but that's not essential here), then you add the fixed effects and a sampling intra-subject error for the i'th subject being given the j'th treatment and a between-subject sampling error for that subject (FDA). In the dataset you will have several observations for one subject (e.g. subj. X receving two times test and two times ref = 4 observations for subject X). The 'added value' for the Subj term must be the same (constant) across all these four or whatever observations sinced this is a between-variability, but the error term is...well... randomly sampled in all cases for that subject. All in all, in addition to the fixed effects you may need two-three normal distributions for the sims: One for the intra-subject error associated with Test. One for the intra-subject error associated with Ref. One for the between-subject error (FDA). Note: In many cases you can forget about this one as well as the fixed effects for Sequence and Period. 1) A starting point is to use the same error for Test as you have ID'ed for Ref. Or you can add a bit or subtract a bit of variance for either depending on your level of optimism, how you have slept, position of Saturn's moons etc. This isn't a bootstrap since you have no replication for Test which you can resample. 2) Answered in the previous line I believe. — Pass or fail! ElMaestro |
|
jag009 ★★★ NJ, 2013-01-31 16:31 (4897 d 20:27 ago) @ ElMaestro Posting: # 9936 Views: 7,873 |
|
|
Thanks ElMaestro, I spoke with my past colleagues and he suggested something similar to what you said but instead of using T/R ratio, he said I can go ahead with using AUC and Cmax. Since I have the intrasubject CV from the reference datasets. I should apply this error to the generation of the 2nd test datasets (to generate AUC, Cmax) from the one test dataset as a starting pt for the 1st simulation. I can vary the intrasubject CVs to generate more scenarios. ie: take Subject 1 data, sample from normal distribution randomly with mean 0, SD= Withinsubject standard deviation. Do the same for subject 2.... Is this similar to what you described? John |
|
ElMaestro ★★★ Denmark, 2013-01-31 18:13 (4897 d 18:45 ago) @ jag009 Posting: # 9937 Views: 7,896 |
|
|
Hi John, ❝ I spoke with my past colleagues and he suggested something similar to what you said but instead of using T/R ratio, he said I can go ahead with using AUC and Cmax. Since I have the intrasubject CV from the reference datasets. I should apply this error to the generation of the 2nd test datasets (to generate AUC, Cmax) from the one test dataset as a starting pt for the 1st simulation. I can vary the intrasubject CVs to generate more scenarios. ❝ ❝ ie: take Subject 1 data, sample from normal distribution randomly with mean 0, SD= Withinsubject standard deviation. Do the same for subject 2.... ❝ ❝ Is this similar to what you described? I am not sure I follow completely, but by and large this sounds correct, although I would probably prefer to simulate all values (both Test and Ref) rather than just re-using those values you had in the previous study. Both approaches could be informative. I don't know how to simulate directly on the T/R for a repliacted study, so I also agree that it should be done for AUC (or Cmax) for each subject and treatment, observing the desired T/R through the relation between the fixed treatment effects. That part is not so hard. — Pass or fail! ElMaestro |
|
jag009 ★★★ NJ, 2013-02-01 16:33 (4896 d 20:25 ago) @ ElMaestro Posting: # 9944 Views: 7,834 |
|
|
Hi ElMaestro, ❝ I am not sure I follow completely, but by and large this sounds correct, although I would probably prefer to simulate all values (both Test and Ref) rather than just re-using those values you had in the previous study. This is the part where I am having trouble with. How do you simulate all values instead of re-using what I have from the study? Would the following be correct? Lets me just focus on a crossover study (forget replicate for now) 1) Obtained the mean and SD for AUC, Cmax of each dataset (log?) 2) Randomly generate from normal distribution AUC and Cmax (based on the mean and SD) 3) introduce the withsubject error (method as mentioned in last post) to each value. Or I better off use T/R ratio for the above instead of using parameters? I am new to this area so any comment would be very helpful. Thanks John |
|
ElMaestro ★★★ Denmark, 2013-02-01 17:50 (4896 d 19:08 ago) @ jag009 Posting: # 9945 Views: 7,934 |
|
|
