RE: FW: OMEGA HAS A NONZERO BLOCK

From: "Serge Guzy" Subject:RE: FW: [NMusers] OMEGA HAS A NONZERO BLOCK Date:Fri, 4 Oct 2002 11:02:14 -0700 I do agree with Ken that a large correlation between parameters should not be thrown out just because it is large. Here are the two scenarios I already investigated using the new MC_PEM (Monte Carlo Parametric Expectation Maximization)algorithm. The first scenario considered a 1 compartment model with only one sample per patient (50 patients).The true correlation was very high (~0.9). -The algorithm retrieved the high correlation as well as the right population PK means and variances - A bootstrap algorithm was used to mimic other samples that could arise from the unknown population - Using the bootstrap 10 times, the correlation stayed high for all 10 new datsets and with a very small standard error and a mean~0.9 - My conclusion was that the correlation was trustable and not an artifact due to the sample design Second scenario The second scenario considered a one compartment model with oral absorption with 3 PK parameters where no correlation existed between PK parameters. - The MC-PEM algorithm was used again for 10 similar datasets - The result was that the correlation between Ka and V ranged from almost -1 to almost +1 - The original dataset had by accident a high positive correlation but the bootstrap algorithm made us thinking that the correlation could go from -1 to 1 and at the same time without changing significantly the other population parameter estimates My conclusion was that there is no much information about the correlation between these two parameters. What to do about that, I don't know exactly but in that case the true correlation was zero. I think we enter in the field of hypothesis testing and ask the following question H0= No correlation HA: correlation The bootstrap algorithm therefore do not allow us to reject the null and we stay with no correlation. It does not mean that there is no correlation but since there is no supportive evidence to reject the null , we stay with the null. I think it is the right thing to do and I would redo a fit forcing the correlation to be zero. I would never fix the correlation to let say 0.5 and use the same other population PK estimates for prediction purposes. By the way , I would be happy to get the data that initiated this very interesting discussion and would try to analyze it using our system Serge Guzy, Ph.D. Head of Pharmacometrics Xoma