Skip to content
Prev 16081 / 20628 Next

spatial autocorrelation as random effect with count data

I was introduced today (by Roger Bivand) to the glmmTMB package that looks
very exciting. As a co-author, I was wondering why you didn't suggest it -
is there a reason it's a no-go in my situation? From a super quick read,
and a very naive thinking, is this not equivalent to the gpmmPQL setup
below?

mod1 <- glmmTMB(Count ~ Stratum + SiteInStratum + ...other predictors +
# random variable
(1 | RoundStart) +
# autocorrelation
exp(site.Easting + site.Northing | RoundStart),

family = nbinom2, # or nbinom1 - I guess decide based on residuals?
correlation = corExp(form=~site.Easting + site.Northing + RoundStart)

where RoundStart is the time of starting sampling along the repeated, set,
20-point sampling grid, easting and northing are the 20 points' coords,
Stratum is the allocation of the 20 sampling points to 5 strata and
SiteInStratum is  the 1:4 allocation within stratum.

This is my current setup:

mod1 <- glmmPQL(Count ~ Stratum + SiteInStratum + ...other predictors,
random = ~ 1 |RoundStart,
family = quasipoisson,
correlation = corExp(form=~site.Easting + site.Northing + RoundStart)

I have INLA and GAMs on my to-do list for this year - sounds like really
helpful ways to go about things, I just haven't gotten there yet...

Thank you so much!
On Wed, Jan 10, 2018 at 5:34 PM, Ben Bolker <bbolker at gmail.com> wrote:
binomial
(MASS),
the
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/015364.html).
4,000
setup?
rep(rnorm(20,