[R-sig-Geo] Multivariate spatio-temporal modelling

[Nicola Batchelor]

Hi Andy,

I was just reading something you wrote about spBayes and was thinking about
emailing you for advice!

The data I currently have is 1000's of village locations with counts of
human disease (and population data, so I could calculate rates/prevalence)
over about 20 years - I'll probably take a sample of these villages for use
in the analysis rather than use all of them as there are probably too many.
I also have a number of locations (far less - probably about 30) with
livestock (reservoir for the parasite) prevalence data from various time
points. So the human data is more flexible - could be used as counts or
rates, but the animal data is a proportion, so would have to be used as a
binomial variable which might constrain the analysis to a binomial GLM.

My ideal scenario would be the ability to create a multivariate
spatio-temporal model with the primary aim of predicting human disease
(spatially, and also perhaps forecasting if possible) based on
environmental/climatic covariates, and also with the input of animal
prevalence information.  I'm only at the grant writing stage at the moment,
so wanted to check this is something which I can do before committing myself
to it!

My supervisor is very keen on the temporal aspect of the work. Personally
I'm more interested in incorporating animal sampling data to improve the
predictions as there is a big problem of under-reporting for this disease
(particularly in areas far from health centres), and I hope that the use of
animal sampling data could improve predictions.  I was hoping that we'd be
able to combine these two aims in one model...is that not possible? Also, do
you know what the limitations of the model fitting is in terms of number of
observation locations?

Sorry if my questions are a bit niave, but I have no experience of
spatio-temporal modelling and am just trying to get my head around what is
possible for this grant!

Thanks very much for your input,
Nicola




-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.


-----Original Message-----
From: Andrew Finley [mailto:finleya at msu.edu] 
Sent: 24 March 2010 18:13
To: N.A.Batchelor at sms.ed.ac.uk
Subject: [R-sig-Geo] Multivariate spatio-temporal modelling

Hi Nicola,
Sorry to answer you off the list. One of my grad students forwarded your 
posting to me.

So based on your description you should be able to use spMvGLM to fit 
this model for any given time point (in its current state this function 
will not be able to fit a full space-time model). So at a given location 
you have disease counts for humans and animals (assuming humans != 
animals ;-)? Or do you have disease rates. If so would you want to fit a 
multivariate spatial binomial regression? Also how many data points do 
you have?

Feel free to answer back on the list (I'm now an official subscriber).
Kind regards-
Andy

-- 
Andrew Finley, PhD
Natural Resources Building
Michigan State University
East Lansing, MI 48824-1222
Phone: 517-432-7219
Fax: 517-432-1143
web: http://blue.for.msu.edu





Dear list.

I've completed a spatial GLM analysis of a dataset with one outcome variable
(binomial) using geoRglm.  I now want to go on to extend this analysis in 2
ways: 1) incorporating a temporal element and 2) using a second (related)
outcome variable.

As far as I am aware, geoRglm doesn't have the functionality for
spatio-temporal modelling, or for multivariate modelling.  Is this correct?
>From a brief look around on the web it seems like spBayes could be the thing
for the job.

Can anyone give me a wee bit of advice as to whether spBayes could
potentially be right for what I want to do (or if there is anything else
which I could use)? Basically, I want to carry out spatio-temporal modelling
of joint distribution data for 2 related disease measures (will be disease
prevalence or counts in humans and animals). The outputs I want (if
possible) are estimates of covariate effects on both human and animal
disease, the relationship between human and animal disease and a
spatio-temporal prediction. 

Please forgive me if I use the wrong technical terms as my first analysis
with geoRglm was my first ever introduction to geostatistics and spatial
modelling...perhaps I am being overambitious!

Many thanks in advance,

Nicola Batchelor
University of Edinburgh and University of Southampton