Multicollinearity with Fixed Effects resulting in inflated VIFs for dummies

[khalilullahmohammad]

Hi,
I have an unbalanced panel for banks with around 2000 observations of bank data from 45 countries.
Each country has a different number of banks.
The regression model has both country dummies 45 and time dummies 5.
I am getting inconsistent signs which is an indicator of multicollinearity.
Since there is no direct test for multicollinearity test that I know of using xt I run a regression using reg and include all the country and time dummies to make it equivalent to a fixed effect regression.

"reg KtoA_1 Profit rate FeetoA hersh LoanLossResGrossLoans Size BtoM DailyAverageVolume Earningpershare CPI gdp inf Typedum2 Countrydum1 Countrydum2 Countrydum3 Countrydum4 Countrydum5 Countrydum6 Countrydum7 Countrydum8 Countrydum9 Countrydum10 Countrydum11 Countrydum12 Countrydum13 Countrydum14 Countrydum15 Countrydum16 Countrydum17 Countrydum18 Countrydum19 Countrydum20 Countrydum21 Countrydum22 Countrydum23 Countrydum24 Countrydum25 Countrydum26 Countrydum27 Countrydum28 Countrydum29 Countrydum30 Countrydum31 Countrydum32 Countrydum33 Countrydum34 Countrydum35 Countrydum36 Countrydum37 Countrydum38 Countrydum39 Countrydum40 Countrydum41 Countrydum42 Countrydum43 Countrydum44 Countrydum45 Countrydum46 Countrydum47 Countrydum48 Countrydum49 Countrydum50 Countrydum51 Y1 Y2 Y3 Y4 Y5 "



After this I use the VIF command to check for the variance inflation factor.
The result comes out to have very large VIFs for the country dummies.

Variable | VIF 1/VIF
-------------+----------------------
Countrydum18 | 107.91 0.009267
Countrydum13 | 57.19 0.017487
Countrydum14 | 37.45 0.026704
Countrydum32 | 29.13 0.034329
Countrydum6 | 28.15 0.035530
Countrydum48 | 26.41 0.037868
Countrydum28 | 25.68 0.038938
Countrydum2 | 23.50 0.042545
Countrydum19 | 20.42 0.048980
Countrydum34 | 19.80 0.050501
Countrydum37 | 19.77 0.050578
Countrydum45 | 19.07 0.052441
Countrydum9 | 13.56 0.073741
Countrydum35 | 13.31 0.075118
Countrydum21 | 12.24 0.081724
Countrydum22 | 11.64 0.085906
Countrydum30 | 10.34 0.096699
Countrydum31 | 9.84 0.101658
Countrydum40 | 9.39 0.106526
Countrydum50 | 8.77 0.113993
Countrydum1 | 8.69 0.115111
Countrydum20 | 8.57 0.116729




But since I'll be using the fixed effects (xtreg ,fe) for the regression where the country dummies get dropped should I care about the large VIF result??

[Jeff Wooldridge]

You shouldn't care about the VIFs for the country dummies; I wouldn't even look at them. If you are planning on a fixed effects analysis at the bank level, which seems sensible, you won't be estimating the country effects, anyway. And if you do it at the country level, those dummies are to account for country-specific heterogeneity. You don't necessarily hope to estimate them precisely (unless that is part of the goal).

[Nick Cox]

A completely separate point is that the syntax

 
reg KtoA_1 Profit rate FeetoA hersh LoanLossResGrossLoans Size BtoM DailyAverageVolume Earningpershare CPI gdp inf Typedum2 Countrydum1 Countrydum2 Countrydum3 Countrydum4 Countrydum5 Countrydum6 Countrydum7 Countrydum8 Countrydum9 Countrydum10 Countrydum11 Countrydum12 Countrydum13 Countrydum14 Countrydum15 Countrydum16 Countrydum17 Countrydum18 Countrydum19 Countrydum20 Countrydum21 Countrydum22 Countrydum23 Countrydum24 Countrydum25 Countrydum26 Countrydum27 Countrydum28 Countrydum29 Countrydum30 Countrydum31 Countrydum32 Countrydum33 Countrydum34 Countrydum35 Countrydum36 Countrydum37 Countrydum38 Countrydum39 Countrydum40 Countrydum41 Countrydum42 Countrydum43 Countrydum44 Countrydum45 Countrydum46 Countrydum47 Countrydum48 Countrydum49 Countrydum50 Countrydum51 Y1 Y2 Y3 Y4 Y5

can surely be made more readable. With modest assumptions about the order of variables in the dataset that looks like

reg KtoA_1 Profit rate FeetoA hersh LoanLossResGrossLoans Size BtoM DailyAverageVolume Earningpershare CPI gdp inf Typedum2 Countrydum1- Countrydum51 Y1-Y5

and perhaps even

reg KtoA_1 Profit rate FeetoA hersh LoanLossResGrossLoans Size BtoM DailyAverageVolume Earningpershare CPI gdp inf Typedum2 Countrydum* Y?

I count 51 country dummies.

The effect is essentially aesthetic, but is still worthwhile.

[Patricia Amandus]

Hi, I am trying to do a similar test as you are, specifically I am trying to test the macro-economic and bank specific determinants of non performing loans in Europe. For that I have a sample of 4150 European banks, in 18 Eurozone countries, with data for the each bank (ROE, ROA, reserve ratio, npl ratio, etc) as well as country level data like GDP, inflation, unemployment.. I am really new at Stata and I am trying to sort my data. I saw you have a similar test and was wondering how you structured your data? I would like to show the results for each country. I would really appreciate your answer. Thank you!