Missing Wald Chi2 with xtnbreg

[Marvin Males]

Hello all,

I´m currently writing a paper about companies CVC activities using Compustat and some other data.
I want to test for an inverted u-shaped influence of CEO tenure on CVC activities.

When I now run my regression, the Wald Chi2 Statistic is not reported. This only happens when including the squared term of CEO tenure, without it the Wald Chi2 statistic gets reported.

. xtnbreg c_cvc_deal ceo_tenure_l1 ceoage_l1 ceo_gen_l1 ln_firm_size_l1_w rdi_l1_w roa_l1_w i.fyear if mysample1==1, fe
note: 3 groups (3 obs) dropped because of only one obs per group
note: 119 groups (1568 obs) dropped because of all zero outcomes

Iteration 0:   log likelihood = -349.00901  
Iteration 1:   log likelihood = -333.84992  
Iteration 2:   log likelihood = -331.12362  
Iteration 3:   log likelihood = -330.55602  
Iteration 4:   log likelihood = -330.44181  
Iteration 5:   log likelihood = -330.42608  
Iteration 6:   log likelihood = -330.42449  
Iteration 7:   log likelihood = -330.42416  
Iteration 8:   log likelihood = -330.42408  
Iteration 9:   log likelihood = -330.42406  

Conditional FE negative binomial regression     Number of obs     =        409
Group variable: gvkey                           Number of groups  =         27

                                                Obs per group:
                                                              min =          6
                                                              avg =       15.1
                                                              max =         17

                                                Wald chi2(22)     =      84.46
Log likelihood  = -330.42406                    Prob > chi2       =     0.0000

-----------------------------------------------------------------------------------
       c_cvc_deal |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
------------------+----------------------------------------------------------------
    ceo_tenure_l1 |   .0697204   .0203694     3.42   0.001     .0297971    .1096437
        ceoage_l1 |  -.0117564   .0148496    -0.79   0.429     -.040861    .0173483
       ceo_gen_l1 |  -17.93889    6690.06    -0.00   0.998    -13130.22    13094.34
ln_firm_size_l1_w |   .7518241   .2157166     3.49   0.000     .3290274    1.174621
         rdi_l1_w |  -.1024028   2.181632    -0.05   0.963    -4.378324    4.173518
         roa_l1_w |   1.140307   1.021765     1.12   0.264    -.8623166     3.14293
                  |
            fyear |
            2003  |  -.0959469   3092.255    -0.00   1.000    -6060.803    6060.612
            2004  |   .0124372   3202.991     0.00   1.000    -6277.735     6277.76
            2005  |  -.1220866   3267.806    -0.00   1.000    -6404.904     6404.66
            2006  |  -.1159723   3191.335    -0.00   1.000    -6255.018    6254.786
            2007  |   19.53205   2350.518     0.01   0.993    -4587.399    4626.464
            2008  |    19.3599   2350.518     0.01   0.993    -4587.572    4626.291
            2009  |   18.77836   2350.518     0.01   0.994    -4588.153     4625.71
            2010  |   18.91742   2350.518     0.01   0.994    -4588.014    4625.849
            2011  |    19.0153   2350.518     0.01   0.994    -4587.916    4625.947
            2012  |    18.9587   2350.518     0.01   0.994    -4587.973     4625.89
            2013  |   18.49565   2350.518     0.01   0.994    -4588.436    4625.427
            2014  |   18.92893   2350.518     0.01   0.994    -4588.003     4625.86
            2015  |   19.18668   2350.518     0.01   0.993    -4587.745    4626.118
            2016  |   18.91033   2350.518     0.01   0.994    -4588.021    4625.842
            2017  |   19.03373   2350.518     0.01   0.994    -4587.898    4625.965
            2018  |   19.53097   2350.518     0.01   0.993      -4587.4    4626.462
                  |
            _cons |  -16.90134   2350.522    -0.01   0.994     -4623.84    4590.037
-----------------------------------------------------------------------------------


. xtnbreg c_cvc_deal ceo_tenure_l1 ceo_tenure_sq_l1 ceoage_l1 ceo_gen_l1 ln_firm_size_l1_w rdi_l1_w roa_l1_w i.fyear if mysample1==1, fe
note: 3 groups (3 obs) dropped because of only one obs per group
note: 119 groups (1568 obs) dropped because of all zero outcomes

Iteration 0:   log likelihood = -345.48588  
Iteration 1:   log likelihood = -331.27649  
Iteration 2:   log likelihood = -325.80395  
Iteration 3:   log likelihood = -324.72132  
Iteration 4:   log likelihood = -324.43138  
Iteration 5:   log likelihood =  -324.3516  
Iteration 6:   log likelihood = -324.32864  
Iteration 7:   log likelihood = -324.32282  
Iteration 8:   log likelihood = -324.32156  
Iteration 9:   log likelihood = -324.32128  
Iteration 10:  log likelihood = -324.32121  
Iteration 11:  log likelihood = -324.32119  

Conditional FE negative binomial regression     Number of obs     =        409
Group variable: gvkey                           Number of groups  =         27

                                                Obs per group:
                                                              min =          6
                                                              avg =       15.1
                                                              max =         17

                                                Wald chi2(22)     =          .
Log likelihood  = -324.32119                    Prob > chi2       =          .

