Re: [Gretl-devel] duration models and Liu-Hong test
On Tue, 20 Dec 2016, Ignacio Diaz-Emparanza wrote:
El 17/12/16 a las 20:19, Allin Cottrell escribió:
> Any duration model experts out there?
>
> I'm trying to get started with implementing Liu and Hong's omnibus
> specification test for duration models with censoring -- see
> http://ajbuckeconbikesail.net/Seminars/Spring_08/Duration_Chapter1.pdf
> -- but I've got stuck on what seems like it should be an elementary point.
>
> The relevant pages of the article are 6-10. Here's my problem: their
> empirical survivor function (p. 10) involves counting cases where
> V_i(\theta) and C_i(\theta), for observations i, are greater than
> duration value t. But V_i and C_i are (if I'm reading the paper right) CDF
> values and therefore limited to [0,1], while the duration t is said on page
> 6 to be distributed on [0, \infty). So I don't see how these two terms can
> be meaningfully compared.
>
> I guess I'm missing some implicit mapping/transformation (maybe of t onto
> [0,1]?). Can anyone help?
>
> Allin
I am not an expert in duration models but have two people here who are.
Talking with one of them he says "I understand it's typical survival notation
and procedures for what could do well to look at the "survival" library in R,
because what you want to estimate are things that I believe are there".
Thanks, Ignacio. I've taken a look at the pdf doc for the survival
package but I'm not seeing any reference to Hong and Liu there.
(Both R and Stata support diagnostics for duration models that involve
plotting generalized residuals against the unit exponential
distribution, but unfortunately these are not very useful when there's
substantial censoring, which is often the case with socioeconomic
data. The attraction of the Hong/Liu approach is that it (seems to)
offer a generalized Kolmogorov-Smirnoff statistic that is valid under
right-censoring -- if only one could make sense of what they're doing!
Their unpublished 2007 paper has been cited a few times in subsequent
publications by others, but AFAICT nobody has gone into any detail on
the matter.)
Allin