PPML with many warnings...

[Klaus Schmidt]

Dear all,

I've been struggling with a ppml-application for a long while. Now it seems to give quite good results - but I am a little confused by all the warnings on "large-values" and "overfitting".
Also, my RESET test fails, so can I trust these results at all? I copy my output below.

What about dummies, should I try to avoid them?

As a robustness check I run a simple OLS (logged dependent variable) - and some of the results really change. So I assume something is not quit ecorrect, right?

Thanks so much for your help!
Klaus

-----------------------------------------------------------------------------------------------------------------------
ppml cum_sizes_concludedb $explan_var1, keep

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling or recentering
WARNING: log_GDP_i has very large values, consider rescaling or recentering
(many more warnings)

Number of regressors excluded to ensure that the estimates exist: 0
Number of observations excluded: 0

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1: deviance = 252057.3
Iteration 2: deviance = 141818
Iteration 3: deviance = 113417.2
Iteration 4: deviance = 107601.2
Iteration 5: deviance = 106889.8
Iteration 6: deviance = 106856.1
Iteration 7: deviance = 106855.3
Iteration 8: deviance = 106855.3
Iteration 9: deviance = 106855.3

Number of parameters: 29
Number of observations: 16940
Pseudo log-likelihood: -54214.637
R-squared: .19015636
Option strict is: off
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0
-------------------------------------------------------------------------------------
| Robust
cum_sizes_conclu~db | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------------+----------------------------------------------------------------
log_distance | -.9528056 .1971872 -4.83 0.000 -1.339285 -.5663257
common_language | 1.455636 .3174794 4.58 0.000 .8333883 2.077884
log_GDP_h | .6108688 .0999441 6.11 0.000 .414982 .8067556
log_GDP_i | .5874121 .0686404 8.56 0.000 .4528793 .7219449
log_GDP_pc_h | -.6919959 .2293051 -3.02 0.003 -1.141426 -.2425662
log_GDP_pc_i | .2883472 .1847046 1.56 0.118 -.0736673 .6503616
net_agric_imports~h | -3.315186 2.274837 -1.46 0.145 -7.773785 1.143414
net_agric_imports~i | -.3133903 .723114 -0.43 0.665 -1.730668 1.103887
log_agric_area_pc_h | 1.388383 .275652 5.04 0.000 .8481147 1.928651
log_agric_area_pc_i | .277707 .1605227 1.73 0.084 -.0369118 .5923258
share_available_l~h | -.0401704 .0126402 -3.18 0.001 -.0649447 -.0153961
share_available_l~i | -.0149193 .0077438 -1.93 0.054 -.0300969 .0002583
log_water_resourc~h | .9159727 .1532531 5.98 0.000 .6156022 1.216343
log_water_resourc~i | .198225 .1343857 1.48 0.140 -.0651662 .4616162
log_agric_product~h | .0325299 .1453248 0.22 0.823 -.2523014 .3173612
log_agric_product~i | .011202 .1072409 0.10 0.917 -.1989862 .2213903
corruption_h | -.0028488 .0084842 -0.34 0.737 -.0194776 .01378
corruption_i | .0162642 .0096734 1.68 0.093 -.0026954 .0352237
n_deals_h | .0446845 .0057439 7.78 0.000 .0334265 .0559424
rta | 1.01743 .4402138 2.31 0.021 .1546272 1.880233
Africa | -.1126869 .3363213 -0.34 0.738 -.7718645 .5464907
Asia | -3.619532 .7332476 -4.94 0.000 -5.056671 -2.182394
America | -1.455312 .5510493 -2.64 0.008 -2.535348 -.3752749
Europe | -.4128112 .4677604 -0.88 0.377 -1.329605 .5039823
America_i | 2.13523 .6506217 3.28 0.001 .8600344 3.410425
Asia_i | 4.363523 .9203239 4.74 0.000 2.559722 6.167325
Africa_i | 2.683462 .8665773 3.10 0.002 .9850013 4.381922
Europe_i | 2.383969 .7021059 3.40 0.001 1.007867 3.760071
_cons | -36.71535 4.897823 -7.50 0.000 -46.3149 -27.11579
-------------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel append
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb //Get fitted values - RESET test according to http://privatewww.esse
> x.ac.uk/~jmcss/reset.do
(13866 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(13866 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var1 fit2, keep // Estimate the model with th
> e additional regressor
WARNING: log_GDP_h has very large values, consider rescaling or recentering
WARNING: log_GDP_i has very large values, consider rescaling or recentering
WARNING: share_available_land_h has very large values, consider rescaling or recenter
> ing
WARNING: share_available_land_i has very large values, consider rescaling or recenter
> ing
WARNING: log_water_resources_h has very large values, consider rescaling or recenteri
> ng
WARNING: log_water_resources_i has very large values, consider rescaling or recenteri
> ng
WARNING: log_agric_productivity_i has very large values, consider rescaling or recent
> ering
WARNING: corruption_h has very large values, consider rescaling or recentering
WARNING: corruption_i has very large values, consider rescaling or recentering
WARNING: n_deals_h has very large values, consider rescaling or recentering
WARNING: fit2 has very large values, consider rescaling or recentering
Number of regressors excluded to ensure that the estimates exist: 0
Number of observations excluded: 0
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0

