Comparing fixed effects coefficients between four groups

[Hale Isaac]

Hi,
I am running fixed effects models across four subsamples. The subsamples are four different countries. My independent variable measures food security (1. High food secure, 2. Moderately food secure, 3. Low food secure). I would like to test which coefficients are statistically different across models. I’ve run the following model:

 xtreg WB i.foodcat2##i.country, fe
note: 504.country omitted because of collinearity
note: 788.country omitted because of collinearity
note: 818.country omitted because of collinearity

Fixed-effects (within) regression               Number of obs     =     16,511
Group variable: Findid                          Number of groups  =      6,454

R-sq:                                           Obs per group:
     within  = 0.0160                                         min =          2
     between = 0.0600                                         avg =        2.6
     overall = 0.0423                                         max =          4

                                                F(8,10049)        =      20.46
corr(u_i, Xb)  = 0.0717                         Prob > F          =     0.0000

----------------------------------------------------------------------------------
              WB |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
        foodcat2 |
            Mod  |    .442859   .1088893     4.07   0.000     .2294143    .6563038
            Low  |   .5037213   .1274802     3.95   0.000     .2538347     .753608
                 |
         country |
        Morocco  |          0  (omitted)
        Tunisia  |          0  (omitted)
          Egypt  |          0  (omitted)
                 |
foodcat2#country |
    Mod#Morocco  |   .1495675   .1284013     1.16   0.244    -.1021248    .4012599
    Mod#Tunisia  |   -.276361    .132398    -2.09   0.037    -.5358876   -.0168345
      Mod#Egypt  |  -.2140328   .1665799    -1.28   0.199    -.5405628    .1124972
    Low#Morocco  |   .3593088   .1499573     2.40   0.017     .0653625     .653255
    Low#Tunisia  |  -.1939231   .1521641    -1.27   0.203    -.4921952     .104349
      Low#Egypt  |  -.1114142    .194939    -0.57   0.568    -.4935336    .2707051
                 |
           _cons |  -.3294008   .0363028    -9.07   0.000    -.4005615   -.2582401
-----------------+----------------------------------------------------------------
         sigma_u |  1.2589113
         sigma_e |  1.3446639
             rho |  .46709923   (fraction of variance due to u_i)
----------------------------------------------------------------------------------
F test that all u_i=0: F(6453, 10049) = 2.06                 Prob > F = 0.0000

. testparm i.foodcat2##i.country

 ( 1)  1.foodcat2 = 0
 ( 2)  2.foodcat2 = 0
 ( 3)  1.foodcat2#504.country = 0
 ( 4)  1.foodcat2#788.country = 0
 ( 5)  1.foodcat2#818.country = 0
 ( 6)  2.foodcat2#504.country = 0
 ( 7)  2.foodcat2#788.country = 0
 ( 8)  2.foodcat2#818.country = 0

       F(  8, 10049) =   20.46
            Prob > F =    0.0000

I think these results mean that we reject the null that the coefficients for all foodcat and country combinations are jointly equal to zero. But I do not think these answer the question of whether (and which country/s) the coefficient of foodcat2 differ across my models (i.e. which ones are statistically significant and which are not).

(I know my within R sq is low, but I am just concerned with testing significance of coefficients across models).

[Carlo Lazzaro]

Isaac:
see -lincom-, being aware that the -fe- estimator wipes out time-invariant predictors.

[Hale Isaac]

Thank you for brining my attention to lincom Carlo. Yes I have indeed only included time varying indepvars in my fe regressions.