Jeff (and Ecologgers):
I'm an ecologist with a specialty in remote sensing, and I wanted to
find out if you had checked the accuracy of the map products you are using?
A lot of people are using horribly inaccurate maps (many are released with
no or buried accuracies), or maps with thematic classes that are not
particularly related to the analyses you are trying to perform (e.g. is
"forest" one of your classes? How do they define forests? The feds define
it as > 25% tree cover, for instance, and the difference between 25% and
100% is significant, particularly in birds). Just an idea.
--j
--
Jonathan A. Greenberg, PhD
NRC Research Associate
NASA Ames Research Center
MS 242-4
Moffett Field, CA 94035-1000
Office: 650-604-5896
Cell: 415-794-5043
AIM: jgrn307
MSN: [EMAIL PROTECTED]
-----Original Message-----
From: Ecological Society of America: grants, jobs, news
[mailto:[EMAIL PROTECTED] On Behalf Of D. Mckenzie
Sent: Thursday, October 27, 2005 11:17 AM
To: [email protected]
Subject: Re: Help: great model but poor prediction
Jeff
I can think of a couple of things offhand.
1) I've had the same experience with very noisy model results when I used
significance tests to choose predictors. Many times a "significant"
model, particularly one with large sample size, still only accounts for a
small proportion of the deviance. AIC is very liberal, such that with
really big samples it always says "more predictors".
2) Sometimes GAMs are very informative but also very "wiggly", and
specifying a nonlinear parametric model, expecially with a discrete
response, can be very tricky.
Spatial autocorrelation in the response may or may not be a factor. You
could compute empirical variograms of the deviance residuals, or more
crudely, just plot the anomalies in geographic space, to see if there's
spatial dependence. If so, there are various methods to incorporate
spatial terms in your model.
HTH
Don
__________________________________________________
Don McKenzie, Fire and Landscape Ecologist
Fire and Environmental Research Applications (FERA)
Pacific Wildland Fire Sciences Lab, USDA Forest Service
Affiliate Assistant Professor
College of Forest Resources
University of Washington
[EMAIL PROTECTED]
[EMAIL PROTECTED]
(206)732-7824; fax (206)732-7801
__________________________________________________
On Thu, 27 Oct 2005, Jeffrey Stratford wrote:
> Greetings,
>
> I'm attempting to model the spatial distrubution of avian species
> richness (SR) across a gradient of distrubance. SR was divided into
> several groups (all, residents, short-, and long-distance migrants) and
> analyzed separately. I used point counts (0.8 km) apart to get the
> species richness data and I used CAPTURE to get estimated SR. I'm
> linking this to remote sensed landscape data using point count stops as
> centers. Landscape data come from 100, 200, 1000 m around points so
> there is considerable overlap in the explanatory variables around points
> BUT the response variable (SR) should be independent from each other.
> I used generalized additive models and used the graphics output
> available in SAS to subjectively select variables to use in subsequent
> Poisson regression models. All the variables that looked meaningful
> (95% CI departures from 0 at some point) were curvilinear so I used
> quadratic forms. I also ran a few models with variables that were not
> correlated (Spearman r < 0.25).
>
> I then ran Poisson or negative binomial models with a portion (80%
> randomly selected) and I used AIC to sort the models. The top models
> had "significant" values for the explanatory variables and were markedly
> different from null models (difference in AIC > 10). Model fit was also
> satisfactory (var/df ~ 1). I used the betas from the heuristic data set
> to get predicted values for a hold out sample.
>
> Plotting predicted vs. observed I get complete scatter!
>
> Is poor prediction a result of autocorrelation? Anything else to
> consider (other than the obvious problems of poor data)?
>
> Thanks,
>
> Jeff
>
> ****************************************
> Jeffrey A. Stratford, Ph.D.
> Postdoctoral Associate
> 331 Funchess Hall
> Department of Biological Sciences
> Auburn University
> Auburn, AL 36849
> 334-329-9198
> FAX 334-844-9234
> http://www.auburn.edu/~stratja
> ****************************************
>
