Re: st: Outlier diagnostics for tobit (postestimation)

[Nick Cox <njcoxstata@gmail.com>]

My #3 was

3. You should be clear that you really do have a tobit problem and not
one analogous to a logit or probit problem.

We see on this list examples in which -tobit- or -intreg- is being
applied to problems in which responses outside an interval [a, b] are
impossible in principle. The arguments from at least some of us are
that that kind of problem is usually much better treated as a logit or
probit or beta regression problem. (Clearly any [a, b] can be scaled
to [0, 1].)

You are repeating your question on whether you can pretend that your
problem is a standard regression problem and want to be told "right or
wrong". Who can tell? Perhaps it would be a decent approximation;
perhaps it would be lousy. We don't have your data to try it out. But
my view is that if you have a bounded response, you should be
respectful of those bounds in all your analysis. I wouldn't be
impressed at that kind of treatment if I were a reviewer of your paper
(examiner of your thesis, whatever).

Nick

(I am deliberately avoiding "censored" or "truncated" as terms of art,
partly because of the first point I made.)

On Fri, Oct 19, 2012 at 10:48 AM, Timo Beck <timoworldwide@gmx.de> wrote:
> Dear Nick and Jay,
>
> Thank you for your help.
>
> @ Nick: I already checked cases for clear outliers, e.g., implausible values (and also simulated different versions). Further I used logarithmic transformation for specific variables which also helped. Still I wanted to use some "established" method for a further check (not for the main analysis, but rather as a robustness check). Not sure, what you mean by number 3) though.
>
> @ Jay: Thank you for the hint, I will definitely look into that.
>
> Once again quickly re my other question, maybe you also have an opinion on whether, just as a robustness test, I could fit OLS as an approximation of the tobit model and use outlier diagnostics thereafter and then simulate the tobit without these identified cases? Or would I be doing something completely wrong? According to Wooldridge a linear model is a good approximation for E(y) in a corner solution model which is what I am looking at. That's why I am thinking that way.

*
*   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/