RE: posthoc step

From: "Ludden, Thomas (MYD)" luddent@iconus.com Subject: RE: [NMusers] posthoc step Date: Wed, December 15, 2004 5:32 pm Hi Nick, Thanks for providing the link to your Excel spreadsheet. It provides an excellent visual display of the objective function. I was finally able to find the place in NONMEM V where I could use a debugger to change the initial estimate for the ETA search when MAXEVAL=0 POSTHOC was specified. Initial estimates of ETA from 0 to -2.750 (initial est. of K = 10 to 7.25) yielded K=9.78. Initial estimates of ETA from -2.768 to -9.9 (initial est. of K = 7.232 to 0.1) yielded K=0.944. Initial estimates of ETA from -2.755 to -2.767 (initial est. of K = 7.245 to 7.233) yielded values of K intermediate between 9.78 and 0.944. Over this very small range of initial values, the step size for the ETA search is small and the algorithm exceeds the maximum number of iterations that is allowed. I was curious about how the "Data" Method missed the local minimum so I used the debugger to trace the search for the "Data" method using the control stream you provided, except that I delete the SLOW option. Starting with an initial estimate of 10 the series of K values and the associated OFV's are tabulated below. Note that the THETA search (different from the code used for the ETA search) "jumps" well over both minima but then returns to the region of the global minimum and rather quickly terminates at this minimum. The ETA search assumes it is starting in the correct region whereas the THETA search makes no such assumption. In this case, that is an obvious advantage. The THETA search never detects the shallow local minimum near a K value of 10. K OFV 10.010 439.794 -20.000 3.2605X10^55 -5.000 2.6717X10^16 2.5 306.355 2.51 307.674 -5.000 2.6717X10^16 -1.25 4895847 0.625 134.856 0.623 125.167 0.802 30.721 0.979 14.1438 0.989 14.6477 0.962 13.5448 0.944 13.333 0.954 13.405 0.934 13.398 0.944 13.33 termination Tom