I am struggling with a ZIGLMM in R. I have a data set on freshwater plant propagules (response variable) and the relation with the ecological state of ponds. Pond state is a categorical factor with three levels. I have three measurements (propagule counts per liter sediment) per pond, hence the mixed model with pond as a random factor. Based on histograms and the assumption of dispersal barriers creating excess zeros, I suspect a zero-inflated model suits my data best.

I am using package glmmadmb as described by Bolker et al. 2012 (“Owls example”):

Propagules <- glmmadmb(formula = Propagules ~ Category + (1|Pond), data = `ZIGLMM.input`, family = "nbinom", zeroInflation = TRUE)

AIC: 483.9 

                     Estimate Std. Error z value Pr(>|z|)    
(Intercept)             3.355      0.237   14.13   <2e-16 ***
CategoryClear/low SM   -2.292      0.648   -3.54   0.0004 ***
CategoryTurbid        -17.116    316.930   -0.05   0.9569    
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

I have a number of questions:

  • Could AIC be used to compare different models with identical factor complexity? I mean, to select between e.g. a Poisson or Negative Binomial distributed model? Or is AIC only relevant when shopping through models with different complexity in terms of number of included factors? Also, would QAIC be a better choice?
  • Graphically (see below), I am not convinced there is no significant difference between ecological state ‘Clear/High SM’ (i.e. my reference) and ‘Turbid’ state. Am I seeing this wrong? I guess the correction for zero-inflation makes the use of box plots less recommended? I have the impression that the large Standard Error for ‘Turbid’ results from a few (ecologically relevant) outliers.
  • How can I change my reference level in order to compare ‘Clear/low SM’ with ‘Turbid’? Should I use Tukey tests as included in the multcomp package? I’m guessing this is not a valid approach for ZIGLMM. I tried adjusting the category names to switch the alphabetical order, but it didn’t work.

Boxplots and histograms

Thank you very much for any kind advice.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.