Skip to content

spatial autocorrelation as random effect with, count data

3 messages · Highland Statistics Ltd, Rich Shepard

#
Try R-INLA.....much easier.


Alain




------------------------------

Message: 4
Date: Wed, 10 Jan 2018 16:59:02 -0400
From: Sima Usvyatsov <ghiaco at gmail.com>
To: R-sig-mixed-models at r-project.org
Subject: [R-sig-ME] spatial autocorrelation as random effect with
	count data
Message-ID:
	<CAFGTTqT8cDbDeKg2cikgdBkjAGz0spXakoEne7EmA28D3vVjgg at mail.gmail.com>
Content-Type: text/plain; charset="UTF-8"

Hello,

I am working on a spatially autocorrelated dataset with a negative binomial
(count) response variable. I have been using the glmmPQL approach (MASS),
but I seem to have a hard time fitting the fixed effects. I came across the
mention that one could build the spatial autocorrelation into a random
effect (https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/015364.html).


I've done some searching but could not find a straightforward example of
this practice. I have 20 sampling locations (sampled repeatedly to a 4,000
point dataset) and I know that there is spatial autocorrelation between
them (by looking at autocorrelation plots of a naive model). The 20 grid
points are clustered into 4 strata, and I am interested in the strata
effects (so would like to keep the strata as fixed).

How would I go about expressing the spatial autocorrelation in this setup?
In the future I'd like to explore GAMs for this application, but for now
I'm stuck with a GLM approach... I would love to be able to use glmer()
with a random effect that expresses spatial autocorrelation.

Here's a fake dataset.

library(MASS)

df <- data.frame(Loc = as.factor(rep(1:20, each = 5)), Lat = rep(rnorm(20,
30, 0.1), each = 5), Lon = rep(rnorm(20, -75, 1), each = 5), x =
rnegbin(100, 1, 1), Stratum = rep(1:5, each = 20))

Thank you so much!
#
On Wed, 10 Jan 2018, Highland Statistics Ltd wrote:

            
+1

   I highly recommend Blangiardo & Cameletti's book, "Spatial and
Spatio-Temporal Bayesian Models with R-Inla."

Rich
#
On Wed, 10 Jan 2018, Rich Shepard wrote:

            
Forgot to mention that they cover hierarchical models which look
appropriate for the subject of this thread.

Rich