Closed rBatt closed 8 years ago
During simulation testing, I'm not noticing a lot of benefit to adding a ton of 0 species. Mainly, it just seems to slow things down. @mtingley what's your approach to choosing 0's?
Depends. Some analyses I don't use data augmentation. In a multi-year study where you have a good idea of the true meta community it seems to make sense to keep augmentation limited.
Fair enough. That at least isn't a sharp contrast with my impression.
I'm going to try my hand at the dynamic model next week. I still don't quite understand this distinction between colonizing, occupying, and persisting. I mean, the distinction in common language is clear to me, but I guess I don't fully understand why you would bother to model them as separate processes.
As I say that, I guess it's because you might think different processes affect occupancy depending on whether or not it was there the previous time step. If your species is a boulder, occupancy probability is low if the boulder wasn't there the last time step (boulders don't just show up out of nowhere very often), and the boulder has a high probability of occupancy if it was there the previous time step, because boulders don't just disappear.
I digress. Thanks.
On Friday, December 4, 2015, Morgan Tingley notifications@github.com wrote:
Depends. Some analyses I don't use data augmentation. In a multi-year study where you have a good idea of the true meta community it seems to make sense to keep augmentation limited.
— Reply to this email directly or view it on GitHub https://github.com/rBatt/trawl/issues/86#issuecomment-162136950.
In my testing/ playing I've been keeping the augmentation tiny. Like 2 species.
I'm currently getting good convergence with 50 'zero' species. Omega is not bumping up against 1.
It was important to scale all of the covariates to mean 0 and sd 1.
To determine richness you don't have to sum Z over the augmented indices. Data augmentation is used not just to get asymptote (sum Z over all indices), but to just get stability in the parameter distributions. Might be worth reducing augmentation to see if I can get the same answer (reducing augmentation has the advantage of reducing the computation demand of model fitting).