Estimating covariate-specific and overall detection probabilities - distsamp/gdistsamp

843 views
Skip to first unread message

Andrew Cox

unread,
May 30, 2013, 2:51:54 PM5/30/13
to unma...@googlegroups.com
Hello fellow unmarked fans,
  Myself and others have previously asked questions on this matter but after reading through the posts and vignettes I'd like to get some final clarification on estimating detection probabilities before I go off and publish something that is not done correctly.  My questions are these:

1.  Is it okay to calculate overall detection probabilities using getP(model_name) when I have detection covariates?  My thought is to use this:

 mean(rowSums(getP(model_name)))

to calculate the mean value of each point's detection probability (each of which is the sum of detection probabilities for each distance bin)

2. To estimate covariate effects on detection rates, Richard has suggested we use the predict function to get values of sigma and then plot them using code such as this:

plot(function(x) gxhn(x, sigma=19.1), 0, 80, xlab="Distance (m)",
+      ylab="Detection prob", cex.lab=0.7,
+      cex.axis=0.7, las=1)

where 19.1 is a predicted value and 0-80 is my range of distances.

If I am using a hazard key function it will look like this:

plot(function(x) gxhaz(x, shape=19.1, scale=exp(coef(model_name,type="scale"))), 0, 80, xlab="Distance (m)",
+      ylab="Detection prob", cex.lab=0.7,
+      cex.axis=0.7, las=1)

Note that I am exponentiating the scale parameter.
Does that all look correct?

Thanks very much!

Andrew

Sarah Kendrick

unread,
Jun 7, 2013, 11:11:38 AM6/7/13
to unma...@googlegroups.com
Great questions!  I'm working through the same issues; any insight on these would be really helpful and appreciated.

Thanks in advance, unmarked group!
Sarah

Richard Chandler

unread,
Jun 27, 2013, 3:25:56 PM6/27/13
to Unmarked package
Hi Andrew,

If detection probability is a function of covariates, then I'm not sure what "overall detection probability" you are after. If p is a function of something like "date", then you could compute p for ever day in the season and then take the mean. 

Your plot command looks right to me.

Richard


--
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.
 
 

Andrew Cox

unread,
Jul 3, 2013, 3:58:47 PM7/3/13
to unma...@googlegroups.com


Hi Richard,
  I think what I am looking for is analogous to the global detection probability Program Distance provides as part of its standard output. You mentioned previously that you don't have much experience with the program, but it somehow provides a single overall detection probability for a given model.

  Using unmarked, I've estimated densities for four species.  The raw counts for the species range from ~200 to ~700 but the density estimates are fairly similar because of differing detection probabilities for each species.  These results make total sense (e.g., one is a skulking, relatively quiet species, whereas another you can see from a mile away) but I want to frame the results within the context of detection probabilities.  My first thought was to use "predict" and hold all detection covariates at their mean values.  This is somewhat akin to your suggestion re: date, but I have observer as a covariate and that cannot be averaged.  So then I thought that perhaps I could treat observer as a dummy variable but model convergence was really problematic when I tried that.  Finally, I thought that because getP provides point-level estimates of detection probabilities (i.e., detection probabilities dependent upon the particular combination of detection covariates at each point - I think), perhaps I could just average those and come up with a global detection probability.
  
Hopefully this clarifies why I keep referring to "overall" detection probabilities.  Any further thoughts you have are appreciated.
  
Andrew  
Reply all
Reply to author
Forward
0 new messages