search

ctesta01 / covid.gradient.estimation

Using generalized additive models to analyze COVID-19 US county level mortality data
2 stars 1 forks source link

A 2d TPRS is not appropriate for lat/lon spatial data #2

Open gavinsimpson opened 1 year ago

gavinsimpson commented 1 year ago

In the manuscript you say

https://github.com/ctesta01/covid.gradient.estimation/blob/90033b7df432cafb9b0d8a26432cbbf12fbb4401/manuscript/manuscript.rmd#L481-L483

but over an large scale like the US, 1 degree of latitude isn't the same thing spatially as 1 degree of longitude. So, while they might technically be in the same units, we certainly would make a strong argument for not modelling them with an isotropic 2D TPRS as you suggest in this section.

I would either:

  1. convert the lat/lon data to a local coordinate system (using a EPSG code suitable for the mainland US) and then continue to use the 2D isotropic TPRS smooth for space, or
  2. keep the data as lat/lon but model them as an anisotropic tensor product smooth of univariate marginal smooths.
ctesta01 commented 1 year ago

Hi Gavin, thanks for this! I had been thinking about this and wanted to address it. This is really helpful.

I've been using EPSG:2163 (US National Atlas Equal Area). My impression is that because this is an equal area projection designed to distort the US relatively little, it should be a good choice, but I'd love to hear it if you have other suggestions.