collinearity - incorrect sign?

[liamplayt]

I read from a book that the collinearity (or multicollinearity) could possibly lead the signs of the regression coefficients to being incorrect.
That is, some coefficient, which should have been positive (negative), is negative (positive) due to the presence of the collinearity.
I, however, can barely find examples from the internet.
Does anyone ever have such a problem?

Thanks in advance.

 

[spunky]

i have seen it but not in the context of collinearity but in the context of suppression... and one of the forms of suppression (net suppression) could probably fall into collinearity depending on the correlations among the variables... nonetheless, the mechanics are still the same (the correlation between one of your X's and the Y's is substantially lower than the correlation of that particular X and another one, so the regression coefficient ends up switching signs)
thing is there are ways to fix multicolloinearity (centering being the most popular) but suppression kind of stays...

 

[liamplayt]

Do you have any source of it? Article or link?

Thank you, Spunky.

 

[spunky]

start by looking at Cohen's et. al. "Applied Multiple Regression/Correlation Analysis for the Behavioural Sciences". page 70 of the 2003 edition discusses a nice (but brief) introduction to suppression and this thing about the sign of the regression coefficients changing. or you know what? just google suppression... or go to your favourite online database and look for suppression. there's TONS of info out there

 

[liamplayt]

Thanks a lot, Spunky. I've got some.