2

I have found some mixed messages in terms of the effective sample size to shoot for (ranging from 2,000 to 10,000) here.

I'm wondering if anyone could give me more information on choosing a guideline for effective sample size?

In particular, I am using a No-U-Turn Sampler.

Dave
  • 1,641
  • 2
  • 14
  • 27
  • 2
    It depends on what you are trying to estimate. I suggest you think about the Monte Carlo standard error of the quantity you are trying to estimate instead of the effective sample size. For example using the `mcse` function in the `mcmcse` R package. – jaradniemi Oct 26 '17 at 00:20
  • Thanks! As far as I know the package I am using ("brms"; which is an R interface for Stan) can't calculate this. But I will look into it. – Dave Oct 26 '17 at 01:09
  • 10,000 for stable HDI limits. See details and justification in Section 7.5.2 of DBDA2E. – John K. Kruschke Oct 26 '17 at 01:20
  • Thanks for the info! I should have mentioned that the package I'm running uses a No-U-Turn Sampler–does that change things? – Dave Oct 26 '17 at 02:37
  • There is no and cannot be any absolute value for the ESS, it all depends as Jarad wrote. In addition, the ESS is not foolproof for assessing convergence, being based on crude approximations. – Xi'an Oct 26 '17 at 05:40

0 Answers0