Re: Dealing with correlated predictors
- From: IRISHSTAT <dave@xxxxxxxxxxx>
- Date: Tue, 9 Dec 2008 12:06:49 -0800 (PST)
On Dec 9, 11:24 am, Dim <vati...@xxxxxxxxxxx> wrote:
I have set of 17 variables (x1..x17), which are significantly
correlated. I use them as predictors in a linear regression. The
results indicate that only 2 of them (x7 and x11) are significant,
rsq=0.322.
At a second step, I calculate a new set of 17 variables (nx1..nx17) ;
each one is calculated removing the effect of the others, i.e. x1 is
regressed on x2...x17 and the residuals represent nx1 and so on.
I perform a new regression using nx1...nx17 and I get a slightly
reduced r-sq (0.314) . This time, however, the number of significant
predictors is 9 (the two most significant are nx7 and nx11).
I guess that if the correlations among x1..x17 were lower the loss of
information resulting from the use of nx1...nx17 would be higher (and
consequently r-sq would be further reduced).
But I' sure I miss something. What are the problems of this method?
Thanks
Dim
is this time series data ?
dave r
.
- References:
- Dealing with correlated predictors
- From: Dim
- Dealing with correlated predictors
- Prev by Date: Re: What is the meaning of Spectral analysis
- Next by Date: Re: Dealing with correlated predictors
- Previous by thread: Dealing with correlated predictors
- Next by thread: Re: Dealing with correlated predictors
- Index(es):
Relevant Pages
|
Loading