. test fit2=0 //Test the significance of the additional regressor (this is equivalent
> to a t-test on fit2)

( 1) fit2 = 0

chi2( 1) = 5.06
Prob > chi2 = 0.0245

[Said Jafar]

Dear Klaus,

For warning messages, pls refer to ppml command's help file

 help ppml

Your results is not readable. Pls follow advice here that we can assist you .
This might help for the Reset test.
And results from ols almost never similar to ppml. The Log of Gravity paper may help in this.

Best.

[Joao Santos Silva]

Dear Klaus,

Just to add to Dias's excellent advice, since you were able to get convergence, you can ignore the warnings about large values.
The warning about overfitting, however, is important. You should estimate the model using a different base category for your dummies. Better, you should estimate the model including the dummies for all categories and let Stata choose which category to exclude.

All the best,

Joao

[Klaus Schmidt]

Dear Dias, dear Joao,

Thanks so much for your excellent advice - I'll try to stick to the advice as best as I can.

I followed your advice and used all dummies and let Stata chose which ones to exclude (also nicely described here: http://personal.lse.ac.uk/tenreyro/Pisch.do)

Concerning this:

• Should I drop the constant then? I read here that this is recommended, but I don't understand why?
• My results get much better by doing this - but only if I also include the "strict" option. The ppml help file howvere says one should be very careful when using it. When is it ok to use?
• For one of my specifications I still get the overfitting warning - this is the specification with very few variables. What does it mean, should I add more variables, or would clustering help (I thought about clustering, but using my Country-Pairs is probably too many (16,786), only investors (155) or the receivers (109) does not make sense). I also get the overfitting warning when I do the RESET test and include the additional regressor. Is this a problem?

Thanks so much for your help, it is much appreciated!!

Best wishes,
Klaus

[Joao Santos Silva]

Dear Klaus,

Please post the results you get with and without the -strict- option so that we can see what is going on.

All the best,

Joao

[Klaus Schmidt]

Dear Joao,

Thanks so much - I admit I am a bit lost as the ppml is still unfamiliar to me.

The results with the strict option (and noconstant) are as follows:

.  ppml cum_sizes_concludedb $explan_var _D* _E*, keep noconst strict

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering

Number of regressors excluded to ensure that the estimates exist: 2
Excluded regressors:  _EGEO_i_2 _EGEO_i_5
Number of observations excluded: 0

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1:   deviance =  359096.2
Iteration 2:   deviance =  194441.8
Iteration 3:   deviance =  150712.6
Iteration 4:   deviance =  142496.2
Iteration 5:   deviance =  141789.4
Iteration 6:   deviance =    141780
Iteration 7:   deviance =    141780
Iteration 8:   deviance =    141780

