Variance inflation factor problem

[Andreas Head]

Hello everyone,
I am facing a problem with -estat vif- after an OLS regression which includes continous, categorical and factor variables.
I never had any problems with computing the vif of a model before. However, this time something weird happened.
I obtained the follwoing output:

. estat vif

    Variable |       VIF       1/VIF  
-------------+----------------------
    villages |
          1  |      2.07    0.483266
          3  |         .           .
          4  |         .           .
          5  |      2.30    0.434749
          6  |         .           .
          7  |      1.91    0.522719
          8  |         .           .
          9  |      2.28    0.439065
         10  |         .           .
         age |      1.89    0.527762
         sex |         .           .
      income |      2.19    0.456496
accesscredit |         .           .
        educ |
          1  |      2.09    0.479603
          3  |         .           .
   healthHoH |      2.71    0.368845
      HHsize |         .           .
  hhaffected |      2.69    0.371174
  fjaddsteps |      1.41    0.710444
  rolemodels |      1.18    0.850576
optimistic~s |      1.31    0.762461
-------------+----------------------
    Mean VIF |         .

Why are some cells empty? Could anyone please pint me into the right direction here?

Tank you,
Andreas

[Richard Williams]

I would suggest showing the regression command and output that came before this.

Also sometimes if you click on a dot you get a message that may clarify things.

[Andreas Head]

Thanks Richard. I just ran the same lines again in order to copy the command and output to this thread. I am feeling a bit silly, because the earlier mentiond issue -for any reason- is not appearing any more.

Hence, this thread can be closed.

[Andreas Head]

The poblem occured again. I have no idea what causes that.
As Richard recommended, I copy the command and output of the regression model here.

.   fvset base 3 educ

.   fvset base 2 villages

.   reg adaptappraisal i.villages age sex income  accesscredit  i.educ healthHoH  HHsize fjaddsteps  rolemodels   

      Source |       SS       df       MS              Number of obs =     213
-------------+------------------------------           F( 19,   193) =    3.01
       Model |   143.61416    19     7.55864           Prob > F      =  0.0001
    Residual |  484.937483   193  2.51262945           R-squared     =  0.2285
-------------+------------------------------           Adj R-squared =  0.1525
       Total |  628.551643   212  2.96486624           Root MSE      =  1.5851

------------------------------------------------------------------------------
adaptappra~l |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
    villages |
          1  |    .140429   .5234871     0.27   0.789    -.8920613    1.172919
          3  |  -1.334822   .5133613    -2.60   0.010     -2.34734   -.3223028
          4  |  -.6341599   .5502348    -1.15   0.251    -1.719405    .4510857
          5  |  -.3851816   .5143835    -0.75   0.455    -1.399717    .6293533
          6  |  -.5212758   .5997871    -0.87   0.386    -1.704255    .6617032
          7  |  -.4006783   .5610591    -0.71   0.476    -1.507273    .7059162
          8  |   -.079398   .5258514    -0.15   0.880    -1.116551    .9577555
          9  |  -.8271979   .5531408    -1.50   0.136    -1.918175    .2637793
         10  |   .0762667   .4926919     0.15   0.877    -.8954851    1.048019
             |
         age |   .0031305   .0089679     0.35   0.727    -.0145573    .0208182
         sex |   .2275542   .2349708     0.97   0.334    -.2358861    .6909944
      income |   .1332306   .0919382     1.45   0.149    -.0481021    .3145632
accesscredit |  -.3557261   .3297639    -1.08   0.282     -1.00613    .2946777
             |
        educ |
          1  |  -1.042849   .4142495    -2.52   0.013    -1.859887    -.225812
          2  |  -.6375354   .3734279    -1.71   0.089    -1.374059    .0989883
             |
   healthHoH |  -.0556691   .0988497    -0.56   0.574    -.2506334    .1392952
      HHsize |   .1121931   .0467182     2.40   0.017     .0200493    .2043369
  fjaddsteps |   .4352882   .1629628     2.67   0.008     .1138716    .7567048
  rolemodels |   .0109248   .0059042     1.85   0.066    -.0007202    .0225699
       _cons |    .967726   1.018594     0.95   0.343     -1.04128    2.976732
------------------------------------------------------------------------------

.
end of do-file

. estat vif

    Variable |       VIF       1/VIF  
-------------+----------------------
    villages |
          1  |      1.98    0.505951
          3  |         .           .
          4  |         .           .
          5  |      2.23    0.447702
          6  |         .           .
          7  |      1.88    0.530526
          8  |         .           .
          9  |      2.24    0.445924
         10  |         .           .
         age |      1.87    0.533990
         sex |         .           .
      income |      2.17    0.461249
accesscredit |         .           .
        educ |
          1  |      2.08    0.480022
          2  |         .           .
   healthHoH |      2.59    0.385624
      HHsize |         .           .
  fjaddsteps |      2.63    0.380652
  rolemodels |      1.39    0.721139
-------------+----------------------
    Mean VIF |         .

Any advise or hint what could provoke this vif output?

Thank you,
Andreas

[Andreas Head]

I was able to narrow done the "failure source" that causes the VIF output to be strange.
I ran the same model without changing the reference categories of the two factor variables (fvset), and suddenly
the vif output appeared without dots. Still, I am wondering why changes to the reference category
leads to "failure" in the vif-output.
Has anyone any explanation for this?

[Richard Williams]

I'm guessing that the Ns for some of the categories are very small. I suggest running frequencies for your categorical variables, preferably limited to the 213 cases in your analysis.

Also, given the small N, you may want to reduce the number of categories. For example, if it makes substantive sense, you might want to create a dichotomy that is village 3 versus not village 3.

As a sidelight, you don't have to do fvset. You could instead say something like

reg adaptappraisal ib2.villages age sex income accesscredit ib3.educ healthHoH HHsize fjaddsteps rolemodels

Normally fvset would save you a little trouble, but given that you are having problems you might want to try different reference categories.

[Andreas Head]

Thank you Richard for your helpful opinion. The education variable is indeed unequally distributed (primary N=70; seondary N=117; tertiary=26). The village variable is relative homogenously distributed, only small variations in the N exist but of course all these categories are small (as you guessed). Is it that reference categories should be preferably larger categories compared against the other variable categories?

Your codes works fine and seems to be an excellent alternative, thanks for that.