Basic questions about Bayesian Spatio-temporal Analysis-INLA

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|

Basic questions about Bayesian Spatio-temporal Analysis-INLA

Quiner, Claire
Greetings,
I am in the process of trying to teach myself how to perform a Bayesian spatio-temporal analysis using INLA in R. I am reading papers and following a number of tutorials but there is one, somewhat basic thing that I can't seem to figure out from my readings.
I have a series of raster stacks of a variety of climatic data, each layer of a stack represents the value from a week in a year. These data will become the prior distributions in my analysis, as I understand it. I was originally under the impression that INLA would read these raster files but I see that the program actually requires tabular data. I can easily transform these raster stacks, getting summary values over each county, by week. However, it is unclear to me what part of the analysis that I should do that in. I would like to prepare a correlation matrix to address multicollinearity, followed by PCA to further eliminate redundant variables. My understanding is that both of these analyses can be done on either raster files or from tabular data. For these preliminary analyses to Bayesian analysis, should I opt for spatial data? Also, how do I handle the temporal nature of this data, which will obviously be correlated, but still may be necessary to maintain?
Any advice would be appreciated.
Thank you,

Claire


        [[alternative HTML version deleted]]

_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
Reply | Threaded
Open this post in threaded view
|

Re: Basic questions about Bayesian Spatio-temporal Analysis-INLA

Virgilio Gómez-Rubio
Hi Claire,


Not sure what type of model or data you are trying to fit. If you have raster data, it would make sense to use them as covariates and not as priors. If you definitely want to fit a spatio-temporal model with INLA  you should check this book:

http://eu.wiley.com/WileyCDA/WileyTitle/productCd-1118326555.html

Also, please check these course materials that I prepared for the GEOSTAT 2017 summer school about spatial model fitting with INLA:

https://www.dropbox.com/s/lb9f7eagmmzou5k/materials.zip?dl=0


In a nutshell, the inla() function works similarly as the glm() or gam() functions: you define your model in a formula (which may include random effects) and use a data.frame to pass the data.

Hope this helps.

Virgilio
_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
Reply | Threaded
Open this post in threaded view
|

Re: Basic questions about Bayesian Spatio-temporal Analysis-INLA

Quiner, Claire
Hi Virgilio,
Thank you for your response and wealth of resources that you sent!
I had originally planned on using the raster files as covariates (went to great pains to get them!) but was swayed away from that approach at some point.

Here is what I have: a number raster stacks of various climatic and geophysical data across a country. Each stack is in the same resolution and each layer represents either the mean, min, max or median of that variable, for a given week. There is a layer for each week of the year, for each variable. As the outcome, I have weekly counts of an infectious disease, aggregated across each county (admin 2) of that country, in each week of the year (weeks that disease counts are aggregated into match the weeks that climatic data is assembled into).
I want my model to predict if there is an association of these climatic variables and the risk of the disease (disease count normalized by the number of residents in that county) and if this association various by different parts of the country...ie different climatic predictors for coastal vs inland. Further, I am interested in measuring if there is a time lag that is predictive of risk increase: i.e. x mm of precipitation predicts an increase in risk of disease 2 weeks later.  

Based on your resources, I believe that a space time geostatistical model would help me to answer these questions- although it is unclear to me if this would work since my outcome is aggregated counties and not points.
Any thoughts on this?
Thank you!
Claire.


-----Original Message-----
From: VIRGILIO GOMEZ RUBIO [mailto:[hidden email]]
Sent: Sunday, October 08, 2017 12:18 AM
To: Quiner, Claire
Cc: [hidden email]
Subject: Re: [R-sig-Geo] Basic questions about Bayesian Spatio-temporal Analysis-INLA

Hi Claire,


Not sure what type of model or data you are trying to fit. If you have raster data, it would make sense to use them as covariates and not as priors. If you definitely want to fit a spatio-temporal model with INLA  you should check this book:

http://eu.wiley.com/WileyCDA/WileyTitle/productCd-1118326555.html

Also, please check these course materials that I prepared for the GEOSTAT 2017 summer school about spatial model fitting with INLA:

https://www.dropbox.com/s/lb9f7eagmmzou5k/materials.zip?dl=0


In a nutshell, the inla() function works similarly as the glm() or gam() functions: you define your model in a formula (which may include random effects) and use a data.frame to pass the data.

Hope this helps.

Virgilio

_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
Reply | Threaded
Open this post in threaded view
|

Re: Basic questions about Bayesian Spatio-temporal Analysis-INLA

Virgilio Gómez-Rubio
Hi,


I think that these model is covered in the book by Michela and Marta. You can also check the book by Banerjee et al. on Bayesian spatial models. I think that this will give a better idea of the different models that you could use.

Best,
Virgilio

_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo