Huggins Closed Capture- Removal Model for Pt Ct data

[db]

I am trying to run a Huggins- Closed Capture analysis on avian point count data in which the data were collected based on a removal design, with 5 min counts divided into 5 1-min intervals, and 2 groups and 1 co-variate (distance class). Do I put dots in the encounter history after the first detection like this for a removal model or do I put 0s?

/*2 5*/ 1.... 1 0 75;
/*2 12*/ 1.... 1 0 175;
/*2 18*/ 001.. 1 0 175;
/*4 1*/ 1.... 1 0 75;
/*4 1*/ 0001. 1 0 375;
/*4 2*/ 1.... 1 0 75;
/*4 3*/ 1.... 1 0 175;
/*4 8*/ 01... 1 0 175;
/*4 8*/ 0001. 1 0 75;
/*4 13*/ 001.. 1 0 175;

Also, I assume I would want to fix c=0, because there are no recaptures in this removal method- is that correct?

Finally, when I run these data with c=0, p comes out to be = 0.012 in the dot dot model, which seems extremely low for the data considering that we often detected the species in the first interval.

Where am I going wrong? Even though I attended the MARk course, I don't think we ever got in this deep!

[gwhite]

Just another note about this model. In general, the Huggins model does not work all that well with removal data. Typically, the full likelihood model that includes N in the likelihood does better -- seems to be much more stable. However, this model has to be coded with zeros and with negative values for the counts.

Gary

[gwhite]

I told you wrong in the previous response. You should use zeros and do not use the negative frequencies in the Huggins model. You do have to fix c=0. This model will give roughly the same results as the data type with N included in the likelihood.

Run a test example to demonstrate this for your self.

Gary

[db]

Thanks, Gary. That helps clarify things!

[murray.efford]

I'd strongly suggest you do not use a removal model to analyse point count data. This is almost certainly useless, despite some published advice to the contrary. See Efford & Dawson 2009 Auk 126: 100-111.

[murray.efford]

On reflection, let me tone that down: given your distance covariate that should read 'very likely ineffective'

[jlaake]

I haven't followed the removal pt count literature but in practice it shouldn't be any different than the double observer line transect literature. You should be able to analyse the removal data to get an estimate of detection probability at distance 0 and then use that with a standard pt transect analysis of all the unique sightings using p(0)=1-(1-p'(0))^2 where p'(0) is the estimate of detection probability for a single observer from the removal data. This is discussed in ch 6 of the Advanced distance sampling book and the 2006 Borchers et al article in Biometrics.

--jeff

[murray.efford]

Jeff
That sort of sophistication hasn't yet penetrated the point count literature! Our simulations in Auk merely addressed the existing basket of capture-recapture and removal models applied to point counts, including distance used as a covariate. It would be good to develop composite methods further. I think the key is to demonstrate their robustness with the types of data actually likely to be collected.
Murray

[db]

I appreciate all of the good advice! As a "by the way" to further explain my motives, we actually have collected a variety of data with a variety of methods that I am analyzing for comparative purposes here, including the removal model! I have indeed read Murray's paper in Auk with great interest!

[cooch]

On reflection, let me tone that down: given your distance covariate that should read 'very likely ineffective'

Indeed - for a minute there, I thought I was mistakenly reading the R maillist...