topic NLMIXED -- (projected) gradient is greater than 1e-3 in Statistical Procedures

NLMIXED -- (projected) gradient is greater than 1e-3

BISTGP — Thu, 01 Sep 2011 18:07:58 GMT

I am fitting a mixed model using NLMixed and while the procedure converges and gives what appear to be reasonable parameter estimates, the log contains the following message.

NOTE: At least one element of the (projected) gradient is greater than 1e-3.

I am not sure how to interpret this, if I should I be concerned about it, or how to "fix" this? Any comments and suggestions will be appreciated.

Greg

NLMIXED -- (projected) gradient is greater than 1e-3

Rick_SAS — Thu, 01 Sep 2011 18:47:27 GMT

At an extremum (minimum or maximum) of the objective function ,the gradient vanishes. The NOTE is telling you that the convergence to an extremum is dubious. You can adjust some of the "Termination Criteria Specifications" to see if you can achieve convergence. Or the function might be extremely flat near the extremum.

Re: NLMIXED -- (projected) gradient is greater than 1e-3

BISTGP — Mon, 12 Sep 2011 19:28:23 GMT

Hi Rick,

Can you say a little more about which specific termination criteria might be adjusted? Maxiter and maxfunc are not a problem since I have "GCONV convergence criterion satisfied" well before the maximum number of iterations and functions are reached. I have set gconv=1E-9[2] but with no effect. NLMIXED is like a machine with many nobs to adjust, and I am not sure where to really start.

Greg