Issue with LMA and SSD model

[DeirdreLoughnan]

@lizzieinvancouver I have been looking at the model output from the joint model of the BC trait data and while some traits produce reasonable model estimates, I am having issues with the LMA and SSD model. I have been trying to trouble shoot it on my own, but am a little stuck.

Leaf mass area model: The lma model runs well, producing good n_eff and Rhat, with no warning messages. But this surprised me since pairs plot looked weird to me, with a very constrained sigma_site:

I tried adding an ncp to site and again the model ran fine, but produced an even stranger pairs plot:

I also reexamined the priors. The priors seemed off when compared to the posterior distribution for chilling and forcing cues and the related beta;s. But when I tried to update the priors, making them wider, the posterior just gets wider as well. I have never had this happen before where the posterior changes so much with a change in the prior and I can't find any indication of others having this problem on the stan forum.

I am not sure what the solution to this issue is. But I am fairly sure it has something to do with the site variation. Of all the traits, it seems to be smallest for LMA:

[DeirdreLoughnan]

Stem specific density model: Running this model has been the most challenging, often producing 4-7 divergent transitions, low ESS, and Rhat values of 1.05-1.12. But when the model does run with no divergent transitions, the model had huge error around the grand mean, despite the mean values being very similar:

In looking at the model estimates for specific species, the model seems to be estimating the same posterior distribution for each species (first panel below), but there is some variation in the values of the raw data for the same species (middle panel.

The pairs plot for the SSD model has similar shapes in it's pairs plot:

@lizzieinvancouver do you have any suggestions for what might be causing these issues? I would be happy to chat about them over skype if that would be easier. Thanks!

[lizzieinvancouver]

@DeirdreLoughnan

When the posterior changes so much with the prior it suggests you do not have any information coming from your data, which I agree seems weird. I assume you think the model is working on other trait data (like height). So, a few queries:

(a) Do you have this model with the same type of sites and number of sites and number of species running on test data? If not, that would be where to start.

(b) If yes to (a), have you tried running your test data with values of the trait similar to what you have for LMA? I worry you may have numerical issues so this would check that.

(c) If you have done a-b, I would next drop site from the model and see if it starts behaving.

Let me know how it goes!

[DeirdreLoughnan]

@lizzieinvancouver I have been working on this quite a bit, but not making much progress.

a) the test data kinda falls apart when it has the same number of sites (4) and species (28). I have tried to dial it back to smaller components, but with those values, even just the trait part of the model generates warning messages of low ESS and/or exceeding the max tree depth. But it does do an ok job of estimating the true values (which are 20, 10, 5, 0.5) .