Bioequivalence and Bioavailability Forum

Hi ElMaestro,

❝ […] occasionally even Austrians say something meaningful...

True. But this is a bad example.

❝ The whole idea about 0.052 as a kind of clinically relevant limit which popped up in Potvin's paper is taken out of the clean blue Canadian air.

Well, North-America’s air. The majority of authors are from the US. Irrelevant employers like FDA, USP…

❝ If we want to control the type I error rate at 5% then my interpretation is we should mean 5% and this should not so much be a discussion of decimals or even the 0.05036 binomial limit at one million sims.

Nope. Everything ≤0.05036 can be pure chance (at the 5% level). If you don’t accept such values in 10⁶ sim’s you would have to adjust α further down in such a way that the “true” risk I is expected to be <0.05 (~0.04964). In other words, you would require two-stage designs to be more conservative than fixed sample designs. Why?

Every year my book shelf gains about 1 kg since I’m a subscriber to Biometrics. Many, many simulations of incomprehensibly complicated models. Quite often n=10,000 (sign. limit 0.0537). If a value of 0.064 pops up, in the discussion it is called ‘minimal inflation’. These methods are applied in phase III. Regulators don’t care. Slowly I get tired about this double standard. Consider going for a hike to Similaun.

❝ eu regulators want the CI's to be within 80.00% to 125.00% - four or five significant digits or two decimals? If the latter is the case…

Seems so. IMHO this entire rounding business is bullshit. According to the (admittedly arbitrary) definition we accept a ∆ [sic] of 20% as clinically not relevant. After long discussions about the distribution / transformation the limits were set to AL_lo = 1-∆ = 0.8 and AL_hi = (1-∆)^-1 = 1.25 precisely. Why not AL_hi = 1+∆ = 1.20 and AL_lo = (1+∆)^-1 = 83.3? Because in the former case the numbers look nicer and are easier to remember (seriously: this was the winning argument in the mid 1980s). That’s what I call a proper justification. Every software could check whether CI_lo < AL_lo and CI_hi > AL_hi in full precision. Phoenix/WinNonlin in its standard format reports the CI to four decimals and gives a verbatim statement (based on the full precision comparison) in the next line. Looks like this:

CI  90% = (   79.9950,  114.1593)
Failed to show average bioequivalence for confidence=90.00 and percent=20.0.

Hurrah!
Canadians regulators have a special love affair with nice numbers and set the upper limit for NTIDs instead of 111.11 to 112.0% (one decimal)…
Should we do triple rounding in ABEL? OK, I have learned that k (derived from ln(1.25)/√ln(0.3²+1 = 0.760128298…) should be treated as 0.76 (two decimals). For CV_WR 50% the scaled limits are 69.84–143.19% (rounding #2). Now assume different lower CLs.
                      AL (exact)  AL (rounded)  AL (0.76)  AL (0.76 rounded)
                      69.83255%      69.83%     69.83678%       69.84%
69.83254% ≥ AL (“BE”)   FALSE         TRUE        FALSE          FALSE
↓ to 69.83%             FALSE         TRUE        FALSE          FALSE
69.83255%               TRUE          TRUE        FALSE          FALSE
↓ to 69.83%             FALSE         TRUE        FALSE          FALSE
69.83677%               TRUE          TRUE        FALSE          FALSE
↑ to 69.84%             TRUE          TRUE        TRUE           TRUE
69.83679%               TRUE          TRUE        TRUE           FALSE
↑ to 69.84%             TRUE          TRUE        TRUE           TRUE
Crap. I guess EMA wants the triple rounding (red in the 4^th column). Red in the 1^st column shows a disagreement with the exact comparison.

❝ … then we could try to argue accordingly that 0.054999 is not inflation. Yuck

Don’t get it.

❝ 0.05 rules.

If so, not even ‘Method B’ (maximum inflation 0.0504 would be acceptable. Are you suggesting to throw away all – including your own – papers?

Happy simulating,
Hötzi