topic Re: PROC MCMC: WARNING: There is still significant autocorrelation after 50 lags ... in Statistical Procedures

PROC MCMC: WARNING: There is still significant autocorrelation after 50 lags ...

fengyuwuzu — Wed, 07 Nov 2018 21:29:13 GMT

When I tried to use the code below to run Proc MCMC, I got two warnings:

WARNING: There is still significant autocorrelation after 50 lags, and the effective sample size for the parameter pm_a might not
be estimated accurately.

WARNING: There is still significant autocorrelation after 50 lags, and the effective sample size for the parameter pm_b might not
be estimated accurately.

I tried removing "thin=25" or changing the value of thin, I still got these warnings. Any suggestions?

title 'Gamma Survival Model';
ods graphics on;
proc mcmc data=adtte outpost=expsurvout nmc=5025 thin=25 seed=4861
   diag=(mcse ess);
   ods select PostSumInt TADpanel ess mcse;

   parms pm_a = 1.2 pm_b = 200;
   prior pm_a ~ gamma(0.01, is=0.01);
   prior pm_b ~ gamma(0.1, is=0.001);

   llike = (1-cnsr)*logpdf("gamma", aval, pm_a, pm_b) +
             cnsr*logsdf("gamma", aval, pm_a, pm_b);

   model general(llike);
run;

Re: PROC MCMC: WARNING: There is still significant autocorrelation after 50 lags ...

SAS_Rob — Thu, 08 Nov 2018 12:54:04 GMT

When you have high autocorrelation among the samples, you need to run the chain much longer to obtain a few thousand independent samples. This is what the NMC= option does. Try increasing that

Re: PROC MCMC: WARNING: There is still significant autocorrelation after 50 lags ...

fengyuwuzu — Thu, 08 Nov 2018 14:13:20 GMT

Thank you. But when I increased nmc=5000000 I still got this same warning.

Re: PROC MCMC: WARNING: There is still significant autocorrelation after 50 lags ...

Clg — Tue, 06 Aug 2024 15:29:13 GMT

Hello,

Did you finally find a solution to this warning?