Open marksorel8 opened 4 years ago
Or maybe it is simpler and better to just model the unexplained (by covariates) annual errors with a multivariate normal with unstructured covariance matrix. I don't think I want to get to deep into evaluating coherence of the streams and life histories in this paper.
Or decomposing the variance into synchronous and asynchronous components by stream, life history, and year along the lines of Thorson et al. 2013. @conversesj, what do you think? Multivariate normal errors, variance decomposition, DFA...
Hi @marksorel8. Sorry for the delayed response on this.
[D]o you think it makes sense to use a dynamic factor analysis approach for modeling shared inter-annual process variance? Do you think it would be effective with 3 streams, 4 life histories, and about 20 years?
I don't recommend this b/c presumably you'd like to have a mean-reverting process with constant variance for the errors.
Is it straightforward to simulate from a DFA (e.g., for a PVA)?
Yes, you can definitely simulate from a fitted DFA model, but I don't think that's what you want here.
Buhle et al. (2018) used a common annual error term (phi_t) with a form
log(phi_t) = X_t Beta + v_t vt ~ N(rho v{t-1}, sigma_phi)
to which you could indeed add some (possibly correlated) site-specific errors.
good idea. One of the things here is that we have this hierarchy where we have 3 streams and 4 life histories within each stream. What about having common errors, as you mention above, but also stream-specific errors, life-history specific errors, and then stream by life history errors? I wonder if this would be difficult to fit?
Can you write down what you're thinking? We're talking about annual errors, right: eps[year,lifehist,stream] ~ ? In particular, how many sigma terms are you thinking about estimating? There could be 12 which seems like a lot.
Yep, I was talking about annual errors, but I have similar questions about the other model parameters. Here is a document I've been working on. It includes thoughts about the other model parameters as well.
More musing about modeling the error structure here... Let me know if you have any thoughts @conversesj @ebuhle @mdscheuerell
Hey @mdscheuerell and @ebuhle, The next stage of this project after estimating the numbers of juvenile emigrants is to model the transitions from spawners to emigrants. I will be modeling 3 streams and 4 "life histories" (a.k.a windows when fish emigrate). I'm not sure what functional form(s) of model(s) I will use, it could be a Beverton-Holt for example, but I will need to model the process errors regardless. I want to include covariates related to flow like @mdscheuerell did in his recent Skagit steelhead IPM paper. I think it also makes sense to model correlation between life histories and streams in some way. Maybe you remember when Kevin Whalen asked me about this at the Cooperators meeting last year :) And finally my question.... do you think it makes sense to use a dynamic factor analysis approach for modeling shared inter-annual process variance? Do you think it would be effective with 3 streams, 4 life histories, and about 20 years? Is it straightforward to simulate from a DFA (e.g., for a PVA)?
Thanks!