semicontinuous variables: what likelihoods are available?

Hi Aurelie,

I am probably more seeking assurance from the gurus than trying to answer your question, as I'm also playing with a data set in a similar situation. (I am looking at area of leaf consumed by insect herbivores.) Can you use a delta-distribution/approach for your data? That is, run binomial models on the presence-absence data of your response variable, and log-normal models on the positive (non-zero) portion of your continuous data. 

I know this approach is fairly common for linear models, e.g.
http://r-project.markmail.org/search/?q=delta%20Tweedie#query:delta%20Tweedie+page:1+mid:gnzpixld5zkl5sig+state:results

and I imagine it's equally applicable for (G)LMM's. I'll let the more knowledgeable members of this list correct me if not. I didn't see any attachment in your last message, so I don't know how your data are distributed, but this approach seemed to work well for my data (with ~70% zeros).

HTH,

George


On Mon, Mar 18, 2013 at 4:25 PM, Aurelie Cosandey Godin <GodinA at dal.ca> wrote:
Thank you Ben and others,

Apologize for not being very precise!
My response variable is measured both in weight (kg) and counts and is very zero-inflated i.e., 91% of my data.
I previously ran models on the count data using a suit of  likelihoods: 2-parts zero inflated poisson & 2-parts zero inflated negative binomial. The latter were the best.
Now I would like to run the same models but with my response variable in kg, but I don't know how to model my positive (truncated or just positive weight data?). See  figure attached of the distribution of my weight data.

Many thanks in advance!!
Aurelie






On 2013-03-18, at 4:30 PM, Ben Bolker wrote:

Aurelie Cosandey Godin <GodinA at ...> writes:


[snip]

I need to run spatio-temporal models for a semicontinuous response
variable (weight in kg).  I am not familiar with the available
semicontinuous likelihood functions available in R and was wondering
if some of you may be able to point me in the right direction for
information.

Can you say any more about exactly what a semicontinuous response variable
is?  Poking around (e.g <http://lpsolve.sourceforge.net/4.0/semi-cont.htm>)
doesn't make it entirely clear: are these data that are

truncated, i.e. values <= a lower threshold are absent from the data set;
censored, i.e. values <= a lower threshold are recorded as
  "less than threshold"?
positive, i.e. values <0 don't even exist?
are the data non-negative (i.e. >=0) or are they positive (>0)?

 The simplest of these cases is positive data, which you
can model fairly easily by log transformation (i.e. assume
a lognormal distribution), or with slightly more difficulty
using a Gamma distribution ...  if you have censored or truncated
data, or data that include zeros, it gets a little harder ...