Re: st: comparing coefficients across 2 models

[francesco.mazzi@unifi.it]

Hi Maarten,

thanks for your help and suggestions. I really appreciate the fact  
that you kind of "ranked" the solutions for dealing with  
multicollinearity.

You also wrote "multicolinearity (high VIFs) is not a problem, it is  
just a description of an unfortunate state of the world, or at least,  
your data".
Given that I got high VIFs for the interacted variables  and  
interaction term, do you think that my model would be biased if I do  
not deal with multicollinearity (e.g. using centering)?
Would a model with high VIFs in interaction terms sound statistically correct?

Thanks for your help,
Francesco





----- Message from maartenlbuis@gmail.com ---------
    Date: Mon, 26 Nov 2012 11:34:52 +0100
    From: Maarten Buis <maartenlbuis@gmail.com>
Reply-To: statalist@hsphsun2.harvard.edu
 Subject: Re: st: comparing coefficients across 2 models
      To: statalist@hsphsun2.harvard.edu


> On Mon, Nov 26, 2012 at 11:08 AM,  <francesco.mazzi@unifi.it> wrote:
> > I am struggling with a model and I am interested in understanding if an
> > independent variable (x1) changed its impact on the dependent variable (y)
> > following an event (prepost=0 if pre event; prepost=1 if post event).
> >
> > I know that I could run a model considering the interaction of the two.
> > xi: regress y c.x1##i.prepost x2 x3, robust cluster(firmcluster)
> > The problem is that I got really high VIFs (above 30!!)
> >
> > I investigate the possibility of centering or orthogonalising the variables
> > (I saved the residuals of regress x1_prepost_interaction x1 prepost). I got
> > quite nice results.
> > Do you think this is ok and acceptable? Can you suggest other methods for
> > overcoming the problem of multicollinearity?
> > Moreover, I thought to split the model and run the following two commands
>
> Modeling is always a judgement call, so we cannot give you our
> "statistical blessing" and say your model is correct. That is in the
> end your decision and your decision alone. Moreover, multicolinearity
> (high VIFs) is not a problem, it is just a description of an
> unfortunate state of the world, or at least, your data.
>
> Having said that, centering is usually an excellent idea; it often
> helps interpretability of results and reduces numerical problems, so
> that is a win-win situation. Orthogonalising variables can reduce
> numerical problems, but often makes it harder to interpret the
> results, so that is a trade-off you need to make. Estimating separate
> models as a "solution" to solve multicolinearity is a horrible idea;
> it does not make the multicolinearity go away, it just makes it harder
> to detect.
>
> Hope this helps,
> Maarten
>
> ---------------------------------
> Maarten L. Buis
> WZB
> Reichpietschufer 50
> 10785 Berlin
> Germany
>
> http://www.maartenbuis.nl
> ---------------------------------
> *
> *   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/


----- End message from maartenlbuis@gmail.com -----




*
*   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/