Expected or Realized N?

questions concerning anlysis/theory using program DENSITY and R package secr. Focus on spatially-explicit analysis.

Expected or Realized N?

Postby sixtystrat » Mon Feb 22, 2016 11:00 am

I have obtained estimates of N based on a density surface model with percent forest as a covariate. Therefore, I used region.N to estimate N rather than multiply the density by the area. In fact, I had to estimate density by dividing N by the area because the estimate the model gave me was for the base value of the forest covariate. All that being said, the Expected N and realized N's were pretty close and I used the Realized N. However, I also wanted to estimate N within a small subset of the study area which had been previously estimated by another researcher. In that case, Expected and Realized N's were very different (Realized N much larger). When I incorporated the density surface into a GIS and summed the density values for each pixel within that small area, the total was very close to the Expected N. So why the difference? What estimate should I use? Thanks!
Joe
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am

Re: Expected or Realized N?

Postby murray.efford » Tue Feb 23, 2016 12:55 am

Hi Joe
I'm not quite following...
In fact, I had to estimate density by dividing N by the area because the estimate the model gave me was for the base value of the forest covariate


'predictDsurface' does most of what you need there, e.g.
Code: Select all
# Suppose we want to predict density across a new region defined by 'possumarea'
new.mask <- make.mask(possumarea, type='poly', poly = possumarea)
# generate the required covariate (rough, for demoonstration only)
covariates(new.mask) <- data.frame(d.to.shore = distancetotrap(new.mask, possumarea))
# predict density at each point in new mask
possumsurface <- predictDsurface(possum.model.Ds, mask = new.mask)
# plot the surface
plot(predictDsurface(possum.model.Ds, mask = new.mask))
# average density over new.mask
mean(covariates(possumsurface)$D.0)
[1] 1.653324


But that wasn't really your question. I think the problem is that 'realized N' is defined only for regions large enough to include all detected animals. The problem with a smaller region is that some of the 'n' detected animals, a hard-wired component of realized N, may actually belong outside the small region. Without telemetry data you have no hope of splitting n into 'inside' and 'outside' components, and you must rely on a model prediction of the entire local population i.e. the expected N. If you do have telemetry data, the game changes...

Murray
murray.efford
 
Posts: 686
Joined: Mon Sep 29, 2008 7:11 pm
Location: Dunedin, New Zealand

Re: Expected or Realized N?

Postby sixtystrat » Tue Feb 23, 2016 3:38 pm

Thanks Murray. I was hoping that was the case. I was thinking that that is what you were saying in the Efford and Fewster paper, but I wasn't completely sure. Also, thanks for the help on getting mean density. That's a lot easier!
Joe
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am

Re: Expected or Realized N?

Postby sixtystrat » Thu Mar 17, 2016 3:45 pm

One more question (or maybe 2)...
We estimated density in a large area with a large buffer using 3x3 trap clusters arranged systematically across the landscape. Based on a null model (D~1, g0~1, sigma~1), I get
estimate SE.estimate lcl ucl n
E.N 659.1114 57.48421 555.7278 781.7277 217
R.N 659.1122 51.43271 569.2422 771.9114 217
With a more complex model (D~1, g0~bk+h2, sigma~h2, hcov="sex"), I get:
E.N 1075.968 140.9322 833.2609 1389.368 217
R.N 1156.670 137.0618 924.0834 1465.763 217
Why do the E.N. and R.N. differ for this model if the surface is homogeneous? Which do I report?
And with a habitat covariate (D~hab, g0~bk+h2, sigma~h2, hcov="sex"), I get:
E.N 1132.337 155.6777 865.9582 1480.657 217
R.N 1214.100 151.9974 957.8438 1558.995 217
Which do I report for this model? Thanks!
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am

Re: Expected or Realized N?

Postby murray.efford » Thu Mar 17, 2016 4:16 pm

Hi Joe

I don't claim any special insight on this. E.N and R.N are different estimators, so I don't think it's surprising they produce different answers, but I cannot explain why this is true when the difference is only in the detection part of the model. I'd like to investigate further, but don't have time just now.

I think the decision to use expected N or realized N is one that should be made globally in relation to your study objectives. In this case you are sampling only a small fraction of the population, so the R.N estimate comprises just a few detected animals and a lot of hypothetical others -- there was never going to be a big gain in precision. My (not-very-informed) choice would be to stick with E.N and avoid the complications.

Murray
murray.efford
 
Posts: 686
Joined: Mon Sep 29, 2008 7:11 pm
Location: Dunedin, New Zealand

Re: Expected or Realized N?

Postby sixtystrat » Thu Mar 17, 2016 4:38 pm

That helps. I am using sex as a heterogeneity covariate (hcov) and the manuals says group effects are not incorporated into region.N. Could that be the reason they are different? I'll go with expected N as you suggest, even for the homogeneous (D~1) model.
Thanks.
Joe
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am

Re: Expected or Realized N?

Postby murray.efford » Thu Mar 17, 2016 4:57 pm

The reference in ?region.N is to groups in the narrow sense specified with the 'groups' argument of secr.fit and model = D ~ g. 'hcov' should be OK.
Murray
murray.efford
 
Posts: 686
Joined: Mon Sep 29, 2008 7:11 pm
Location: Dunedin, New Zealand

Re: Expected or Realized N?

Postby sixtystrat » Fri Mar 18, 2016 7:56 am

Finally, one last question... We are doing a density surface based on habitat covariates but there is a sizeable part of one study area for which we have no habitat data. How does secr handle no data? Do we need to code it in a special way or can we use the mean from the parts of the study area where the data are available? Should we omit that covariate altogether? Thanks much Murray!
Joe
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am

Re: Expected or Realized N?

Postby murray.efford » Fri Mar 18, 2016 11:27 am

You need habitat data at each point. One way is to fill in missing values with the nearest non-missing value. See example in ?skink for the general idea. The function 'nearesttrap' doesn't care if the target is a traps object or a mask object.
Murray
Code: Select all
library(secr)
tmpmask <- possummask

# mask with missing covariate values
covariates(tmpmask)$d.to.shore[300:500] <- NA
plot(tmpmask, cov = 'd.to.shore', dots = FALSE)

# test (fails)
plot(predictDsurface(possum.model.Ds, tmpmask))

# infer covariate from nearest non-missing value
tmpmask.OK <- subset(tmpmask, !is.na(covariates(tmpmask)$d.to.shore))
OK <- nearesttrap(tmpmask, tmpmask.OK)
covariates(tmpmask)$d.to.shore <- covariates(tmpmask.OK)$d.to.shore[OK]

# re-test (works)
plot(predictDsurface(possum.model.Ds, tmpmask))
murray.efford
 
Posts: 686
Joined: Mon Sep 29, 2008 7:11 pm
Location: Dunedin, New Zealand

Re: Expected or Realized N?

Postby sixtystrat » Fri Mar 18, 2016 2:34 pm

That helps. Thanks Murray.
sixtystrat
 
Posts: 115
Joined: Wed Jun 01, 2011 9:19 am


Return to analysis help

Who is online

Users browsing this forum: No registered users and 12 guests