Bioequivalence and Bioavailability Forum

Dear Joshi!

❝ We have WinNonlin and trying to measure with 90% CI.

In the workbook you need the following columns: subject, sequence, treatment, Tmax.
Tools > Crossover Design... drag subject, sequence, treatment to the named fields and Tmax to Response Calculate
WinNonlin has an unreported bug, giving differences always in alphabetical order... (Just a side note for people who like software 'validation' by developers: if I remember it correctly the 'validation pack' costs about 75% of the the initial price of the software or 4× the annual licence - but nonparametrics are not a strength of WinNonlin - to say it politely).
If your treatments are coded 'test' and 'reference' WinNonlin will come up with:
Treatment_Diff_(reference - test) (!!)
in the column Median: the point estimate
in the columns CI_Lower and CI_Upper you find the confidence interval - in the second row the 90% Confidence_Level.
Again, don't forget to change the sign of these three values, if the coding of your test treatment is in lexical order before the code of the reference treatment!
Example:

• Test coded 'T' or 'test', reference coded 'R' or 'reference'
  → wrong, because T>R: change signs of PE and CL!
  
• Test coded 'A', reference coded 'B'
  → correct, because A<B...

• Don't be surprised that the point estimate is not the one you probably expect; many people think it's just the difference of medians of test and reference, but actually it's the median of all pairwaise differences (for details see the reference given already).
  
• The confidence level given in the 4^th column (Heading Exact) is rarely what it seems to be, namely exact. I had many conversations with Pharsight's support over the years and simply gave up, because I don't use WinNonlin for nonparametrics myself.
  E.g., for 12 subjects (balanced 6/6) WinNonlin comes up with 0.9000; correct is 0.9069…

The underlying test statistic (Mann-Whitney U) is discrete (different from parametric ones, which are continuous); therefore it's rather unlikely that α is exactly 0.05 (p=1-2α=0.90). Since the test is conservative, the nonparametric confidence interval is slightly wider if applied to normal distributed data – but tighter, if applied to nonnormal distributed ones.

If you know FORTRAN or somebody with a little experience in programming, you can compute exact values of U for any combination of subjects within sequences (very useful, because the table given by Hauschke et al. covers only up to 24 subjects).*

❝ But what exactly meant by clinically determined rage?

This has nothing to do with statistics; you must ask a specialist in the medical field. For instance a difference of 30 minutes may be clinically relevant for a pain-killer (I want my headaches to go away – now!); for venlafaxine in steady state even hours may be irrelevant (but I'm neither a physician nor a patient – so I can't tell).

---

• Dinneen LC, Blakesley BC. Algorithm AS 62: A Generator for the Sampling Distribution of the Mann- Whitney U Statistic. J Roy Stat Soc C. 1973;22(2):269–73. doi:10.2307/2346934.