Getting the different variables in VIF analyse compared with variables used for regression analyse

Laura Witlox

Join Date: May 2020
Posts: 4

Getting the different variables in VIF analyse compared with variables used for regression analyse

26 May 2020, 13:26

Hi Everyone,

After my regression results I want to test for multicollinearity with the function VIF. However, VIF does not show the same variables as used in my regression analysis. The variables under six2fixed in the VIF analyse does not even exist. Does someone know what is going on?

I start with the following regression

Code:

regress abs_dca earlyea size loss growth inventory sd_sales sd_cfo bm fees i.fyear i.sic2fixed, robust

The output looks like this:

Code:

Linear regression                               Number of obs     =      1,127
                                                F(36, 1090)       =      42.90
                                                Prob > F          =     0.0000
                                                R-squared         =     0.4906
                                                Root MSE          =     .16089

------------------------------------------------------------------------------
             |               Robust
     abs_dca |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     earlyea |   .0139611   .0126574     1.10   0.270    -.0108746    .0387968
        size |    .003257   .0098638     0.33   0.741    -.0160972    .0226112
        loss |   -.009179   .0191993    -0.48   0.633    -.0468508    .0284927
      growth |  -.0008725   .0004332    -2.01   0.044    -.0017224   -.0000226
   inventory |   .1039879   .1085149     0.96   0.338    -.1089338    .3169097
    sd_sales |   4.60e-06   1.77e-06     2.59   0.010     1.12e-06    8.07e-06
      sd_cfo |  -.0000155   8.05e-06    -1.92   0.055    -.0000313    3.39e-07
          bm |   .0306091   .0246821     1.24   0.215    -.0178208     .079039
        fees |  -.0080096   .0120266    -0.67   0.506    -.0316076    .0155884
             |
       fyear |
       2016  |  -.0323195   .0138803    -2.33   0.020    -.0595546   -.0050844
       2017  |   .0107309   .0148073     0.72   0.469    -.0183233     .039785
       2018  |  -.0263443   .0151248    -1.74   0.082    -.0560214    .0033327
       2019  |  -.0986281   .0472746    -2.09   0.037    -.1913876   -.0058687
             |
   sic2fixed |
         20  |    .007519   .0289802     0.26   0.795    -.0493444    .0643823
         23  |  -.1068286   .0290454    -3.68   0.000    -.1638199   -.0498374
         26  |  -.1332961   .0282909    -4.71   0.000    -.1888069   -.0777853
         28  |   .4013773   .0316068    12.70   0.000     .3393602    .4633944
         29  |  -.2151065   .0419823    -5.12   0.000    -.2974818   -.1327311
         30  |  -.1281859   .0324364    -3.95   0.000    -.1918307   -.0645411
         34  |  -.1044339   .0339218    -3.08   0.002    -.1709933   -.0378745
         35  |   .0051912   .0283465     0.18   0.855    -.0504288    .0608112
         36  |   .0531267   .0342207     1.55   0.121    -.0140192    .1202726
         37  |  -.1002941   .0274196    -3.66   0.000    -.1540953   -.0464929
         38  |   .0828333   .0294362     2.81   0.005     .0250752    .1405914
         39  |  -.0913435   .0315329    -2.90   0.004    -.1532156   -.0294713
         45  |   -.125567   .0239606    -5.24   0.000    -.1725812   -.0785529
         48  |   .0575326   .0362255     1.59   0.113    -.0135469    .1286122
         49  |  -.1038844   .0238229    -4.36   0.000    -.1506284   -.0571404
         50  |   .1548413   .0853878     1.81   0.070    -.0127017    .3223843
         51  |   .1588181   .0921888     1.72   0.085    -.0220694    .3397056
         55  |  -.1566967   .0429625    -3.65   0.000    -.2409953   -.0723982
         56  |  -.1292806   .0329217    -3.93   0.000    -.1938778   -.0646834
         58  |  -.0861951   .0426981    -2.02   0.044    -.1699748   -.0024153
         59  |  -.0584408   .0446967    -1.31   0.191    -.1461421    .0292605
         73  |   .1492332   .0274652     5.43   0.000     .0953425     .203124
         80  |   .1703843   .0710612     2.40   0.017      .030952    .3098166
             |
       _cons |   .2450955   .1252664     1.96   0.051    -.0006952    .4908861
------------------------------------------------------------------------------

Now I want to test for multicollinearity, but I get other sic2fixed number than shown in my regression.

