Bioequivalence and Bioavailability Forum

Hi Simon!

❝ ❝ WinNonlin has an unreported bug, giving differences always in alphabetical order...

❝ In Pharsight's case that is through [...]

❝ http://pharsight.com/support/support_home.php [...]

Yeah, that's a funny website telling me that I never posted before 2007, and some IDs give in the field 'Solutions' the comment 'None Found', while I received one by e-mail… (see this post or that one).

❝ If in this case you meant undocumented then yes I guess that is an oversight.

Oh, yes undocumented is correct! I reported it a couple of times up to WNL v4 (first time according to my files for v3.3 on 08 July 2002) and then gave up. I even provided Pharsight's support with a reference of FORTRAN-code to calculate the CDF. I didn't really need it myself (I use StatXact or EquivTest/PK for nonparametrics)… The output – besides the lexical order – is flawed in another sense:
In the 'Crossover Design Workbook' – 'Confidence Intervals' tab the column 'Confidence_Level' gives the aimed one, but numbers given in the Column 'Exact' are wrong (WinNonlin 5.2.1 always comes up with something like 'Confidence_Level 0.90, 'Exact' 0.9000). Below is a table of exact confidence levels for a couple of balanced sequences n₁=n₂ (N=n₁+n₂):

 N   alpha    p{U n1,n2}      N    alpha    p{U n1,n2}
 6   0.0500    0.9000       36   0.0485    0.9029
 8   0.0286    0.9429       38   0.0482    0.9036
10   0.0476    0.9048       40   0.0482    0.9036
12   0.0465    0.9069       42   0.0486    0.9028
14   0.0487    0.9027       44   0.0491    0.9018
16   0.0415    0.9170       46   0.0499    0.9002
18   0.0470    0.9061       48   0.0486    0.9028
20   0.0446    0.9108       50   0.0497    0.9006
22   0.0440    0.9121       52   0.0490    0.9020
24   0.0444    0.9113       54   0.0485    0.9030
26   0.0454    0.9092       56   0.0499    0.9002
28   0.0469    0.9061       58   0.0498    0.9004
30   0.0488    0.9025       60   0.0498    0.9004
32   0.0469    0.9062       62   0.0499    0.9002
34   0.0493    0.9013       64   0.0487    0.9026


Due the discrete nature of the U-distribution α is ≤0.05, or in other words, nonparametric confidence intervals are generally slightly wider (>0.90) than ones based on normal theory (if the underlying 'true' distribution is normal).