Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Re: A fractional logit can't converge. Anyway, backed up results are useful for mfx? how?

From   Maarten Buis <[email protected]>
To   [email protected]
Subject   Re: st: Re: A fractional logit can't converge. Anyway, backed up results are useful for mfx? how?
Date   Thu, 8 Mar 2012 14:29:11 +0100

On Thu, Mar 8, 2012 at 1:14 PM, Roger Cremades Rodeja
<[email protected]> wrote:
>  I am using fractional tobit "glm depvar indepvar, link(logit)
>   family(binomial)  vce(robust)". I successfully used it in other two models,
>  but in this case after some iterations Stata says "backed up" and cannot
>  converge, then I create a table with the results (outreg), and still appears
>  a resulting table with results that are consistent with my other two models,
>  do those results makes statistical sense?

They make no sense, so don't look at them and certainly don't try to
publish anything with them.

>  I want to create then elasticities with "mfx" but I can't because the
>  model did not converge, if those results produced with outreg (previous
>  line) make a bit of sense, how can I create elasticities with them using
>  Stata?

Don't do that.

>  Also, after typing the command appear "note: perc_mod_com has noninteger
>  values", I do not understand why, as I think proportion variables should be
>  0 to 1.

That message makes sense: remember that -glm- is a general purpose
program that can be "tricked" into estimating a fractional logit.

> Following previous emails in Statalist, I tried delete some
>  variables and I still couldn't solve it, OLS do not reject any variable, the
>  model do not work trying to delete indep variables one by one, and I am
>  trying more and more combinations, but it still does not work.

I would look for some form of perfect prediction. This becomes more
likely if you have a very large spike in the dependent variable
(typically at 0 or 1) or a small sample size.

>  Please somebody knows if this analysis can be carried on using another
>  sofware? I need these results..

I think it is a problem with your data not your software.

-- Maarten

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index