Re: st: Re: Probability threshold for inclusion in ZINB count model

[Stas Kolenikov <skolenik@gmail.com>]

Jeremy,

That's not how this model works; you have a rather mechanistic
understanding of it.

ZINB is a joint model that has a two-part likelihood, one for
identical zeroes, and one for negative binomials, that are weighted
according to the probability from the inflation equation. This is very
clearly shown in -zinb- documentation. What you are thinking about is
a sort of two-step model: first, separate the zeroes; second, estimate
the negative binomial for the non-zeroes. But the first step cannot be
performed perfectly, as the negative binomial can produce random
zeroes, so the second step will not make much sense, either.

On Sat, Feb 25, 2012 at 3:23 PM, Jeremy Bowling
<jlbkq6@mail.missouri.edu> wrote:
> According to http://www.ats.ucla.edu/stat/stata/output/Stata_zinb.htm, the
> "negative binomial model is generated predicting the counts for those
> [observations] who are not certain zeros."  Since the inflation/logit model
> determines whether or not an observation is a certain zero by probability,
> is there a probability threshold by which the observations are dropped?
> Or... are each unique observations' probabilities included in the negative
> binomial equation to control for the probability of being a certain zero,
> similar to when predicted counts are created in which probabilities are
> included in the equation.  Are observations dropped due to meeting a certain
> zero probability threshold, or is the negative binomial (count model)
> altered in some way to account for the probability that an observation may
> be a certain zero?  If there is a probability threshold and dropped
> observations in the count model, I would like to know.  If not, by what
> process does the zinb count model control for an observation being in the
> certain zero group?
>
> -----Original Message----- From: Nick Cox
> Sent: Saturday, February 25, 2012 4:46 AM
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: Probability threshold for inclusion in ZINB count model
>
> Perhaps you explain what threshold you expect there to be on relation
> to the manual entry on -zinb- or expand on how you think -zinb- should
> use such a threshold.
>
> Nick
>
> On 25 Feb 2012, at 05:05, Jeremy Bowling <jlbkq6@mail.missouri.edu>
> wrote:
>
>> I am having a hard time finding the probability threshold that zinb  uses
>> to include/limit observations in the count model.  Is there a  set
>> probability threshold?  If not how do I find the threshold?
>>
>> After running zinb for a number of count models, I was seeking to  run a
>> simple logit model by dichotomizing my dv instead of having a  count.  I
>> would like to limit my observations by excluding those  that would be in the
>> certain zero group.  Of course the significance  of the logit coefficients
>> would be affected by the threshold I set  (which observations are included).
>> I would like to use the logit  model to seek not what causes multiple
>> occurrences (count) but what  factors lead to the likelihood of an
>> occurrence, from zero to one,  for those observations not in the certain
>> zero group.
>>
>> If there is not a set probability threshold or it can not be found,  what
>> would be the best probability to use?  50% cutoff? Mean  probability
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/



-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/