Multicollinearity and VIF

[Darcy Hill]

I am using the following code
xtreg recycling loginc logpopden gcses alevels md11 md12 md13 md14 md15 md16 md17 md18 md19 md20 md21 md22 md23 md24 md25 md26 md27 md28 md29 md291 wasteavg dryavg quarter2 quarter3 quarter4 year2 year3 year4 year5, fe vce(cluster acode)

There is a pwcorr of 0.44 between alevels and gcses, so I suspect no multicollinearity.

However when I run

vif, uncentered

I get the following results. I don't know how to interpret these results and what to do with them. I have never used vif command before. Thank you

Variable | VIF 1/VIF
-------------+----------------------
loginc | 236.36 0.004231
gcses | 188.65 0.005301
logpopden | 13.28 0.075327
alevels | 10.53 0.094996
md17 | 7.59 0.131731
md12 | 6.67 0.149955
md15 | 6.44 0.155223
dryavg | 5.48 0.182631
md11 | 3.92 0.254828
md27 | 3.85 0.259701
wasteavg | 2.88 0.347274
md28 | 2.73 0.366574
md22 | 2.51 0.398896
md14 | 2.29 0.436294
year5 | 2.28 0.438546
year4 | 2.17 0.459991
year3 | 2.12 0.470777
year2 | 2.08 0.480658
md18 | 2.08 0.481430
md24 | 2.02 0.494153
quarter3 | 2.00 0.500285
quarter4 | 2.00 0.500285
quarter2 | 2.00 0.500285
md23 | 1.86 0.537648
md26 | 1.84 0.544478
md291 | 1.69 0.592504
md21 | 1.49 0.669333
md29 | 1.31 0.765507
md13 | 1.30 0.769849
md16 | 1.30 0.770104
md19 | 1.29 0.775184
md20 | 1.20 0.830814
md25 | 1.08 0.930047
-------------+----------------------
Mean VIF | 15.95

[Carlo Lazzaro]

Darcy:
actually -estat vif- gives you an idea about quasi-extreme multicollinearity (when multicollinearity is perfect, Stata omits variables).
Goldberger's outstanding textbook (http://www.hup.harvard.edu/catalog.p...ontent=reviews) devotes the entire Chapters 23 on multicollinearity.
Quasi-extreme multicollinearity can be considered differently in terms of its effects on model estimation. Following Goldberg's idea, some authoritative contributors of this list consider it nothing more than a nuisance, being the confidence interval of the coefficient the gauge to look at to get an idea whether quasi-exterme multicolinearity is actually biting that hard. The Stata .pdf manual reports two rule of thumbs for the largest VIF (more than 10 or 30 means multicollinearity) along with the threshold of 1 for the mean VIF.
In your case, I would take a look at -loginc- and -gcses- data, as their leverage on the mean VIF is remarkable, no matter the idea that you might have about multicollinearity.

[Darcy Hill]

Thank you Carlo!