R help - Mar 2011 - glm: modelling zeros as binary and non-zeroes as coming from a continuous distribution

Hello,

I'd like to implement a regression model for extremely zero-inflated
continuous data using a conditional approach, whereby zeroes are
modelled as coming from a binary distribution, while non-zero values
are modelled as log-normal.

So far, I've come across two solutions for this: one, in R, is
described in the book by Gelman & Hill
(http://www.amazon.com/dp/052168689X), where they just model zeros and
non-zeros separately and then bring them together by simulation. I can
do this, but it makes it difficult to assess the significance of
regression coefficients wrt to zero and each other.

Another solution I have been pointed at is in SAS:
http://listserv.uga.edu/cgi-bin/wa?A2=ind0805A&L=sas-l&P=R20779,
where they use NLMIXED (with only fixed effects) to specify their own
log-likelihood function.
I'm wondering if there's any way to do the same in R (lme can't deal
with this, as far as I'm aware).

Finally, I'm wondering whether anyone has experience with the COZIGAM
package - does it do something like this?

Many thanks,
Mikhail

Hi:

You might want to consider hurdle models in the pscl package.

HTH,
Dennis

On Wed, Mar 30, 2011 at 2:41 AM, a11msp <absorbtion@gmail.com> wrote:
> Hello,
>
> I'd like to implement a regression model for extremely zero-inflated
> continuous data using a conditional approach, whereby zeroes are
> modelled as coming from a binary distribution, while non-zero values
> are modelled as log-normal.
>
> So far, I've come across two solutions for this: one, in R, is
> described in the book by Gelman & Hill
> (http://www.amazon.com/dp/052168689X), where they just model zeros and
> non-zeros separately and then bring them together by simulation. I can
> do this, but it makes it difficult to assess the significance of
> regression coefficients wrt to zero and each other.
>
> Another solution I have been pointed at is in SAS:
> http://listserv.uga.edu/cgi-bin/wa?A2=ind0805A&L=sas-l&P=R20779,
> where they use NLMIXED (with only fixed effects) to specify their own
> log-likelihood function.
> I'm wondering if there's any way to do the same in R (lme can't
deal
> with this, as far as I'm aware).
>
> Finally, I'm wondering whether anyone has experience with the COZIGAM
> package - does it do something like this?
>
> Many thanks,
> Mikhail
>
> ______________________________________________
> R-help@r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
	[[alternative HTML version deleted]]

On Mar 30, 2011; 11:41am Mikhail wrote:
>> I'm wondering if there's any way to do the same in R (lme
can't deal
>> with this, as far as I'm aware).
You can do this using the pscl package.

Regards, Mark.

--
View this message in context:
http://r.789695.n4.nabble.com/glm-modelling-zeros-as-binary-and-non-zeroes-as-coming-from-a-continuous-distribution-tp3417718p3417857.html
Sent from the R help mailing list archive at Nabble.com.