Bioequivalence and Bioavailability Forum

Dear HS,

thanks for sharing your first impression.
Here are my first thoughts about the paper.

1. It seems, that the ARS was not restricted to the timepoints later than tmax.
At least not for creating figure 2.
No one would implement the algorithm in that way!

2. Beside your concern about EXCELs random number generator I argue:
The simulation is not that what we see in real world for different subjects.
It is the simulation of thousands of c-t curves of one subject with the 'true' parameters given for the concentration time curve (one-compartment or two compartment).
Thus it answers the question: How good is the estimation of one lamda_z and / or one AUC(0-inf).
I am concerned if we can draw any conclusion from that.

3. Although it is true that scientifically we should choose that method with best properties in bias and / or variability.

But I cannot see any practical implications of the reported differences, especially in the context of bioequivalence studies.
A bias of 5% for residual area amounts to max 1% absolute if you has planned your sampling times propperly thus that AUC(0-tlast) is at least 80% or greater. It makes only a difference if you are borderline.

I argue that the bioequivalence test as the last end is not at all affected in any way.

4. The statement, that both methods deliver statistical significant results is a fundamental misunderstanding what statistics is for. It is easy to obtain significance for ten thousands of values with low variability.
But significance is not relevance.
I think no one would argue that AUC values for instance of 296.59 or 297.22 (AUCinf from study A) are different.

5. I cannot verify the claim that TTT should not applied if the concentration time curve is biphasic after Cmax, at least not for the reported results.
The bias in lamda_z for models B and C are -1.15% or 1.39% (according to table 1). This is not different to 'Study A', for which TTT is recommended.
What's the point?

6. The suggestion that low N=3, which occures relative often in the ARS algorithmn, is associated with a higher varaibility, is a mis-interpretation of the extra simulations. Here all calculations were stopped at N=3 and so on. Thus it can only regarded as a criticismn to choose a low predefined fixed number of points.

To my experience (with only a limited number of trys), the ARS stops only, if the first 3 points fit is superior.
If this lead to a higher variability remains open.

7. Looking forward to your simulations.
I suggest that the TTT method and any Fit statistics algorithmn (ARS or AICc or whatever) should be combined to get the best of.
TTT could be used to restrict the number of points used by Fit statistics algorithms.

A side question to your simulations:
You report a positive bias for the ARS method.
What is the source for that?
The simulations where ARS goes not beyond N=3?