multi-method model: availability across methods

[giancarlo]

Greetings list;

It's been awhile but I was hoping for some insight from the PRESENCE-masters (Jim? Darryl? et al.?) regarding the single-season multi-method models and approach of Nichols et al. 2008.

The way I understand the multi-method model (and similar to the single-occupancy w/ correlated detection model), the theta parameter(s) are estimated to account for varying (often reduced) availability probabilities across secondary sampling units given presence at the scale of the primary sampling unit. As employed in the Nichols et al. 2008 paper and others I've seen (e.g. Jeffress et al 2011 JWM), these models are primarily employed for secondary units that are spatial segments or sub-units of the larger site (primary sampling unit).

I haven't dug into all the other examples (63 by Google Scholar's count), but it seems reasonable to apply this method to multi-method temporally-repeated secondary sampling units if availability for detection is expected to vary across these sampling occasions.

Assuming I'm not off-base with that assertion, my question is whether it would be appropriate to employ the theta parameter(s) to account for availability-for-detection **across methods** rather than across secondary sampling units. For example, in temporally-repeated secondary sampling without correlated observations (i.e. the standard single-season model), the underlying assumption is that availability-for-detection does not vary within the single season. However, if using multiple methods, this availability could vary by various unknown factors (and which could be useful to account for).

If this is appropriate, I could see it requiring data to be entered in the order of method-sampling unit rather than the sampling unit-method as indicated in the manual.

Thanks for any insight!

Giancarlo

[darryl]

If I'm following your question correctly, the skunk example in Nichols et al. (2008) is doing what you're asking: there the surveying was over a 2-week period with checks of the devices about every 3 days, so the theta relates to a temporal availability probability.
Darryl

[giancarlo]

Thanks Darryl, that answers my first question (guess I'll go back and rtfp!).

Regarding my second question, does it sound reasonable to get/account for a method-specific availability probability by reordering the observation columns, ie:

surv1meth1 surv2meth1 surv1meth2 surv2meth2

Rather than

surv1meth1 surv1meth2 surv2meth1 surv2meth2

expected for a temporally-varying availability model?

Thanks again!

Giancarlo

[darryl]

Knee-jerk reaction woud be yes, but temporal availability would now be confounded with detection

[giancarlo]

Thanks Darryl. I welcome any further thoughts on this. I have a few survey (date, time), weather (wind), and habitat covariates I expect influenced detection but I assume availability didn't change over time (all surveys within a month).

Giancarlo

[giancarlo]

I should have noticed this earlier, but looking back at the help file raises some confusion:

http://www.mbr-pwrc.usgs.gov/software/doc/presence/presence.html#multi_method

Help file states:

ψ - probability that the area is occupied by the species,
θx - probability that individuals are available for detection using method x, given presence,
pxi - probability of detecting species using method x in survey i

However, in examples in Nichols et al. 2008, θ is either treated as temporally-constant (θ) or temporally-varying (θt), but not as varying among methods (θx), which is what the help file indicates PRESENCE is estimating. To clarify my previous comments, the help file also notes:

Data input for this model is as follows:

site1 h1,1 h1,2 h1,3... h2,1 h2,2 h2,3...
site2 h1,1 h1,2 h1,3... h2,1 h2,2 h2,3...
site3 h1,1 h1,2 h1,3... h2,1 h2,2 h2,3...
:
:

where hi,j = 1 if detection at site for survey i, method j; hi,j = 0 if no detection

Given the order of detection histories indicated above, I'd appreciate if anyone can confirm that PRESENCE is, in fact, estimating θx and not θt. This, of course, is only an issue when θ varies (when number of occasions per season < number of columns).

Many thanks. Giancarlo

[giancarlo]

After some experimentation with some simulated data, I can confirm that θ in the multi-method model is θt (also employed this way in Nichols et al. 2008 and others)... assuming that detections are entered in the survey-major, method-minor order and No. methods per survey (aka No. Occ/season) is entered correctly. This is consistent with the default multi-method model (psi,theta(t),p(.)), assuming "t" is time. It is, however, inconsistent with the help file

θx - probability that individuals are available for detection using method x, given presence,

If one was interested in estimating differences in the probability of availability for detection (not just detection probability) over different methods (assuming constant availability for detection over time/surveys), it sounds (from Darryl's) response, that reordering the detection history as method-major, survey-minor would get you there.

Giancarlo