Offset term in pcount model statement for variation in area sampled

[Matt Giovanni]

Dear Unmarked users-

Before this google group was initiated, I was given the recommendation
of using an "offset" term in my pcount model statements to account for
variation in area sampled among sites when counting birds. So I
specified models as, for example,

mng=pcount(~obs ~mng, umf, offset(log(area))),

but this model, without the offset term,

mng=pcount(~obs ~mng, umf),

provides identical model results, even though our sites ranged widely
in area sampled. I suspect this term is in fact not accounting for
sampling effort as I had hoped. Does anyone have any insights or
recommendations on how to account for variation in sampling area per
site?

Thanks for your assistance.

Matt

[Richard Chandler]

Hi Matt,

I think you want

pcount(~obs ~mng + offset(log(area)), umf)

See the example on the glm help page for using offsets in formulas.
There are also some examples in the R scripts here:

https://sites.google.com/site/hierarchicalmodelingcourse/home/r-scripts

Have a good weekend,
Richard

[Jeffrey Royle]

hi Matt,
it would also be worth putting log(area) in as a covariate, i.e.,
allowing for the estimation of its coefficient. If the estimated
coefficient is very different from 1 then it suggests the effort per
unit area was not proportional and you should use the covariate model
instead of the offset.
regards
andy

On Fri, Sep 17, 2010 at 7:44 PM, Matt Giovanni
<matthew...@gmail.com> wrote:

[Jeffrey Royle]

when I run the minkfrog example from Richard's class notes it seems to
give a sensible result that suggests the offset is doing the right
thing.....see below
regards
andy

> (null <- pcount(~1~1, minkFrog))

Call:
pcount(formula = ~1 ~ 1, data = minkFrog)

Abundance:
Estimate SE z P(>|z|)
-0.172 0.0805 -2.14 0.0325

Detection:
Estimate SE z P(>|z|)
0.157 0.0856 1.84 0.0658

AIC: 1383.131
>
> (null2 <- pcount(~1~offset(log(pondArea)), minkFrog))

Call:
pcount(formula = ~1 ~ offset(log(pondArea)), data = minkFrog)

Abundance:
Estimate SE z P(>|z|)
-1.94 0.0806 -24.0 9.3e-128

Detection:
Estimate SE z P(>|z|)
0.139 0.087 1.60 0.111

AIC: 1369.915
>

On Fri, Sep 17, 2010 at 10:51 PM, Giovanni,Matthew [Reg]
<matthew....@ec.gc.ca> wrote:
> Hi Richard, Andy-
>
> Thanks for your responses.  It must be Friday afternoon because I should've typed exactly what Richard suggested (adding the offset term to my covariate for management), which is what I have been doing.  Sorry for the typo and confusion.
>
> But, I'm definitely confused because I still get the same summaries from my models, and I also get identical model summaries when I run the script directly and fit the n-mix models to the minkfrog data from class 6 for Richard's hierarchical modeling course:
>
> (null <- pcount(~1~1, minkFrog))
>
> Call:
> pcount(formula = ~1 ~ 1, data = minkFrog)
>
> Abundance:
>  Estimate     SE     z P(>|z|)
>   -0.172 0.0805 -2.14  0.0325
>
> Detection:
>  Estimate     SE    z P(>|z|)
>    0.157 0.0856 1.84  0.0658
>
> AIC: 1383.131
>> (null2 <- pcount(~1~offset(log(pondArea)), minkFrog))
>
> Call:
> pcount(formula = ~1 ~ offset(log(pondArea)), data = minkFrog)
>
> Abundance:
>  Estimate     SE     z P(>|z|)
>   -0.172 0.0805 -2.14  0.0325
>
> Detection:
>  Estimate     SE    z P(>|z|)
>    0.157 0.0856 1.84  0.0658
>
> AIC: 1383.131
>
>
>
> I also fit a model with "log(area)" as a covariate, and there was no particular correlation in either direction, but I find this hard to believe given the large range of areas among our sampling sites.
>
> If you have a moment, I would be tremendously grateful if you could have a gander at my R and csv files (attached).
>
> Guys, thanks again and have a great weekend.
>
> Matt
>
> ______________________________________
> Matt Giovanni, Ph.D.
> Postdoctoral Visiting Research Fellow
> Canadian Wildlife Service
> 2365 Albert St., Room 300
> Regina, SK S4P 4K1
> 306-780-6121 work
> 402-617-3764 mobile
> Research website:
> http://sites.google.com/site/matthewgiovanni/

[Richard Chandler]

I'm not sure if this has been resolved, but I will look at it on
Monday. One issue is that it is important that you are using a recent
version of unmarked because offsets were ignored in early versions.

Richard

[Matthew Giovanni]

Richard is correct, in that I get the same coefficients when I fit the minkfrog null models without and with offset terms with unmarked 0.8-1, but I get different results (same as Andy produced) when I use 0.8-7.

Thanks for helping me figure this one out, guys.  Big relief!!!  Cheers,

Matt

____________________________________
Matt Giovanni

Postdoctoral Visiting Research Fellow
Canadian Wildlife Service
2365 Albert St., Room 300
Regina, SK S4P 4K1
306-780-6121 work
402-617-3764 mobile
Research website:
http://sites.google.com/site/matthewgiovanni/