Code:

. vif

    Variable |       VIF       1/VIF
-------------+----------------------
     earlyea |      1.40    0.714037
        size |      5.83    0.171586
        loss |      1.53    0.652766
      growth |      1.29    0.776751
   inventory |      3.46    0.288736
    sd_sales |      3.56    0.280609
      sd_cfo |      2.71    0.368631
          bm |      1.64    0.610074
        fees |      3.99    0.250430
       fyear |
       2016  |      1.50    0.667212
       2017  |      1.58    0.633030
       2018  |      1.61    0.619921
       2019  |      1.10    0.908733
   sic2fixed |
          5  |      2.65    0.377903
          6  |      1.67    0.598078
          8  |      1.56    0.639988
         10  |      3.50    0.285662
         11  |      2.04    0.491027
         12  |      1.37    0.728192
         14  |      1.44    0.693666
         15  |      2.68    0.373003
         16  |      2.46    0.406148
         17  |      2.14    0.466498
         18  |      3.04    0.328443
         19  |      1.28    0.780565
         21  |      1.44    0.693522
         22  |      1.71    0.583440
         23  |      3.46    0.289082
         24  |      1.92    0.522031
         25  |      2.00    0.501147
         26  |      1.99    0.503408
         27  |      2.20    0.454065
         28  |      1.37    0.731688
         29  |      2.00    0.499786
         30  |      3.80    0.263003
         33  |      1.44    0.692617
-------------+----------------------
    Mean VIF |      2.23

Why are there other sic2fixed numbers displayed? Thanks in advance

Last edited by Laura Witlox; 26 May 2020, 13:29.

Tags: fixed effect, Multicollinearity, VIF

William Lisowski

Join Date: Dec 2014

Posts: 10150
#2

26 May 2020, 20:07

Welcome to Statalist. Thank you for the clear presentation of your problem and well-formatted sample output, sorrowfully infrequent occurrences in first posts.

I had some guesses as to what is going wrong, but the guesses didn't hold up when I tested them. Your question has been up for a while and nobody else has had any success, either, apparently. So maybe we need a little more information from you.

First, did Stata produce any output between the regress command and the beginning of the results table? If so it would be useful to see that.

Secondly, can you post the results of

Code:

describe sic2fixed codebook sic2fixed

to give us a little more insight into your variable?

Finally, what version of Stata are you using?

Let me post one of my dead ends to save others some time. The output of help vif tells us that the vif command, though it continues to work, has been out-of-date since Stata 9, and tells us that estat vif is the replacement, and to see the output of help regress postestimation for more details. With that said, on my toy example I found that estat vif gave results identical to those from vif in all my tests, which led me to look behind the scenes and find that the estat command for regress postestimation just calls the ("out-of-date") vif command.
1 like
Comment

Laura Witlox

Join Date: May 2020
Posts: 4

27 May 2020, 02:17

Hi William,

Thanks for your response!

To your first question: Stata did not produce any output between the regress command and the beginning of the result tables. Everything works fine.
To your second question:

Code:

. describe sic2fixed

              storage   display    value
variable name   type    format     label      variable label
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
sic2fixed       long    %8.0g      sic2fixed

Code:

. codebook sic2fixed

--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
sic2fixed                                                                                                                                                                              (unlabeled)
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

                  type:  numeric (long)
                 label:  sic2fixed

                 range:  [2,33]                       units:  1
         unique values:  24                       missing .:  0/1,127

              examples:  10    28
                         16    36
                         22    48
                         27    56

Furthermore, I tried the following code. What gives me levels of sic2fixed which do not exist when I browse my sample.

Code:

. levelsof sic2fixed
2 5 6 8 10 11 12 14 15 16 17 18 19 21 22 23 24 25 26 27 28 29 30 33

Comment

Carlo Lazzaro

Join Date: Apr 2014

Posts: 17703
#4

27 May 2020, 02:30

Laura:
what happens if you type:

Code:

regress abs_dca earlyea size loss growth inventory sd_sales sd_cfo bm fees i.fyear i.sic2fixed, robust estat vif

My guess is that the old-fashioned -vif- command (that I was not aware of, admittedly) internally renumbers the level of some categorical variables and give back those renumbered levels in its output.
I did not experience such an issue with -estat vif-.

