mikemahoney218
commented
1 month ago On a first read-through, I think this is a very fair example! I think the quote from the tidymodels book is trying to distinguish the Ames data from data where observations are remeasured, because rsample provides group_* resampling methods for those situations. That said, while I don't think they were trying to deny any spatial autocorrelation in measured values, a straight reading definitely gives that impression.
The only notes I'd have regarding spatialsample is that I tend to have a strong predictive modeling focus, where I'm much more interested in predictive accuracy than in inference or model coefficients. This is why I tend to focus on machine learning approaches, and why our paper focuses on RMSE estimates (though Mila et al 2022 does a much better job of this). In that mindset, spatial CV is mostly useful to assess models that will need to extrapolate outside of the area data was collected (or more abstractly, will need to predict on new data that wasn't generated by the same spatially autocorrelated process). The argument then is that spatial CV will give a better estimate of model performance than a single hold-out set could, as it's impossible to know if performance on a hold-out set is actually representative of general performance if we're assuming that our model is misspecified to the extent that we have spatially autocorrelated residuals. I think the ideal is that you then follow up with a map accuracy assessment, using an independent probability sample to assess mapped values, so hopefully your spatial CV estimates are useful guidance through the model development process, before your final assessment stage.
I've actually found waywiser -- and specifically the autocorrelation metrics we wrap from spdep -- as more useful for improving model specification. A big part of our model development workflow is mapping local spatial autocorrelation estimates, and using that visual guide to help identify areas that are "weird" for one reason or another. We then use our knowledge of the study area to try and guess what variables might explain the difference; as ecologists, we're often able to look at the local autocorrelation map and say "that looks like elevation" or "that looks like rainfall" (or, once, "that looks like where the National Forest comes over the border, I bet those trees are larger than the surrounding area"). It's not a useful approach if you're interested in inference, but for predictive modeling this can be a very useful workflow!
I'm not aware of any preprints in this space that would be useful; I honestly haven't kept up with the space recently while I've been focused on other projects.
The 4-way split is definitely interesting! I'm wondering how it might help in sizing blocks for spatialsample -- a common complaint about our approach to spatial CV is that there's not clear guidance on how big blocks should be. I wonder if there'd be a way to dynamically determine block size, so that you use the largest blocks that minimize trte?
Small note on slide "spatial diff-in-diff" -- small typo in bullet 2: "alaysis"
Thanks for looping me in -- is the talk being recorded? Would love to watch later if so :smile:
rsbivand
topepo
@mikemahoney218 Could I ask whether you could give me a view on whether the Ames data set examples make any sense in this draft talk (I'm presenting this on Monday afternoon my time, so a bit late to elicit input now)? I've tried to be fair wrt
waywiserandspatialsample, but as far as I can see, althoughspatialsamplemitigates information flows between data subsets, I don't think we know how to improve the specification of spatial models, or predict for test/validation sets.Are there other arxiv papers I've missed that you know of? The rendered version of the beamer slides is at: https://rsbivand.github.io/nem24_talk/. Any comments very welcome!
Might you the 4-way Moran split be useful in
waywiserin its current restricted form (row-standardised weights)? Using a permutation test is a bit trivial, but at least it's something.