Change in detection function

[N.Coallier]

Hi ,

I run into a curious problem.
I've tried to run the following model:

secr.fit(MesicBrown2011,model=list(D~session,g0~session+b,sigma~session),buffer=100,detectfn=0,trace=TRUE)

But I get

Checking data
Preparing detection design matrices
Preparing density design matrix
Finding initial parameter values...
Initial values D = 7.80193, g0 = 0.04695, sigma = 17.98337
Maximizing likelihood...
Eval Loglik D D.sessionLG1PP2 D.sessionLG1PP3 g0 g0.sessionLG1PP2 g0.sessionLG1PP3 g0.b sigma sigma.sessionLG1PP2 sigma.sessionLG1PP3
1 NA 2.0544 0.0000 0.0000 -3.0105 0.0000 0.0000 0.0000 2.8894 0.0000 0.0000
2 NA 2.0544 0.0000 0.0000 -3.0105 0.0000 0.0000 0.0000 2.8894 0.0000 0.0000
3 NA 2.0544 0.0000 0.0000 -3.0105 0.0000 0.0000 0.0000 2.8894 0.0000 0.0000
4 NA 2.0544 0.0000 0.0000 -3.0105 0.0000 0.0000 0.0000 2.8894 0.0000 0.0000


So I've read about this kind of problem and some suggest that it could be du to an animal with crazy movement. In that case I could change the detection function to a exponential (2) to be able to get correct Density estimate !

I've done that and it seems to work ...

Checking data
Preparing detection design matrices
Preparing density design matrix
Finding initial parameter values...
Initial values D = 7.86567, g0 = 0.04802, sigma = 17.7744
Maximizing likelihood...
Eval Loglik D D.sessionLG1PP2 D.sessionLG1PP3 g0 g0.sessionLG1PP2 g0.sessionLG1PP3 g0.b sigma sigma.sessionLG1PP2 sigma.sessionLG1PP3
1 -1447.786 2.0625 0.0000 0.0000 -2.9868 0.0000 0.0000 0.0000 2.8778 0.0000 0.0000
2 -1447.786 2.0625 0.0000 0.0000 -2.9868 0.0000 0.0000 0.0000 2.8778 0.0000 0.0000
3 -1447.786 2.0625 0.0000 0.0000 -2.9868 0.0000 0.0000

I get proper density estimate but I'm not sure the result are ok since I've used halfnormal detection function for the rest of my data (long-term data).

Am I doing this right ?

Thanks a lot !!

Nicolas

[murray.efford]

'secr' fits the model by numerically maximizing the likelihood;that requires starting values of the parameters, and the values provided automatically are not always close enough. Bad starting values can cause the likelihood evaluation to fail - my guess is that that is the reason for the NA values in your case. You can give whatever values you like in the 'start' argument, so I suggest you try something like

Code: Select all

secr.fit(..., start=list(D = 8, g0=0.05, sigma=25))

Although your density estimates with a negative exponential detection function should be comparable (it is possible the EX function is a better fit if you have a few extreme movements) my feeling is that it is tidier not to let the start value problem dictate your model.
Murray

P.S. If your models have session-specific values for all parameters it is far quicker to fit separately to each session.

[N.Coallier]

Hi M.Efford,

Thanks a lot for your quick answer !

It did work when I specifiy starting value. Therefore, I am wondering if I should use the exponential detection function anyways... I'll investigated on that because I have another example of data that seems to be really sensitive to the detection function used....

Indeed, it gives me density of either 8, 17 or 30 depending on the detection function that I used (0,1,2)... it is kind of crazy. A density of 30 is not probable in my system but still curious. I've tried to repeated model with different starting value to check for a local minimum but it dosen't change anything so that dosen't seem to be the problem. Let me know if you have any leeds.


Thanks a lot !

Nicolas

[murray.efford]

The exponential and hazard-rate detection functions have long tails, so they are sensitive to buffer width (if the habitat mask is not limited by the distribution of suitable habitat). That seems the likely explanation - there must be one!
Murray