Kind regards,
Carlo
(Stata 19.0)
Comment

Laura Witlox

Join Date: May 2020
Posts: 4

27 May 2020, 02:41

Hi William,

I tried something more and I kind of solved the problem. sic2 is my original variable, what was a string variable. However to do my regression I have destringed sic2 and created a new variable: sic2fixed. However, I got problems with labeling during my VIF analysis with sic2fixed (as you saw in my previous post).

Now I started again and destringed sic2 (instead of creating a new variable).

Code:

destring sic2, replace

followed by my regression analyses, what is still gives the same outcome as previous:

Code:

. regress abs_dca earlyea size loss growth inventory sd_sales sd_cfo bm fees i.fyear i.sic2, robust

Linear regression                               Number of obs     =      1,127
                                                F(36, 1090)       =      42.90
                                                Prob > F          =     0.0000
                                                R-squared         =     0.4906
                                                Root MSE          =     .16089

------------------------------------------------------------------------------
             |               Robust
     abs_dca |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     earlyea |   .0139611   .0126574     1.10   0.270    -.0108746    .0387968
        size |    .003257   .0098638     0.33   0.741    -.0160972    .0226112
        loss |   -.009179   .0191993    -0.48   0.633    -.0468508    .0284927
      growth |  -.0008725   .0004332    -2.01   0.044    -.0017224   -.0000226
   inventory |   .1039879   .1085149     0.96   0.338    -.1089338    .3169097
    sd_sales |   4.60e-06   1.77e-06     2.59   0.010     1.12e-06    8.07e-06
      sd_cfo |  -.0000155   8.05e-06    -1.92   0.055    -.0000313    3.39e-07
          bm |   .0306091   .0246821     1.24   0.215    -.0178208     .079039
        fees |  -.0080096   .0120266    -0.67   0.506    -.0316076    .0155884
             |
       fyear |
       2016  |  -.0323195   .0138803    -2.33   0.020    -.0595546   -.0050844
       2017  |   .0107309   .0148073     0.72   0.469    -.0183233     .039785
       2018  |  -.0263443   .0151248    -1.74   0.082    -.0560214    .0033327
       2019  |  -.0986281   .0472746    -2.09   0.037    -.1913876   -.0058687
             |
        sic2 |
         20  |    .007519   .0289802     0.26   0.795    -.0493444    .0643823
         23  |  -.1068286   .0290454    -3.68   0.000    -.1638199   -.0498374
         26  |  -.1332961   .0282909    -4.71   0.000    -.1888069   -.0777853
         28  |   .4013773   .0316068    12.70   0.000     .3393602    .4633944
         29  |  -.2151065   .0419823    -5.12   0.000    -.2974818   -.1327311
         30  |  -.1281859   .0324364    -3.95   0.000    -.1918307   -.0645411
         34  |  -.1044339   .0339218    -3.08   0.002    -.1709933   -.0378745
         35  |   .0051912   .0283465     0.18   0.855    -.0504288    .0608112
         36  |   .0531267   .0342207     1.55   0.121    -.0140192    .1202726
         37  |  -.1002941   .0274196    -3.66   0.000    -.1540953   -.0464929
         38  |   .0828333   .0294362     2.81   0.005     .0250752    .1405914
         39  |  -.0913435   .0315329    -2.90   0.004    -.1532156   -.0294713
         45  |   -.125567   .0239606    -5.24   0.000    -.1725812   -.0785529
         48  |   .0575326   .0362255     1.59   0.113    -.0135469    .1286122
         49  |  -.1038844   .0238229    -4.36   0.000    -.1506284   -.0571404
         50  |   .1548413   .0853878     1.81   0.070    -.0127017    .3223843
         51  |   .1588181   .0921888     1.72   0.085    -.0220694    .3397056
         55  |  -.1566967   .0429625    -3.65   0.000    -.2409953   -.0723982
         56  |  -.1292806   .0329217    -3.93   0.000    -.1938778   -.0646834
         58  |  -.0861951   .0426981    -2.02   0.044    -.1699748   -.0024153
         59  |  -.0584408   .0446967    -1.31   0.191    -.1461421    .0292605
         73  |   .1492332   .0274652     5.43   0.000     .0953425     .203124
         80  |   .1703843   .0710612     2.40   0.017      .030952    .3098166
             |
       _cons |   .2450955   .1252664     1.96   0.051    -.0006952    .4908861
------------------------------------------------------------------------------

If I use the VIF code now, I don't get problems with the wrong labeling anymore.

