R help - Jul 2005 - nlme, MASS and geoRglm for spatial autocorrelation?

Hi.

I'm trying to perform what should be a reasonably basic analysis of some
spatial presence/absence data but am somewhat overwhelmed by the options
available and could do with a helpful pointer. My researches so far
indicate that if my data were normal, I would simply use gls() (in nlme)
and one of the various corSpatial functions (eg. corSpher() to be
analagous to similar analysis in SAS) with form = ~ x+y (and a nugget if
appropriate). However, my data are binomial, so I need a different
approach. Using various packages I could define a mixed model (eg using
glmmPQL() in MASS) with similar correlation structure, but I seem to
need to define a random effect to use glmmPQL(), and I don't have any.
Could this requirement be switched off and still use the mixed model
approach? Alternatively, it may be possible to define the variance
appropriately in gls and use logits directly, but I'm not quite sure how
and suspect there's a more straight-forward alternative. Looking at
geoRglm suggests there may be solutions here, but it seems like it might
be overkill for what is, at first appearance at least, not such a
difficult problem. Maybe I'm just being statistically naive, but I think
I'm looking for a function somewhere between gls() and glmmPQL() and
would be grateful for any pointers.

Thanks very much,

Colin Beale

...
	[[alternative HTML version deleted]]

You seem to want to model spatially correlated bernoulli variables.
That's a difficult task, especially as these are bernoulli and not 
binomial(n>1).  With a much fuller description of the problem we may be 
able to help, but I at least have no idea of the aims of the analysis.

glmmPQL is designed for independent observations conditional on the 
random effects.

On Wed, 13 Jul 2005, Beale, Colin wrote:
> Hi.
>
> I'm trying to perform what should be a reasonably basic analysis of
some
> spatial presence/absence data but am somewhat overwhelmed by the options
> available and could do with a helpful pointer. My researches so far
> indicate that if my data were normal, I would simply use gls() (in nlme)
> and one of the various corSpatial functions (eg. corSpher() to be
> analagous to similar analysis in SAS) with form = ~ x+y (and a nugget if
> appropriate). However, my data are binomial, so I need a different
> approach. Using various packages I could define a mixed model (eg using
> glmmPQL() in MASS) with similar correlation structure, but I seem to
> need to define a random effect to use glmmPQL(), and I don't have any.
> Could this requirement be switched off and still use the mixed model
> approach? Alternatively, it may be possible to define the variance
> appropriately in gls and use logits directly, but I'm not quite sure
how
> and suspect there's a more straight-forward alternative. Looking at
> geoRglm suggests there may be solutions here, but it seems like it might
> be overkill for what is, at first appearance at least, not such a
> difficult problem. Maybe I'm just being statistically naive, but I
think
> I'm looking for a function somewhere between gls() and glmmPQL() and
> would be grateful for any pointers.
>
> Thanks very much,
>
> Colin Beale
>
> ...
> 	[[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
>
-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

My data are indeed bernoulli and not binomial, as I indicated. The
dataset consists of points (grid refs) that are either locations of
events (animals) or random points (with no animal present). For each
point I have a suite of environmental covariates describing the habitat
at this point. I was anticipating some sort of function that could run:

function(present ~ env1 + env2 + env3 + x + y, correlation corSpher(form=~x+y),
family = binomial)

where env1 to env3 are the habitat covariates, x & y the grid refs. If
my data were normal, I undertand I would use gls() with exactly this,
but drop the family requirement. As my data are bernoulli this is
clearly not possible, but I was hoping the analysis may be analagous?
The eventual aim is to firstly understand which environmental covariates
are important in determining presence and then to use habitat maps to
identify the areas expected to be most important.

Colin

-----Original Message-----
From: Prof Brian Ripley [mailto:ripley at stats.ox.ac.uk] 
Sent: 13 July 2005 11:30
To: Beale, Colin
Cc: r-help at stat.math.ethz.ch
Subject: Re: [R] nlme, MASS and geoRglm for spatial autocorrelation?

You seem to want to model spatially correlated bernoulli variables.
That's a difficult task, especially as these are bernoulli and not
binomial(n>1).  With a much fuller description of the problem we may be
able to help, but I at least have no idea of the aims of the analysis.

glmmPQL is designed for independent observations conditional on the
random effects.

On Wed, 13 Jul 2005, Beale, Colin wrote:
> Hi.
>
> I'm trying to perform what should be a reasonably basic analysis of 
> some spatial presence/absence data but am somewhat overwhelmed by the 
> options available and could do with a helpful pointer. My researches 
> so far indicate that if my data were normal, I would simply use gls() 
> (in nlme) and one of the various corSpatial functions (eg. corSpher() 
> to be analagous to similar analysis in SAS) with form = ~ x+y (and a 
> nugget if appropriate). However, my data are binomial, so I need a 
> different approach. Using various packages I could define a mixed 
> model (eg using
> glmmPQL() in MASS) with similar correlation structure, but I seem to 
> need to define a random effect to use glmmPQL(), and I don't have any.
> Could this requirement be switched off and still use the mixed model 
> approach? Alternatively, it may be possible to define the variance 
> appropriately in gls and use logits directly, but I'm not quite sure 
> how and suspect there's a more straight-forward alternative. Looking 
> at geoRglm suggests there may be solutions here, but it seems like it 
> might be overkill for what is, at first appearance at least, not such 
> a difficult problem. Maybe I'm just being statistically naive, but I 
> think I'm looking for a function somewhere between gls() and glmmPQL()
> and would be grateful for any pointers.
>
> Thanks very much,
>
> Colin Beale
>
...

> -----Original Message-----
> From: r-help-bounces at stat.math.ethz.ch
> [mailto:r-help-bounces at stat.math.ethz.ch]On Behalf Of Beale, Colin
> Sent: 13 July 2005 10:15
> To: Prof Brian Ripley
> Cc: r-help at stat.math.ethz.ch
> Subject: Re: [R] nlme, MASS and geoRglm for spatial autocorrelation?
> 
> 
> My data are indeed bernoulli and not binomial, as I indicated. The
> dataset consists of points (grid refs) that are either locations of
> events (animals) or random points (with no animal present). For each
> point I have a suite of environmental covariates describing 
> the habitat at this point. I was anticipating some sort of function that 
> could run:
> 
> function(present ~ env1 + env2 + env3 + x + y, correlation >
corSpher(form=~x+y), family = binomial)
> 
> where env1 to env3 are the habitat covariates, x & y the grid refs. If
> my data were normal, I undertand I would use gls() with exactly this,
> but drop the family requirement. As my data are bernoulli this is
> clearly not possible, but I was hoping the analysis may be analagous?
> The eventual aim is to firstly understand which environmental 
> covariates are important in determining presence and then to use habitat
maps to
> identify the areas expected to be most important.
This could be done with geoRglm. I did something similar last week, but without
covariates, only the spatial coordinates (i.e. my spatial process had
expectation
equal to a constant). If you are willing to sacrifice some spatial resolution
you
can create cells in your spatial data (say 100 m x 100 m) and in each cell count
the number of successes in observing your spatial process and the number of
trials.
This will be a binomial problem and it seems to me to be the spatial equivalent
of
logistic regression where the predictor continuous variable is structured in
bins
and then events are counted in those bins. You can move to the R-sig-geo list
if you have questions about geoRglm
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
Btw, this can also be done in SAS using the glimmix macro.
Ruben