[R] overdispersed zero inflated continuous data

[Ben Bolker]

Jenushka <jhazlehu <at> tulane.edu> writes:

> 
> I'm a beginning R user. 
> 
> The data: Volume of nectar in flowers under 4 different treatments,
>  nested
> for individual (measures were taken mutliple times from different 
> flowers of
> the same individual- never the same flower). 

  This is really more of a statistical question than an R question,
and might get more informed answers at r-sig-ecology at r-project.org,
but:

> 
> Specs: 54% of the data = 0. Variance=3.89. Mean=1.03. Sample size per
> treatment varies, all are >20. 
> 
> I am trying to determine if treatment had any impact on nectar volume. 
> 

   Since it will be hard to transform the data to normality (since
you have a big pile of responses equal to exactly zero), one possibility
would be a two-stage model: fit a binomial GL(M)M for zero vs non-zero,
then fit a linear (mixed) model to the (probably log-transformed) positive
responses.

  Alternatively you could ignore the distribution and use a randomization
(permutation or bootstrap) approach to get reasonable p-values/confidence
intervals, although you'll have to be careful to do the randomization
respecting the grouping by individual.

  Ben Bolker