Re: multiple linear regression

On May 13, 11:52 am, Richard Ulrich <Rich.Ulr...@xxxxxxxxxxx> wrote:
On Tue, 13 May 2008 09:14:37 -0700 (PDT), JW <wallis...@xxxxxxxxxxx>
wrote:

On May 13, 8:39�am, Ray Koopman <koop...@xxxxxx> wrote:
On May 13, 8:26�am, JW <wallis...@xxxxxxxxxxx> wrote:

[snip, some]
JW > >





IV1 naturally ran from 1 to 5. IV2 was dummy coded as 1 and 2.
I standardized the DV and IV before running the regression.
Ray > >
Were these the degrees of freedom for the two analyses?

� � � � �Anova � Regression
� IV1 � � �4 � � � 1
� IV2 � � �1 � � � 1
IV1*IV2 � �4 � � � 1- Hide quoted text -

- Show quoted text -
JW >
For the Anova, yes that's exactly right. For the regression, I'm not
sure. I assume those are the degrees of freedom, but I can't see it
reported anywhere (I'm using MATLAB).

However, I think I might have done something stupid - I standardized
the interaction term. In other words, I multiplied IV1 and IV2 and
then standardized the result. If I do it the other way round, i.e.
standardize IV1 and IV2 and then multiply to return the interaction
term, my regression results look a lot more like the ANOVA results.
Might this be the explanation for what I'm doing wrong?

I sounds like there were two separate problems in
translating the ANOVA to the regression.

Ray points out that using 5 categories is inherently
a different test (with 4 d.f.)  from using one continuous
variable (with 1 d.f.).

Also, as above, "centering" the terms for the interaction
before multiplying them is a vital step that gives a
different result.  When positive-signed numbers are
multiplied, the result is correlated fairly well with each
of the multipliers.  Because of that, the result will be
"confounded with" the other predictors.  They account for
some of the same variance.  That's usually not a desirable
way to look at interactions.

--
Rich Ulrich

http://www.pitt.edu/~wpilib/index.html- Hide quoted text -

- Show quoted text -


Hi Ray and Rich,

Thank you very much for your help - I really appreciate it. Just to be
clear I wasn't trying to replicate the ANOVA exactly, but using it
more as a sanity check. I could see from plotting it that there was a
(linear) effect of IV1, no effect of IV2 and no interaction (and the
ANOVA confirmed this). But I was getting strange results when trying
the same thing with multiple linear regression, specifically that it
wasn't as 'significant' as it looked from the graph.

It seems that "centering" was key. When I centered the variables it
made little difference whether I standardized the product term or not,
but when they were uncentered it made a big difference. So I guess my
final question is, generally speaking do people standardize their
product term or not?

Thanks again!
.

Relevant Pages

  • Re: multiple linear regression
    ... I standardized the DV and IV before running the regression. ... standardize IV1 and IV2 and then multiply to return the interaction ... before multiplying them is a vital step that gives a ... made little difference whether I standardized the product term or not, ...
    (sci.stat.consult)
  • Re: multiple linear regression
    ... I standardized the DV and IV before running the regression. ... standardize IV1 and IV2 and then multiply to return the interaction ... my regression results look a lot more like the ANOVA results. ... before multiplying them is a vital step that gives a ...
    (sci.stat.consult)
  • Re: multiple linear regression
    ... I'm having trouble interpreting the results of a multiple linear ... When I run the analysis through a 2-way ANOVA I ... value for the overall regression equation. ... standardize IV1 and IV2 and then multiply to return the interaction ...
    (sci.stat.consult)
  • Re: theta = (XX)^-1*Xy
    ... First of all, in Ridge Regression, the X'X matrix is ... The ordinary Ridge subtracts a fixed lambda from the ... Multiplying the diagonals of XX' by 1+delta, ...
    (sci.stat.math)
  • Re: R^2 and beta coefficients in multiple regression
    ... vector of standardized regression coefficients in ... regression coefficients. ... the Beta vector should equal R2. ... standardize my principal components. ...
    (sci.stat.math)