None of the models pass GOF

176 views
Skip to first unread message

Ana

unread,
Oct 18, 2017, 7:13:11 PM10/18/17
to distance-sampling
Hi everyone,
I'm trying to estimate density of Thrushes is south America. I conducted transects in forest and coffee during three years in three study sites. So, I have three study sites for each habitat for three years.
I tried analyzing all my data together but after several attempts I decided running separated analysis for forest and for coffee.

Right now I'm analyzing the forest data. I have 153 transects and 1125 birds detected over the three years in the three sites. I'm interested in the density in forest during each year (I'll do the same for the coffee data). Most of my detections were by ear since Thrushes are hard to see (in both habitats).

After looking at the histograms I truncated the data at 36m. The data is spiked every 5 meters (5,10,15,20,25,35) so I created these binds: 0-6,6-12, 12-18, 18-24, 24-30, 30-36. I run models only with key functions (but adjustments were auto selected), and with site as covariate assuming that detection changes with site. However, when I run the GOF tests (chi squared) it's significant for all my models. For one of my models I couldn't run the test because did not have enough df (message "No degrees of freedom for test"). I have tried different binning and truncation and in all cases the models don't pass the GOF or I get the degrees of freedom message. 

Also, I thought that by having year (2013, 2014, 2015) in the column "Region.Label" I'd get the density estimates with the function summary(bestmodel). Unfortunately it doesn't seem to work that way and I'm quite lost.

I'll appreciate your advice with these data.

Many thanks,
Ana


Eric Rexstad

unread,
Nov 8, 2017, 9:25:49 AM11/8/17
to Ana, distance-sampling

Ana

You cover lots of topics in this email; hard to know where to begin.

It is unfortunate your data are spiked and doubly unfortunate that this was not diagnosed until the end of the three years of data collection.  Nevertheless, the purpose of binning is to get the detection into the correct "bin" if it was not in the correct "exact" distance because of rounding.  To get data into the correct bin, it seems the "preferred" distance (multiple of 5) should be in the centre of the bin, not at the edge; so I would use cutpoints different to the ones you have used.  I might also believe (but you know better because the data are yours) that rounding was less of a problem at small distances than at large distances--suggesting that you have narrower bins at small distances--there is no reason that the bins need to be of the same width: perhaps 0-1, 1-2, 2-3, 3-4, 4-6,...

Because you have binned the data, you do not have the Kolomogorov-Smirnov or Cramer von-Mises tests at your disposal and you must use the chi-square test (which is sensitive to the location of the cutpoints for binning).  The degrees of freedom of the chi-square test is u-q-1 where u is the number of bins (6 in the example you provide) and q is the number of parameters in your model.  If you fit a hazard rate model with 2 parameters and include site as a factor covariate (adding 2 more parameters to be estimated), I think you will end up with 6-4-1=1 degrees of freedom left for your test.  Should an adjustment term appear in the fitted model you will see the "no degrees of freedom for test" message.

If you are in this situation, then you are in the uncomfortable situation of making inference about thrush density based on a detection function model inconsistent with the data you observed.  You may say this lack of fit was caused by rounding of data in the field, but your inference will be somewhat tarnished.

I believe to get year-specific estimates you would need to include Region.Label in your model; give that a try.

--
You received this message because you are subscribed to the Google Groups "distance-sampling" group.
To unsubscribe from this group and stop receiving emails from it, send an email to distance-sampl...@googlegroups.com.
To post to this group, send email to distance...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/a282500c-c0bd-49d9-be87-d3b9b14afa0d%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

-- 
Eric Rexstad
Research Unit for Wildlife Population Assessment
Centre for Research into Ecological and Environmental Modelling
University of St. Andrews
St. Andrews Scotland KY16 9LZ
+44 (0)1334 461833
The University of St Andrews is a charity registered in Scotland : No SC013532

Ana

unread,
Apr 7, 2018, 8:34:10 PM4/7/18
to distance-sampling
Hi Eric,
Thanks so much for your response, it was very helpful.
At the end I pooled the data from the three years and stratified by habitat, so I got the density in Forest and Coffee.
The best cutpoints for the bins in my data were 0,10,18,24,29,38.
For some models I got a warning message that "the number of adjustment parameters allowed were reduced due to my limited number of intervals" but in all cases the LRT showed that no more than one adjustment term was needed. When looking at the chi square results, all models had a good fit.   
Thanks again for your time and advice. 
Cheers,
Ana 
Reply all
Reply to author
Forward
0 new messages