Questions to discuss in New York (October 2025) - vvandermeersch/climategrowthshifts GitHub Wiki

Notes before the meeting with Mike

Can we separate out predictive power of GPs versus climate predictors? (see also #48, #62 "How do species-level effects compare to stand-level effects? We would like to compare two possible extreme realities: one where all species at a site (and surrounding sites) are doing poorly or actually certain species are doing well across space?")
And we can separate out the two GPs? (see also #23)
(If we could separate them ...) Could we estimate how different a single species' response is in different locations?
Variation not captured by climate variables see this issue
How to make predictions with GPs? We're thinking of +drought -moisture +GDD. Should we do it on new fake stands? Or use the same stands? etc.
(Temporal vs spatial climate signal in the model): The model captures climate responses due to year to year variation within individuals only? Or does it capture responses due to year to year variation within individuals and also across space? I think this would be a good launching point to discuss how our model may handle the type of climate filtering issues discussed by St George and Ault 2014, which relates to Rob Guy's query about whether we just see the biggest climate effects from trees that exist in certain places?
Relatedly, how should we feel about extrapolating our linear climate models? (Err, Lizzie is adding this the day after a meeting with Ben and Victor and I think I have not captured exactly what we wanted....)
How to speed up since 1900s model (#60, chain parallelization)?
Add a third GP (for allometry ... so our original length scale)
Let's think about drought legacy effects (like here, but better?)
Can we look at effects in different decades and say anything? (Such as: responses in 2000-2025 were much greater than in 1980-2000)
How to identify which climatic predictor is the main driver of climate-growth trends for each species/site?
What to do with gamma and kappa
2-dimensional (space and phylogeny) GP for competition model?

Lizzie moved some HMM questions

Notes from meeting with Mike (14 October 2025)

The world does not decompose into independent parts! It's all convoluted.

We have ...

Hierarchical GP for each stand (each stand is a draw from rho and gamma), and each stand is scaled by a species-specific $\kappa$
GP for each species
Principled centering (meaning we have a baseline we center around)

A few bits:

Negative slopes for betas means the growth is less than the baseline behaviour of the GP (for that species)
The model cannot currently compensate for species that grow faster than other species, exactly. So the GP has to soak it up, or the climate covariates. We could add a species-level baseline for the model to estimate. (We could visualize this possibility in some of the stand-level plots.)
We should think about the model as relative (proportional) behaviours (not absolute behaviours). For example, if the gamma on GP is log(2) larger than a beta, than the multiplicative change is twice as large.

Shock model!

The GP encodes rigidity in the functional behaviour. To capture kinks it needs to go down to a short-length scale. This is called 'ringing.' Our GPs for the stand look like shocks. For these, you do not want to use a GP because what happens one year is not predictive the next year. A better model for that would be to do a two-level latent mixture model, as see in Mike's sparsity chapter. Each year is predicted in the model and pushes most years to zero effect.
This would not help us pick up legacy effects, but you could think about adding an HMM to the shock model to make the shocks related to one another and learn the relationships. (But no one has tried this, 'it would be wild' says Mike.) States would be basically depressed growth versus staying depressed growth. Mike says: 'It's not going to be productive to worry about this right now.'
We could then look at how many shocks you have in each decade etc.
For this move the stand-level GP priors to 1-10 years.
Do the shock model first, then if you see nearby stands doing similar shocks, then maybe you can figure out a covariate to put them together (but I doubt it).

Thinking about separating out parts of the model

We would like to compare two possible extreme realities: one where all species at a site (and surrounding sites) are doing poorly or actually certain species are doing well across space?

... if we mean all the species at one site are doing poorly for a few years, then only the stand-level GP can accommodate that.
Seems like we need to use hypotheticals (counterfactuals) to figure out the 'contributions' -- because climate and stand GP vary by stand, you would to work with these together -- you cannot isolate stand-level GP from climate variables.
Maybe one is that you extrapolate climate a lot and see how much it matters, that tells you, for climate, how much it mattered.
Could do hypothetical where the shocks turn off (or add an autoregressive so the shocks could be increasing).
Would it be useful to build a model without stand-level GPs and shocks to show that they are not sucking up climate variation? No. Would be more useful to trust the model is soaking up variation as the model defines them and develop hypotheticals to understand the model.
Some times you can think of REASONABLE hypotheticals where you end up setting parameters to zero, but they have to make sense.
Read the new case study
Consider changing the stand-level GP to be a baseline with the stand-level GP being deviations from the baseline ... right now we have common hyperparameters and simulate new GPs for each stand, an alternative model is (more aligned with a regular hierarchical model) where you can have fixed hyperparameters, you have a baseline GP for all stands and each stand has a correction (or offset). This will couple the functional behaviour but you have to answer all the annoying questions to answer to do this.
For the above, we could look at our stand-level GPs and see if there is commonalities worth trying to capture. The model we have now is more flexible, so maybe that is what we want (says Lizzie).

Climate signal in the model?

The model captures climate responses due to year to year variation within individuals only? Or does it capture responses due to year to year variation within individuals and also across space?

It captures variation in response to climate within individuals and across stands and how species grow .... We are modeling how individual trees grow, and how that growth changes with different climate predictors and stand-specific and we are also considering how similar trees within the same species grow.
You can separate out stand-level variation due proportional change, but the individual-tree level variation is there too.
The stand contribution is always relative to how each individual tree is growing. If you really wanted to say this, you could imagine maybe making a tree grow a lot more versus moving it to a new stand ... the hypotheticals are sort of uncomfortable which is probably a sign.
beta $\times$ soil moisture is time-dependent component of the model, just like the GPs. If the time-series for the soil moisture is a 5 year slow change, then the stand-GP may pick it up -- they both are consistent with the data and you should see it in the posteriors.

We need speed!

See ratings case study to check geometry vs model evaluation as issue (if it is the model evaluation, then you invest in parallelization etc.; otherwise you work on making the geometry better).
Fit recent data by itself versus old data by itself to see if you can help diagnose the problem. If they only get worse when you add them together then the model must be missing some changes the model needs to have added in.
Maybe it's non-linear climate responses that show up, maybe it's the species-level GP changing ...
Whatever is wrong, you should most likely add it to the short-term models also, but if it's a shift in the predictors (you're in a different linear phase) then that could be okay.

On the fence about adding a non-linearity?

A model is more sensitive to adding in the non-linearity (e.g., a quadratic in response to climate) than just looking for non-linearity in a retrodictive check.
Basically, when in doubt, add in the non-linearity and look at the posterior for the parameter.

Wrapping up questions

Can we look at effects in different decades and say anything? (Such as: responses in 2000-2025 were much greater than in 1980-2000)

Mike says, sure, of course.

But frame this as a population summary (see also the epidemiology case study again): you can look at a reconstructed tree, and you want to aggregate across time. You have an inferred tree width of every year, construct a histogram of a tree width across years, then represent the distribution of histograms. Now, you want to look at multiple trees ... take all the widths of all trees for that species for that stand, get a histogram for each MCMC sample, then summarize that.
(But 'it's no a good idea' he said to divide up the data and fit the model, because everything is coupled. But I never meant to suggest that!)
Later we came back to this, and Victor took more notes in his notebook (and I was tired so I did not try to take notes)

Notes before the meeting with Ben

Is there an interesting VPD x SM angle?
Can we solve the divergence problem now?
Let's model drought legacies! (see here)
Variation not captured by climate variables see this issue