Bioequivalence and Bioavailability Forum

Dear NewInPK,

sorry for the late response; you raised some tricky questions.

❝ I've been looking for information regarding the correct way of reportingaverages of PK parameters. From what I've found so far, it seems that for log-transformed data compared with a t-test, it is recommended to report the geometric mean and 95% CI.

In the English Wikpedia your statement would be tagged, like: "... it is recommended^{[by whom?]} to report the geometric mean and 95% CI."

If you compare PK-metrics I would report the 90% CI of test vs. ref.
If you want to give descriptive statistics separately for treatments, I would report geometric means ± 1 geometric SD.

For BE-Studies FDA requires Geometric Least Squares Means, Ratios, and 90% CIs (Formatting of Bioequivalence Summary Tables, Table 3, January 2007). No SD or any other measure of dispersion at all.

❝ If log-transformed data does not follow a normal distribution, then a nonparametric test is used to compare 2 samples and the median and interquartile range is used to report the data.

It's almost impossible to justify the underlying distribution given the common sample size in PK. Log-transformation for AUC and C_max is based upon prior knowledge (multiplicative due to AUC=D×F/CL and error propagation in bioanalytics due to serial dilutions). See this example slides 43-44. Pooled AUC values from 405 subjects. It's clear that lognormal (right pannel) fits data better than normal (left pannel), both with a Shapiro-Wilk test or QQ-Plots. In the next slide you see data from a subset of 12 subjects (MPH has very small CVs, and so are the BE studies). No way to 'select' the 'correct' distribution. If you have opted for lognormal in your protocol, it will be very difficult to convince regulators about the use of a noparametric method.

• EMA: impossible
  
• FDA: close to impossible
  
• Canada: If Eric Ormsby is your assessor, fine; otherwise, possible, but still difficult
  
• Japan: possible
  
• ANVISA: almost impossible

❝ For consistency, I would prefer to use only one way of reporting. Would it be acceptable to report median and IQ range for the log-transformed data?

Without back-transformation IMHO that's a perfect method to confuse assessors.

❝ Alternatively, would it be acceptable to report geometric mean and 95% CI for the samples which were analyzed using a nonparametric test? I think the geometric mean is misleading in this case and also the 95% CI doesn't describe the data well.

1. IMHO no. 2. Agree.