Number of parameters: 16
Number of observations: 21250
Pseudo log-likelihood: -71726.238
R-squared: .15738383
Option strict is: on
---------------------------------------------------------------------------------
                |               Robust
cum_sizes_co~db |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
   log_distance |  -.7644989   .1797362    -4.25   0.000    -1.116775   -.4122224
common_language |   1.294402   .3216139     4.02   0.000     .6640507    1.924754
      log_GDP_h |   .3859054   .1050693     3.67   0.000     .1799734    .5918374
      log_GDP_i |   .5502893   .0625846     8.79   0.000     .4276257    .6729529
   log_GDP_pc_h |   -.407014   .2978928    -1.37   0.172    -.9908731    .1768452
   log_GDP_pc_i |    .436156   .1170062     3.73   0.000     .2068281    .6654838
      n_deals_h |   .0450022   .0041509    10.84   0.000     .0368667    .0531378
            rta |   1.056946     .34619     3.05   0.002     .3784261    1.735466
      _DGEO_h_1 |   -18.7301   2.174117    -8.62   0.000    -22.99129   -14.46891
      _DGEO_h_2 |  -18.35923   2.333426    -7.87   0.000    -22.93266    -13.7858
      _DGEO_h_3 |  -21.51532   2.341257    -9.19   0.000     -26.1041   -16.92654
      _DGEO_h_4 |  -17.91337   2.290921    -7.82   0.000     -22.4035   -13.42325
      _DGEO_h_5 |  -17.06806   2.515984    -6.78   0.000     -21.9993   -12.13683
      _EGEO_i_1 |  -.0388593   .4431417    -0.09   0.930     -.907401    .8296824
      _EGEO_i_3 |   1.505537   .3115474     4.83   0.000     .8949157    2.116159
      _EGEO_i_4 |   .0420963   .3682248     0.11   0.909    -.6796111    .7638036
---------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel replace
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb //Get fitted values - RESET test according to http://privatewww.essex.ac.uk/
> ~jmcss/reset.do
(9556 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(9556 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var _D* _E* fit2, keep  noconst strict  // Estimate th
> e model with the additional regressor
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering
WARNING: fit2 has very large values, consider rescaling  or recentering
Number of regressors excluded to ensure that the estimates exist: 2
Number of observations excluded: 0

. test fit2=0 //Test the significance of the additional regressor (this is equivalent to a t-t
> est on fit2)

 ( 1)  fit2 = 0

           chi2(  1) =    9.53
         Prob > chi2 =    0.0020


. ppml cum_sizes_concludedb $explan_var1 _D* _E*, keep noconst strict

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering

Number of regressors excluded to ensure that the estimates exist: 2
Excluded regressors:  _DGEO_h_5 _EGEO_i_5
Number of observations excluded: 0

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1:   deviance =  296038.8
Iteration 2:   deviance =  165420.1
Iteration 3:   deviance =  132075.3
Iteration 4:   deviance =  125898.1
Iteration 5:   deviance =  125295.4
Iteration 6:   deviance =  125284.3
Iteration 7:   deviance =  125284.3
Iteration 8:   deviance =  125284.3

