unable to remove spatial autocorrelation from a binomial gam

Hello Olga

Thanks a lot for your response. It is very helpful.

Yes, my data is presence/absence because I'm observing the occurrence of
bear damaging apiaries in a particular region. Since there is a
compensation system that is running for a long time we can assume that
almost all damage is included in the database. So perhaps a few absences
could be presences (a beekeeper not claiming the damage) but I'm
pretty sure that it'd be marginal. I have also read what you say about
environmental data not being always an issue that should be removed from a
model. But in some books and articles, it is written that properly
accounting for autocorrelation is necessary for obtaining reliable
statistical inference (

http://highstat.com/index.php/mixed-effects-models-and-extensions-in-ecology-with-r
 see also here
https://esajournals.onlinelibrary.wiley.com/doi/10.1002/ecy.1674 ). What
should I follow? So far my approach is more conservative and I try to
remove since I imagine reviewers asking me to do so.

I knew about the possibility of subsampling to avoid autocorrelation but
I've read that it's not the best solution. That's why I was trying to use
correlation structures. I have got the advice to use the function gamm that
allow such correlations and check if the model fit is more ore less similar
to the one of a gam model. I am in the middle of that now and waiting for
the gamm to finish as it is computationally costly (it may take a few
days).
I didn't know about the package that you recommended so I will take a
look at it. Maybe the weightCases() function will be a good solution to my
problem.

Thank you so much once again for your help.

All the best,
Carlos

On Fri, 10 Apr 2020 at 12:04, Olga Boet <formigareina at gmail.com> wrote:

Hi Carlos,

Excuse me, I don't sure that I can help you, I know little about GAM. I
don?t understand your script and variogram, I work different. I hope
someone else gives you a better answer than mine. But if it can help,

here

are some considerations.

Spatial data is often correlated, but it must be evaluated if it is a
problem or not. For exemple, some species are distributed by stains as
frogs, fihes or some plants species (this correlation should not be
eliminated).

I think the smooothing function in GAM is to smooth the curves, that is,
it softens (less abrupt) the effect of environmental variables (not the
coordinates, since the coordinates are not environmental variables in a
spatial model).

However, in Dimo package, there are two interesting functions: balancing
weights function and thinning function.

Balance function is weightCases(), and it is used when the background is
very large with respect to the number of presences. So that the values of
the variables in the presence points have more weight in the model

despite

the lower number.

Thinning function removes points that are too close to each other (or in

a

space where variable data is not available). It is used when there are
points that are too clustered as a result of sampling (but it does not
correspond to the actual distribution). In this function you can

determine

the minimum distance between the points.

thinning() is from package spThin (URL:
https://cran.r-project.org/web/packages/spThin)


Finally, are your data really presence/absence data? did you go to at

3355

cells and detect presence/absence of the species? spatial models are
different if we have absences, pseudoabsences or backround. The type of
absence data is important for choosing a model.


I'm sorry I couldn't answer your questions



Kind regards,


Olga Boet
Documentalista de la col?lecci? de cordats. CMCNB
*Myrmex*


Missatge de Carlos Bautista <carlosbautistaleon at gmail.com> del dia dj.,

9

d?abr. 2020 a les 17:52:

Dear list members,

I am using gam (from mgcv package in R) to model presence/absence data

in

3355 cells of 1x1km (151 presences and 3204 absences). Even though I
include a smooth with the spatial locations in the model to address the
spatial dependence in my data, the results from a variogram show spatial
autocorrelation in the residuals of my gam (range=6000 meters). Since I

am

modelling a binary response, using a gamm with a correlation structure

is

not advisable because it "performs poorly with binary data", neither

gamm4

because (although is supposed to be appropriate for binary data) it has
"no
facility for nlme style correlation structures".

The alternative I have found is to fit my model using the function magic
from the same mgcv package. Because I found no examples of how to use
magic
for spatially correlated data I have adapted the ?magic example for
temporally correlated data. The results of the output change the
coefficients of the model but do not remove the spatial autocorrelation
and
the smooth plots show the same effect.
You can find find the output from my models and figures of the

variograms

and plots of the smooth effects in the following link

https://stackoverflow.com/questions/61110762/gam-with-binomial-distribution-and-with-spatial-autocorrelation-in-r

Could someone tell me if there is something wrong in my script? Does
anyone
know another alternative to remove the residuals' spatial

autocorrelation

from a binomial gam?

Thank you very much.
Kind regards,
Carlos
--
Carlos Bautista
Institute of Nature Conservation
Polish Academy of Sciences
Mickiewicza 33
31-120 Krakow, Poland
www.carpathianbear.pl
www.iop.krakow.pl

        [[alternative HTML version deleted]]

--
Carlos Bautista
Institute of Nature Conservation
Polish Academy of Sciences
Mickiewicza 33
31-120 Krakow, Poland
www.carpathianbear.pl
www.iop.krakow.pl

        [[alternative HTML version deleted]]