c-Hat and Pradel Robust design

[ctlamb]

I have searched the forum history for this, and found two answers:

viewtopic.php?f=1&t=2492&p=7850&hilit=chat+robust#p7850

Where Cooch recommends reducing all secondary sessions into single sessions and analyzing the fit of these data using median c-hat, and I assume the CJS model, not the Pradel Robust given that we have no secondary sessions now.

I was going to do this, but I'm curious how lumping all my secondary sessions will affect resulting c-Hat, given that my secondary sessions are composed of sessions where different trap-types are used, which have much different capture probabilities, hence I include a trap-type covariate in the probability model. The issue being that if I lump all these sessions together I can no longer account for this variance in the data, and c-hat may be inflated?

A secondary post, in which Gary White responded:
viewtopic.php?f=1&t=2487&p=7838&hilit=chat+robust#p7838

"The only lack of fit that you can detect in the open models (Jolly-Seber, Pradel, Link-Barker) is from the recapture portion of the likelihood (i.e., the Cormack part of the CJS) -- something Ken Burnham has stressed. So, GOF tests from RELEASE are generally what we've recommended in the past"

Which further stresses my concern about lumping all my heterogeneous capture sessions.

I tried to use the RELEASE call in RMARK on my data as they were in the Pradel Robust phi and f, and I got results for one file, but not the other file, which had the same structure, minus the time intervals. I'm struggling with c-hat estimation and could really use some advice on how to proceed.

Code: Select all


##Import captures, specify column classes
SR<- read.csv("SR_Clip_DATA_Short.txt" ,
                 colClasses=c( Individual="factor",ch="character", Sex="factor", freq="numeric"),
                 strip.white=FALSE)

HWY<- read.csv("HWY_Clip_DATA_Short.txt" ,
               colClasses=c( Individual="factor",ch="character", Sex="factor", freq="numeric"),
               strip.white=FALSE)

##################STEP 1
##Process Data
SR.proc=process.data(SR, model="RDPdfHuggins", groups="Sex", begin.time=2006,time.intervals=c(0,0,1,0,0,1,0,0,1,0,0,1,0,0,1,0,0,1,0,0,0,1,0,0))

HWY.proc=process.data(HWY, model="RDPdfHuggins", groups="Sex", begin.time=2007,time.intervals=c(0,0,1,0,0,1,0,0,1,0,0,1,0,0,1,0,0,0,1,0,0))

####GOF using RELEASE
> release.gof(SR.proc)
RELEASE NORMAL TERMINATION
      Chi.square df      P
TEST2     4.4228  9 0.8814
TEST3     7.2771 11 0.7762
Total    11.6999 20 0.9260

> release.gof(HWY.proc)
RELEASE NORMAL TERMINATION
Error in (x3 + 4):length(out) : argument of length 0


[cooch]

There isn't a de facto standard, robust GOF test for robust design models of any flavor. Period. Or open population models in general. Period. If GOF testing isn't presented in an individual chapter of the MARK book, then no such test exists for that data type.

The best you can do are (i) consider the various suggestions you cited in your post, or, (ii) try assessing the sensitivity of your results to manual changes to c-hat. This is discussed in section 5.9.1 of Chapter 5 of the MARK book (particularly item 4): here.

[ctlamb]

Thanks, cooch. I appreciate the reply. I used the method you initially recommended - collapse secondary sessions and calculate median c hat. I did this and found no evidence for over dispersion in these data.. Median chat was 1.2, with 95% CIs being 0.95-1.45.

I will also take your second piece of advice and inflate chat and ensure model order is robust. Thanks.