Re: [unmarked] Bayes version of Dail-Madsen Model, overestimating Lambda?

430 views
Skip to first unread message

Richard Chandler

unread,
Feb 21, 2013, 4:22:57 PM2/21/13
to Unmarked package
Hi Shannon,

For some reason neither WinBUGS or OpenBUGS work for this model. You have to use JAGS, and it needs to be one of the more recent versions of JAGS.

Richard


On Thu, Feb 21, 2013 at 1:26 PM, Shannon Knapp <shannon.k...@gmail.com> wrote:
 
A coworker has some repeated point counts over time and wants my help to apply the Dail-Madsen model to this data.  Abundance is of some interest, but the primary interest is in trend in abundance.  There are several sites and it's been suggested that a site-effect is going to be appropriate, so I'm thinking a random site effect will be in order, which means going the Bayes route.
 
To prepare for all this, I wanted to make sure I could code the simple analysis, so I simulated some data and ran the analysis using WinBUGS (R2WinBUGS) and pcountOpen in unmarked. 
 
I noticed that I kept getting severe overestimates of lambda (and underestimates of p, though pretty good estimates of gamma - the trend).  "Severe overestimates" meaning overestimating by 30-60% and CrIs not including the true values.  However, the pcountOpen estimates were very good.
 
Wondering if anyone else has noticed this problem?
 
I used this script as a reference (for simulating the data as well as for the unmarked and Bayes analyses):
script-dail-madsen.R
 
Initially I was using a large lambda (like 50), 5 years, a pretty severe gamma (like 0.5 - so a sharply decreasing population) and a decently high p (like 0.7) and 100 sites.  I then tried the values in script-dail-madsen.R (lambda of 5, gamma around 1.01, p = 0.5, and 10 years of data).  Same basic outcome:  the Bayes analysis severely overestimated lambda, underestimated p, (gamma was good), but the pcountOpen estimates were great.
 
The priors I had on lambda should have, if anything, resulted in a tendency to underestimate, so that shouldn't be it.
Also the posterior distribution of lambda (and the other parameters) didn't show any obvious influence of the priors and it looked like things had converged just fine.  But I was getting this on a few-several replicate simulations.  I also checked the average of the N's from the first year and it wasn't that the true population size was just higher than "expected".
 
Also, I tried this both with one survey per year and with 3 surveys per year.  Same result.  But when I did 3 surveys per year I also did a N-mixture model, that had an expoential trend in lambda for each year, but totally ignored any "autoregressive" tendency in the data (that is, the lambda for a site at a year is the same for all sites in that year - regardless of the specific N for that site the previous year) -- this model actually did give me a very good estimate of lambda (and everything else).
 
Anyway, just wondering if other people have run into this issue.
 
Shannon

--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

Shannon Knapp

unread,
Feb 22, 2013, 9:56:50 AM2/22/13
to unma...@googlegroups.com
Well that's deeply disturbing, but good to know!  And something I **never** would have guessed!
 
Thank you!
Shannon
Reply all
Reply to author
Forward
0 new messages