Report loss of R^2 when terms are dropped

[Aariq]

I had previously read Wood 2013 and my main takeaway was "yes, you can trust the p-values output by summary(gam)". However, Wellington et al. 2023 cite this paper and interpret it as "the very large dataset size effectively voids the hypothesis testing procedure adapted to GAMs". So, need to re-read the Wood paper to double-check that this is the case.

In Wellington et al. 2023, they do not report p-values or effect sizes, but rather the lose of % deviance explained (R^2?) when model terms are dropped. (importantly, when dropping 'year', they also dropped interactions with year because it doesn't make sense to have interactions in a model without main effects).

We might consider a similar more "qualitative" approach if the goal is to say "yes, there are temporal trends" or "yes, there are spatial trends".

How does this translate to the p-values used to shade out areas in the average slope plots? I don't think it does—I think those p-values are still trustworthy.