hello,
here is chapter 3 (I hope the final one ) of my attempt to make a spatial
model of fallow deer distribution using line transect data (cap 4 of
advanced distance sampling book). To give a framework to the reader I've
reported below my previous questions.
I order to check if the results were reasonable, I compared them with the
results of CDS using the same LT data.
The CDS yielded a cluster density of 15.161 per km2 (CV=9.59) and an animal
density of 41.068 (CV=11.3)
The spatial model yielded an average cluster density of 11.1 and an animal
density of 31.6.
This result might be reasonable because probably the dense woods where
fallow deer are less likely to be found are also less likely to be sampled
by the observer, despite we tried to distribute transects at random.
The question now is how to compute a CV value for the spatial model (with a
special regards to cluster density)
Is the problem solved of I have to discuss this result in a purely
qualitative way without making any statistical comparison between the CDS
and "spatial" estimates?
thanks
Stefano
Stefano Focardi
Istituto Nazionale per la Fauna Selvatica
via Ca' Fornacetta, 9
40064 Ozzano dell'Emilia
Italy
tel.
++390516512223
fax
++39051796628
e-mai:l
stefano...@infs.it
============= 2 chapter
=============================================================
> > with reference to my previous question dated 29/11/04 I was able to write
a software to apply a methods described in chapter 4 (section 4.3) of the
book Advanced distance sampling by Buckland et al.
> >
> > and that system seems working appropriately in the sense that the model
predicts the distribution of radiocollared fallow deer in the area.
> >
> > However some doubts remain.
> >
> > 1. I've used 4 surveys to estimate local fallow deer density. To do that I
used the ESW estimated by pooling the data of the 4 surveys and the waiting
distance I used all points of the 4 surveys. By instance
> >
> > A------B-----A------------C (translated transect of fig 4.1)
> > x1.....x2...x3...........x4 coordinates
> >
> > I computed l_i=x2-x1 between A and B, between B and A (x3-x2) and so on
where A,B,C are detection points from the first, second,.. survey.
> >
> > I argued that the density obtained is the "average density" of the 4
surveys but I worry whether or I've to divide by 4 (or for some other value)
to obtain it?
Yes, by pooling, you're increased the encounter rate fourfold, so you will
need to divide the density estimate by 4.
> > 2. Fallow deer occurs in clusters. For each point of the grid used to
extrapolate the model, I used the following equations derived by eq. (4.1)
> >
> > D_hat=s_hat/2*mu_hat*l_hat
> >
> > where s_hat is the average cluster size per habitat (we detected no size
bias). Is this correct or there are smarter way to modelise s_hat?
That should be fine. You could also fit a spatial model for s - see p68
of the advanced distance sampling book.
=====================1 chapter
==========================================================
Stefano, we hope to have the simple count method in Distance within the
next year. More sophisticated methods, including methods that model
autocorrelation, are a bit further down the line, but should be in
Distance in under 3 years. Meantime:
> > In wish to perform a spatial distance sampling analysis in order to
> > develop a spatially explicit map for density in a roe deer population of
> > conservation interest. The aim is to identify high density areas where
> > to adopt "expensive" conservation measures.
> >
> > During the last years a number of classical distance sampling surveys
> > were made through the study area, trying to estimate population size but
> > where the exact location of the animals was also recorded
> >
> > I've studied cap 4 of "Advanced distance sampling" book for spatial
> > line transects method, but before developing a procedure in SAS to do
> > the job I'd like to be sure about some points which I've not well
> > understood.
> >
> > all OK until eq 4.4 . I've not understood what f((u1....) is.
This is the probability density function of detection locations (u,v),
evaluated at each location that you detected an animal, and the product
taken across the detected animals.
> > section 4.3
> >
> > the GAM has to be fitted to spatial covariates z. I presume that in the
> > most simple case one should use the x coordinates, or both the x and y
> > coordinates?
Yes, in the simplest case, z would probably be x and y. The x and y
distances would then be proxies for whatever it is that causes variation
in animal density within your survey region.
> > pag 56. is mu_hat the effective strip width estimated from CDS on the
> > whole study area?
Yes. We're working on methods to integrate estimation of mu with the
spatial modelling, but we're not there yet.
> > it is not clear to me how to adjust the waiting distance l . I do not
> > understand the difference between D_hat (x,0) and D_hat (xi,0).
> >
> > I presume that D_hat (xi,0) is the average density estimated by CDS
> > but what I missed what really means D_hat (xi,0)is the value obtained
> > from eq.
> >
> > (2mu_hat l)^-1
Having fitted the density surface D^(x,y), then D^(x_i,0) is the value of
this function at x=x_i (the x value of detection i) and y=0. D^(x,0) is
still this function evaluated at y=0, but now with no value substituted in
for x. Hence the 2-dimensional surface is now a one-dimensional function,
which we integrate between x=x_{i-1} and x=x_i.
> > I imagine that what I'd do is to compute the mean density for the whole
> > study area using the values estimated for each segment and "oblige" the
> > program to set that CDS density and spatial DS density be the same over
> > the whole study area.
That is not what we intend, as the spatial method can give you a more
precise estimate of overall density. If you rescale it to match the CDS
estimate, you will lose the extra precision.
> > Finally, what can be a good criterion to establish that iteration has
> > converged? Maybe that
> >
> > absolute value (mean of D_hat (xi,0) - CDS density) <0.0001?
It depends on the units you're using to estimate density, but it converges
quickly, so any small value is likely to work well. One possibility might
be that l'_i in one iteration differs from the previous iteration by less
than 0.0001*l'_i for each i.
--
distance-sampling list home page:
http://www.jiscmail.ac.uk/lists/distance-sampling.html
To unsubscribe, go to the home page and click "Join or leave the list",
or email
jisc...@jiscmail.ac.uk with the following message body:
leave distance-sampling
Before posting to the list, read the list guidelines at
http://www.jiscmail.ac.uk/files/distance-sampling/
distance-sampling.welcome
Any questions, email the list owner at
distance-sam...@jiscmail.ac.uk
--