Clamping Yet Again

[David Kidd]

I have spent some time reading the discussion on clamping and remain
confused about the recommended best practice for clamping and GUI
settings.

In Version 3.0, "The flag "fadebyclamping" causes the prediction at a
point to decrease when there is clamping there, with the amount of
decrease determined by the amount of clamping. The "dontextrapolate"
option causes the prediction to be zero wherever there is clamping."

I have generated my output using the GUI, but cannot find any
information as to what the default clamping settings are.

And then there is the advice to "subtract off the clamping values from
the logistic predictions. That's a reasonably conservative approach
to limit predictions when environmental conditions are outside the
training range."
http://groups.google.com/group/Maxent/browse_thread/thread/3dcc7eabbf29f5b2/a7058c193eef606f?lnk=gst&q=clamping#a7058c193eef606f

I find that the clamping value is often larger than the projection
value so I get negative values when I subtract the former from the
latter?

Are my projected values already "faded", and is fading just the
subtraction described.

Thanks
Dave

[Marcelo Lima]

Hi all,
I´m very happy to share with all of you this paper that will be in this month´s edition of the International Journal of Primatology. Thanks to Dr. Jean Boubli´s extensive field work, we modeled the distribution of four new world monkeys in the Amazon.

One of our goals was to show how Maxent can be used for conservation planning and further policy making by the Brazilian government.

Far from calling it perfect, we intend to continue to enhance the output the best as we can using other variables.

Hope you enjoy our work!

All the best
Marcelo and Jean
--
Marcelo Gonçalves de Lima
Biologist, PhD
www.flickr.com/cegonha
http://blogdocegonha.blogspot.com/
http://lattes.cnpq.br/1539538568877382

www.pequi.org.br
Pequi - Pesquisa e Conservação do Cerrado
SCLN 408 bloco E sala 201, Brasília-DF
Cep: 70.856-550
www.pequi.org.br

[asha_mc]

Hi Dave

I was lucky enough to get an answer from Steven Phillips about this
question finally.

The default option is neither fadebyclamping nor dontextrapolate, it
just leaves values as is, and Steven says to go with these.

The clamping info simply indicates where the species is being
predicted into novel environmental space. I think that in itself is an
important aspect of your models and shouldn't be ignored.

And of course, you must be cautious about interpretation in the
clamped areas (perhaps be skeptical of large areas of predicted
presence in particular).

Im not sure however of why values become negative when subtracted off
(i had small areas of the same thing, but didn't ask Steven about it)
but because im not using this method any more i'm not going to waste
my time with it, sorry.

Hope this helps,

Asha

On Apr 2, 11:47 pm, David Kidd <d...@nescent.org> wrote:
> I have spent some time reading the discussion on clamping and remain
> confused about the recommended best practice for clamping and GUI
> settings.
>
> In Version 3.0, "The flag "fadebyclamping" causes the prediction at a
> point to decrease when there is clamping there, with the amount of
> decrease determined by the amount of clamping. The "dontextrapolate"
> option causes the prediction to be zero wherever there is clamping."
>
> I have generated my output using the GUI, but cannot find any
> information as to what the default clamping settings are.
>
> And then there is the advice to "subtract off the clamping values from
> the logistic predictions.  That's a reasonably conservative approach
> to limit predictions when environmental conditions are outside the

> training range."http://groups.google.com/group/Maxent/browse_thread/thread/3dcc7eabbf...

[Steven Phillips]

Hi Dave and Asha,

I'm afraid there isn't really a "recommended best practice" in this
regard, and there are a number of reasonable and justifiable approaches
one might take. The most conservative would be not even trying to
predict to variables outside the range seen during training. The second
most conservative would be to zero out predictions whenever variables
are outside the training range - that's the "dontextrapolate" option.
Maxent's default is to treat variables outside the training range as if
they were at the edge of their training range, and you can see the
effect of that by noting the little horizontal parts at the left and
right edge of all response curves; this is what we call "clamping". A
fourth option, of intermediate conservativeness, is to subtract off from
the prediction the effect of the clamping, i.e., reduce the prediction
by the absolute difference between the prediction with and without
clamping (and setting the result at zero if the difference is negative);
that's the "fadebyclamping" idea.

An example of the clamping amount being greater than the prediction: if
the prediction without clamping is 0.8 and the prediction with clamping
is 0.3, then the amount of clamping is |0.8-0.3|=0.5, which is greater
than the prediction.

It's worth noting that the clamping picture should be thought of as
providing only a lower bound on uncertainty due to variables being
outside their training range. Areas with lots of clamping should be
treated with caution, but that doesn't mean that you don't need to worry
about areas with little or no clamping. For example, if you use only
threshold features clamping has no effect, so the clamping picture will
always be zero, but you should still be cautious about predicting into
climatic conditions with no analog in the training data.

-- Steven