Problem with Stata 14

[FernandoRios]

Dear all,
I encounter a problem while running Stata, and have not found any other post related to it. For replication of the problem, Im including here a toydataset of my main data, for which the problem persists.

sum

    Variable |        Obs        Mean    Std. Dev.       Min        Max
-------------+---------------------------------------------------------
         sex |      4,319    1.463302    .4987092          1          2
     citizen |      4,319           0           0          0          0
         wtf |      4,319    3125.397    1704.387   270.8347   13390.37
       tran4 |      4,319    .0754804    .2641956          0          1

.  reg tran4 i.citizen sex  [pw=wtf]
(sum of wgt is   1.3499e+07)

Linear regression                               Number of obs     =      4,319
                                                F(1, 4316)        =          .
                                                Prob > F          =          .
                                                R-squared         =     0.0003
                                                Root MSE          =      .2612

------------------------------------------------------------------------------
             |               Robust
       tran4 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     citizen |
        NIU  |   .0091304   .0016563     5.51   0.000     .0058833    .0123775
         sex |  -.0084173   .0090477    -0.93   0.352    -.0261555    .0093208
       _cons |   .0767286   .0073079    10.50   0.000     .0624013    .0910559
------------------------------------------------------------------------------

As you can see, when using the i.citizen, even i find a regression for which the variable is not dropped, as it should be, since there is no variation in citizen.

While i can avoid the problem just excluding this variable, i think it requires some attention.
Thank you
Fernando
DATA:
Data_small.dta

[Clyde Schechter]

I can duplicate your results. I think you should report this to technical support.

The weights seem to be confusing Stata somehow: if we drop the pweights, then it runs as expected and citizen gets omitted.

[FernandoRios]

Thank you, Will do.

[Richard Williams]

I have problems too but at least the coefficient for citizen is not coming up as significant:

. reg tran4 i.citizen sex  [pw=wtf]
(sum of wgt is   1.3499e+07)

Linear regression                               Number of obs     =      4,319
                                                F(1, 4316)        =          .
                                                Prob > F          =          .
                                                R-squared         =     0.0003
                                                Root MSE          =      .2612

------------------------------------------------------------------------------
             |               Robust
       tran4 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     citizen |
        NIU  |  -.0654856     786432    -0.00   1.000     -1541811     1541811
         sex |  -.0084173   .0090477    -0.93   0.352    -.0261555    .0093208
       _cons |   .1513446     786432     0.00   1.000     -1541811     1541811
------------------------------------------------------------------------------

That to me suggests some sort of collinearity problem, where the program calculates a coefficient as being slightly different than zero when it should be zero.

Interestingly, glm does what you would expect it to do:

. glm tran4 i.citizen sex  [pw=wtf]
note: 0.citizen omitted because of collinearity

Iteration 0:   log pseudolikelihood =  -48594314  

Generalized linear models                         No. of obs      =      4,319
Optimization     : ML                             Residual df     =      4,317
                                                  Scale parameter =   213.1883
Deviance         =  920333.7181                   (1/df) Deviance =   213.1883
Pearson          =  920333.7181                   (1/df) Pearson  =   213.1883

Variance function: V(u) = 1                       [Gaussian]
Link function    : g(u) = u                       [Identity]

                                                  AIC             =   22502.58
Log pseudolikelihood = -48594313.89               BIC             =   884197.1

------------------------------------------------------------------------------
             |               Robust
       tran4 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     citizen |
        NIU  |          0  (omitted)
         sex |  -.0084173   .0090456    -0.93   0.352    -.0261464    .0093118
       _cons |    .085859   .0144293     5.95   0.000     .0575781    .1141399
------------------------------------------------------------------------------

[Joe Canner]

I got the right answer (citizen omitted because of collinearity). What version/update of Stata are you using, Fernando?

[Clyde Schechter]

For me it was:

. about

Stata/MP 14.1 for Windows (64-bit x86-64)
Revision 19 May 2016
Copyright 1985-2015 StataCorp LP

Total physical memory:     8269900 KB
Available physical memory: 2238232 KB

Single-user 2-core Stata perpetual license:
       Serial number:  [redacted]
         Licensed to:  Clyde Schechter
                       Albert Einstein College of Medicine

. update query
(contacting http://www.stata.com)

Update status
    Last check for updates:  08 Jun 2016
    New update available:    none         (as of 08 Jun 2016)
    Current update level:    19 May 2016  (what's new)

Possible actions

    Do nothing; all files are up to date.

