region.N confusion

[mobear]

I am using secr 2.5.0 in Rstudio on Windows 7.

I estimated density of a black bear population using DNA-based encounter history data. I am interested in deriving population size in the mask region for management purposes. I used region.N() with the top ranked model (D~1, g0~bk, sigma~1) and the mask used in the fitted model for ‘region=’. My mask excludes areas of non-habitat. My fitted models specified groups = ‘Sex’ as M and F.

When I run region.N() I get the following result:

bk.regN <- region.N(bk, region = subBigMask, se.N = T, alpha = 0.05)

> bk.regN

estimate______SE.estimate____lcl__________ucl_________n
E.N 125.9206___24.93855______85.73047 ___184.9516___90
R.N 170.9223___22.27129______137.65019___227.4268___90

When I call N from my fitted model, I get:

> bk$N
_____F________M
P 125.9206__125.9206


So is my region.N() estimate a sex-specific estimate? The secr manual says “Group-specific N has yet to be implemented”, so I was under the impression that the region.N() estimate was for the total population size. But now the separate ‘F’ and ‘M’ estimate has me questioning what these population estimates actually represent.

Also, D*mask area = 251.8

which makes me think region.N() only generated population size for one sex.

Any clarification on how R(N) and E(N) are calculated using region.N is greatly appreciated! Thanks!

[murray.efford]

Hi mobear

I can clarify some, but maybe not solve the whole problem.
1. If you really want you can see how region.N works by looking at the code (type region.N at the prompt). I agree that's not very helpful!
2. You have fitted a model in which groups do not appear in the formula for D, which implies equal densities of males and females. Why bother?
3. I think the manual should read 'Group-specific N has yet to be implemented. Do not expect sensible results if you have defined groups'. This is a loose end I should get back to. The discrepancy between expected and realised N is an indication that something is wrong.
4. For now you are better to bypass region.N and simply multiply the density and its SE by the area of the mask. This gives expected N and its SE, unless you have specified details=list(distribution='binomial'), in which case the SE is equivalent to that for realised N in the area of integration (see App 1 of Efford & Fewster 2013 Oikos).
5. This works if the area of interest is the region of integration as specified in the mask. And it's much clearer than relying on region.N.

I hope this helps
Murray

Code: Select all

library(secr)
## add phoney sex covariate
covariates(captdata) <- data.frame(sex = sample(c('F','M'), size = nrow(captdata),
 replace = T, prob = c(0.5,0.5)))
## fit model with fixed N (i.e. binomial n)
tmp <- secr.fit(captdata, groups = 'sex', model = D~g, details = list(distribution = 'binomial')
## Female realised N
predict(tmp)[[1]][1,2:3] * nrow(tmp$mask) * attr(tmp$mask,'area')
  estimate SE.estimate
D 60.63946    5.920332
## Male realised N
predict(tmp)[[2]][1,2:3] * nrow(tmp$mask) * attr(tmp$mask,'area')
  estimate SE.estimate
D 54.52511    5.597342


[mobear]

I am using secr 2.7.0 in Rstudio on Windows 7.

I am revisiting the above post from me, but now with a slight twist to the issue...

After some restructuring of my analysis, I now have D ~ x + y, g0 ~ bk, sigma ~ h2 as my most supported model.

fitted model:

fit.10 <- secr.fit(CH2012, model=list(D ~ x + y, g0 ~ bk, sigma ~ h2), hcov = "Sex", mask = FullMask2012)

In my previous post (ie, when D~1), specifying CL=TRUE and using region.N() for some reason fixed the estimation problems between E.N. and R.N.

But of course now I can’t use CL=TRUE while modeling a density trend and I am stuck with incorrect population size estimates from region.N().

region.N(fit.10, region = FullMask2012, se.N=T, alpha=0.05)

_____Est.__SE___lcl____ucl
E.N.__863__236__509__1461
R.N.__608__234__312__1297

These estimates are wildly different from estimating E.N. by D*mask area (EN = 307 bears, 95% CI = 231-426). And 863 bears is certainly an overestimate.

I am happy obtaining E.N. by multiplying D and its SE by the area of my mask as you previously suggested. My worry is that something bigger is going on that is beyond my understanding (which is a simple one). Perhaps something with the density surface or even the use of ‘hcov’ in my model is causing the issue? Or both?

I have read some other Phidot posts where you seem to caution against extrapolating N from a density surface model (eg, “help with N” by sixtystrat), but couldn’t discern whether this was a case-specific or overall recommendation.

So I am wondering if population size should be interpreted differently, used more cautiously, or not at all when modeling spatial trends in density? Is it still safe to calculate E.N. by multiplying D*mask area?

Thanks for your help and apologies for the lengthy lead up to a fairly straightforward question!

Clay

[murray.efford]

Hi Clay
With a fitted density surface there is always a risk that N extrapolated across the mask area will be implausible, especially using the default log link for density. I wonder also what D you are using for your manual extrapolation (the value reported by default by predict.secr is for the first point in the mask). I can probably diagnose the issue more precisely if you send me a fitted model to check - just save(fit.10, file = 'fit10.Rdata'). You're not the only person to have trouble with this in the last few weeks, so at the very least I should improve the documentation.
Murray