Bioequivalence and Bioavailability Forum

Hi libaiyi,

❝ I can't get the same results from winnonlin and nonmem.

Well, we are dealing with nonlinear mixed effects models with all the bells and whistles. The software tries to find the optimum in an n+1 dimensional parameter-space. NONMEM and Phoenix/NLME employ maximum likelihood, whereas WinNonlin least squares. See this post for details and the presentation given at the PAGE meeting 2005 about comparing PopPK software in a blinded manner. Also interesting: Bauer et al.¹, Colby and Bair². The latter deals with Phoenix.

❝ Is there any difference in algorithm in two softwares?

Well, in both there are more than one. It would be very difficult to get exactly the same results even if you use the same algorithm, initial estimates, parameter constraints, convergence criteria, max. number of iterations,
We can only hope that an algo converges to the global – and not a local – optimum.
That’s the reason why in BE only NCA is acceptable.

❝ Or does it exist some mistakes in code?

Always possible (see this presentation). The software industry generally considers one defect / 2,000 lines of code ‘stable’. Oh dear! Package PowerTOST has currently ~10,200 LOC – I hope, we did better. Of course, higher standards are used in certain fields (see what I wrote about white-box validation in this post).

---

1. Bauer RJ, Guzy S, Ng C. Survey of Population Analysis Methods and Software for Complex Pharmacokinetic and Pharmacodynamic Models with Examples. AAPS J. 2007;9(1):E60–83. doi:10.1208/aapsj0901007. PMC Free Full text.
   
2. Colby E, Bair E. Cross-Validation for Nonlinear Mixed Effects Models. J Pharmacokinet Pharmacodyn. 2013;40(2):243–52. doi:10.1007/s10928-013-9313-5. PMC Free Full text. supplementary material.