RE: OMEGA HAS A NONZERO BLOCK

From:"Steve Duffull" Subject:RE: [NMusers] OMEGA HAS A NONZERO BLOCK Date:Mon, 7 Oct 2002 11:31:22 +1000 Hi All I, also, have enjoyed all the comments. My last email was just to point to other difficulties in the matrix - that do not stem from the "apparent" correlation of 1 {Cov(par1,par4)} that is quite apart from the true matrix problem. I should add a comment that I specifically did not include the parameter names (i.e. attributing the omegas with CL's etc) since I was wondering why we could not take the output from NONMEM and use it as an input when we can mathematically show that the matrix is invertible and the determinant can be computed (both required by NONMEM's objective function). The practical solution for us was to "pull back" and accept a more parsimonious model. The full model was structurally identifiable but was probably not deterministically identifiable. What I mean by this is that in theory all the parameters of the model were globally identifiable in theory - but the data does not seem to support their estimation as accurately as we would like. I think the 100% correlation (which was a red herring wrt the problem) - is artificial and was induced by a poor data set. I would not be happy about setting the cov to 1 (but as Lewis indicates setting to to zero is tantamount to the same assumption). This data could very easily be modelled using a different method - such as MCMC. However, where the data is not "speaking to us" I doubt that we would be able to get around the correlation problem. I have never tried to add in an informative prior on the covariance terms using BUGS. In fact, given the specification of the prior for the inverse of the VC matrix (typically a Wishart) I do not know if it would even be possible to specify a different strength prior for different components of the VC matrix while preserving the entire VC structure. To be more specific the only way I know to add strength to the prior for a Wishart is to increase the degrees of freedom which would affect all parameters presumably equally. Perhaps someone has some ideas on this? Kind regards Steve =================================== Stephen Duffull School of Pharmacy University of Queensland Brisbane 4072 Australia Tel +61 7 3365 8808 Fax +61 7 3365 1688 http://www.uq.edu.au/pharmacy/duffull.htm University Provider Number: 00025B