Number of parameters: 28
Number of observations: 16720
Pseudo log-likelihood: -63420.144
R-squared: .17288283
Option strict is: on
------------------------------------------------------------------------------------------
                         |               Robust
    cum_sizes_concludedb |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
            log_distance |  -1.201793   .1801519    -6.67   0.000    -1.554884   -.8487016
         common_language |   1.138324   .4179556     2.72   0.006     .3191461    1.957502
               log_GDP_h |   .3913882   .1393914     2.81   0.005     .1181861    .6645903
               log_GDP_i |   .4823699   .0643716     7.49   0.000     .3562039     .608536
            log_GDP_pc_h |  -.6416376   .3619771    -1.77   0.076      -1.3511    .0678245
            log_GDP_pc_i |  -.1496204   .2003057    -0.75   0.455    -.5422123    .2429716
  net_agric_imports_pc_h |  -.2959744    2.40482    -0.12   0.902    -5.009335    4.417386
  net_agric_imports_pc_i |  -.7711362   .8465933    -0.91   0.362    -2.430429    .8881562
     log_agric_area_pc_h |   1.357242   .2203435     6.16   0.000     .9253767    1.789107
     log_agric_area_pc_i |  -.0698345   .2100094    -0.33   0.739    -.4814454    .3417763
  share_available_land_h |  -.0265334   .0123975    -2.14   0.032    -.0508319   -.0022348
  share_available_land_i |  -.0011298   .0060209    -0.19   0.851    -.0129305    .0106709
   log_water_resources_h |   .4094063   .1347031     3.04   0.002     .1453931    .6734194
   log_water_resources_i |  -.1177472   .1148511    -1.03   0.305    -.3428512    .1073567
log_agric_productivity_h |  -.4160468   .1383597    -3.01   0.003    -.6872268   -.1448669
log_agric_productivity_i |  -.1575241   .1354979    -1.16   0.245    -.4230951    .1080469
            corruption_h |  -.0111685   .0088038    -1.27   0.205    -.0284236    .0060866
            corruption_i |   .0340047   .0099084     3.43   0.001     .0145845    .0534248
               n_deals_h |   .0473115    .006525     7.25   0.000     .0345227    .0601003
                     rta |   .5814468   .5075564     1.15   0.252    -.4133453    1.576239
               _DGEO_h_1 |  -2.315515   .9760781    -2.37   0.018    -4.228593   -.4024373
               _DGEO_h_2 |  -1.866438   .6921475    -2.70   0.007    -3.223022   -.5098537
               _DGEO_h_3 |   -4.78573   .9438335    -5.07   0.000     -6.63561   -2.935851
               _DGEO_h_4 |  -1.629991   .8536095    -1.91   0.056    -3.303035    .0430527
               _EGEO_i_1 |  -4.033844   .8303639    -4.86   0.000    -5.661327    -2.40636
               _EGEO_i_2 |  -2.812537    .660103    -4.26   0.000    -4.106315   -1.518759
               _EGEO_i_3 |  -1.216558   .8424364    -1.44   0.149    -2.867703     .434587
               _EGEO_i_4 |  -2.868658   .7179761    -4.00   0.000    -4.275865   -1.461451
------------------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel append
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb //Get fitted values - RESET test according to http://privatewww.essex.ac.uk/
> ~jmcss/reset.do
(14086 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(14086 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var1 _D* _E* fit2, keep noconst strict // Estimate the
>  model with the additional regressor
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering
WARNING: fit2 has very large values, consider rescaling  or recentering
Number of regressors excluded to ensure that the estimates exist: 2
Number of observations excluded: 0
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0

. test fit2=0 //Test the significance of the additional regressor (this is equivalent to a t-t
> est on fit2)

 ( 1)  fit2 = 0

           chi2(  1) =    3.06
         Prob > chi2 =    0.0800

And without the strict option:

.  ppml cum_sizes_concludedb $explan_var _D* _E*, keep noconst

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering

Number of regressors excluded to ensure that the estimates exist: 1
Excluded regressors:  _EGEO_i_5
Number of observations excluded: 0

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1:   deviance =    365271
Iteration 2:   deviance =  196395.2
Iteration 3:   deviance =  151197.4
Iteration 4:   deviance =  142551.7
Iteration 5:   deviance =  141759.5
Iteration 6:   deviance =  141732.9
Iteration 7:   deviance =    141727
Iteration 8:   deviance =  141724.9
Iteration 9:   deviance =  141724.1
Iteration 10:  deviance =  141723.8
Iteration 11:  deviance =  141723.7
Iteration 12:  deviance =  141723.7
Iteration 13:  deviance =  141723.6
Iteration 14:  deviance =  141723.6
Iteration 15:  deviance =  141723.6
Iteration 16:  deviance =  141723.6
Iteration 17:  deviance =  141723.6
Iteration 18:  deviance =  141723.6
Iteration 19:  deviance =  141723.6
Iteration 20:  deviance =  141723.6
Iteration 21:  deviance =  141723.6
Iteration 22:  deviance =  141723.6
Warning:  variance matrix is nonsymmetric or highly singular

