Dealing with Occupancy models that fail

[birdnerd]

Hello all,

I'm using MARK to examine occupancy of a few dozen forest dwelling birds in relation to timber harvests. I've built a model set of 7 relevant veg covariates, plus treatment, that I'm running for all species (122 models). My goal is to determine what is having a larger effect on occupancy, treatment or the vegetation characteristics. Detection probabilities and occupancy range from just over 0 to just under 100%, and I've had to drop many of the species because I can't get the models to converge (all species with less than 20% detection probability).

However, I have several species with between 20-30% detection probability that I'm having trouble getting the models to work for as well. One or two models for these species will fail (MARK only estimates detection and fails to estimate any other parameters). My most common species, with 85% detection probability and almost 100% occupancy, also has the same problem.

I've a few questions: Can I use these species in my analysis? Should I just remove the 1 or 2 models that fail, or would that invalidate my results? I've found that when I check the "use alt. opt. method", the models will usually work. Would I then have to run all 120 other models that way as well?

Thanks for your time
David

[bacollier]

Hello all,

I'm using MARK to examine occupancy of a few dozen forest dwelling birds in relation to timber harvests. I've built a model set of 7 relevant veg covariates, plus treatment, that I'm running for all species (122 models). My goal is to determine what is having a larger effect on occupancy, treatment or the vegetation characteristics. Detection probabilities and occupancy range from just over 0 to just under 100%, and I've had to drop many of the species because I can't get the models to converge (all species with less than 20% detection probability).

I assume you mean a 20% naive detection rate? I guess the first question to ask is whether or not you have enough data. Home many 'sites' are you sampling/collecting P/A data on? How many sampling occasions are you using? what do your enc. histories look like in general for each site/species combination (e.g., are they 1101101, or 0000100)?

However, I have several species with between 20-30% detection probability that I'm having trouble getting the models to work for as well. One or two models for these species will fail (MARK only estimates detection and fails to estimate any other parameters). My most common species, with 85% detection probability and almost 100% occupancy, also has the same problem.

What models are failing? Is it the same model every time?

I've a few questions: Can I use these species in my analysis? Should I just remove the 1 or 2 models that fail, or would that invalidate my results? I've found that when I check the "use alt. opt. method", the models will usually work. Would I then have to run all 120 other models that way as well?

Not enough information yet to say yes or no on dropping stuff. I would not remove models until you get a feel for why the error is occurring. Can you provide information that is specific to what you are doing, what data looks like, and what models are failing?

\bret

[birdnerd]

I assume you mean a 20% naive detection rate? I guess the first question to ask is whether or not you have enough data. Home many 'sites' are you sampling/collecting P/A data on? How many sampling occasions are you using? what do your enc. histories look like in general for each site/species combination (e.g., are they 1101101, or 0000100)?

I sampled 120 sites 3 times in a season. My detection histories vary, but for the species I'm having trouble with, there's a lot of 100 or 001 (and a lot of 000's as well). I'm guessing this is part of the problem, having a low number of sampling occasions.

What models are failing? Is it the same model every time?

It's not the same model every time, and for many species, it seems to be random. For one species, it was one variable that seemed to be the problem, but others only the fully saturated model would fail, or just 2 or 3 seemingly random model.

I've a few questions: Can I use these species in my analysis? Should I just remove the 1 or 2 models that fail, or would that invalidate my results? I've found that when I check the "use alt. opt. method", the models will usually work. Would I then have to run all 120 other models that way as well?

Not enough information yet to say yes or no on dropping stuff. I would not remove models until you get a feel for why the error is occurring. Can you provide information that is specific to what you are doing, what data looks like, and what models are failing?

\bret

Here's an example for a species I'd especially like to use (Yellow-bellied Sapsucker):

TPA is trees/acre, and TREAT is a dummy variable whether indicating a forestry treatment was conducted at that site.

In the dot model, p=.35, phi=.65
The only model that fails is {P+PHI+TPA+TREAT+TPAXTREAT}.

Here's the output I get:

YBSA13

LOGIT Link Function Parameters of {P+PHI+TPA X TREAT}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1:P -0.8212437 0.1679888 -1.1505018 -0.4919856
2:PHI 1.0050269 0.6877274 -0.3429189 2.3529727
3:TPA -0.9525849 0.5493958 -2.0294006 0.1242309
4:TREAT 306.62691 3826.5921 -7193.4938 7806.7477
5:INTERACTION -382.72479 725.20227 -1804.1213 1038.6717

95% Confidence Interval
Parameter Estimate Standard Error Lower Upper
-------------------------- -------------- -------------- -------------- --------------
1:p 0.3054997 0.0356421 0.2403974 0.3794259
2:Psi 1.0000000 0.0000000 1.0000000 1.0000000

All the other models seem to work fine, including ones with TREAT and interactions with TREAT. The encounter history file is a mix of 100, 101 and a few 111, and maybe half of the sites 000.

Thanks for taking a look!

David

[bacollier]

Here's an example for a species I'd especially like to use (Yellow-bellied Sapsucker):

TPA is trees/acre, and TREAT is a dummy variable whether indicating a forestry treatment was conducted at that site.

In the dot model, p=.35, phi=.65
The only model that fails is {P+PHI+TPA+TREAT+TPAXTREAT}.

Here's the output I get:

YBSA13

LOGIT Link Function Parameters of {P+PHI+TPA X TREAT}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1:P -0.8212437 0.1679888 -1.1505018 -0.4919856
2:PHI 1.0050269 0.6877274 -0.3429189 2.3529727
3:TPA -0.9525849 0.5493958 -2.0294006 0.1242309
4:TREAT 306.62691 3826.5921 -7193.4938 7806.7477
5:INTERACTION -382.72479 725.20227 -1804.1213 1038.6717

95% Confidence Interval
Parameter Estimate Standard Error Lower Upper
-------------------------- -------------- -------------- -------------- --------------
1:p 0.3054997 0.0356421 0.2403974 0.3794259
2:Psi 1.0000000 0.0000000 1.0000000 1.0000000

All the other models seem to work fine, including ones with TREAT and interactions with TREAT. The encounter history file is a mix of 100, 101 and a few 111, and maybe half of the sites 000.

Thanks for taking a look!

David

OK, without seeing the data/dm there is obviously a issue with your treatment variable. Do the sapsuckers only occur in one (or occur in all) of the categories (e.g., you don't hear them in untreated for instance)?

This might be easier to fix offline so as not to flood folks email with our back and forth, if you want to zip up and send me the sapsucker .inp/fpt/dbf files I will take a look at see if I can help you out (you can email me: bret at tamu.edu).

\bret