Re: st: Heckprob problems

 From "Fernando Rios Avila" To Subject Re: st: Heckprob problems Date Mon, 07 Feb 2011 20:53:49 -0500

```Well, I ve been checking ur results and also simulating some examples with other data. From what i see in your last output, it seems that you have to recheck again the nature of your data.
If u see carefully, u have a lot of problems with the model convergence, which seems to be explained because one of the parameters, "rho", is constantly reaching the lower bound of its possible values (-1). From my experience it could be that you are doing too much of a good job explaining the selection model.
What i suggest is to check the correlation of all the variables and see if there is any suspiciously high or low correlations between your explainatory variables.
Best regards

>>> Qing Gong Yang <qinggong.yang@gmail.com> 02/07/11 8:05 PM >>>
Thanks for this.

I did use
heckprob y1 x1 x2 x3 x4, sel (y2= x1 x5 x6).

There are about 300 observations, a little bit over 200 are censored.

for one regression, I estimate heckprob y1 x1 x2 x3 x4, sel(y2= x5 x6
x1) just for the convenience of output to table, yet the estimation
does not work any more.

Here is an example

.heckprob slc1 x1 leader bte counter efficiency  vig divestment
coordination   if efficiency<2, sel(ncomp=  x1 x2  lnms lnsmall )

Fitting probit model:

Iteration 0:   log likelihood = -52.340138
Iteration 1:   log likelihood = -30.883946
Iteration 2:   log likelihood = -28.066265
Iteration 3:   log likelihood = -27.746348
Iteration 4:   log likelihood = -27.740815
Iteration 5:   log likelihood = -27.740812

Fitting selection model:

Iteration 0:   log likelihood = -176.79924
Iteration 1:   log likelihood = -126.10593
Iteration 2:   log likelihood = -123.48405
Iteration 3:   log likelihood = -123.44962
Iteration 4:   log likelihood = -123.44961

Comparison:    log likelihood = -151.19042

Fitting starting values:

Iteration 0:   log likelihood =  -58.91751
Iteration 1:   log likelihood =   -29.3684
Iteration 2:   log likelihood = -23.529911
Iteration 3:   log likelihood = -21.781926
Iteration 4:   log likelihood = -21.505145
Iteration 5:   log likelihood = -21.497738
Iteration 6:   log likelihood = -21.497731

Fitting full model:

Iteration 0:   log likelihood = -232.35861  (not concave)
Iteration 1:   log likelihood = -199.37133  (not concave)
Iteration 2:   log likelihood = -188.63839
Iteration 3:   log likelihood = -156.78336  (not concave)
Iteration 4:   log likelihood = -148.90122
Iteration 5:   log likelihood = -145.49084
Iteration 6:   log likelihood = -144.70601
Iteration 7:   log likelihood = -144.55228
Iteration 8:   log likelihood = -144.53009
Iteration 9:   log likelihood = -144.52176
Iteration 10:  log likelihood = -144.51786
Iteration 11:  log likelihood = -144.51461
Iteration 12:  log likelihood = -144.51368
Iteration 13:  log likelihood = -144.51337
Iteration 14:  log likelihood = -144.51303  (backed up)
Iteration 15:  log likelihood = -144.51293
Iteration 16:  log likelihood = -144.51286  (not concave)
Iteration 17:  log likelihood = -144.51282  (not concave)
Iteration 18:  log likelihood = -144.51282  (not concave)
Iteration 19:  log likelihood = -144.51281  (not concave)
Iteration 20:  log likelihood = -144.51281  (not concave)
Iteration 21:  log likelihood =  -144.5128  (not concave)
Iteration 22:  log likelihood =  -144.5128  (not concave)
Iteration 23:  log likelihood =  -144.5128  (not concave)
numerical derivatives are approximate
nearby values are missing
Iteration 24:  log likelihood =  -144.5128  (not concave)
numerical derivatives are approximate
nearby values are missing
Iteration 25:  log likelihood =  -144.5128  (not concave)
numerical derivatives are approximate
nearby values are missing
Iteration 26:  log likelihood =  -144.5128  (not concave)
numerical derivatives are approximate
nearby values are missing
Iteration 27:  log likelihood =  -144.5128

Probit model with sample selection              Number of obs      =       294
Censored obs       =       209
Uncensored obs     =        85

Wald chi2(8)       =  13603.11
Log likelihood = -144.5128                      Prob > chi2        =    0.0000

------------------------------------------------------------------------------
|      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
slc1         |
x1    |   .0050248    .012672     0.40   0.692    -.0198118    .0298614
leader |    .595136   .4217066     1.41   0.158    -.2313938    1.421666
bte |   .8001224   .2163814     3.70   0.000     .3760227    1.224222
counter |  -.4242046   .0052946   -80.12   0.000    -.4345818   -.4138273
efficiency |  -.3950646   .0125419   -31.50   0.000    -.4196463   -.3704829
vig |   .2503349   .4029704     0.62   0.534    -.5394726    1.040142
divestment |   .0444567    .432335     0.10   0.918    -.8029043    .8918178
coordination |   .8577901   .3261174     2.63   0.009     .2186118    1.496968
_cons |  -1.192662   .5165695    -2.31   0.021     -2.20512   -.1802043
-------------+----------------------------------------------------------------
ncomp        |
x1      |   .5470875   .0137725    39.72   0.000     .5200939    .5740811
x2 |   -.060799   .2821573    -0.22   0.829    -.6138171    .4922191
lnms |   .0202606    .001606    12.62   0.000     .0171128    .0234084
lnsmalls |   .0329228   .0008017    41.06   0.000     .0313514    .0344941
_cons |  -2.623737   .1526243   -17.19   0.000    -2.922875   -2.324599
-------------+----------------------------------------------------------------
/athrho |  -16.16538   12092.39    -0.00   0.999    -23716.82    23684.49
-------------+----------------------------------------------------------------
rho |         -1   4.38e-10                            -1           1
------------------------------------------------------------------------------
LR test of indep. eqns. (rho = 0):   chi2(1) =    13.36   Prob > chi2 = 0.0003

when I estimate
heckprob slc1  x1 leader bte counter efficiency  vig divestment
coordination   if efficiency<2, sel(ncomp= lnms lnsmall x1 x2   )

it does not work any more.

On Tue, Feb 8, 2011 at 1:01 PM, Fernando Rios Avila
<ecofrax@langate.gsu.edu> wrote:
> Well, my only suggestion is to be sure that your y1 variable is appropriatly censored, since it seems that you are not impossing any censorship on the selection equation. Perhaps what u want to do is:
>
> heckprob y1 x1 x2 x3 x4, sel (y2= x1 x5 x6)
> assuming that y2 =1 when the selection holds, and y2=0 when its not observed
> Finally, i wonder which are the variables that you were switching possitions.
>
> Best regards
>
> Fernando Rios Avila
>
>
>>>> Qing Gong Yang <qinggong.yang@gmail.com> 02/07/11 6:54 PM >>>
> I am trying to estimate a probit model with selection using Stata 9 and 10's
> Heckprob command in the following form:
> Heckprob y1 x1 x2 x3 x4, sel (y2 x1 x5 x6).
>
> I came across several problems:
> first, the results are not stable. They change even when I just change the
> order of the variables.
> second,  funny results sometimes with very big Wald Chi2 value and very big
> Z values.
> third, The maximum log likelihood estimation seems to be critised by some as
> inconsistent too.
>
> Any suggestions?
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```