Number of parameters: 17
Number of observations: 21250
Pseudo log-likelihood: -71698.063
R-squared: .15788584
Option strict is: off
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0
---------------------------------------------------------------------------------
                |               Robust
cum_sizes_co~db |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
   log_distance |   -.763162          .        .       .            .           .
common_language |   1.293424          .        .       .            .           .
      log_GDP_h |   .3859217          .        .       .            .           .
      log_GDP_i |   .5476548          .        .       .            .           .
   log_GDP_pc_h |  -.4063192          .        .       .            .           .
   log_GDP_pc_i |   .4349815          .        .       .            .           .
      n_deals_h |   .0450038          .        .       .            .           .
            rta |   1.053489          .        .       .            .           .
      _DGEO_h_1 |  -36.47161          .        .       .            .           .
      _DGEO_h_2 |  -36.10625          .        .       .            .           .
      _DGEO_h_3 |  -39.25592          .        .       .            .           .
      _DGEO_h_4 |  -35.65588          .        .       .            .           .
      _DGEO_h_5 |  -34.80601          .        .       .            .           .
      _EGEO_i_1 |   17.76028          .        .       .            .           .
      _EGEO_i_2 |   17.81901          .        .       .            .           .
      _EGEO_i_3 |   19.31081          .        .       .            .           .
      _EGEO_i_4 |   17.85156          .        .       .            .           .
---------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel replace
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb //Get fitted values - RESET test according to http://privatewww.essex.ac.uk/
> ~jmcss/reset.do
(9556 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(9556 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var _D* _E* fit2, keep  noconst   // Estimate the mode
> l with the additional regressor
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering
WARNING: fit2 has very large values, consider rescaling  or recentering
Number of regressors excluded to ensure that the estimates exist: 1
Number of observations excluded: 0
Warning:  variance matrix is nonsymmetric or highly singular
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0

. test fit2=0

 ( 1)  fit2 = 0
       Constraint 1 dropped

           chi2(  0) =       .
         Prob > chi2 =         .


 ppml cum_sizes_concludedb $explan_var1 _D* _E*, keep noconst

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering

Number of regressors excluded to ensure that the estimates exist: 1
Excluded regressors:  _EGEO_i_5
Number of observations excluded: 0

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1:   deviance =  273234.5
Iteration 2:   deviance =  152042.2
Iteration 3:   deviance =  120736.9
Iteration 4:   deviance =  114467.6
Iteration 5:   deviance =  113764.5
Iteration 6:   deviance =  113741.3
Iteration 7:   deviance =  113738.5
Iteration 8:   deviance =  113737.5
Iteration 9:   deviance =  113737.1
Iteration 10:  deviance =  113736.9
Iteration 11:  deviance =  113736.9
Iteration 12:  deviance =  113736.9
Iteration 13:  deviance =  113736.9
Iteration 14:  deviance =  113736.9
Iteration 15:  deviance =  113736.9
Iteration 16:  deviance =  113736.9
Iteration 17:  deviance =  113736.9
Iteration 18:  deviance =  113736.9
Iteration 19:  deviance =  113736.9
Iteration 20:  deviance =  113736.9
Iteration 21:  deviance =  113736.9
Iteration 22:  deviance =  113736.9
Warning:  variance matrix is nonsymmetric or highly singular

Number of parameters: 29
Number of observations: 16720
Pseudo log-likelihood: -57646.44
R-squared: .19437534
Option strict is: off
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0
------------------------------------------------------------------------------------------
                         |               Robust
    cum_sizes_concludedb |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
            log_distance |  -.9636159          .        .       .            .           .
         common_language |   1.380642          .        .       .            .           .
               log_GDP_h |   .6158404          .        .       .            .           .
               log_GDP_i |   .5687462          .        .       .            .           .
            log_GDP_pc_h |  -.6418756          .        .       .            .           .
            log_GDP_pc_i |   .3423702          .        .       .            .           .
  net_agric_imports_pc_h |  -3.089397          .        .       .            .           .
  net_agric_imports_pc_i |  -.4997463          .        .       .            .           .
     log_agric_area_pc_h |   1.494319          .        .       .            .           .
     log_agric_area_pc_i |   .2836532          .        .       .            .           .
  share_available_land_h |  -.0432753          .        .       .            .           .
  share_available_land_i |  -.0165432          .        .       .            .           .
   log_water_resources_h |   .9540423          .        .       .            .           .
   log_water_resources_i |   .2379842          .        .       .            .           .
log_agric_productivity_h |   .0275815          .        .       .            .           .
log_agric_productivity_i |  -.0219905          .        .       .            .           .
            corruption_h |   -.005127          .        .       .            .           .
            corruption_i |   .0174166          .        .       .            .           .
               n_deals_h |   .0450861          .        .       .            .           .
                     rta |   .9991179          .        .       .            .           .
               _DGEO_h_1 |  -53.00853          .        .       .            .           .
               _DGEO_h_2 |  -54.51339          .        .       .            .           .
               _DGEO_h_3 |  -56.62995          .        .       .            .           .
               _DGEO_h_4 |  -53.55057          .        .       .            .           .
               _DGEO_h_5 |  -52.69731          .        .       .            .           .
               _EGEO_i_1 |   17.73117          .        .       .            .           .
               _EGEO_i_2 |   17.43558          .        .       .            .           .
               _EGEO_i_3 |   19.76491          .        .       .            .           .
               _EGEO_i_4 |   17.65751          .        .       .            .           .
------------------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel append
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb
(14086 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(14086 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var1 _D* _E* fit2, keep noconst // Estimate the model
> with the additional regressor
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering
WARNING: fit2 has very large values, consider rescaling  or recentering
Number of regressors excluded to ensure that the estimates exist: 1
Number of observations excluded: 0
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0

