The GLIMMIX Procedure

 

Iteration History

The "Iteration History" table describes the progress of the estimation process. In singly iterative methods, the table displays the following:

  • the iteration count, Iteration

  • the number of restarts, Restarts

  • the number of function evaluations, Evaluations

  • the objective function, Objective

  • the change in the objective function, Change

  • the absolute value of the largest (projected) gradient, MaxGradient

Note that the change in the objective function is not the convergence criterion monitored by the GLIMMIX procedure. PROC GLIMMIX tracks several convergence criteria simultaneously; see the ABSCONV=, ABSFCONV=, ABSGCONV=, ABSXCONV=, FCONV=, or GCONV= option in the NLOPTIONS statement.

For doubly iterative estimation methods, the "Iteration History" table does not display the progress of the individual optimizations; instead, it reports on the progress of the outer iterations. Every row of the table then corresponds to an update of the linearization, the computation of a new set of pseudo-data, and a new optimization. In the listing, PROC GLIMMIX displays the following:

  • the optimization count, Iteration

  • the number of restarts, Restarts

  • the number of iterations per optimization, Subiterations

  • the change in the parameter estimates, Change

  • the absolute value of the largest (projected) gradient at the end of the optimization, MaxGradient

By default, the change in the parameter estimates is expressed in terms of the relative PCONV criterion. If you request an absolute criterion with the ABSPCONV option of the PROC GLIMMIX statement, the change reflects the largest absolute difference since the last optimization.

If you specify the ITDETAILS option in the PROC GLIMMIX statement, parameter estimates and their gradients are added to the "Iteration History" table. For ODS purposes, the name of the "Iteration History" table is "IterHistory."