Re: What exactly are 'model residuals'?

[John Baumgartner]

Hi,

A search for "Maxent autocorrelation residuals" (without quotes) on scholar returns this paper: 

Spatial analysis improves species distribution modelling during range expansion

They state: "Spatial autocorrelation in model residuals (i.e. observed occur-rence–probability of occurrence given by MAXENT at each cell) was investigated using Moran's I coefficients (Dormann et al. 2007)."

Could be worth chasing up Dormann 2007. 

John

-- 
John Baumgartner

Sent with Sparrow

On Tuesday, 12 February 2013 at 7:18 PM, Maxent Newbie wrote:

Hello,

I might have some issues with spatial auto-correlation of my occurrence data. In such cases, some authors recommend to examine if residuals of Maxent models are spatial auto-correlated or not. 

However, after many, many hours of paper reading and google searching, I'm still not sure what exactly model residuals are in the context of Maxent. I figure they somehow express how HSI values vary at occurrence data points, is this right?

If yes, this could be easily calculated with a GIS or R... but what would be the next step in examining Model residuals? A visual assessment if residuals show a spatial pattern?

For reference, here (page 108) is about the only visual representation of model residuals I've found so far. 

I would be very thankful for any help 

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[Maxent Newbie]

John, thanks for your reply!

Unfortunately, I read both papers before and couldn't find an answer to my question. 

So far I've found 5-10 papers that examine spatial autocorrelations in model residuals of Maxent models, however without any explanation whatsoever what is considered as model residuals.

Actually, I think the definition given by de Marco is the most detailed so far: "observed occurrence–probability of occurrence given by MAXENT at each cell"... but unfortunately I've got no idea what this means exactly. Do you? :)

[Alvar Carranza]

Folks, a model residual = observed - predicted values. the problem
here is that observed values are our presence only points, a binary
variable. And that predicted values are the maxent output. It is not
trivial how to get a maxent model residuals. i.e. variation in
observed occurrence not accounted for in the model (may be logistic
regression?)
...question open.

Once we can assign a residual value for each cell (how much it
deviates for model expectations), we can easilly test for spatial
autocorrelation, using Moran or something....

goooood topic!!!

best

Alvar

--
Dr. Alvar Carranza

Profesor Adjunto
Polo de Desarrollo Universitario "Grupo de Investigación y formación
de recursos humanos en biodiversidad"
Centro Universitario Regional Este - CURE
Sede Maldonado
Universidad de la República Oriental del Uruguay

Investigador Asociado
Área Biodiversidad y Conservación
Museo Nacional de Historia Natural
Montevideo, Uruguay

http://www.wix.com/alvarcarranza/alvar-carranza

[Francisco Rodriguez Sanchez]

Hi,

I'm not at all sure, but my interpretation is that the residuals are calculated for the presence points only (observed occurrences). So for every presence you calculate the residual as 1 - Maxent predicted suitability at that cell.

I also wanted to bring this post to the list: some big news about Maxent http://methodsblog.wordpress.com/2013/02/20/some-big-news-about-maxent/. Quite relevant I think

Cheers,

Paco

El 20/02/2013 23:04, Maxent Newbie escribi�:

John, thanks for your reply!

Unfortunately, I read both papers before and couldn't find an answer to my question.�

So far I've found 5-10 papers that examine spatial autocorrelations in model residuals of Maxent models, however without any explanation whatsoever what is considered as model residuals.

Actually, I think the definition given by de Marco is the most detailed so far: "observed occurrence�probability of occurrence given by MAXENT at each cell"... but unfortunately I've got no idea what this means exactly. Do you? :)

On Wednesday, 20 February 2013 12:58:36 UTC+1, John Baumgartner wrote:

Hi,

A search for "Maxent autocorrelation residuals" (without quotes) on scholar returns this paper:�

Spatial analysis improves species distribution modelling during range expansion

They state: "Spatial autocorrelation in model residuals (i.e. observed occur-rence�probability of occurrence given by MAXENT at each cell) was investigated using Moran's I coefficients (Dormann et al. 2007)."

Could be worth chasing up Dormann 2007.�

John

--�

John Baumgartner

Sent with Sparrow

On Tuesday, 12 February 2013 at 7:18 PM, Maxent Newbie wrote:

Hello,

I might have some issues with spatial�auto-correlation�of my occurrence data. In such cases, some authors recommend to examine if residuals of Maxent models are spatial auto-correlated or not.�

However, after many, many hours of paper reading and google searching, I'm still not sure what exactly model residuals are in the context of Maxent. I figure they somehow express how HSI values vary at occurrence data points, is this right?

If yes, this could be easily calculated with a GIS or R... but what would be the next step in examining Model residuals? A visual assessment if residuals show a spatial pattern?

For reference,�here�(page 108) is about the only visual�representation of model residuals I've found so far.�

I would be very thankful for any help�

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.

�
�

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To post to this group, send email to max...@googlegroups.com.
Visit this group at http://groups.google.com/group/maxent?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.

�
�

-- 
Dr Francisco Rodriguez-Sanchez
Forest Ecology and Conservation Group
Department of Plant Sciences
University of Cambridge
Downing Street 
Cambridge CB2 3EA
United Kingdom
http://sites.google.com/site/rodriguezsanchezf

[Maxent Newbie]

Hello, 

Here’s some new input from Franklin & Miller “Mapping species distributions: spatial inference and prediction”:

“Residuals are checked in order to determine whether a potential SAC problem exists with respect to model formulation (Wagner & Fortin, 2005;Wintle & Bardos, 2006; Zhang et al., 2008; Elith & Leathwick, 2009), presuming that any SAC in the response variable may be “taken care of” by a spatially structured predictor variable, i.e., in the case of exogenous SAC. Mapping the residuals (plotting their values on a map and examining visually) can be particularly useful for interpreting potential causes of spatially structured errors (Dormann et al., 2007; Osborne et al., 2007; Zhang et al., 2008).”

Well, it really seems that residuals can indeed only be calculated at presence points, and that they express basically the Maxent prediction at each presence point. 

This can be easily done, but then what? 

> Once we can assign a residual value for each cell (how much it deviates for model expectations), we can easily test for spatial autocorrelation, using Moran or something.... 

We could use an interpolation method like kriging to estimate residuals at each cell of our study area, based on the Maxent prediction at each presence point. However, the quality of any form of spatial interpolation depends on the distribution of points; if they are clustered in some regions and absent in other regions, kriging would result in bad interpolations. 

Alternatively, we could calculate Moran's I based on the point pattern of species presence points, which could be considered as a marked point process. However, as far as I know, global statistics such as Moran's I assume a stationary/homogeneous point process, which is rather the exception in biological data used for presence-only SDM… but maybe I’m wrong in that point?

Anyway, for now I’ll limit my analysis to “plotting their values on a map and examining visually”. Thanks for the input and happy modelling! :)