. test fit2=0

 ( 1)  fit2 = 0

           chi2(  1) =    3.30
         Prob > chi2 =    0.0692

Hope this time it is sort of understandable. Thanks so much!

All the best,
Klaus

[Joao Santos Silva]

Thanks, Klaus. Are you using the option "keep" for any special reason? Please run the regressions without this option, OK? Also, are variables like _EGEO_i_2 dummies?

All the best,

Joao

[Klaus Schmidt]

Thanks for your patience, Joao! I am using the "keep" option because otherwise many observations are dropped (about 1500) . Is there any way to find out which ones? I have very few positive values and worried that dropping observations would mean I drop too many of these non-zero observations....

The EGEO-variables are dummies (regional dummies), yes. Results don't change much without the keep options - still overfitting without the strict option (with the strict option it is fine), and low p-values in the RESET test (about 0.07).

Best wishes,
Klaus

[Joao Santos Silva]

Klaus,

The observations that are dropped are all equal to zero and have no information about the parameters of the model because they are perfectly predicted. So, please do not use "keep".

All the best,

Joao

[Klaus Schmidt]

Ok, fine, then I do it without keep - sorry, I wasn't aware of this!
But can I sue the strict option? And should I use the noconst with all dummy variables?

Thanks so much!!

[Joao Santos Silva]

Klaus,

Do not use the -strict- option either; just include all the dummies with or without constant.

All the best,

Joao

[Klaus Schmidt]

Thanks Joao, I have run the regressions without keep and strict and my results seem fine - the only remaining problem is the overfitting warning for some of the specifications. I assume I have to work on the specifications, then.

Thanks again and all the best,
Klaus

[Joao Santos Silva]

Dear Klaus,

Please post the results for one of the cases with the warning.

Best regards,

Joao

[Klaus Schmidt]

Thanks Joao, here it is (including the RESET tes):

ppml cum_sizes_concludedb $explan_var1 _D* _E*

note: checking the existence of the estimates
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering

Number of regressors excluded to ensure that the estimates exist: 0
Number of observations excluded: 0

note: _DGEO_h_5 omitted because of collinearity
note: _EGEO_i_1 omitted because of collinearity

note: starting ppml estimation
note: cum_sizes_concludedb has noninteger values

Iteration 1:   deviance =  276914.3
Iteration 2:   deviance =  153590.1
Iteration 3:   deviance =  121413.3
Iteration 4:   deviance =  114845.7
Iteration 5:   deviance =  114107.1
Iteration 6:   deviance =  114089.3
Iteration 7:   deviance =  114089.2
Iteration 8:   deviance =  114089.2