Hi John, ❝ This is the part where I am having trouble with. How do you simulate all values instead of re-using what I have from the study? Would the following be correct? Lets me just focus on a crossover study (forget replicate for now) ❝ 1) Obtained the mean and SD for AUC, Cmax of each dataset (log?) ❝ 2) Randomly generate from normal distribution AUC and Cmax (based on the mean and SD) ❝ 3) introduce the withsubject error (method as mentioned in last post) to each value. Yes that's almost it (and yes to logs ). You just simulate the Test and Ref values on basis of the ratio of treatment marginal means (point estimate) or expectations you have. So if the T/R ratio was 0.98 then you can just set Ref to 1 (zero on the log scale) and sample Test from a normal dis. with mean log(0.98) and whatever st.dev. you like. Sokme people (Detlew?) sometimes do a small correction due to the assymetric distribution on the ordinary scale, I believe. It is solely the T/R ratio and the variation that affects the conclusion towards BE so for most applications you can simply forget about between-subject variation and the Per and Seq effects. Therefore, I'd skip pt. 2 above. ❝ ❝ Or I better off use T/R ratio for the above instead of using parameters? ❝ I am new to this area so any comment would be very helpful. Potvin showed how to simulate the T/R directly (i.e. not Test and Ref. separately for each subject) for a normal 2,2,2-scenario. You arrive at exactly the same conclusion if you do it this way, of course. But I don't know how to simulate T/R directly in the replicated situation, so there I would just do it the harder way as described above. — Pass or fail! ElMaestro |
|
jag009 ★★★ NJ, 2013-02-01 21:36 (4896 d 15:22 ago) @ ElMaestro Posting: # 9946 Views: 7,817 |
|
|
Hi ElMaestro, ❝ You just simulate the Test and Ref values on basis of the ratio of treatment marginal means (point estimate) or expectations you have. So if the T/R ratio was 0.98 then you can just set Ref to 1 (zero on the log scale) and sample Test from a normal dis. with mean log(0.98) and whatever st.dev. you like. Sokme people (Detlew?) sometimes do a small correction due to the assymetric distribution on the ordinary scale, I believe. Let me make sure if I am on the right path... The st.dev you referred to is the std.dev of the T/R ratios? What I mean is --> T &R Cmax data from 24 subjects, calculate T/R ratio per subject, caluate mean and SD for T/R ratio, then use the mean and SD from this T/R ratio data to generate the T and R values as you listed above? Once I get this set of data, I can then randomly assign errors (sample from normal distribution with mean 0 and Ste.dev from within subject standard deviation) and add to the data? Thanks John |
|
ElMaestro ★★★ Denmark, 2013-02-02 03:21 (4896 d 09:37 ago) @ jag009 Posting: # 9948 Views: 7,799 |
|
|
Hi John, ❝ Let me make sure if I am on the right path... The st.dev you referred to is the std.dev of the T/R ratios? What I mean is --> T &R Cmax data from 24 subjects, calculate T/R ratio per subject, caluate mean and SD for T/R ratio, then use the mean and SD from this T/R ratio data to generate the T and R values as you listed above? Strictly the st.dev. comes from an anova's mean sq. error (model fixed effects Per, Seq, Treatment, Subj; the latter can be a random as well without impact on the outcome in the non-repliacted case) if you have one such. That way you have taken the fixed effects into account for its calculation. — Pass or fail! ElMaestro |
|
jag009 ★★★ NJ, 2013-02-04 16:06 (4893 d 20:52 ago) @ ElMaestro Posting: # 9963 Views: 7,834 |
|
|
Hi ElMaestro, ❝ Strictly the st.dev. comes from an anova's mean sq. error (model fixed effects Per, Seq, Treatment, Subj; the latter can be a random as well without impact on the outcome in the non-repliacted case) if you have one such. That way you have taken the fixed effects into account for its calculation. Thanks. Got it  John |
|
d_labes ★★★ Berlin, Germany, 2013-02-04 10:05 (4894 d 02:53 ago) @ ElMaestro Posting: # 9957 Views: 7,761 |
|
|
Hi Grosser Meister, ❝ ... Sokme people (Detlew?)  Turkish?❝ sometimes do a small correction due to the assymetric distribution on the ordinary scale, I believe. BTW: IMHO you guys should ultimately change the subject line. — Regards, Detlew |
|
ElMaestro ★★★ Denmark, 2013-02-04 12:16 (4894 d 00:42 ago) @ d_labes Posting: # 9960 Views: 7,728 |
|
|
Hi Detlew, ❝ ❝ ... Sokme people (Detlew?) ❝ I apologise - I meant "some". I shall immediately fire my secretary who had the delegated task of writing this post on behalf of me.  — Pass or fail! ElMaestro |
Ing. Helmut Schütz