Help: Non-positive semi-definite message

From: stuart@c255.ucsf.edu Date: Thu, 16 Aug 2001 11:09:05 -0700 (PDT) Regarding some discussion ensuing from Alan Xiao's question about the message he is getting that the R matrix is non-positive definite: Discussion about how to get around this message, by using other options on the $COV record to obviate the need to compute the R matrix, is off the mark. After obtaining a final parameter estimate that appears to be reasonable, it is a good idea to compute the R matrix. (This is easily done if a MSF was used when the parameter estimate was computed.) With the R matrix, and only with the R matrix, can one check that the estimate is not a saddle point. If, as in Alan's case, the estimate is a saddle point, then it probably is not a suitable estimate, and standard error estimates (no matter how they may be computed) and the like are irrelevant. By default, if the R matrix turns out to be positive definite, and the S matrix to be nonsingular, then NONMEM computes the estimate of the variance-covariance matrix of the estimator by using both the R and S matrices (viz. as (Rinv)S(Rinv)). Stuart Beal