[R-sig-ME] Modelling zero-inflated count data in nested, experiments

[Highland Statistics Ltd]

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

Message: 3
Date: Thu, 7 Feb 2013 11:55:23 +0000
From: Eva Muiruri <emuiruri25 at gmail.com>
To: r-sig-mixed-models at r-project.org
Subject: [R-sig-ME] Modelling zero-inflated count data in nested
	experiments
Message-ID:
	<CAFcdEfNTU7cxoPjUuoaZQZ8qwWQf3C_-_f7+Q-FWn8984j79AA at mail.gmail.com>
Content-Type: text/plain

Hello,

I'm trying to analyse the effect of tree species richness within a plot on
the number of insect herbivores per tree.
Unfortunately, a huge proportion of the herbivore count data has zero
values. I have tried all the transformations I have come across in the
literature and have even tried ranking the data but to no avail.

Due to the nested nature of our experiment, I initially ran models using
lmer and the poisson family, specifying AREA and PLOT as random effects.
> m1 <- lmer(HERBIVORE ~ RICHNESS, (1|AREA/PLOT), family="poisson")

The residuals were not normally distributed so I tried to run the same
model using the AD model builder, accounting for zero inflation and trying
both the poisson family or the negative binomial family but still, there is
decreasing variance in the residuals for higher fitted values.
>m2 <- glmmadmb(HERBIVORE ~ RICHNESS, (1|AREA/PLOT), family="nbinom",
zeroInflation=TRUE)

My only other option, as far as I can see, would be to report analysis
based on the presence/absence of herbivores on each tree (using a binomial
error distribution in lmer) but I would be unable to discuss any effect of
tree species richness on herbivore densities.

Is there a further method of analysis I can try in R?







Just fit a zero inflated GLMM in JAGS, WinBUGS or OpenBUGS. I would use JAGS.
Not sure what you mean with more than 3 random effects.
You need plenty of plots and plenty of areas.
Yes..that is MCMC....

Residuals do not need to be normally distributed. It is lack of fit you need to investigate. Because
you used a negative binomial distribution (or Poisson) for small values, the distribution is likely to be skewed...hence the
reason you have more negative than positive residuals....

zero inflated negative binomial is quite a complicated distribution.
Try a ZIP GLMM..assuming your data is large enough, and you have enough levels for the plots and areas.

Alain







  Could the MCMC method be used with this dataset? (From what I understand,
it is only to be used when you have 3 or more random effects but, as I am
new to this, I may be wrong.)

Thanks, in advance.

Eva

	[[alternative HTML version deleted]]



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

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


End of R-sig-mixed-models Digest, Vol 74, Issue 11