[R-sig-ME] Modeling for a zero-inflated continuous response

[Erin Latham]

Hi R list,

I'm attempting to model a continuous response (CHANGEM) using three
variables (SLOPE, ASPECT, ELEVATION).

> str(sheep)
'data.frame':   419 obs. of  10 variables:
 $ SITE     : Factor w/ 4 levels "Gibbs","Langley",..: 1 1 1 1 1 1 1 1 1 1
...
 $ AREACODE : Factor w/ 20 levels "Aspect1","Aspect2",..: 2 2 2 2 2 2 2 2 3
3 ...
 $ GRID     : int  0 3 4 5 6 1 2 7 0 4 ...
 $ SLOPE    : num  40.8 30 59.7 44.8 56.1 56.7 60.7 47.2 36.6 25.1 ...
 $ ASPECT   : num  0.7597 -0.4677 0.0963 0.2272 -0.2194 ...
 $ ELEVATION: int  2576 2757 2521 2659 2562 2725 2627 2657 2221 2242 ...
 $ LATITUDE : Factor w/ 2 levels "n","s": 1 1 1 1 1 1 1 1 1 1 ...
 $ CHANGEM  : num  0 0 0 0 0 ...
 $ CHANGEP  : num  0 0 0 0 0 ...
 $ VEG      : Factor w/ 18 levels "","BA","BL","BM",..: 1 1 1 1 1 1 1 1 1 1
...

I performed a change analysis with an unbalanced hierarchical sample design.
I interpreted usually 8 grids (GRID) (sometimes less) from each AREACODE of
each SITE. There are 4 sites with 8 to 19 sub-plots (AREACODEs) for a total
of 419 total grids analyzed.
My problem is that the response has many zeroes (260/419 grids). The
response was calculated as a change detection between two dates, and the
units are square meters. Technically I performed the change based on a grid,
so it could be considered a proportion or change as percent cover, as well
(CHANGEP). Besides the zeros, the data is normally distributed.

I found this helpful post by Ben Bolker, but I'm uncertain how to go about
modeling the zeroes as part of the mixture process when the distribution is
otherwise normal.
http://permalink.gmane.org/gmane.comp.lang.r.lme4.devel/3454
I'm been reading Bolker (2007), but I'm a bit stuck translating the
discussion to code.

Any help is greatly appreciated.
Thank you.

-- 
Erin Latham, M.GIS