spatial autocorrelation in NPMR

[esgarbanzo]

Bruce et al.,
I'm modeling spawning site selection in river systems and my models in
Hyperniche are showing promise, but some of the predictors are
spatially autocorrelated. Specifically, depth is a strong predictor of
spawning, but variograms of residuals from a regression of redds~depth
suggest that there is spatial autocorrelation. Is there an approach
that you would recommend for dealing with this issue within
Hyperniche?

[Bruce McCune]

Spatial autocorrelation is a potential problem for NPMR models (and
for other regression methods) if you are emphasizing hypothesis
testing -- e.g. randomization tests. But I don't see it as a serious
problem if you are using the model for more descriptive or
exploratory purposes. I would be interested if others have different
views on this.

If you have an abundance of sites in your data, such that you can
afford to drop some, you might be able to reduce the spatial
autocorrelation by subsampling your data, for example not taking
sites within a certain distance of each other.

-Bruce

[Heather Lintz]

Another way to remove spatial autocorrelation would be to fit a model of
your predictor to spatial coordinates such as longitude and latitude and
possibly elevation in HyperNiche. Then take the residuals of the response
variable with respect to the predicted values from that model and use those
as the predictor. However, if what you care about biologically involves a
spatially auto-correlated process, then you may just want to accept that
spatial autocorrelation in the data.

[Heather Lintz]

Another way to remove spatial autocorrelation would be to fit a model of

your predictor (which becomes a response here) to spatial coordinates such

[esgarbanzo]

Thanks for your thoughts. Our study sites are continuous throughout
100 km of a river system, which explains the presence of
autocorrelation. We'd prefer to avoid dropping sites because they are
valuable and the continuous nature of our study is part of what makes
it special. We may need to accept the autocorrelation or try dropping
our most autocorrelated variable and see what Hyperniche replaces it
with in the models. Does HyperNiche produce residuals as an output?



On Feb 18, 10:28 am, "Heather Lintz" <hli...@coas.oregonstate.edu>
wrote:

[Bruce McCune]

Yes, it does. Highlight your desired model, then select Graph |
Response Points | and the residuals (along with estimates, etc.) will
automatically be written into the result file. You can also graph
them if you wish with the response points menu item.

You can easily import these into a spreadsheet by selecting the table
in the output, pasting it into a spreadsheet, then selecting Data |
Text to columns | Fixed width.

-Bruce

[esgarbanzo]

Thanks again for all of your help. I have now looked at variograms
from NPMR residuals of 1 predictor models and they don't show spatial
autocorrelation like residuals from simple linear models did. Is this
due to the underlying model structure?

[Heather Lintz]

Wow, that is VERY INTERESING.
Does that mean that the areas of state space (or features of the response
pattern) that NPMR was able to model and linear regression wasn't are also
tied to specific spatial domains?

[Heather Lintz]

Also, do your predictors vary with spatial coordinates?

-----Original Message-----
From: hyper...@googlegroups.com [mailto:hyper...@googlegroups.com] On

Behalf Of esgarbanzo
Sent: Thursday, February 23, 2012 10:48 AM
To: HyperNiche and NPMR

[esgarbanzo]

Heather,
We have one predictor that had shown some spatial autocorrelation that
was prominent in the NPMR models. Because we're working in a river
system, we use river km as our spatial reference. The rest of our NPMR
model predictors do not show spatial dependence according to
variograms from regression residuals.

I investigated a few more NPMR models at different grains sizes. At
broader scales there still appears to be spatial autocorrelation,
though not as strong and only if I manipulate lag distance and lag
class interval to find it.


On Feb 23, 1:21 pm, "Heather Lintz" <hli...@coas.oregonstate.edu>
wrote:

[Bruce McCune]

Interesting. I don't know the answer to this. If one of the
predictors is something like "river km", then I could see it removing
the autocorrelation, otherwise, it's not obvrious to me.

-Bruce

[esgarbanzo]

Could it be that the relationship is non-linear and is better
described by NPMR than by linear regression, thus making the residuals
not show spatial autocorrelation?

We did not use river km as a predictor.

Thanks!

[Bruce McCune]

Yes, that is quite possible. If one or more important predictors are
autocorrelated, then fitting the response to them could leave you
with non-autocorrelated residuals.
-Bruce

[Peter Nelson]

Heath',
Sorry I didn't get to returning your call. I am in the final throes of
preparing for field season. I will be in transit all day tomorrow so
it would be a good day to talk. My first flight is 245 Pacific Time.
We could talk before that or in the evening.
Looking forward to hearing about what your scheming. CHile is rad.
-Ptr

[Peter Nelson]

Whoops. Sent this to the wrong address. Please disregard.

Peter