76  Use of SIGL and NSIG with the new methods

Introduction to NONMEM 7.6.0

For the new analysis methods, SIGL is also used to set up forward-difference or central difference gradients as needed. Such finite difference gradients need to be set up for sigma parameters and THETAs not MU modeled to etas, or where OMEGA values of etas to which the THETAs are MU associated are set to 0.

NSIG is used only with the iterative two stage method, among the new methods. The iterative two stage is not Monte Carlo, and has a more deterministic, smooth trajectory for its parameter movements with each iteration. In this case, NSIG is used as follows: The average of the last CITER/2 parameters are evaluated and compared with the average of the next to last CITER/2 parameters. If CITER is odd valued, (CITER+1)/2 will be used. For example, for CITER=5, at iteration 102, iterations 97-99 are compared with iterations 100-102. If they differ by no more than NSIG significant digits, then this parameter is considered to have converged. When this is true for all parameters tested, optimization is completed.