topic Re: Proc genmod warning relative Hessian convergence criterion in Statistical Procedures

Proc genmod warning relative Hessian convergence criterion

jojo — Fri, 17 Aug 2018 21:34:32 GMT

　I got a warning when I ran the following code. I have tried to use aggregate option, but it is not working. appreciate it for any inputs/suggestion on the reason and solutions.

PROC GENMOD data=data descending;

Class trt classvars;

MODEL avalc = trt classvars contvars/dist=binomial link=Identity alpha=0.10;

Estimate 'xx'trt 1 -1/alpha=0.05;

run;

WARNING: The relative Hessian convergence criterion of 0.0027102422 is greater than the limit of 0.0001. The

convergence is questionable.

WARNING: The procedure is continuing but the validity of the model fit is questionable.

Re: Proc genmod warning relative Hessian convergence criterion

PaigeMiller — Sat, 18 Aug 2018 00:23:51 GMT

Google "Hessian convergence is questionable"

Lots of discussion on this topic, including

https://groups.google.com/forum/m/#!topic/comp.soft-sys.sas/uXDWNF_oL6U

Which says "The implication is that PROC GENMOD has reached a local maximum in this function rather than a global maximum."

Re: Proc genmod warning relative Hessian convergence criterion

jojo — Sun, 19 Aug 2018 21:26:29 GMT

Thank you! I have gone through the discussion, but it does seems to provide any solutions. They tried many solutions, but none of them can fixe the issue.

My goal is to get risk difference and CI between the treatment group from Estimate statement (link=identify). I tried link = logit, the warning disappeared, but need to convert OR to RD and get the confidence interval. Does anyone have a reference for calculating RD and CI from OR?

Re: Proc genmod warning relative Hessian convergence criterion

StatDave — Mon, 27 Aug 2018 15:39:34 GMT

See this note which describes using the NLMeans macro to estimate the risk difference. The identity link does not assure that the predicted values from the model are valid probabilities, so errors like this are common when not using the canonical link (logit).