Re: st: Outlier diagnostics for tobit (postestimation)

["JVerkuilen (Gmail)" <jvverkuilen@gmail.com>]

On Thu, Oct 18, 2012 at 3:44 PM,  <timoworldwide@gmx.de> wrote:
> Dear All,
>
> I hope someone can help me with the following question on regression diagnostics for tobit. So far I've only used regress and for outlier diagnostics normally cooksd, rstudent and dfbeta. As these are not available for tobit postestimation I wondered if anything comparable exists for tobit that I could use (and have not found so far). As I deal with a two-limit tobit, I am mainly interested in outliers of the independent variables (i.e., cooksd and dfbeta).>

I didn't think of it last week when Ebru Ozturk was asking but I guess
you could consider Cook's likelihood displacement ideas. Most of the
usual outlier detection statistics are essentially analytic
approximations of them.

It would take some programming but you could jackknife each case and
consider how much it contributes to the log-likelihood. Large
likelihood displacement indicate that the case in question moves the
likelihood substantially, and thus is potentially influential. Um, the
Google sez:

http://childpsych.columbia.edu/brainimaging/hongtu/pdffile%5C910579.pdf

Zhu, H. & Zhang, H. (2004). A diagnostic procedure based on local
influence. Biometrika, 91, 3, 579-589.



> If nothing exists, would it be possible (only for regression diagnostic purposes) to just fit OLS and use postestimation tools thereafter instead of after tobit?
>
> Due to the large dataset graphical solutions do not work. I use Stata 10 (Stata/SE 10.1 for Windows, Born 01 Oct 2009).

It's not just the large dataset. There's no really good reason to
suppose that the usual graphical methods will work at all.

Jay

*
*   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/