-----------------------------------------------------------------------------------
       c_cvc_deal |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
------------------+----------------------------------------------------------------
    ceo_tenure_l1 |   .2106574   .0465719     4.52   0.000     .1193781    .3019367
 ceo_tenure_sq_l1 |   -.010805   .0032116    -3.36   0.001    -.0170996   -.0045104
        ceoage_l1 |  -.0139913   .0146007    -0.96   0.338    -.0426081    .0146254
       ceo_gen_l1 |  -18.99538   10823.96    -0.00   0.999    -21233.57    21195.58
ln_firm_size_l1_w |    .871011   .2095682     4.16   0.000     .4602649    1.281757
         rdi_l1_w |  -.7506713   2.122157    -0.35   0.724    -4.910023     3.40868
         roa_l1_w |    1.25552   1.018909     1.23   0.218    -.7415045    3.252544
                  |
            fyear |
            2003  |  -.3496922   5236.553    -0.00   1.000     -10263.8     10263.1
            2004  |  -.3321408   5379.453    -0.00   1.000    -10543.87     10543.2
            2005  |  -.4833295   5414.189    -0.00   1.000     -10612.1    10611.13
            2006  |  -.4962658   5246.153    -0.00   1.000    -10282.77    10281.77
            2007  |   20.24154    3720.51     0.01   0.996    -7271.823    7312.307
            2008  |   20.08695    3720.51     0.01   0.996    -7271.978    7312.152
            2009  |   19.51208    3720.51     0.01   0.996    -7272.553    7311.577
            2010  |   19.62558    3720.51     0.01   0.996    -7272.439    7311.691
            2011  |   19.70619    3720.51     0.01   0.996    -7272.359    7311.771
            2012  |   19.71911    3720.51     0.01   0.996    -7272.346    7311.784
            2013  |   19.30338    3720.51     0.01   0.996    -7272.762    7311.368
            2014  |   19.76745    3720.51     0.01   0.996    -7272.298    7311.832
            2015  |   19.98728    3720.51     0.01   0.996    -7272.078    7312.052
            2016  |   19.68505    3720.51     0.01   0.996     -7272.38     7311.75
            2017  |   19.77708    3720.51     0.01   0.996    -7272.288    7311.842
            2018  |   20.21707    3720.51     0.01   0.996    -7271.848    7312.282
                  |
            _cons |  -7.976564   3746.677    -0.00   0.998    -7351.328    7335.375
-----------------------------------------------------------------------------------

. 
end of do-file

My supervisor said I have 3 options:
1. I just don´t mention the statistic at all. If I go with this option, is there something else I can report in order to show that the model is significant?
2. I mention the Wald Chi2 statistic. Then I have to come up with an explanation on why there is no output in this particular regression but in all the others. What can be the reason for this?
3. I need to find another way to compute the statistic manually for this specific regression. Is there any?

[Clyde Schechter]

There is something wrong here. The missingness of the overall model chi square is not a problem in any case--you shouldn't care about it in the first place as it is a meaningless test of a null hypothesis that nobody cares about. There is no reason to "show that the model is significant" as it makes no difference to your research goals whether it is or not, unless your pre-specified research goal is to test the joint null hypothesis that all of the model coefficients (including the year variables) are zero simultaneously.

But there is something definitely wrong with these outputs. There are several situations where the model chi square cannot be calculated. But these models do not meet any of those descriptions: you are not using robust vce, and if there are singleton problems in the second model, they would also be present in the first because all of you done is add a new continuous variable to the model, and one which doesn't even change the estimation sample through missing values. Second, your second model contains 23 predictor variables, but the output is showing 22 df. That's wrong. So I wouldn't trust anything else in it either.

To be clear, what disturbs me here is the juxtaposition of the two sets of results. The missing F-statistic in the second model could be attributable to any of the usual reasons for a missing F-statistic. In fact, the 1 df shortfall is strongly suggestive of being due to a singleton indicator variable. But if that were the case, we should see the same thing in the model without the quadratic term. (Actually, that raises an idea: how did that quadratic term get calculated? Are you sure it's correct and not mistakenly 0 in all but one observation? Check that variable before you do anything else.)

I would start by rebooting the computer and then doing an -update all- on your Stata. (If you are fully updated already, do -update, all force-) Then re-run the analyses. If that doesn't fix the problem, wait a day and see if anybody else on this Forum comes up with an explanation for what's going on. And if that doesn't happen, I recommend bumping this to StataCorp technical support.

[Marvin Males]

Thank you very much for your answer.

This shows how I calculated the variable:

. gen ceo_tenure_sq = ceo_tenure * ceo_tenure
(90 missing values generated)

. gen ceo_tenure_l1 = l1.ceo_tenure
(279 missing values generated)

. gen ceo_tenure_sq_l1 = l1.ceo_tenure_sq
(279 missing values generated)

Rebooting and updating Stata didn't fix the problem.
How does this problem affect my outputs and my results? Can I still use them?

EDIT: I did some testing with my regression excluding one of the variables each time:
If I run the regression without ln_firm_size_l1_w but with all other variables, the F statistic shows.
If I run the regression without rdi_l1_w but with all other variables, the F statistic shows.
If I run the regression without i.fyear but with all other variables, the F statistic shows.
If I run the regression without one of the other variables, the F statistic is missing, as described above.
If I run the same regression with the sqared term on a different dependent variable (Number of Patents), the F statistic gets reported.
Does this information help in any way?

[Clyde Schechter]

I think the information you provide is probably a good clue as to what is going on, but I do not know, myself, what to make of it.

I would not, at this point, trust the results. Something is causing the VCE matrix of the second model (and the others that have a missing F-statistic) to be singular. And it isn't one of the usual problems of singletons, or more variables than clusters because a) the reduced model would have the same problem, and b) you are not using cluster robust variance. I am especially puzzled that the F-statistic gets reported when you change the dependent variable: none of the causes of missing F statistic that I am familiar with care what the dependent variable is. So I'm going to have to bow out of trying to help at this point. I will continue to follow this thread in the hope that somebody else who understands what's going responds and we can both learn what's happening here. If you do not get a response from somebody else soon, I again suggest taking this up with StataCorp technical support.