Especially in that probability range I would be very weary of using a
linear regression on a proportion, especially if you intend to do an
extrapolation. Extrapolations only work if the model is true, and at
that probability range you know that the effect of covariates cannot be
linear as is implied in linear regression, because in that range the
effect of variables has to level off to ensure that the probability
remains larger than zero.
You will just need to think very carefully what is you want: the
difference in probabilities for persons with typical values on the
other covariates, or the difference in typical probabilities. In the
latter case you need to make assumptions about the distributions of the
other covariates as well.
Hope this helps,
Maarten
--- Jn <[email protected]> wrote:
> Thanks for that helpful quote. I did notice if I were to run a
> standard linear regression before working with -adjust- command, my
> new predicted values were in the right range (around 0.10). But if I
> use -logit-, I get values around (0.05) which is way off and makes no
> sense. Given that I am trying to use the future projected mean for
> some of my x-variables (which is why I am doing running these
> postestimation commands in the first place), I don't think there is a
> way around fixing my problem if I were using a logit regression. Do
> you think it would be far too incorrect for me to run a standard
> linear regression just for this purpose only (forecasting future
> probability of a positive outcome)? At least I get reasonable
> predicted values that way..
>
>
> - student
>
> On Tue, Mar 25, 2008 at 7:09 AM, Maarten buis
> <[email protected]> wrote:
> > This is discussed in Buis (2007) "predict and adjust with logistic
> > regression", The Stata Journal, 7(2), pp. 221-226.
> > http://www.stata-journal.com/article.html?article=st0127
> >
> > The reason for the difference is that -logit- implies a non-linear
> > transformation, so it makes a difference whether you first create
> > predicted values and than compute the mean, or when you first
> compute
> > the means of explanatory variables and than compute a predicted
> value.
> > To quote from the article: "It is the difference between a typical
> > predicted probability for someone within a group and the predicted
> > probability for someone with typical values on the explanatory
> > variables for someone within that group."
> >
> > Hope this helps,
> > Maarten
> >
> >
> > --- Jn <[email protected]> wrote:
> > > I am trying to get at the magnitude of a change in Pr(y=1|x) by
> > > replacing each explanatory variable with its sample average,
> save for
> > > my variable of interest, which I was hoping to use a future
> projected
> > > distribution (I'm trying to see how this change in distribution
> of
> > > this certain binary independent variable changes the probability
> of
> > > y=1). I had no problem doing this with linear regressions
> (replace
> > > all
> > > variables with its sample mean, except use projected
> distribution for
> > > my variable of interest, do a linear prediction, note the
> > > difference).
> > > However, when I try to carry out the same procedure in a logit
> > > regression, I am running into problems. I was under the
> impression
> > > that, if I were to replace ALL of my independent variables with
> its
> > > sample mean and then run -predict-, I should get the same
> predicted y
> > > value as if I were to just run a normal regression without
> replacing
> > > my x-var with its sample mean. Am I wrong? I hope I am making
> myself
> > > clear..
> >
> >
> > -----------------------------------------
> > 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/
> > -----------------------------------------
> >
> >
> >
> > ___________________________________________________________
> > Rise to the challenge for Sport Relief with Yahoo! For Good
> >
> > http://uk.promotions.yahoo.com/forgood/
> >
> >
> > *
> > * 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/
> >
> *
> * 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/
