Questions for Mike (Summer‐Fall 2026) - vvandermeersch/climategrowthshifts GitHub Wiki

  • Individual GPs: microenv., allometry, and shifts in length scale estimates... Can one GP do the job of two GPs (if it seems to capture the same functional behavior)
  • WTF is Pareto min. sample size? (from StanCon)
  • What's the best way to check prior and likelihood sensitivity? (e.g. powerscale stuff: $p(\theta | y) = p(\theta)^\alpha \times p(y | \theta)$, from StanCon)
  • normal x half_normal conjugate already discussed this on Zoom, indeed it's a skew normal!
  • Where to put climate in the shocks? Right now on phi (p(event)), could climate impact omega, p(shock | event)?
  • Treedepth issue when going to 2nd order Taylor approx.
  • 2nd order estimates are =/0, coherent effect on big values, but people not happy with small values already discussed this on Zoom, take the log(predictors)!
  • Is taking mean(multiple cores) for each tree ok (we did that in Banff, and we never changed)? I think you had a geometrical justification
  • How can we reconstruct past climate? For example, could VPD and precipitation be a latent variable when unobserved?
  • How to show the importance of shocks? Run a version without shocks and compare posteriors (model with and without shocks), then you can argue that the inferences you're getting without shocks are problematic. ... or (asks Victor) should we just show some predictions without shocks (in some particular years where it is obvious)? Or we could compare model with and without shocks fit on half the data, and predict forward (hold out data).