glmmTMB negbinom not working with spatial autocorrelation

Did you really try to fit a model with

+ ziformula = ~.

in the formula? Or do you mean that you fitted a zero-inflated negative
binomial (i.e.

  y ~ ....,  ziformula=~., family=nbinom1, ... ) ?

 This is surprising behaviour.

 * Does it work in a clean R session?
 * Does your model work with any alternative families (e.g. nbinom2) or
simplifications (e.g. leave out the zero-inflation, or make ziformula =
~1 ) ?
 * Can you post a minimal reproducible example? The code below works
(sort of -- convergence problems with the second model, but these are
related to the data and don't sound like what you described).


===
set.seed(101)
library(glmmTMB)
library(lme4)
n <- 400
dd <- expand.grid(population=letters[1:20],
                  species=LETTERS[1:10],
                  rep=seq(n/(10*20)))
dd <- transform(dd,
                b1=rnorm(n),
                b2=rnorm(n),
                x=rnorm(n),
                y=rnorm(n))
dd$eta <- simulate(~poly(b1,2)+poly(b2,2)+
                       (1|population/species),
                   family=gaussian,
                   newdata=dd,
                   newparam=list(beta=rep(1,5),
                                 theta=rep(1,2),
                                 sigma=1))[[1]]
dd$resp <- rnbinom(n, mu=exp(dd$eta), size=1)
dd <- transform(dd,
                pos=numFactor(x,y),
                group=factor(1))
model1 <- glmmTMB(resp ~ poly(b1,2) + poly(b2,2) + exp(pos + 0|group ) +
            (1|population/species), data = dd,
        family=poisson,
        verbose=TRUE)

model2 <- update(model1, family=nbinom1)

Hello all,

I had been running a mixed model with poisson distribution of the following type, with a spatial autocorrelation term, which works fine:

Y(count data) ~ x1 + square(x1) + x2 + square(x2) + exp( ) + (1|population/species)

I realized that my dataset has a lot of small values (mostly 1 and 2) and some large values, so that the data is highly skewed and over dispersed. So I tried to run the following negbinom1 model:

Y(count data) ~ x1 + square(x1) + x2 + square(x2) + exp( ) + (1|population/species) + ziformula = ~.

This time the model doesn?t run and says it cannot find one of the independent variables in the dataset. If I remove that variable from the model then it says so for another variable and so on. If I remove the factor for spatial autocorrelation, the model seems to work fine. Can anyone tell me what?s happening and if what I am doing is appropriate for a highly skewed and over dispersed dataset?

Thank you
Udita Bansal
Project Assistant
Centre for Ecological Sciences
Indian Institute of Science
India


	[[alternative HTML version deleted]]

_______________________________________________
R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models