Hi @Carlos_Zarzar and @cmcd. Thank you both. I was taking the time to make some additional tests and research with the informaion that Carlos shared. Michael’s comment is really illuminating, but it seems to contradic this old thread, “If WAIC is derived from cross-validation it does not need the assumption of independent errors”. Aki Vethari also says “WAIC is sane choice although it seems that it can break sometimes when using very flexible models (like Gaussian process or Markov random field spatial models)”. My models seems to fall in the last category of this post, actually spatial models are mentioned too, but, is there a way to correct them? Or, can I ignore the warning and still use WAIC or PSIS loo as a criteria to select one spatial model over another based on this information?
3 Likes