Projection choice and residual checks with sdmTMB

[seananderson]

Question via email:

I was curious what projections would you recommend creating the mesh with. I have my data set both in decimal degrees and according to the WGS84 World Mercator (unit: metres). While my DHARMa results and sanity checks came back just fine for a coarse mesh (cutoff=10 for the DD), using the World Mercator projection yielded violating sanity checks and residuals. Once I reduced the cutoff, all DHARMa results came back significant.

Furthermore, I assessed the Moran's I using DHARMa and realised that my data set is still spatially autocorrelated. Therefore, I was curious whether that is negligible when using the spatial = "on" command within the sdmTMB interface.

[seananderson]

An equal-distance projection with units such that the range parameter (distance at which correlation is effectively independent) is a reasonable size (not too big or small). UTMs in km are common if the area isn't so large as to cross many UTM boundaries. See sdmTMB::add_utm_columns() or do it yourself using the sf or sp packages.

I wouldn’t get too caught up in the DHARMa residuals (or the regular residuals you get out of residuals() here). The Laplace approximation error can dominate in some cases and this can be mesh specific. I think there are also issues with DHARMa when the random effects have spatial structure to them (ongoing research). So, even a perfect model can have residuals that look off with DHARMa. residuals(fit, type = 'mle-mcmc') will be the best residuals, but slow to calculate. You can reduce the number of iterations.

Sometimes convergence can depend on the mesh. See the above about units though. The PC prior on the spatial Matérn parameters (?sdmTMBpriors) can help with that, but I’d start with the above.