Skip to content

Variance Inflation Factor VIC() with a matrix

3 messages · Martin H. Schmidt, John Fox, Michael Friendly

Martin H. Schmidt
# 
Hi everyone,

Running the vif() function from the car package like

----------------------------------------------------
 > reg2 <- lm(CARsPur~Delay_max10+LawChange+MarketTrend_20d+MultiTrade, 
data=data.frame(VarVecPur))
 > vif(reg2)
     Delay_max10       LawChange MarketTrend_20d      MultiTrade
        1.010572        1.009874        1.004278        1.003351
----------------------------------------------------

gives a useful result. But using the right-hand variables as a matrix in 
the following way doesn't work with the vif() function:

----------------------------------------------------
 > reg  <- lm(CARsPur~VarVecPur)
 > summary(reg)

Call:
lm(formula = CARsPur ~ VarVecPur)

Residuals:
      Min       1Q   Median       3Q      Max
-0.72885 -0.06461  0.00493  0.06873  0.74936

Coefficients:
                           Estimate Std. Error t value Pr(>|t|)
(Intercept)              -0.037860   0.006175  -6.131 9.25e-10 ***
VarVecPurDelay_max10      0.003661   0.001593   2.298   0.0216 *
VarVecPurLawChange        0.004679   0.006185   0.757   0.4493
VarVecPurMarketTrend_20d  0.019015   0.001409  13.493  < 2e-16 ***
VarVecPurMultiTrade      -0.005081   0.003129  -1.624   0.1045
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Residual standard error: 0.1229 on 6272 degrees of freedom
Multiple R-squared: 0.03021,    Adjusted R-squared: 0.02959
F-statistic: 48.84 on 4 and 6272 DF,  p-value: < 2.2e-16

 > vif(reg)
Error in vif.lm(reg) : model contains fewer than 2 terms

----------------------------------------------------
Is there a solution or a way to work around?

Thank you very much in advanced.
John Fox
# 
Dear Martin,
Not with vif() in the car package, which wants to compute generalized variance inflation factors (GVIFs) for multi-df terms in the model. Single-df VIFs are pretty simple, so you could just write your own function. Alternatively, there are other packages on CRAN, such as DAAG, that compute VIFs, so you might try one of these.

I hope this helps,
 John

-----------------------------------------------
John Fox
Senator McMaster Professor of Social Statistics
Department of Sociology
McMaster University
Hamilton, Ontario, Canada
Michael Friendly
# 
You've stumbled across the answer to your question --

while lm() supports y~X formulas without a data=argument
and y~ X1+X2+X3 formulas with one, you can't depend on
all contributed functions to do the same.

As John pointed out, the advantage of car::vif over other
implementations is that it correctly handles the cases
of factors, polynomial terms, etc. for which generalized
VIF is more useful, and this is most easily accommodated
with the formula interface.

The matrix interface takes less typing, but sometimes
leaves you wondering later what you actually had in VarVecPur.

-Michael
On 9/20/2012 8:52 AM, Martin H. Schmidt wrote: