Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Supported by

convergence in JASP

I was wondering: what would JASP show me if I run a Bayesian repeated measures ANOVA and for some reason it can't converge nicely? Would it give very large error% in the table? Give me some error? Crash?
I can't seem to find any trace plots, or indication how many iterations were needed, so I am uncertain how to claim convergence other than trusting JASP to tell me if anything was wrong, and redoing the analysis a couple of times to check if results don't vary too much.

Comments

  • For many regression/ANOVA analyses the integral only goes over one parameter, so full MCMC is not needed. I am not sure what the BayesFactor package does for repeated measures ANOVA though -- should be in the documentation. But yes, BayesFactor assesses the error and when it is too large you can increase the number of samples (under the Advanced Options tab). I'll ask Richard about the BayesFactor.
    E.J.

  • thank you! Will it tell me if the error is too large, or should I just manually check whether is goes above -say- 10% and increase the amount of samples if it does? I thought the samples option indicates how many samples to take from the posterior, not how many iterations are done in total? But I am a bit uncertain what the repeated measures ANOVA is exactly doing on a sampling level as well, so my comprehension is a bit rather fuzzy here.

  • You should check yourself, there is no gold standard; it also depends on the numbers themselves -- you may not care whether the BF is one or two billion.
    I am not sure about the difference between "samples from the posterior" and "iterations". Richard says:
    "No, BayesFactor does not use MCMC; it uses MC using a rough heuristic to choose between simple MC integration and a custom importance sampling method (you can also use a Laplace approximation, if you set it manually). Usually big error rates mean model problems or something strange."

Sign In or Register to comment.