[Clyde Schechter]

The problem in #14 is revealed by the line:

(Std. Err. adjusted for 9 clusters in Industry11_num)

Now let's count your predictor variables. You have ib3.Industry11_num, and since we know there are 9 values of Industry11_num, that gives you 8 variables. There is then i.advisor_num--I don't know how many of these there might be, but there's at least one. That brings us to 9. Then from the output table I see two more shown: epsv55 and tsrv411. That makes 11. There may be more that aren't shown, I can't tell. Anyway, with 9 clusters you can't have more than 8 predictor variables and still be able to get an overall model chi square test. So that's the reason the chi square statistic is given as a missing value. You have more than exhausted your available degrees of freedom.

As for whether you should be concerned about it, that depends on your research goals. In most situations, the overall model chi square is irrelevant, and the inability to calculate it is of no importance (the coefficient standard errors and CIs and p-values are all unaffected by this). Only if the overall model chi square is for some unusual reason actually important to answering your research question would this be a matter of concern.

All of that said, 9 clusters is too few to use the cluster robust vce anyway. Its properties are based on asymptotics and require an ample number of clusters for validity. In samples with only a few clusters, as here, it can actually be worse than the ordinary vce. There is no consensus about how many is enough, but I think everyone would agree that 9 is too few.

[andy macrobarty]

Clyde. Many thanks for your answer. May I ask any reason for high chi square in a code when I use a robust option? Moreover, is it common practise to use robust option in multinomial logit model?

[Clyde Schechter]

I don't attach any particular meaning to the high chi square with the robust option. The robust statistics are often rather different from the ordinary ones, and the difference can be in either direction, and large or small in magnitude.

The robust statistics are just as valid in multinomial logit models as in any other, as far as I know. I wouldn't know whether it's a common practice or not; I haven't done any kind of survey of the literature. In my world, the use of the multinomial logit model itself is not very common.

[andy macrobarty]

----------------------------------------------------------------------------------
            TSR1 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
1                |
          edsf55 |   .1546826   .0275477     4.32   0.000     .0650901    .1730751
          tsrv44 |  -1.104019   1.3233271    -2.17   0.030    -5.913179    -.294859
            PBB3 |   .0004925   .0055063     0.09   0.929    -.0102997    .0112848
              FE |   2.334565   .12357994     4.36   0.000     .7352089    1.933921
    Insadnt1 |    .024105   .0055655     4.33   0.000     .0131968    .0350132
          LasdS |  -.213642826   .012380848    -6.27   0.000    -.4781267   -.2504385

[Marcos Almeida]

As a matter of fact, numbers are part and parcel of the daily routine in this forum.

That said, language is a crucial aspect as well.

Without a claryfing message, I fail to unveil the core message of # 19.

[andy macrobarty]

 
  Variable
Ln (Cash Compensation)

ROA%
2.826

TSR
0.91

Board
-0.022**

Nonexec
0.237***

Sales
-0.0481

Institutional ownership
-0.014