re: RE: st: Hurdle model

["Ariel Linden, DrPH" <ariel.linden@gmail.com>]

Neil,

Your justification for choosing a hurdle model does not sound appropriate to
me. 

You argue that "all individuals" have the same chronic illness and " there
is little/no chance of them having an event without it being recorded."

How does this logic differentiate between those having/not having a
hospitalization (the binary component), and those having multiple admissions
(the count component)? The idea is that patients not having an admission
(ie., not crossing the hurdle) will have certain characteristics that would
differentiate them from those having multiple admissions (that's why you
estimate two models here). There is nothing in your statements that would
answer that point. 

I would argue that, in this instance, the main reason for crossing that
hurdle is illness severity, and that would more reasonably be estimated as a
covariate in a count model, as Jeph suggested.

More importantly, hospital admissions are rare events, even in chronically
ill populations. Thus, it doesn't make even intuitive sense to me, for you
to model these data separately, since even after the hurdle is crossed, I
would expect the data to be highly skewed.

It is important for a researcher to think through these issues thoroughly,
and thoughtfully, and not immediately go to the "fancy statistics".
Ultimately, I expect that you'll have to defend your choice of analysis, and
that should rely on content knowledge as well as sufficient knowledge about
statistical approaches.

Ariel
 

Date: Fri, 13 Sep 2013 10:04:16 +0100
From: Neil Hewitt <mcxnh4@nottingham.ac.uk>
Subject: RE: st: Hurdle model

Thanks Jeph,

I have used a hurdle model not a zero inflated model as I believe all my
zeros are real and not structural. All individuals in my panel have the same
chronic condition, the outcome is related to that chronic condition and
there is little/no chance of them having an event without it being recorded.

That is how I understand the difference between the two though will not
profess to be an expert.

Thanks,

Neil 

- -----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Jeph Herrin
Sent: 13 September 2013 09:51
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: Hurdle model

It sounds like what you have is a "zero inflated Poisson" model, not a
hurdle model. Either way, I don't think there is a panel data version
available (official or user written) for either in Stata. Almost certainly
not for the hurdle model
- - I needed to estimate some hurdle models for panel data recently and dug
around a lot before finally/reluctantly running them in SAS.

Jeph


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