atsa-es / tvvarss Goto Github PK

Just got an email from github saying I was granted push access to this repo. Was that accidental? Maybe you meant to add someone else?

There is an error in the specification of the prior for vecBdev. It is intended to be the process error for the RW on B, but the specification on line 31 limits it to the interval [0,1] instead of [-c,c] as on line 101. This won't affect models with time-invariant B, though.

Right now MASS is called from within simTVVAR(), so we should instead add that to the list of others required at load.

It would seem that we could safely delete the /simulated_data/ folder and simulated_data_example.R because we're trying to use the same generating & estimation models?

I was talking to Buhle about the if() statements causing us a bit of comp time, and he and I came up with this soln. At the point the user specifies the food web topology, they can also optionally specify the range of interactions strengths to consider for each type. I imagine we would have could have some defaults (eg, top-down are [-1,0]), but that people could opt for the possibility of, for example, specifying density dependent effects of [-1,1] instead of [0,1] to allow for 2-pt oscillations.

To do so, we would change the R code to track not only the position of each of the "td", "bu", etc elements, but also the [lower,upper] priors for the interactions. Then, instead of the if() statements necessary to know whether we are multiplying by and/or adding a constant, we simply do the math for every param. For example, now we have something like

if(top-down)
b_t = -1 * [inv-logit(logit(b_{t-1}) + error_t)]
if(competitive)
b_t = 2 * [inv-logit(logit(b_{t-1}) + error_t)] - 1

Instead we would have for all b
b_t = (upper - lower) * [inv-logit(logit(b_{t-1}) + error_t)] - lower

where [upper,lower] are passed to Stan for all b.

It would be good to include the flexibility for adding covariate effects to the process model.

We have 2 copies of an example from Ives et al: one in the base (Apr 19) and one in the /ives/ folder (Apr 3). Can we delete one of them?

It would be nice to have an option for count-based obs model (e.g., negative binomial), and it would make sense to include it as an option in the R wrapper.

A totally minor detail, but can we add another "V" to the name for Time Varying Vector AutoRegressive State-Space (ie, TVVARSS) in order to match the abbreviation of Ives & Dakos?

tvvarss requires the data to be passed as an array with dims (sites, years, species), but simTVVAR outputs the states (fake data) as we normally think about them in MARSS models: a matrix with dims (species, years). I would be in favor of changing tvvarss to match the way we typically envision the data: (species, years) or optionally (species, years, sites).