SAS nlmixed for count data with random effect
in [SAS]
|
Prev: perl REGEX -- What does it mean that regexes are greedy? How
Next: Command that runs at DOS prompt does not run when using X
From: Nitish on 9 Mar 2010 12:32 Hi, I am trying to estimate a very simple model with a small twist. Counts are distributed by Poisson arrival rate, along with a normal distribution. proc nlmixed data=countdata; parms logsig 0 alpha1 1 beta1 1; eta = (alpha1 + beta1*lnW)+ e; lambda = exp(eta); model count ~ poisson(lambda); random e ~ normal(0,exp(2*logsig)) subject=id; run; I get the error "Optimization cannot be completed". How can this model be suitably reimplemented in NLMIXED? Should I be looking at some other procedure? Appreciate any comment. Regards Nitish
From: Robin R High on 9 Mar 2010 14:20 Nitish, It would help to know more about the data set, but with 9.2 you can do the same analyses with GLIMMIX: PROC GLIMMIX method=laplace ; * or method= quad; model count = lnW / dist=poisson solution; random int / subject=id; run; And you might start with dist=negbin, or at least compare the results. That the NLMIXED is not converging could be due to a number of reasons, though knowing more about what data are like that go into the proc -- how many id's, excess 0's? small counts, etc. -- helps one to better understand what does or does not come out of it. Robin High UNMC From: Nitish <nitish.ranjan(a)GMAIL.COM> To: SAS-L(a)LISTSERV.UGA.EDU Date: 03/09/2010 12:34 PM Subject: SAS nlmixed for count data with random effect Sent by: "SAS(r) Discussion" <SAS-L(a)LISTSERV.UGA.EDU> Hi, I am trying to estimate a very simple model with a small twist. Counts are distributed by Poisson arrival rate, along with a normal distribution. proc nlmixed data=countdata; parms logsig 0 alpha1 1 beta1 1; eta = (alpha1 + beta1*lnW)+ e; lambda = exp(eta); model count ~ poisson(lambda); random e ~ normal(0,exp(2*logsig)) subject=id; run; I get the error "Optimization cannot be completed". How can this model be suitably reimplemented in NLMIXED? Should I be looking at some other procedure? Appreciate any comment. Regards Nitish
From: Dale McLerran on 9 Mar 2010 14:41
--- On Tue, 3/9/10, Nitish <nitish.ranjan(a)GMAIL.COM> wrote: > From: Nitish <nitish.ranjan(a)GMAIL.COM> > Subject: SAS nlmixed for count data with random effect > To: SAS-L(a)LISTSERV.UGA.EDU > Date: Tuesday, March 9, 2010, 9:32 AM > Hi, > > I am trying to estimate a very simple model with a small twist. Counts > are distributed by Poisson arrival rate, along with a normal > distribution. > > proc nlmixed data=countdata; > parms logsig 0 alpha1 1 beta1 1; > eta = (alpha1 + beta1*lnW)+ e; > lambda = exp(eta); > model count ~ poisson(lambda); > random e ~ normal(0,exp(2*logsig)) subject=id; > run; > > I get the error "Optimization cannot be completed". How can this model > be suitably reimplemented in NLMIXED? Should I be looking at some > other procedure? Appreciate any comment. > > Regards > Nitish > Nitish, First of all, you might fit a fixed effect model to get better initial values for alpha1 and beta1. So, first run the code: proc nlmixed data=countdata; parms alpha1 1 beta1 1; eta = (alpha1 + beta1*lnW); lambda = exp(eta); model count ~ poisson(lambda); run; That should converge without problem. (If that doesn't converge, then you have some really serious issues to deal with.) Now, run the random effects model initializing alpha1 and beta1 to values obtained from the model just fitted. I would suggest that you also examine the model likelihood using several initial values for the random effect variance. I would further suggest that you initialize the random effect variance to a value which is smaller than 1. (Note that V(e) = exp(2*logsig)=1 when logsig=0.) Thus, given fixed effects alpha1=<alpha1> and beta1=<beta1>, try running the code: proc nlmixed data=countdata; parms logsig -5 to 0 by 1 alpha1 <alpha1> beta1 <beta1>; eta = (alpha1 + beta1*lnW)+ e; lambda = exp(eta); model count ~ poisson(lambda); random e ~ normal(0,exp(2*logsig)) subject=id; run; where <alpha1> and <beta1> are replaced by specific values. Of course, there is no guarantee that this will converge, either. However, I would be surprised if such a simple model did not converge. But if you still run into problems, then you can try different estimation methods. The NLMIXED procedure has quite a few estimation methods available (quasi-Newton, Newton-Raphson, Newton-Raphson with ridging, conjugate gradient, double dogleg, Nelder-Mead-Simplex, and trust region) Moreover, the different estimation methods can be fine-tuned with optional parameters. You can evaluate numerical derivatives of the parameters using inexpensive forward differences or you can evaluate numerical derivatives using more expensive central differences. There are just a ton of options that you might consider. It is not feasible to discuss here the different options. My guess is that with better initial estimates of the parameters, you won't need to specify any options to aid convergence. Dale --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra(a)NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 --------------------------------------- |