In brouwern/avesdemazan: What the Package Does (One Line, Title Case)

knitr::opts_chunk$set(echo = TRUE)

Jun 24, 2021, 12:42 PM

....

Most of the models have the random effects as

... + (time_cts + 0|Specie.Code:Location) +         
ar1(as.ordered(time_cts) + 
0|Specie.Code:Location)     # key line

but the negbin-2 model (which was the best model) was

(time_cts + 0|Specie.Code:Location) +         
ar1(as.ordered(time_cts) +
0|Specie)                   # key line

which has no ":Location" in the ar1() statement

This is the only model in the code I looked at was like this.

Updating the negbin-2 model with ":Location" to be like the others results in a model that has some issues

• no errors thrown when running the model
• "dispersion parameter" results by summary() is HUGE: 2.78e+06
• (the dispersion parameter accessed through the model object is 14.83931; 1 to 2 is normal I think ; not sure if 2.78e+06<-->14.83931 is a transformation;
• Bolker's experimental Hessian tests say the dispersion parameter is bad
• I think if I adjust the optimizer I can get an error thrown re. the hessian when I run the model, but haven't reproduced that yet using your original code

Your poisson model with AR1, in contrast to the negbin behaves nicely (negbin 1 behaves just as poorly). Poisson is slightly overdispered (dispersion = 1.801, p-value = 0.064) according to DHARMa (super cool glmm diagnostics package I always forget to use....), but so are all the other models.

The poisson model (ok I have to link to this https://www.youtube.com/watch?v=UoJxBEQRLd0 ) also seems to let me toss in some more relevant random effect.

The resulting poisson models are better than the (problematic) negbin models (I don't yet yet understand why negbin is so much more problematic than poisson in all contexts).

Anyway, the model I just arrived at is below. Differences from your best model

• Poisson, not nb2
• 1 new random effects to represent within-year correlations (year:session , year:session:species)
• has different ar1 structure from your best negbin model (but same as your Poison and other models)

glmmTMB(N ~ 1 +
Location +
time_cts +
(1|Specie.Code) + (1|year:session) + # new rf (1|year:session:Specie.Code) + # new rf (1|Specie.Code:Location) +
(time_cts + 0|Specie.Code:Location) + ar1(as.ordered(time_cts) + 0|Specie.Code:Location) + # has location
offset(log(tot_net_hours)), family = poison, # Poisson data = ecuador)

Any thoughts on this?

The only thing that gives me pause is that I always think of Poisson as an ideal that is never met, but these data seem Poisson enough.

I'm curious to see if this actually changes anything. There's an interesting dynamic ecology blog post on model (mis-) specification and how we need to make sure models are robust to specification because we can never know if we're fitting the right model, or can't fit the right model b/c of compuational limitations https://dynamicecology.wordpress.com/2021/06/21/the-death-knell-for-the-one-right-way-approach-to-statistics-and-the-need-for-sturdy-statistics/

brouwern/avesdemazan documentation built on July 26, 2022, 8:38 p.m.