Thanks for the suggestions @junpenglao. I will re-check the LOO.
I did a little bit of investigation. It seems, you are correct. When calculating the WAIC; computed lppd_i are getting smaller as we increase the number of samples drawn from the target distribution.
However, when we compute the var(logp), this result in some large variance values and that cause the large WAIC.
So what is the alternative, if we can’t trust WAIC or LOO for this case. Can we just use lppd values?