Bioequivalence and Bioavailability Forum

Dear yicaoting!

I’m not surprised; WinNonlin (all versions up to the current release of Phoenix/WinNonlin) con­tains a bug (filed as incident 00133350 on 22 Mar 2011, QC 10856). See my white paper.

❝ Can anyone give me the result of PNX WNL 6.0 or later version?

Phoenix/WinNonlin 6.2.1.51 (Bioequivalence object)
Average Bioequivalence:

Difference:   -0.028652092233559
Diff_SE:       0.055693090418743
Ratio_%Ref:   97.1754486595894
CI_90_Lower:  88.3108418878651
CI_90_Upper: 106.929881091872

LSM Differences:

Reference - Test:  0.028652092233559
StdError:          0.055693090418743
T_critical:        1.71714434835526
Lower_CI:         -0.066980983221424
Upper_CI:          0.124285167688542

If we change signs (lexical order in LSM!) and antilog we get:

90%lo:  88.3127965410502
90%hi: 106.9275143812607

… matching your manual calculation (antiloged: 88.3127964130425 – 106.927514536250) to 8-9 significant figures, and your antiloged LSM-CI exactly.^*) The discrepancy is caused by lacking precision of the t-value only to a minor part. According to Pharsight (correct) output of the LME-engine is taken over to the BE wizard. Pharsight is suspecting that results are not used in full precision – but are internally rounded (!!)…
So either use the workaround I suggested in the whitepaper or abandon the BE wizard entirely and set up the analysis in LME, backtransforming afterwards.

❝ So I need Daer HS's help to give out the result of R's Tinv(0.1,22).

Here you are (R 2.14.0):

options(digits=16)
qt(1-0.05,22)
1.717144374380243

As expected matching Gnumeric (according to Gnumeric’s documentation the same algo is used).

---

^*) Proving that the bug is consistent within WNL 5.1.1 and PHX 6.2.1.51…