Open population analysis in secr?

[charlotte...@gmail.com]

Hello,

I have a data set with 238 continuous trap nights which I would like to divide into 3 sessions to meet the demographic closure assumption. However, if I do that the sessions will not be independent. I also have 56 trap nights from the previous year that I would like to incorporate as well.
Seems like I should be doing an open population analysis with repeated sampling instead. I think you can do this in Spacecap but was wondering if there is a way of doing it in secr?


Thank you,
Charlotte

[Murray Efford]

Open population models are a way to deal with this, but I'm not entirely convinced. Does Spacecap do open-population models? I hadn't caught up with that. 'secr' does not itself do open populations, but if you can wait 2-3 months there is other software in development that will fill that gap.

I suggest plowing on with a multi-session secr analysis: if the constant density estimated from a multi-session model is the same as density from a pooled single-session model then I think you can ignore turnover. CI from the multi-session model will be too short due to non-independence, so use those from the single-session model.  Or just relax: I suspect that your sample sizes are small enough that these minor analysis issues will be overwhelmed by sampling error.

Murray

[Ben Augustine]

I agree with Murray that you're probably best off doing multiple sessions. Spacecap does not do open populations. The only software I know of to do open population SCR other than using JAGS is the likelihood-based

https://github.com/r-glennie/openpopscr

and MCMC-based

https://github.com/benaug/OpenPopSCR

--
You received this message because you are subscribed to the Google Groups "secr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to secrgroup+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Chris Sutherland]

Hi Charlotte,

Check out Ben Augustines package 'OpenPopSCR':

https://github.com/benaug/OpenPopSCR

It will get you what you need withouut the wait.

Chris

[charlotte...@gmail.com]

Thanks for the quick replies!
I have continued trying to use secr but will move on to OpenPopSCR if I can't get secr to work.
I first did a pooled single-session model and then a multi-session (3 sessions, 80 days each) which seem to work but I do get this warning message:

Warning message:

In autoini(ch, msk, binomN = details$binomN, ignoreusage = details$ignoreusage) :
  'autoini' failed to find g0; setting initial g0 = 0.1

My main concern is that the density estimates seem unreasonably small (

4.897127e-04 vs 4.888377e-04) unless I'm interpreting the output incorrectly. I'm new to secr so I'm suspecting that I'm doing something wrong but I can't work out what. 

Pooled single-session
CH10 <- read.capthist('Cap_10days.txt', 'Detectors_3sessions.txt', detector = 'count')

fit3 <- secr.fit(CH10, buffer = 15000, method="Nelder-Mead")

Fitted (real) parameters evaluated at base levels of covariates 
       link     estimate  SE.estimate          lcl          ucl
D       log 4.897127e-04 7.852386e-05 3.583598e-04 6.692117e-04
g0    logit 9.999998e-01 9.162790e-06 3.232398e-39 1.000000e+00
sigma   log 2.978438e+03 8.028579e+01 2.825193e+03 3.139996e+03


Multi-session (3 sessions, 80 days each) 

CH3S <- read.capthist('SECR_capture3S.txt', 'Detectors_3sessions.txt', detector = 'count')
summary(CH3S)

fit4 <- secr.fit(CH3S, buffer = 15000, method="Nelder-Mead")

link     estimate  SE.estimate           lcl          ucl
D       log 4.888377e-04 5.997816e-05  3.846945e-04 6.211741e-04
g0    logit 1.000000e+00 6.289403e-06 1.187767e-128 1.000000e+00
sigma   log 1.875961e+03 7.899571e+01  1.727412e+03 2.037284e+03

This is what the capthist look like for the single-session model:

Object class       capthist 
Detector type      count 
Detector number    19 
Average spacing    2651.127 m 
x-range            435683 460289 m 
y-range            8122335 8141074 m 

Counts by occasion 
                   1  2  3  4   5   6  7  8  9 10 Total
n                 14 12 18 17  15  17  9 13 19 12   146
u                 14  5  4  7   4   3  0  1  2  1    41
f                 12  4  7  4   7   3  0  1  2  1    41
M(t+1)            14 19 23 30  34  37 37 38 40 41    41
losses             0  0  0  0   0   0  0  0  0  0     0
detections        48 75 58 35 379 105 23 25 41 43   832
detectors visited  9 10 11 10  10  13  8  8  9 11    99
detectors used    19 19 19 19  19  19 19 19 19 19   190

Many thanks,
Charlotte

[Murray Efford]

1. I would have said you could ignore the autoini warning, but you also have a problem with the g0 estimates themselves hitting the boundary at 1.0. (i) are you using a recent version of secr? (ii) for count data, especially, it makes sense to use one of the hazard-based detection functions (e.g. detectfn = 'HHN' not 'HN') for which the intercept is lambda0 not g0, (iii) I assume your traps object has non-null usage - be aware that will determine the scaling of g0 or lambda0. Your occasions in the single-session analysis seem to be about one month long, which is fine, but if you want a daily g0 or lambda0 then values in the usage matrix should be 30 rather than 1.
2. Density is expressed by 'secr' in animals per hectare. There are 100 hectares per sq km. 5 jaguar/100 km^2 seems alright to me - were you expecting more?
3. The SE of D-hat from the multi-session model is spuriously low because of non-independence as discussed before.
Non-closure seems like a non-issue, but you'll want to sort out the g0 issue before trusting the density estimates.
Murray