Anindya Sankar Dey wrote/hat geschrieben on/am 30.12.2015 07:35:
Hi All,
The fmsb package has a function called Variance Inflation Factor and it
states the definition of the function as follows:-
"To evaluate multicolinearity of multiple regression model, calculating the
variance inflation factor (VIF) from the result of lm(). If VIF is more
than 10, multicolinearity is strongly suggested.
"
The function computes VIF of a model as 1/(1-R^2) where R^2 is the
coefficient of determination.
Now nowhere in literature I have come across this definition of VIF, as VIF
is always computed at individual variable level. Though the structure is
almost the same, R^2 in theoretical VIF is the partial correlation
coefficient.
I only came aware when lots of freshers from non statistics background I
interviewed for analytics position answered that the only definition of VIF
they know is 1/(1 - Coeff. of Determination), and there is a R package
which calculates VIF like that.
After researched I found that such a function indeed exist in fmsb package.
Please help me understand has an alternate definition of Variance Inflation
Factor has ever emerged in theory? Does it really make sense to have VIF at
a model level, as it does not help in solving the problem of
multicollinearity during model building.
And if I am right, what steps I should do about it.
Dear Anindya,
to me it seems clear from the example on the help page that VIF() is not
intended to be applied to the model of interest, but to separate models
for each covariable.
The model of interest in the example is
# the target multiple regression model
res <- lm(Ozone ~ Wind+Temp+Solar.R, data=airquality)
The VIF is calculated on submodels for each covariate.
# checking multicolinearity for independent variables.
VIF(lm(Wind ~ Temp+Solar.R, data=airquality))
VIF(lm(Temp ~ Wind+Solar.R, data=airquality))
VIF(lm(Solar.R ~ Wind+Temp, data=airquality))
Does that agree with your usual definition of a variance inflation factor?
best regards,
Heinz
______________________________________________
R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see
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.
