RE: st: Moderation effect by splitting the sample

["David Radwin" <dradwin@mprinc.com>]

Rebecca,

Aside from the simplicity argument outlined below, I can't think of a
reason.

David

--
David Radwin
Senior Research Associate
MPR Associates, Inc.
2150 Shattuck Ave., Suite 800
Berkeley, CA 94704
Phone: 510-849-4942
Fax: 510-849-0794

www.mprinc.com


> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-
> statalist@hsphsun2.harvard.edu] On Behalf Of Rebecca Pope
> Sent: Wednesday, January 02, 2013 8:33 AM
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: Moderation effect by splitting the sample
> 
> Thanks David. Sorry for the delay responding. I've been away for the
> holidays.
> 
> Your earlier post is missing from my inbox. All I have is the chain
> with Maarten's response. I went back and read the archives for context
> on your response to my query, but I appreciate you reposting the
> paper.
> 
> I see the value of using the graphs presented for conveying
> information to non-technical audiences. I'm still not convinced after
> reading this paper that discretizing a continuous regressor is a good
> idea when conducting inferential analysis. Reading your earlier posts,
> the ones you cited, and the subtext of the paper, I am left with the
> impression that this is the general consensus.
> 
> I'm going to rephrase my original question to my intent rather than
> what it strictly said:
> Is there any econometric (or statistical if you prefer) reason to
> choose to conduct a "split" analysis unless you have natural groups
> and a strong theoretical reason to not force equality in their
> variances?
> 
> Thanks,
> Rebecca
> 
> 
> 
> On Fri, Dec 21, 2012 at 11:09 AM, David Radwin <dradwin@mprinc.com>
wrote:
> > Rebecca,
> >
> > Sometimes you want to present a result in a simpler or less technical
> way,
> > perhaps to a non-expert audience. It is often easier and more
> parsimonious
> > to compare two groups, whether verbally or in a table or graph. The
cost
> > is some loss in power. But it may be possible to present the
continuous
> > relationship, too, perhaps in an appendix or some other less prominent
> > fashion.
> >
> > For an example of how income (a continuous variable that could be
split
> > into two groups for simplicity) is related to voting in US
presidential
> > elections, please see the work I referred to earlier:
> >
> > Gelman, A., & Park, D. K. (2009). Splitting a predictor at the upper
> > quarter or third and the lower quarter or third. The American
> > Statistician, 63(1), 1-8.
> > http://www.stat.columbia.edu/~gelman/research/published/thirds5.pdf
> >
> > David
> > --
> > David Radwin
> > Senior Research Associate
> > MPR Associates, Inc.
> > 2150 Shattuck Ave., Suite 800
> > Berkeley, CA 94704
> > Phone: 510-849-4942
> > Fax: 510-849-0794
> >
> > www.mprinc.com
> >
> >
> >> -----Original Message-----
> >> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-
> >> statalist@hsphsun2.harvard.edu] On Behalf Of Rebecca Pope
> >> Sent: Thursday, December 20, 2012 1:24 PM
> >> To: statalist@hsphsun2.harvard.edu
> >> Subject: Re: st: Moderation effect by splitting the sample
> >>
> >> Maarten wrote: "Splitting a sample means that you added an
interaction
> >> term with all variables. This is typically not what you want, and
> >> often leads to a severe loss of power."
> >>
> >> My understanding is that you would only do this when you have natural
> >> groups and a strong theoretical reason to not force equality in their
> >> variances. Is there any other situation where this approach is
> >> warranted?
> >>
> >>
> >> Thanks,
> >> Rebecca
> >>
> >>
> >>
> >>
> >> On Thu, Dec 20, 2012 at 1:53 PM, Maarten Buis
<maartenlbuis@gmail.com>
> >> wrote:
> >> > On Thu, Dec 20, 2012 at 8:42 PM, Ebru Ozturk wrote:
> >> >> For non-linear models, I want to test the moderation effect of X
> >> variable. Can I test this moderation effect by spliting the sample
> >> according to X variable (moderator)?
> >> >
> >> > That is typically inefficient. Moderation is just an interaction
> >> > effect. Splitting a sample means that you added an interaction term
> >> > with all variables. This is typically not what you want, and often
> >> > leads to a severe loss of power. It is even worse if your variable
x
> >> > is continuous and you are splitting the sample by first making it
> >> > categorical by splitting it at some arbitrary number (e.g. the
median
> >> > from your previous question). That is a very bad idea, as you would
> >> > loose even more information that way. Instead you should just add
> your
> >> > interaction effect and interpret it correctly. Various examples are
> >> > given here:
> > <http://www.maartenbuis.nl/publications/interactions.html>.
> >> >
> >> > -- 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/