Log in
Register|
AngusMcLean ★★ USA, 2013-01-11 03:19 (4491 d 05:43 ago) Posting: # 9823 Views: 6,261 |
|
|
Good Evening (Guten Abend): We have completed a parallel group PK design of a drug formulation (test versus reference). So there is only 1 period and 2 formulations evaluated in 2 groups of 12 subjects. I have the program WinNonlin and another program that does ANOVA (NCSS) To use the data and make a sample size estimate for a future parallel group study. I believe I need to have the inter-subject variance of the PK paramater. I use the LN values of the PK paramater. So I think a one way ANOVA is appropriate and what is the correct name of the inter subject variance paramater given by the ANOVA output? I am on the LN scale so I think I take the square root of the intersubject variance paramater and multiply by 100 to get the CV as percentage. Please can I be advised of the terminology and if I am doing the correct calculation. Best Regards, Angus |
|
Helmut ★★★   Vienna, Austria, 2013-01-11 04:04 (4491 d 04:58 ago) @ AngusMcLean Posting: # 9824 Views: 5,611 |
|
|
Dea-maidin Angus! ❝ I have the program WinNonlin… Forget it (will tell you later why). ❝ …and another program that does ANOVA (NCSS) OK, could do the job. ❝ I believe I need to have the inter-subject variance of the PK paramater. Nope. You have the total variance. See this post. ❝ I use the LN values of the PK paramater. So I think a one way ANOVA is appropriate and what is the correct name of the inter subject variance paramater given by the ANOVA output? Haven’t used NCSS for years. Maybe residual variance? ❝ I am on the LN scale so I think I take the square root of the intersubject variance paramater and multiply by 100 to get the CV as percentage. That’s an approximation. Use \(100\sqrt{e^{MSE}-1}\) like you would do in a Xover. ❝ Please can I be advised of the terminology and if I am doing the correct calculation. If you using the formula above, no problem. Now a warning (see this presentation; some more stuff here). Though the t-test is fairly robust if the two groups show unequal variances, it is sensitive against imbalance.* The outcome will be liberal (CI too narrow). So you asked one question (the CV) and get another answer: WinNonlin (up to the latest release of Phoenix 6.3) gives the CI based on the t-test for equal variances – even if you select “Satterthwaite degrees of freedom” in the options. That’s a bug. ![[image]](img/search.png) Search the forum for some background information.
— 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 |
|
AngusMcLean ★★ USA, 2013-01-11 18:34 (4490 d 14:29 ago) @ Helmut Posting: # 9828 Views: 5,345 |
|
|
Thank you Helmut: this is a confusinbg field. It seems that the English language description of the output you get is poor. Both WinNonlin and NCSS are from here. The lack of clarity is the major problem and of course statistics is a confusing subject especially when the terms used are not well described or ccan be equivalent. I compared the output from WinNonlin and NCSS: the variance term is exactly the same. WinNonlin calls the variance term "variance residual" NCSS provides a row in a Table and calls it the Error term of the 2 treatments A and B and it refers to it as the Mean Square Error (MSE) in another part of the output. They are in my example exactly the same number. So it seems that indeed it is the total error or total variance not the inter subject variance we have. So that means that the value we have is a convolution of the intersubect error and intrasubject error? Or is there another factor I am overlooking? And as previously referred to is higher than intersubject variance so using it provides a higher estimate of sample size. Angus |
|
Helmut ★★★ Vienna, Austria, 2013-01-11 18:59 (4490 d 14:03 ago) @ AngusMcLean Posting: # 9829 Views: 5,353 |
|
|
Hi Angus! ❝ It seems that the English language description of the output you get is poor. Both WinNonlin and NCSS are from here. You are absolutely right. Sometimes manuals increase confusion. PHC/WNL states in one place “variability” meaning standard deviation in one setting and variance in an other.  ❝ The lack of clarity is the major problem and of course statistics is a confusing subject especially when the terms used are not well described or ccan be equivalent. You are not alone. ❝ I compared the output from WinNonlin and NCSS: the variance term is exactly the same. WinNonlin calls the variance term "variance residual" NCSS provides a row in a Table and calls it the Error term of the 2 treatments A and B and it refers to it as the Mean Square Error (MSE) in another part of the output. The latter is the most commonly one given by other software. ❝ They are in my example exactly the same number. OK, fine. I was concerned about the way WNL calculates the confidence interval – which assumes equal variances. Only the degrees of freedom of the t-test have to corrected by the Satterthwaite approximation. If you want to estimate the sample size for another study you need just the residual variance, aka mean square error. ❝ So it seems that indeed it is the total error or total variance not the inter subject variance we have. Yep. ❝ So that means that the value we have is a convolution of the intersubect error and intrasubject error? I would rather call it a combination. Convolution smells of ∫… ❝ And as previously referred to is higher than intersubject variance so using it provides a higher estimate of sample size. Almost. You could separate inter- and and intrasubject variance only in a cross-over study (see here). If you have data of a parallel study and want to plan another one simply use the error term. Call it CVtotal if you are a nitpicker like me. — 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 |
|
AngusMcLean ★★ USA, 2013-01-12 16:31 (4489 d 16:31 ago) @ Helmut Posting: # 9832 Views: 5,291 |
|
|
Helmut: Thank you; that is an excellent slide you refer me to. Yes convolution is not good. I do agree combination is better. The term shown in the slide MSEB is the Between subject mean square error term from a crossover study and denotes Between subject mean square error term from the ANOVA of that study. This ANOVA term is a misnomer, since it contains a combination of between subject error mean square error and within subject mean square error terms. That is why you have to make the subtraction for inter that you show in the slide. Do you agree with misnomer. Angus |
Ing. Helmut Schütz