[R] RES: Linear multivariate regression with Robust error

[Mike Marchywka]

________________________________
> Date: Fri, 10 Jun 2011 16:50:24 -0300
> From: filipe.botelho at vpar.com.br
> To: friendly at yorku.ca; bkkochar at gmail.com
> CC: r-help at r-project.org
> Subject: [R] RES: Linear multivariate regression with Robust error
>
>
>
> --Forwarded Message Attachment--
> Subject: RES: [R] Linear multivariate regression with Robust error
> Date: Fri, 10 Jun 2011 16:50:24 -0300
> From: filipe.botelho at vpar.com.br
> To: friendly at yorku.ca; bkkochar at gmail.com
> CC: r-help at r-project.org
>
>
> Hi Barjesh,
>
> I am not sure which data you analyze, but once I had a similar situation
> and it was a multicolinearity issue. I realized this after finding a
> huge correlation between the independent variables, then I dropped one
> of them and signs of slopes made sense.
>
> Beforehand, have a glance at your correlation matrix of independent
> variables to see the relationship between them.

I guess "look at the data" ( LATFD ) seems to be the recurring solution.
Certainly with polynomial regression, there's a tendency to think that
the fit parameter is a pure property of the system being examined. With
something like a Taylor series, and you have a slope at a specific point,
maybe you could think about it that way but here the coefficient
is just "whatever optimizes your (arbitrary ) error function for the data
you have." A linear approaximation to a non-line could be made at
one point and maybe that property should remain constant but 
you have people claiming " past authors samples some curve near x=a
and got a different slope than my work largely smapling f(x) around
x=b what is wrong with my result?" It may be interesting to try to
write a taylor series around some point and see how those coefficients
vary with data sets for example( you still need arbitrary way to
estimate slopes and simply differencing two points may be a bit noisy LOL 
but you could play with some wavelet families maybe ). 

If you try to describe a fruit as a linear combination of vegetables,
you may get confusing but possibly useful results even if they don't correspond to
properties of the fruits so much as a specific need you have. For example, if you
are compressing images of fruits and your decoder already has a dictionary
of vegetables, it may make sense to do this.  This is not much different from 
trying to compress non-vocal music with an ACELP codec that attempts to fit the sounds
to models of human vocal tract. Sometimes this may be even be informative
about how a given sound was produced even if it sounds silly. 












>
> Not sure how helpful my advice will be, but it did the trick for me back
> then ;)
>
> Cheers,
> Filipe Botelho
>
> -----Mensagem original-----
> De: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org]
> Em nome de Michael Friendly
> Enviada em: sexta-feira, 10 de junho de 2011 12:51
> Para: Barjesh Kochar
> Cc: r-help at r-project.org
> Assunto: Re: [R] Linear multivariate regression with Robust error
>
> On 6/10/2011 12:23 AM, Barjesh Kochar wrote:
> > Dear all,
> >
> > i am doing linear regression with robust error to know the effect of
> > a (x) variable on (y)other if i execute the command i found positive
> > trend.
> > But if i check the effect of number of (x.x1,x2,x3)variables
> > on same (y)variable then the positive effect shwon by x variable turns
> > to negative. so plz help me in this situation.
> >
> > Barjesh Kochar
> > Research scholar
> >
> You don't give any data or provide any code (as the posting guide
> requests) , so I have to guess that you
> have just rediscovered Simpson's paradox -- that the coefficient of a
> variable in a marginal regression can have an opposite sign to that in
> a joint model with other predictors. I have no idea what you mean
> by 'robust error'.
>
> One remedy is an added-variable plot which will show you the partial
> contributions of each predictor in the joint model, as well as whether
> there are any influential observations that are driving the estimated
> coefficients.
>
>
> --
> Michael Friendly Email: friendly AT yorku DOT ca
> Professor, Psychology Dept.
> York University Voice: 416 736-5115 x66249 Fax: 416 736-5814
> 4700 Keele Street Web: http://www.datavis.ca
> Toronto, ONT M3J 1P3 CANADA
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
> "This message and its attachments may contain confidential and/or
> privileged information. If you are not the addressee, please, advise
> the sender immediately by replying to the e-mail and delete this
> message." "Este mensaje y sus anexos pueden contener información
> confidencial o privilegiada. Si ha recibido este e-mail por error por
> favor bórrelo y envíe un mensaje al remitente." "Esta mensagem e seus
> anexos podem conter informação confidencial ou privilegiada. Caso não
> seja o destinatário, solicitamos a imediata notificação ao remetente e
> exclusão da mensagem."
> ______________________________________________ R-help at r-project.org
> mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do
> read the posting guide http://www.R-project.org/posting-guide.html and
> provide commented, minimal, self-contained, reproducible code.