Is WAIC well suited to estimate out of sample predictive accuracy for spatial problems? In Gelmans paper he suggests:
WAIC relies on a data partition that would cause difficulties with structured models such as for spatial or network data. -section 7.3
What does he mean here? Does he mean that WAIC is a point wise estimate and if one assumes non independence between nearby points WAIC is ignoring this? Don’t hierarchical models make a similar assumption by making the lower level groups come from a group level distribution?