F value is missing in FE model

[Geralt Ji]

Hi Statalists,
I'm a Stata beginner and currently dealing with an unbalanced panel data in Stata. After running fixed-effect model, I'm not sure how to fix missing F-statistics value and what makes it happen?
waveid variable is individual identifier, which means every respondent has their own unique value for waveid variable.
And all variables in this model have 48,610 observations except for hp, ex, bt. These three variables have 45,671 observations.

. xtreg oop hp ex bt $xvars_fe, fe i(waveid) vce(robust)

Fixed-effects (within) regression               Number of obs     =     45,671
Group variable: waveid                          Number of groups  =     21,236

R-sq:                                           Obs per group:
     within  = 0.0573                                         min =          1
     between = 0.0846                                         avg =        2.2
     overall = 0.0815                                         max =          3

                                                F(19,21235)       =          .
corr(u_i, Xb)  = -0.4441                        Prob > F          =          .

                                   (Std. Err. adjusted for 21,236 clusters in waveid)
-------------------------------------------------------------------------------------
                    |               Robust
                oop |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------------+----------------------------------------------------------------
                 hp |   1281.577   109.3248    11.72   0.000     1067.293    1495.862
                 ex |   370.7177   72.40659     5.12   0.000     228.7953    512.6401
                 bt |   1453.749   69.21981    21.00   0.000     1318.073    1589.425
               age1 |   82.52885    16.4526     5.02   0.000      50.2805    114.7772
               age2 |   .3782986   .1484827     2.55   0.011     .0872613    .6693359
             Lowses |   331.4415   57.25253     5.79   0.000     219.2222    443.6608
          Mediumses |   533.9892   61.07942     8.74   0.000      414.269    653.7095
            Highses |   798.7355   68.63578    11.64   0.000     664.2042    933.2668
           Vhighses |   1167.934   84.30342    13.85   0.000     1002.692    1333.175
            Defacto |  -499.3382    106.983    -4.67   0.000    -709.0331   -289.6433
  SeparatedDivorced |  -798.5498   151.8375    -5.26   0.000    -1096.163   -500.9368
            Widowed |   -609.736   172.8746    -3.53   0.000    -948.5832   -270.8887
NeverMarriedDefacto |  -131.2468   113.3167    -1.16   0.247     -353.356    90.86244
    marital_missing |  -1242.613   183.8037    -6.76   0.000    -1602.882   -882.3436
        Certificate |  -174.7789   130.0111    -1.34   0.179    -429.6106    80.05271
               Dipl |  -476.4853    153.694    -3.10   0.002    -777.7371   -175.2335
               Bach |  -588.5644   170.4818    -3.45   0.001    -922.7216   -254.4073
        edu_missing |  -629.1102   140.7475    -4.47   0.000     -904.986   -353.2343
       EmployerSelf |  -88.94381   106.6972    -0.83   0.405    -298.0784    120.1907
 employment_missing |   211.4715   58.08135     3.64   0.000     97.62761    325.3153
              _cons |  -2960.399    428.011    -6.92   0.000    -3799.333   -2121.465
--------------------+----------------------------------------------------------------
            sigma_u |  3298.6995
            sigma_e |  2595.5526
                rho |   .6176198   (fraction of variance due to u_i)
-------------------------------------------------------------------------------------

.
end of do-file

.

[Clyde Schechter]

You have used cluster-robust variance estimation. When you do that, the overall model F-statistic becomes undefined if there are singleton clusters. The output of the regression shows that the minimum number of observations per cluster is 1, so you do have that. That explains why the F statistic is missing.

The only way to "fix" this is to not use cluster robust variance estimation. However, why do you want to "fix" this? It probably isn't broken in the first place. The model F statistic is a test of the hypothesis that all of the regression coefficients are simultaneously 0. That hypothesis is almost never of any interest or importance in real life. In real life, usually some of the variables (in fact, most commonly, most of the variables) are included to adjust for their confounding effects ("control variables"), and nobody cares about their coefficients. So unless your research goals explicitly require you to test the hypothesis that all of these coefficients are simultaneously zero, there is no problem here. Just leave everything alone and move on.

[Carlo Lazzaro]

Geralt:
as an aside to Clyde's excellent explanation, you might be interested in taking a look at -help j_robustsingular-.
More substantively, have you already considered to add -i.wave- in the righ-hand side of your regression equation?
Then you can check -i.wave- joint statistical significance via -testparm-.