Model validation

[Aariq]

Since switching to 10ºC for base temp, the new model outputs are more qualitatively different from pixel-wise linear regression and also seem to vary depending on either data resolution or k or both. (see also cct-datascience/organization#2354)

Might want to develop methods for this stuff, but not actually do it until we decide on thresholds (#5) since some model tuning will be specific to each dataset.

In no particular order (but I should order these before doing them):

• Check that k is appropriately large for all models
• Check residuals—is gaussian error family still appropriate?
• Check (temporal and spatial) autocorrelation—itsadug package or spdep package (for Moran's I permutation test)
• Explore gaussian process smoother #11
• What about NCV? Is it worth trying again? (Explore/implement NCV #1)
• Read papers I have about spatiotemporal analysis with GAMs—what, if anything, do they do differently?

[Aariq]

Temporal correlation is there, but values of rho are negative and relatively small (I think). ACF plots have really no pattern in them. I now know how to deal with temporal autocorrelation if I reviewer asks (on a branch, will push to GH soon)

[Aariq]

Create a .qmd for model validation that can eventually become an appendix

[Aariq]

I added the output from gratia::appraise() in PR #18, but need to evaluate this and k-check results on a per-model (i.e. per GDD threshold) basis.

[Aariq]

Looks like at least some thresholds have leptokurtic residuals—a student T distribution may be more appropriate