Re: st: Tobit regression Model

[Nick Cox <njcoxstata@gmail.com>]

At one time the contributions of a famous Italian Bayesian

http://en.wikipedia.org/wiki/Bruno_de_Finetti

seemed important, although they now seem eccentric in at least one
sense. In the last few decades Bayesian statistics has been slowly
morphing into a branch of numerical analysis, and has lost some of its
sting, or distinctiveness.

As some people know, almost all Stata users' meetings held in London
have been within a few hundred metres of the grave of Thomas Bayes,
although this should be regarded as a coincidence rather than an
indication of any philosophy or policy. (The grave is neither easy to
see nor especially interesting in any way, unfortunately.)

Grave matters aside, I don't think improper priors should be blamed on Bayes.

Nick

On Sun, Dec 2, 2012 at 6:51 PM, Carlo Lazzaro
<carlo.lazzaro@tiscalinet.it> wrote:
> I agree with Jay about the trade-off between research assumptions and
> empirical evidence.
> About Jay's second point, my smattering of the Bayesian framework is limited
> and, at least in Italy, it is not common at all. I usually go (empirical)
> Bayesian for running probabilistic sensitivity analysis (I'm a health
> economist)  and play on the safe side of the matter with conjugated
> distributions. Others may find hard times in dealing with more intimidating
> facets of Rev Bayes' machinery (eg: improper priors)!
>
> Best wishes,
> Carlo
>
> -----Messaggio originale-----
> Da: owner-statalist@hsphsun2.harvard.edu
> [mailto:owner-statalist@hsphsun2.harvard.edu] Per conto di JVerkuilen
> (Gmail)
> Inviato: domenica 2 dicembre 2012 18:44
> A: statalist@hsphsun2.harvard.edu
> Oggetto: Re: st: Tobit regression Model
>
> On Sun, Dec 2, 2012 at 12:00 PM, Carlo Lazzaro <carlo.lazzaro@tiscalinet.it>
> wrote:
>> Dear Jay,
>> thanks for the plug.
>> Yes, I have heard something alike to what you report about handling
>> with care Tobit model.
>
> Tobit depends on normality quite strongly in a way that neither of its
> parent models, ordinary linear regression or probit regression, do.
>
>
>> Erring on the high side of N is always a good practice. Unfortunately,
>> this is often a matter of (tight) research funds.
>
> Well it is, absolutely, but if you don't have the dataset you want you
> either trade off assumptions (e.g., Bayesian estimation with informative
> priors, not an easy thing to do) or run a simpler model than you would like.
>
*
*   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/