By request, I'm posting some more information about my model (Thank you
Timothy Handley).
Here is what I am trying to do. I have a count data set (elk) from the years
1978-2010 based on aerial transect counts in a national park in Canada. The
data is in the form counts per transect. Some years certain transects
weren't flown, some years were not flown at all and after 1995 transects
were extended out of the park. My boss wants me to try and extrapolate
transects to years when they were not flown, also the extended transects -
and them sum them up to get an estimate of population trend. I have no
measure of detectability - no efforts were made on testing that (!), and I
was not able to get landcover data over the years either. So basically I
will just be looking at changes in trend over time.
So my plan so far has been to use a log-linear model (with an overdispersed
poisson distribution for counts) to get count estimates per transect and
then extrapolate those estimates to years where the transects weren't flow.
To do that I have treated transects extended out of the park as separate
transects such that the "new" transects started at the park border and
outwards into the buffer zone. This way I might estimate counts for the
"new" transects for the years where such transects were not flown. I have
created a variable, segment, to retain information on which "old" transect a
"new" transect originally belonged to. And maybe I can run the segment
effect as a random effect? I'm not so sure if that is right though (I tried
running a CAR proper with only fixed intercept and slope, but it didn't
converge at all - I don't know if the CAR proper would account for the "new"
transects and I don't have to run a segment effect??)
Anyways, I'm looking a fixed covariates such as temperature, snow depth,
productivity, hunting and transect area and have been looking at DIC values
for model selection (unfortunately it seems like they all improve DIC, I
would have liked to get rid of more covariates).
With the models I have running now I find that for trend both quadratic
effects and autoregressive effects improve on the DIC, so I have
beta(t-fixedyear), beta(t-fixedyear)^2 and beta((t-1)-fixedyear) in the
model.
So the model is pretty big and is taking quite some time to run - The
problems comes when I try to add random terms such as overdispersion, random
transect term, random year term and random segment term (although I'm not
sure if I should even use this last one). I do realize that I may not be
able to fit that many random terms, but even if I just try to fit the
overdispersion term (epsilon), my model won't converge.
And that is when I was thinking that maybe I need to account for zero
inflation in my model, and that the overdispersed Poisson may not be best
for my data. I did consider using the negative binomial distribution but
I've read that the one in Winbugs in not really the one I would like to use
for ecological data.
I'm running R2Winbugs from R - I have received suggestions switching to
Jags, so I may just look into that.
Thanks,
Lene
______________________________
Lene Jung Kjaer, PhD
NSERC Visiting Fellow
Parks Canada | Parcs Canada
Western and Northern Service Centre
145 McDermot Avenue
Winnipeg, MB, R3B 0R9
office: 204-984-5821
Mobile: 204-951-1290
fax: 204-983-0031
email work: [log in to unmask]
email home: [log in to unmask]
_____________________________
-----Original Message-----
From: [log in to unmask] [mailto:[log in to unmask]]
Sent: Saturday, March 19, 2011 10:45
To: [log in to unmask]
Subject: Re: BUGS Digest - 15 Mar 2011 to 18 Mar 2011 (#2011-40)
Could you repost with more information? My feeling is that several of your
posts have contained not-quite-enough information for folks on this list to
give good answers. Specifically, it would be nice if you could include the
entire model description, as well as a few short paragraphs explaining the
biological context, and the meaning of the variables within your model.
Quick thought: Zeros are different from 1's. With BUGS, you have the
ability to construct a proper zero inflated model with only minor
modifications to your current model. I think this would always be better
than fudging .0001 values. When I have zero-inflated data, I have used one
of the several models models described in chapter 11 of Zuur et al 2009. I
would recommend you look at the chapter, and see if one of those seems
appropriate to your situation.
Tim Handley
Research Assistant
Channel Islands National Park
(Will be working from both CHIS and SAMO)
SAMO Phone: 805-370-2300 x2412(Mon, Tue)
CHIS Phone: 805-658-5759 (Wed, Thu, Fri )
Date: Thu, 17 Mar 2011 22:19:21 -0500
From: Lene Jung <[log in to unmask]>
Subject: Zero-inflation?
Hi Guys.
Sorry for bothering the list with yet some more questions, but as I'm quite
new to Bayesian statistics, I'm still learning.
I have recently asked quite a few questions on the list server regarding
fitting transect count data to a log-linear model in winbugs. I have been
using an overdispersed poisson distribution and have been having a lot of
problems fitting random effects to my model - ie non-convergence, even when
I just added the overdispersion term.
My data is in the form of counts per transects, and I have both missing
transect counts and zero transect counts. Including the missing transects
counts, zero counts make out about 12% of total counts and if I just
compare
to actual counts (without NAs) it amounts to 52%. Have I been going about
this the wrong way? Can I not use the overdispersed poisson when I have
that
many zero counts (I've been substituting zeros with 0.00001 for the log
model as per suggestion from the list server)??? Should I look into
zero-inflated distributions instead?
I've been working on this for a long time now and am getting increasingly
frustrated with the models not converging - I could really use some help
and
suggestions here.
Again, sorry for any inconvenience,
Lene Jung Kjaer
______________________________
Lene Jung Kjaer, PhD
NSERC Visiting Fellow
Parks Canada | Parcs Canada
Western and Northern Service Centre
145 McDermot Avenue
Winnipeg, MB, R3B 0R9
office: 204-984-5821
Mobile: 204-951-1290
fax: 204-983-0031
email work: [log in to unmask]
email home: [log in to unmask]
_____________________________
-------------------------------------------------------------------
This list is for discussion of modelling issues and the BUGS software.
For help with crashes and error messages, first mail [log in to unmask]
To mail the BUGS list, mail to [log in to unmask]
Before mailing, please check the archive at www.jiscmail.ac.uk/lists/bugs.html
Please do not mail attachments to the list.
To leave the BUGS list, send LEAVE BUGS to [log in to unmask]
If this fails, mail [log in to unmask], NOT the whole list
|