Unfortunately the answer is "it depends, and you really need to
understand why the two are different in order to choose". The real
issue is: do you want to control for the marginal distribution, that
is, the baseline odds? If you do, then you want to look add the odds
ratios, otherwise you want to look at the risk differences.
Hope this helps,
Maarten
--- David Zetland <[email protected]> wrote:
> Maartin,
>
> Thanks for your response (my understanding = "there is a reason for
> the difference" :)
>
> What about the *bigger* question?
>
> > > Should I care? What can be done about it? How do I interpret
> results?
>
> If gender is insignificant in MLOGIT but significant in MFX (or vice
> versa), which result should I scream from the rooftops?
>
> Thanks!
>
> David
>
>
> On Wed, Mar 26, 2008 at 7:42 AM, David Zetland <[email protected]>
> wrote:
> > Date: Tue, 25 Mar 2008 11:26:27 +0000 (GMT)
> > From: Maarten buis <[email protected]>
> > Subject: Re: st: mlogit and mfx -- statistical significance
> >
> > The reason for such a difference is that there is a subtle
> difference
> > in what these two measures measure: The output from -mlogit- gives
> you
> > a test whether the odds ratio (*) is significantly different from
> 1,
> > while -mfx- gives you a change in probabilities.
> >
> > Say the odds ratio is 10, but the baseline odds is .0001 success
> for
> > every failure, so we are talking about comparing a odds of .0001
> with
> > an odds of .001, which corresponds to a probabilities .00009999
> and
> > .000999, leading to a risk difference (result of -mfx-) of
> .00089901 .
> >
> > So a big sounding odds ratio of 10 can easily correspond to a very
> > small sounding risk difference of .0009 . The reason is that by
> > computing the ratio of odds, the baseline odds drops out, so whith
> odds
> > ratios you only look groups are different but leave out the fact
> that
> > those differences hardly matter if probability of success is very
> high
> > or very low for everybody.
> >
> > Hope this helps,
> > Maarten
> >
> > (*) Some disciplines don't call the parameters of -mlogit- log
> odds
> > ratios, but they are the log of a ratio of odds, so I think the
> other
> > disciplines are correct in refering to them as log odds ratios,
> for
> > more on this see:
> > http://www.stata.com/statalist/archive/2007-02/msg00085.html
> >
> > Ps. has anybody noticed that I have refrained from repeating my
> > favourite feature request?
> >
> > - --- David Zetland <[email protected]> wrote:
> >
> > > Can anyone explain why some RHS vars are significant in <mlogit>
> but
> > > not in <mfx> (and vice versa)?
> > >
> > > I have several RHS dummies, but the problem also shows up with
> > > <margeff>
> > >
> > > Should I care? What can be done about it? How do I interpret
> results?
> > >
> > > I have searched the FAQs and forums and seen similar questions
> but no
> > > answers
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
__________________________________________________________
Sent from Yahoo! Mail.
More Ways to Keep in Touch. http://uk.docs.yahoo.com/nowyoucan.html
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
