Bioequivalence and Bioavailability Forum

Hi Astea,

❝ Recently I've received a letter from the Customer where it was pointed out that according to the planning regulatory documents we have to include missing data in analyses in a very specific way.

You posted in the category Regulatives / Guidelines. What do you mean by “planning regulatory documents”?

❝ For example in several cases the missing data from one volunteer in one period should be replaced be the mean arithmetic of the other's data.

That’s bizarre. Practically all regulatory documents point out that common PK metrics (except t_max) are assumed to follow a lognormal distribution. Hence, we apply a multiplicative model (log-transformation in order to obtain additive effects). Arithmetic mean is wrong.

❝ Of course this method has a right to exist.

Don’t think so.

❝ It was described in European Pharmacopeia 5.0 p.3.2.6 and in several articles, but it seems to me very strange to use this technic in bioequivalence data analyses.

The technical term is “imputation”. It is common in clinical studies if a patient drops out midcourse (i.e., Last-Observation-Carried-Forward – LOCF). Here it can be (!) conservative since in a superiority test the estimated mean difference will me smaller and if one assumes that the treatment performs better than placebo. But it is not that easy. ICH E9 states in Section 5.3:

Unfortunately, no universally applicable methods of handling missing values can be recommended.

I never ever have seen it in BE and for good reasons. If you impute the missing by the geometric (!) mean of the other subjects:

• The between-subject variance likely will be lower than the one of the dataset before imputation.
  
• The T/R-ratio will be biased to an unknown degree.
  
• The SE of the difference might be lower as well. Hence such a comparison might be anticonservative.

If you want to work with an incomplete data set (i.e., missing period(s)) you can try a mixed effects-model. See this rather lengthy discussion of the issue.
BTW, the EMA’s GL wants only complete subjects (T and R) in simple crossovers and at least one period with each of T and R in replicate designs. Both the EMA’s and the FDA’s SAS-code for reference-scaling drop incomplete subjects.

Impeccable timing as have just had a very similar situation arise myself that I would be extremely grateful for any input on:

I have been scouring the relevant EMA/FDA guidance but haven’t come across anything too useful in relation to assessing the impact of missing samples in a bioequivalence study.

The scenario is:

4 way cross BE study; test vs ref in fed and fasted states

N = 22 HVs

Tmax ~60 mins
T1/2 ~120 min

Sampling: 0, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90, 105, 120, 180, 240, 360, 480 and 720 mins

During one of the study periods a protocol non-compliance has resulted in the 480 min time point not being available for 2 out of 22 subjects.

Given where the 480 h time point lies in the profile, my instinct tells me that the missing samples should have minimal/negligible impact on the study objectives (i.e. to test for bioequivalence) and should not introduce an unacceptable amount of bias into the estimation of AUCt for the 2 subjects in question.

However, I would like to check if there are any guidelines/best practices etc that reinforce my ‘instinct’?

Any thoughts most welcome!