Number of parameters: 29
Number of observations: 16940
Pseudo log-likelihood: -57832.251
R-squared: .19419177
Option strict is: off
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0
------------------------------------------------------------------------------------------
                         |               Robust
    cum_sizes_concludedb |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
            log_distance |  -.9621258   .2204523    -4.36   0.000    -1.394204   -.5300471
         common_language |   1.373658   .3556246     3.86   0.000     .6766465    2.070669
               log_GDP_h |   .6187184   .1024241     6.04   0.000     .4179707     .819466
               log_GDP_i |   .5694416   .0714125     7.97   0.000     .4294755    .7094076
            log_GDP_pc_h |  -.6477244    .260485    -2.49   0.013    -1.158266   -.1371832
            log_GDP_pc_i |   .3404469   .1890304     1.80   0.072    -.0300458    .7109396
  net_agric_imports_pc_h |  -3.072139   2.371134    -1.30   0.195    -7.719476    1.575199
  net_agric_imports_pc_i |  -.4902605   .7200062    -0.68   0.496    -1.901447    .9209258
     log_agric_area_pc_h |   1.499046   .2665528     5.62   0.000     .9766118     2.02148
     log_agric_area_pc_i |   .2834167   .1867459     1.52   0.129    -.0825985    .6494318
  share_available_land_h |  -.0434304   .0123642    -3.51   0.000    -.0676637   -.0191971
  share_available_land_i |  -.0165234   .0080678    -2.05   0.041    -.0323359   -.0007109
   log_water_resources_h |   .9564134   .1871224     5.11   0.000     .5896602    1.323167
   log_water_resources_i |   .2367187   .1422818     1.66   0.096    -.0421485    .5155858
log_agric_productivity_h |   .0262156   .1456809     0.18   0.857    -.2593137     .311745
log_agric_productivity_i |  -.0219723   .1122832    -0.20   0.845    -.2420433    .1980987
            corruption_h |  -.0049471   .0083976    -0.59   0.556    -.0214061    .0115119
            corruption_i |   .0174745   .0099878     1.75   0.080    -.0021012    .0370502
               n_deals_h |   .0450622   .0055958     8.05   0.000     .0340946    .0560298
                     rta |    1.00482   .4719735     2.13   0.033      .079769    1.929871
               _DGEO_h_1 |    -.30352   .9718875    -0.31   0.755    -2.208384    1.601344
               _DGEO_h_2 |  -1.804123   .7840516    -2.30   0.021    -3.340835   -.2674096
               _DGEO_h_3 |  -3.925264   .9891404    -3.97   0.000    -5.863944   -1.986585
               _DGEO_h_4 |  -.8457818    .903183    -0.94   0.349    -2.615988    .9244245
               _EGEO_i_2 |  -.2960225   .6884029    -0.43   0.667    -1.645267    1.053223
               _EGEO_i_3 |   2.031076    .571698     3.55   0.000     .9105687    3.151583
               _EGEO_i_4 |  -.0746455   .5813259    -0.13   0.898    -1.214023    1.064732
               _EGEO_i_5 |  -2.756632   1.241112    -2.22   0.026    -5.189166   -.3240981
                   _cons |  -35.01173   5.870927    -5.96   0.000    -46.51854   -23.50493
------------------------------------------------------------------------------------------

. outreg2 using "Results/ppml_compare", excel append
Results/ppml_compare.xml
dir : seeout

.
.
. predict fit, xb
(13866 missing values generated)

. gen fit2=fit^2 //Square the fitted values
(13866 missing values generated)

. qui ppml cum_sizes_concludedb $explan_var1 _D* _E* fit2 
WARNING: log_GDP_h has very large values, consider rescaling  or recentering
WARNING: log_GDP_i has very large values, consider rescaling  or recentering
WARNING: share_available_land_h has very large values, consider rescaling  or recentering
WARNING: share_available_land_i has very large values, consider rescaling  or recentering
WARNING: log_water_resources_h has very large values, consider rescaling  or recentering
WARNING: log_water_resources_i has very large values, consider rescaling  or recentering
WARNING: log_agric_productivity_i has very large values, consider rescaling  or recentering
WARNING: corruption_h has very large values, consider rescaling  or recentering
WARNING: corruption_i has very large values, consider rescaling  or recentering
WARNING: n_deals_h has very large values, consider rescaling  or recentering
WARNING: fit2 has very large values, consider rescaling  or recentering
Number of regressors excluded to ensure that the estimates exist: 0
Number of observations excluded: 0
WARNING: The model appears to overfit some observations with cum_sizes_concludedb=0

. test fit2=0 //Test the significance of the additional regressor (this is equivalent to a t-t
> est on fit2)

 ( 1)  fit2 = 0

           chi2(  1) =    3.30
         Prob > chi2 =    0.0692

Thanks a million!
Klaus

[Joao Santos Silva]

It looks to me as if the base category of your dummies is a perfect predictor; are you generating the dummies using xi with the noomit option, and including all of them in the model? If you are doing that, then just run it without the constant.

All the best,

Joao