Log in
Register|
Croosov ☆ Germany, 2014-08-26 17:26 (3911 d 00:43 ago) Posting: # 13433 Views: 6,613 |
|
|
Hello, unfortunately I am not an expert in the matter of biostatistics yet, so maybe you can help me with one problem. What test would you use for tmax if you have a 3 period + 3 treatment design or 4 period + 4 treatment design with one treatment as a reference. At the moment I am using Friedman test but would it be appropriate to use generalized mixed effect model if you use LMEM (proc mixed) for Cmax and AUC? Thanks in advance! Edit: Category changed. [Helmut] |
|
d_labes ★★★ Berlin, Germany, 2014-08-26 18:19 (3910 d 23:50 ago) @ Croosov Posting: # 13434 Views: 5,434 |
|
|
Dear Croosov, welcome to the club. ❝ What test would you use for tmax if you have a 3 period + 3 treatment design or 4 period + 4 treatment design with one treatment as a reference. Have a look at the papers cited in this relative old post. They deal with 3-period-3treatment crossover designs with six sequences, but the ideas may be extended to 4x4 designs also. The only thing to worry is the null-distribution of the tests if you need an exact solution. This is not an easy task, but approximate solutions are easily obtained. Search the Forum's category Nonparametrics and you will find some more discussions on this topic. If you have a study for submitting to the European regulatory body EMA you don't need any test for Tmax. Just to cite the EMA guidance: "A statistical evaluation of tmax is not required. However, if rapid release is claimed to be clinically relevant and of importance for onset of action or is related to adverse events, there should be no apparent difference in median tmax and its variability between test and reference product." page 15 Moreover non-parametric test are not honoured by the EMA: "A non-parametric analysis is not acceptable." page 15. What else they expect for the second part of the quote above is not clear. Seems some sort of descriptive statistics. Again search the forum to find numerous discussions on this topic. Hope this helps. BTW: WMW = Wilcoxon-Mann-Whitney test. — Regards, Detlew |
|
Croosov ☆ Germany, 2014-09-24 19:58 (3881 d 22:10 ago) @ d_labes Posting: # 13567 Views: 5,065 |
|
|
Thanks for the answer! While I was reading the literature you recommended I wondered if it isn't possible to use a generalized linear model? If not why?  I am still not so familiar with statistics so please apologize this question (if it is not an "intelligent" one)...  |
|
Helmut ★★★   Vienna, Austria, 2014-09-24 22:00 (3881 d 20:08 ago) @ Croosov Posting: # 13571 Views: 5,174 |
|
|
Dear Croosov, ❝ While I was reading the literature you recommended I wondered if it isn't possible to use a generalized linear model? If not why? Before we apply any statistical method (not only a test, but already reporting the location and dispersion of a sample) we must understand the data-generating process and the original / resulting / observed distribution. Dependent on the data structure particular transformations, tests, etc. are possible – and others are not. Simple example: You had 12 males and 10 females in a study. Would you code them with 1/2 and report “sex = 1.455”? It does not make sense (and statistically speaking it is not even allowed) to calculate the arithmetic mean of categorial data. Although the distribution of the tmax in the population likely is continuous (actually a “ratio scale” with a true zero-point) what we get in a study is coming from a discrete distribution (due to our sampling schedule). Therefore, we can only apply statistics suitable for an ordinal scale (location: median, dispersion: percentiles, tests: Wilcoxon and its “relatives”). The late Carl Metzler once suggested to run a GLM not on the raw data but on their ranks, but I have never seen an example in the wild. ❝ I am still not so familiar with statistics so please apologize this question (if it is not an "intelligent" one)... There are no unintelligent questions. Maybe some answers are.  — 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, 2014-09-24 23:27 (3881 d 18:42 ago) @ Croosov Posting: # 13574 Views: 5,072 |
|
|
Hi Croosov, ❝ While I was reading the literature you recommended I wondered if it isn't possible to use a generalized linear model? If not why? You might be able to do just that. Let's say we model something like: Observed Tmaxi = Treatmentj + error where Tmaxi is the i'th Tmax observation and Treatmentj denotes the j'th treatment given for the i'th Tmax observation. In ordinary linear models we use an error that is normal with mean zero and some variance and the fit corresponds to tweaking the constants so that the variance is minimized (then that's the solution that is most likely). For Tmax, as Helmut says, we probably do not have a normal error and we might not have a qualified guess about its distribution. If you happen to have a good guess for the error distribution then you can specify that distribution as the link function for the model (hence the term generalized: the error can be anything, not just normal). But the next issue will then be what do you do with the model once it has been fitted? Is there a relevant test? If there is, is there a relevant interpretation of the test? It is all beyond me, to be honest. To use a term from Helmut's vocabulary: It's wacky. And to me it is more wacky than useful. — Pass or fail! ElMaestro |
Ing. Helmut Schütz