Code:

. vif

    Variable |       VIF       1/VIF 
-------------+----------------------
     earlyea |      1.40    0.714037
        size |      5.83    0.171586
        loss |      1.53    0.652766
      growth |      1.29    0.776751
   inventory |      3.46    0.288736
    sd_sales |      3.56    0.280609
      sd_cfo |      2.71    0.368631
          bm |      1.64    0.610074
        fees |      3.99    0.250430
       fyear |
       2016  |      1.50    0.667212
       2017  |      1.58    0.633030
       2018  |      1.61    0.619921
       2019  |      1.10    0.908733
        sic2 |
         20  |      2.65    0.377903
         23  |      1.67    0.598078
         26  |      1.56    0.639988
         28  |      3.50    0.285662
         29  |      2.04    0.491027
         30  |      1.37    0.728192
         34  |      1.44    0.693666
         35  |      2.68    0.373003
         36  |      2.46    0.406148
         37  |      2.14    0.466498
         38  |      3.04    0.328443
         39  |      1.28    0.780565
         45  |      1.44    0.693522
         48  |      1.71    0.583440
         49  |      3.46    0.289082
         50  |      1.92    0.522031
         51  |      2.00    0.501147
         55  |      1.99    0.503408
         56  |      2.20    0.454065
         58  |      1.37    0.731688
         59  |      2.00    0.499786
         73  |      3.80    0.263003
         80  |      1.44    0.692617
-------------+----------------------
    Mean VIF |      2.23

I still do not know what went wrong with sic2fixed. However the outcome of the VIF numbers is the same for sic2 and sic2fixed.

I forgot to answer your last question: my version of Stata is: 16.0. Thanks for helping me anyway!

Comment

Laura Witlox

Join Date: May 2020

Posts: 4
#6

27 May 2020, 03:25

Hi Carlo,

When I use the code that you suggested I still get the same labels as with VIF.

However, I am happy that I now have the VIF results I wanted (see the last post). Unfortunately, I still do not know what went wrong with sic2fixed.

Kind regards,
Laura

Last edited by Laura Witlox; 27 May 2020, 03:33.
Comment
William Lisowski

Join Date: Dec 2014

Posts: 10150
#7

27 May 2020, 05:37

I am glad that the exercise I suggested led you to discover that you had a problem with the creation of sic2fixed that you were able to solve.

My hypothesis is that to create sic2fixed you did

Code:

encode sic2, generate(sic2fixed)

As the output of help encode tells us,

Do not use encode if varname contains numbers that merely happen to be stored as strings; instead, use generate newvar = real(varname) or destring; see real() or [D] destring.

The encode command takes the string variable, assigns a number to each distinct value, and then labels the values so that the original string values are displayed rather than the numbers they are assigned. So the string "28" was assigned the value 10, but when you browse your data you see "28" rather than 10. But note that sic2fixed is shown in a different color than either a string variable or a numerical variable without value labels; that is a clue that what you see is not what you get.

Your original results can be explained by regress displaying the value labels but vif displaying the encoded values themselves. I see now that that is exactly what happens, which seems a bit cavalier of the vif command.
Comment
Laura Witlox

Join Date: May 2020

Posts: 4
#8

27 May 2020, 07:03

Hi William,

You are right. I have used encode to create sic2fixed:

Code:

encode sic2, generate(sic2fixed)

In the future I will use destring instead of encode.

Thanks for helping me out!

Kind regards,
Laura
Comment

Announcement

Getting the different variables in VIF analyse compared with variables used for regression analyse

Comment

Comment

Comment

Comment

Comment

Comment

Comment