Recommendations for what to do when k exceeds 0.5 in the loo package?

Modeling
3

My “base” model is the Poisson model I described in this thread.

Saying that my base model is already complex is a fair description, that is why I called “base” instead of “simple.” The reasons why I started with that model are a bit complex and hopefully not relevant for this discussion. That said, I actually want to compare this model with simpler and more complex versions.

A follow-up question, is this plot a good reason to think that this base model is not really bad?

If I understand this correctly, loo indicates that my second model is better than the base model. That said, I still have some problems with k.

> print(loo_2)
Computed from 6000 by 1795 log-likelihood matrix

         Estimate    SE
elpd_loo  -9079.3  84.5
p_loo      1346.5  48.3
looic     18158.7 169.0

Pareto k diagnostic values:
                         Count  Pct 
(-Inf, 0.5]   (good)     721   40.2%
 (0.5, 0.7]   (ok)       676   37.7%
   (0.7, 1]   (bad)      320   17.8%
   (1, Inf)   (very bad)  78    4.3%
See help('pareto-k-diagnostic') for details.
> 
> print(compare(loo_1, loo_2), digits = 3)
elpd_diff        se 
  105.062    22.444

Thanks, I will carefully read this Saturday morning, and probably come back with more questions after that.

Looking forward to loo 2.0. Do you know when are you going to release it?

Thanks a lot for your amazing work and all the help!