Bioequivalence and Bioavailability Forum

Dear dlabes !

ad) values > LLOQ or beginning <LLOQ: I deal with large molecule drugs – I have only measured values or values below LOQ (limit of quantification) I do not have experience with a “fuzzy range”.

ad) geometric mean. Yes of course your right but using exp(mean(log(x))) - it is simply not defined. Omitting values to calculate the mean can give you a terrible wrong picture (personal experience: I had a data set that switched from a one-compartmental to a two-compartmental model - by visual inspection - simply by excluding values below LOQ). Values below limit of detection are informative missing and not missing at random!

Ad textbooks: Most of them are applicable for problems in a perfect world. From a theoretical point of view one has to model values below LOQ as censored observations (like in survival studies). I think when you have a perfect situation the geometric mean is applicable whereas in the case of values below LOQ it can give you a terrible wrong picture. For standardization of figures I would go for the arithmetic mean as 1) figures should give an overall unbiased picture and the arithmetic allows to handle values below <LOQ and 2) is theoretically (i.e. asymptotically and on assumption of an intra-subject correlation of zero) justified (which may not be the case using medians or geometric means).

Best regards

martin

PS.: what do you think on providing boxplots per time point instead of means?