Programming the statistical procedures from SAS

Proc NLMIXED hundreds of parameters

Occasional Contributor
Posts: 12

Proc NLMIXED hundreds of parameters



My question revolves around the estimation of many parameters in Proc NLMIXED (i'll post code below). The reason that I'd like to use Proc NLMIXED is to use Guassian Quadrature, but also take advantage of using all CPUs (Proc Glimmix only uses [4] and NLMIXED will utilize all [8]).


Essentially I have a working model in Proc Glimmix that involves all categorical variables, that in some cases have hundreds of levels. I've gotten the design output from Proc Glimmix, and can manage establishing the parameters; however when it comes to creating an estimation function (i.e. b0 +b1*x1 +b2*x2 +b3*x3 + ... +bn*xn) I'm not sure as to how I should go about that. As there are 800+ parameters. 


I've tried using the [array] steps, but NLMIXED seems to get stuck processing those...i think last time i stopped it after [9] hours, and it never reached the optimization phase. 


Is there any way to construct the estimation function that I've seemed to have missed?


Here is the original Glimmix code:

ods listing close;
				proc glimmix data=modeldata method=laplace inititer=25 empirical=mbn;
				class  	quarterback facility msdrg msdrgsl aprdrg age type_id ppayercode
						los rom soi race sex readmit;
				id Year facility quarterback ppayercode measure id episodeid seqkey cost charge log_charge log_cost ;
				model 	charge =	facility msdrg msdrgsl aprdrg age type_id  
									ppayercode los rom soi race sex readmit
									/dist=gamma link=log ddfm=bw;
				random int /subject = quarterback type=vc s cl;
				parms (.0001,.001,.01,.1,1)(.01,.1,1);
					tech=dbldog instep=.5
					gconv=1e-8 fconv=1e-16
					maxiter=10000 maxfunc=10000;
				output out=test predicted=RelAdjLink predicted(ilink)=RelAdj predicted(ilink noblup)=Fixed 
					Residual=ResidualLink Residual(ilink)=Residual pearson(ilink)=PearsonLink;
				ods output SolutionR=RandomEffects;

Thanks in advance!!

Ask a Question
Discussion stats
  • 0 replies
  • 1 in conversation