[Richard Williams]

I dug out my old copy of Stata 11 and it worked fine:

. reg tran4 i.citizen sex  [pw=wtf]
(sum of wgt is   1.3499e+07)
note: 0.citizen omitted because of collinearity

Linear regression                                      Number of obs =    4319
                                                       F(  1,  4317) =    0.87
                                                       Prob > F      =  0.3522
                                                       R-squared     =  0.0003
                                                       Root MSE      =  .26117

------------------------------------------------------------------------------
             |               Robust
       tran4 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
   0.citizen |  (omitted)
         sex |  -.0084173   .0090467    -0.93   0.352    -.0261534    .0093188
       _cons |    .085859    .014431     5.95   0.000     .0575669    .1141511
------------------------------------------------------------------------------

But in 12, 13 and 14 I had problems.

[FernandoRios]

Dear Joe,
I was using Stata 14 updated to 30 Match 2016.

Stata/MP 14.1 for Windows (64-bit x86-64)
Revision 30 Mar 2016
Copyright 1985-2015 StataCorp LP

Total physical memory:     8345288 KB
Available physical memory: 3450016 KB

Single-user 4-core Stata perpetual license:

I did notice that when I used a logit, 0.citizen is dropped out before it starts estimating the model. What is particularly annoying is that this particular problem comes randomly.
It should also be noticed that in this examples, the coefficient affected is the constant, although it did created some problems with other coefficients in my larger models.
Fernando

[Richard Williams]

This seems to work:

. svyset [pw = wtf]

      pweight: wtf
          VCE: linearized
  Single unit: missing
     Strata 1: <one>
         SU 1: <observations>
        FPC 1: <zero>

. svy: reg tran4 i.citizen sex
(running regress on estimation sample)

Survey: Linear regression

Number of strata   =         1                 Number of obs     =       4,319
Number of PSUs     =     4,319                 Population size   =  13,498,590
                                               Design df         =       4,318
                                               F(   1,   4318)   =        0.87
                                               Prob > F          =      0.3521
                                               R-squared         =      0.0003

------------------------------------------------------------------------------
             |             Linearized
       tran4 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     citizen |
        NIU  |          0  (omitted)
         sex |  -.0084173   .0090456    -0.93   0.352    -.0261513    .0093167
       _cons |    .085859   .0144293     5.95   0.000     .0575702    .1141479
------------------------------------------------------------------------------

[Joao Santos Silva]

Dear All,

I had a similar problem in 2013 and Miguel Dorta from StataCorp told me the following:

...since factor variables were first introduced with Stata 11, -regress- and -_rmcoll- switch the numeric tolerance for omitting nearly perfect collinear variables when factor variables operators are used.

I guess this may be the source of the problem and it may also explain the following behavior (there is no problem if we do not use factor notation):

. reg tran4 citizen sex  [pw=wtf]
(sum of wgt is   1.3499e+07)
note: citizen omitted because of collinearity

Linear regression                               Number of obs     =      4,319
                                                F(1, 4317)        =       0.87
                                                Prob > F          =     0.3522
                                                R-squared         =     0.0003
                                                Root MSE          =     .26117

------------------------------------------------------------------------------
             |               Robust
       tran4 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     citizen |          0  (omitted)
         sex |  -.0084173   .0090467    -0.93   0.352    -.0261534    .0093188
       _cons |    .085859    .014431     5.95   0.000     .0575669    .1141511
------------------------------------------------------------------------------

At the time I also discussed this with Jeff Pitblado because I find it very annoying that Stata is not consistent in the way collinearity is treated, but the problems persist. Ideally, the user should be able to have some control over the choice of threshold that is used to exclude variables. Hopefully this additional example will help persuading StataCorp to do something about this.

All the best,

Joao

[Jeff Pitblado (StataCorp)]

This is specific to regress and occurs when pweights, aweights, or iweights are specified.

We have found the source of the problem and hope to have it fixed in the next Stata 14 update.