RE: $OMEGA blocks and log-likelihood profiling

From: Kowalski, Ken Ken.Kowalski@pfizer.com Subject: RE: [NMusers] $OMEGA blocks and log-likelihood profiling Date: Fri, 2 Jul 2004 09:37:48 -0400 Nick, It sounds like we are not in disagreement about the value of statistical theory just that we have to understand the assumptions we make when applying it...no argument there. Just as there is a wealth of PK/PD terminology and different words are sometimes used to mean the same thing, the same is true in the statistical literature. I equate stability or rather 'instability' with 'ill-conditioning' which can be a result of 'over-parameterization'. Bates and Watts (pp. 86-91) discusses inspecting the correlation matrix of the parameter estimates to help diagnose collinearity among the parameter estimates (general ill-conditioning) and/or over-parameterization (fitting too many parameters). With regards to 'reliability' I'm using the term as per Webster to mean the extent to which results are reproducible. In this context a model fit might be unreliable because it is difficult to reproduce estimates across platforms/compilers, with different starting values that may only differ by 10%, etc. Certainly numerical instability of NONMEM itself may indeed be the explanation for your particular problem, however, if you don't strive to get the COV step to run and inspect the output during model building before performing bootstrapping a reviewer could always question whether your model is stable. Moreover, Pete Bonate's exercise showed that NONMEM differences across platforms/compilers were most likely to occur when one is dealing with an unstable model. Thus, numerical instability of NONMEM and model instability in many cases may just be two sides of the same coin. Regards, Ken _______________________________________________________