[R-sig-ME] Modelling overdispersed, high-zero Abundance data

ONKELINX, Thierry Thierry.ONKELINX at inbo.be
Mon Sep 12 14:31:04 CEST 2011


Dear Gillian,

I would go for MCMCglmm with a zero-inflated poisson model. Use site as random effects. You'll need to think about two models. One explaining the extra zero's and one explanining the abundance. Overdisperion is taken into account by default with MCMCglmm.

Best regards,

Thierry

> -----Oorspronkelijk bericht-----
> Van: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-
> bounces at r-project.org] Namens Gillian Eastwood
> Verzonden: maandag 12 september 2011 11:23
> Aan: r-sig-mixed-models at r-project.org
> Onderwerp: [R-sig-ME] Modelling overdispersed, high-zero Abundance data
> 
> 
> 
> 
> 
> 
> Dear list,
> 
> 
> 
> I was hoping to get some consensus on an analysis I'm doing because at present
> I seem to going round in circles trying different techniques depending on who
> I've last spoken to. I'm also abit confused between the different options of
> mixed models that there are how to test them. To
> explain:
> 
> 
> 
> I'm trying to model species abundance from dataset that has both spatial (site,
> distance to water, distance to village) and temporal (ie.
> trap-sites monitored 5+ times on different dates across 5 years, some more than
> others) elements to it.   The ideal
> outcome would to be able to predict for additional site that weren't sampled,
> and to include a visual map of hi/med/lo abundance i.e. some presentable to
> management.
> 
> 
> 
> 
>  The
>      data is count data - therefore I'm using Poison family of errors (also
>      tried Neg Bin)
>  The data seems OverdispersedThere
>      are a lot of Zero's (since the species often aren't at the same site) -
>      therefore I did try zero-inflated (zinb) models, and also (log(count+1))
>      transforming the outcome.
>  Some
>      covariates seem to have a non-linear relationship to the response var, so
>      I wondered if some of the techniques I've considered aren't suitable  I
>      have 1070 observations, 38 sites, 7 habitats, plus climatic covariates.
>  Note
>      that not all sites were monitored regularly (some very often/some hardly
>      ever). Also I suspect the data is biased depending on the time of year
>      when those few monitoring's took place, i.e. if it took place in summer there
> seem some very large sites averages, and if all in winter, zeros.I have another
> 300 observations kept aside which I was using to check model predictions upon,
> 

snipped

 
> And then GLMER and MCMCglmm seem options also.
> 
> 
> 
> I have a rather short timeframe by which to complete this.  I think my data is
> quite hard to explain on the available factors anyway but I was hoping the list
> might be able indicate which of these I seem to be going in the right direction to
> persevere with.
> 
> 
> 
> Many thanks in advance for any advice that you might offer
> 
> 
> Gillian
> 
> 	[[alternative HTML version deleted]]




More information about the R-sig-mixed-models mailing list