Elena777 ☆ Belarus, 2018-07-18 17:35 (2276 d 04:54 ago) Posting: # 19079 Views: 6,724 |
|
Dear colleagues. I would like to know how pharamcokinetic data is manipulated during statistical analysis. Specifically, I would like to know when we should perform ln transformation: we calculate Geom mean and then transform its value to ln value and do ANOVA or we transform each concentration Cmax to Ln-data and then calculate geom mean. Or maybe it is not relevant? I also would like to know how GMR is usually presented in tables. Whether it is usually GMR of ln-transformed or non-transformed data? I provide the table below as an example and would like to know the origin of its values as I asked above. |
ElMaestro ★★★ Denmark, 2018-07-19 01:19 (2275 d 21:10 ago) @ Elena777 Posting: # 19080 Views: 6,084 |
|
Hi Elena777, I did not sanity-check the figures but at a glance it looks all normal and consistant with a standard BE study outcome. Re. "manipulations": Data is typically log-transformed and then fit with a normal linear model with the four standard factors. If you use WNL with default settings it will fit Subject as random. From the model fit comes a set of effects for the factors, which give rise to treatment LSMeans (on the log scale). These are then backtransformed to the normal scale. — Pass or fail! ElMaestro |
Ohlbe ★★★ France, 2018-07-19 12:31 (2275 d 09:58 ago) @ Elena777 Posting: # 19081 Views: 6,070 |
|
Dear Elena, I am not a statistician, but I should be able to answer in lay language ❝ Specifically, I would like to know when we should perform ln transformation: we calculate Geom mean and then transform its value to ln value and do ANOVA or we transform each concentration Cmax to Ln-data and then calculate geom mean. Log-transform each value and then calculate the arithmetic (not geometric) mean of the log-transformed values. If you calculate the exponential of this arithmetic mean, it will give you the geometric mean on the original scale. ❝ I also would like to know how GMR is usually presented in tables. Whether it is usually GMR of ln-transformed or non-transformed data? I provide the table below as an example and would like to know the origin of its values as I asked above. What you have in your table is the non-transformed data. Just try and calculate the exponential: the values you would get for AUC would be really extreme. If reporting log-transformed data: you would need to calculate a difference, not a ratio. The ratio in your table is not a geometric mean ratio and is therefore rightly not labelled GMR. It is a ratio of LSMEAN, not of geometric means. It makes no difference if your study is balanced (i.e. same number of subjects with sequence TR and sequence RT, in which case the LSMEAN is equal to the geometric mean), but if you have a drop-out and your study gets unbalanced you will get some differences. — Regards Ohlbe |
Helmut ★★★ Vienna, Austria, 2018-07-19 15:01 (2275 d 07:28 ago) @ Ohlbe Posting: # 19083 Views: 6,014 |
|
Hi Elena, extending what Ohlbe wrote… ❝ The ratio in your table is not a geometric mean ratio and is therefore rightly not labelled GMR. It is a ratio of LSMEAN, not of geometric means. It makes no difference if your study is balanced (i.e. same number of subjects with sequence TR and sequence RT, in which case the LSMEAN is equal to the geometric mean), but if you have a drop-out and your study gets unbalanced you will get some differences. See this post for an example of unbalanced sequences and especially the PS. — 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 |
crdreyes ☆ Philippines, 2018-07-25 19:21 (2269 d 03:08 ago) @ Helmut Posting: # 19101 Views: 5,770 |
|
❝ ❝ The ratio in your table is not a geometric mean ratio and is therefore rightly not labelled GMR. It is a ratio of LSMEAN, not of geometric means. It makes no difference if your study is balanced (i.e. same number of subjects with sequence TR and sequence RT, in which case the LSMEAN is equal to the geometric mean), but if you have a drop-out and your study gets unbalanced you will get some differences. ❝ ❝ See this post for an example of unbalanced sequences and especially the PS. Hi Helmut, From your experience, do you recommend showing the geometric means along with the adjusted geometric means (from ls means) and the ratio from the adjusted in a single table? Or it does not make sense to even include the geometric means. My initial thought it would be good to include both just to see if the adjustments from the model. But it may cause confusion and may not be worth it. Thanks, Russel |
Helmut ★★★ Vienna, Austria, 2018-07-26 17:38 (2268 d 04:51 ago) @ crdreyes Posting: # 19106 Views: 5,719 |
|
Hi Russel, ❝ […] do you recommend showing the geometric means along with the adjusted geometric means (from ls means) and the ratio from the adjusted in a single table? Or it does not make sense to even include the geometric means. My initial thought it would be good to include both just to see if the adjustments from the model. But it may cause confusion and may not be worth it. Well, that’s a matter of taste. I mainly deal with European submissions where the tables mentioned in Appendix IV of the BE-GL are mandatory (see esp. Table 3.1). Never give the arithmetic means (we know that AUC and Cmax follow a log-normal distribution) but the geometric means ±CV%. I tried to convince the EMA to use only the geometric least squares means (see there) but didn’t succeed. Geometric means are fine to represent the outcome of subjects under each treatment. In the synopsis of my statistical reports I always give xgeo ± CV%. Only if the study was unbalanced (or a parallel design with unequal group sizes) I give additionally the GLSM ± SD (together with a footnote clarifying why they are different). That’s like you would do. Never got a request for clarification from any agency (either the assessors were clever or the footnote helped). — 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 |