RE: Help: Non-positive semi-definite message

From: "KOWALSKI, KENNETH G. [R&D/1825]" <kenneth.g.kowalski@pharmacia.com> Subject: RE: Help: Non-positive semi-definite message Date: Mon, 20 Aug 2001 18:40:18 -0500 Nick, I share your concern however, you might be surprized how often this occurs. We need to run the full OMEGA model more often as a diagnostic and calculate these correlations...I wish NONMEM would automatically report these correlations rather than just the variances and covariances. I think this issue is the same as when we observe a particular variance component being estimated as zero. We should not infer that the variability in that particular variable is negligible just that the data/design do not provide sufficient information to estimate it. When a full OMEGA has an off-diagonal element going to unity, we shouldn't automatically assume that they are perfectly correlated, just that they are strongly correlated but the data/design don't provide sufficient information to estimate it different from unity. I gave a talk on this at the CDDS workshop on modeling and simulation best practices a couple years ago. In that talk I show how if we ignore these strong correlations and resort to a diagonal OMEGA simply because we can get the COV step to run without the "non-positive semi-definite message" problems, we can end up simulating unrealistic data. If OMEGA is over-parameterized we need to understand why and try to reduce the dimensionality in the most parsimonious way...diagonal OMEGAs usually are not the most parsimonious. Ken