Re: RE: ROUNDING ERRORS (ERROR=134)

From:Leonid Gibiansky Subject:Re: RE: [NMusers] ROUNDING ERRORS (ERROR=134) Date: Wed, 14 May 2003 10:29:46 -0400 Jeri, The main property of multidimensional nonlinear optimization problems is that they do not converge. So this is quite usual for the FOCE method with non-diagonal 6 by 6 omega matrix. If you believe that correlations are important (e.g., looking on the scatter plots of ETAs vs ETAs) you can try several things that would allow you to simplify OMEGA matrix (comparing with the full matrix that does not converge). 1. Start with the new initial values, e.g., those obtained by the model with the diagonal OMEGA matrix. 2. Instead of the full BLOCK(6) matrix, try,e.g., BLOCK(5) + 1 separate effect. 3. Try the band matrix of the type $OMEGA BLOCK(6) .24 .01 .28 .27 .01 .36 0 0.01 .25 .32 0 0 0.1 .10 0.77 0 0 0 0.03 0.07 0.77 (where the diagonal and X number of the element below the diagonal are not zero, and the rest is zero): as far as I understood, then the NONMEM will keep this structure of the OMEGA matrix. 4. I prefer to express OMEGA matrix via THETA parameters, and then control them one-by one, as follows: MyETA1=THETA(1)*ETA(1) MyETA2=THETA(2)*ETA(1)+THETA(3)*ETA(2) .... MyETA6=THETA(19)*ETA(1)+THETA(20)*ETA(2)+...+THETA(21)*ETA(6) ... CL=TCL*EXP(myETA1) V =TV*EXP(myETA2) with $OMEGA 1 1 1 1 1 1 FIXED In this setting you can add or delete correlations one by one. Final OMEGA matrix can be expressed via THETA(1)-THETA(21) parameters. Also, sometimes change of FORTRAN compiler can help to avoid rounding error or other numerical problems. Freeware gcc compiler (ask Bill for details of implementation) sometimes is more stable than anything else. Good luck, Leonid _______________________________________________________