Bioequivalence and Bioavailability Forum

Dear yicaoting,

❝ Precision is only at the lelvel of 0.01(%), Incrediable? Interesting? Strange?

Credible. Somewhat dull. Perhaps 0.001% strange.

The fitting engines must stop at some point when the objective function cannot be improved much further. The software internally keeps track of the improvement at each iteration and sets some criteria for stopping. It can for example be when the Log likelihood (or SS) cannot be improved by more than 0.0001 arbitrary units or something like that.
How to set that limit is actually a little tricky, because if you set it too low you may screw up some other important aspects of the iterations (for example: small numbers and matrices are sometimes a toxic cocktail for numerical reasons) and if you set it too high you just get an apparent lack of precision.
It all depends on how the math are actually implemented and we know nuffin about the source code really. There are many different ways to optimise; some prefer the brutally efficient Nelder-Mead algo, others the softcore Newtonian-based meffuds like Marquardt-Levenberg. Entire books are written about this.

Of course any difference in precision can in theory lead to erroneous conclusions but in your case it would be a very rare event. Before jumping to any conclusions perhaps you should compare the precision in terms of objective function (SS or logLik). At least that might give an indicator of which program is actually doing the best job. But be sure to compare like for like (SS against SS, logLik against logLik, bear in mind if WNL by default fits by REML while SAS etc does it through SS).

❝ Can anyone give me the result of PNX WNL 6.0 or later version? Dear HS or ElMaestro or others warm-hearted?

I don't have that software, sorry. But you're right I am warm-hearted, except when I tie my men to the mast and flog them.

By the way, with R 2.10.1 I get:
format(qt(0.95,22), digits=20)
[1] "1.717144374380243"