Panel data with time-fixed effects

[Jacopo Michelon]

Dear readers,
I am working on a panel data and I am trying to decide between OLS, (one or two-way) fixed effects and random effects.
These are the characteristics of the panel data:

 xtset teamid year
       panel variable:  teamid (unbalanced)
        time variable:  year, 1990 to 2005
                delta:  1 unit

The first thing that I do is to run the Hausman test. The results seem to support for fixed effects. Prob>chi2 < 0.05

xtreg y x1 x2 x3 x4 x5 x6 x7, fe
estimates store fixed
xtreg y x1 x2 x3 x4 x5 x6 x7, re
estimates store random
hausman fixed random

 Test:  Ho:  difference in coefficients not systematic

                  chi2(7) = (b-B)'[(V_b-V_B)^(-1)](b-B)
                          =       19.67
                Prob>chi2 =      0.0063

Secondly I test for time-fixed effects. The results seem to support for fixed-time effects, but not for team-fixed effects.

 xtreg y x1 x2 x3 x4 x5 x6 x7 i.year, fe

Fixed-effects (within) regression               Number of obs     =        423
Group variable: teamid                          Number of groups  =         29

R-sq:                                           Obs per group:
     within  = 0.4175                                         min =          9
     between = 0.9225                                         avg =       14.6
     overall = 0.5200                                         max =         15

                                                F(21,373)         =      12.73
corr(u_i, Xb)  = 0.3649                         Prob > F          =     0.0000

--------------------------------------------------------------------------------------
                   y |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
---------------------+----------------------------------------------------------------
                  x1 |   .2976339   .0514635     5.78   0.000     .1964388    .3988289
                  x2 |  -1.197834   .6025895    -1.99   0.048    -2.382733   -.0129358
                  x3 |  -1.004244   1.703068    -0.59   0.556    -4.353061    2.344574
                  x4 |   14.66767   5.020646     2.92   0.004     4.795354    24.53999
                  x5 |  -2.157621   .4948095    -4.36   0.000    -3.130586   -1.184655
                  x6 |  -.8719717   .3822189    -2.28   0.023    -1.623546   -.1203977
                  x7 |   9.237686   9.208618     1.00   0.316    -8.869627      27.345
                     |
                year |
               1992  |   2.259566    2.68975     0.84   0.401     -3.02941    7.548541
               1993  |    .312609   2.843719     0.11   0.913    -5.279122     5.90434
               1994  |    .029145   2.904487     0.01   0.992    -5.682077    5.740367
               1995  |    .756507   2.997573     0.25   0.801    -5.137754    6.650768
               1996  |   7.127056   3.009425     2.37   0.018      1.20949    13.04462
               1997  |   3.209303   3.092813     1.04   0.300    -2.872233    9.290839
               1998  |    .056752   3.131287     0.02   0.986    -6.100435    6.213939
               1999  |  -8.521511   3.113896    -2.74   0.007     -14.6445   -2.398519
               2000  |   5.357031   3.118691     1.72   0.087    -.7753893    11.48945
               2001  |  -2.571675   3.084095    -0.83   0.405    -8.636068    3.492718
               2002  |  -3.055465   3.086275    -0.99   0.323    -9.124144    3.013214
               2003  |  -3.455438   3.112955    -1.11   0.268     -9.57658    2.665704
               2004  |  -.4980288   3.185596    -0.16   0.876    -6.762007    5.765949
               2005  |  -.3646745   3.818328    -0.10   0.924    -7.872823    7.143474
                     |
               _cons |   39.25244   10.12028     3.88   0.000     19.35249     59.1524
---------------------+----------------------------------------------------------------
             sigma_u |  2.9698973
             sigma_e |  9.4862393
                 rho |  .08926603   (fraction of variance due to u_i)
--------------------------------------------------------------------------------------
F test that all u_i=0: F(28, 373) = 0.98                     Prob > F = 0.5004

. testparm i.year

 ( 1)  1992.year = 0
 ( 2)  1993.year = 0
 ( 3)  1994.year = 0
 ( 4)  1995.year = 0
 ( 5)  1996.year = 0
 ( 6)  1997.year = 0
 ( 7)  1998.year = 0
 ( 8)  1999.year = 0
 ( 9)  2000.year = 0
 (10)  2001.year = 0
 (11)  2002.year = 0
 (12)  2003.year = 0
 (13)  2004.year = 0
 (14)  2005.year = 0

       F( 14,   373) =    4.21
            Prob > F =    0.0000

At this point, it seems that I should keep only the time-dummy variables in my model. Is it so? Do I need to run --regress-- with the option --vce(cluster teamid)-- now? Thank you.

regress y x1 x2 x3 x4 x5 x6 x7 i.year, vce(cluster teamid)

[Carlo Lazzaro]

Jacopo:
as the F-test of alL u_I being=0 does not raech statistical significance, pooled OLS (with the option --vce(cluster teamid)) outperforms -xtreg, fe-.
However, I would keep the -xs-, too in your POLS model.

[Jacopo Michelon]

Thanks for your quick answer, Carlo. Can I ask you what you mean by keeping the -xs-?

[Carlo Lazzaro]

Jacopo:
yes, of course.
I meant your x1....x7 predictors.

[Abdullah Helal]

Hello Carlo and Jacopo,

Could you please tell us how did you conclude that individual-fixed effect is not appropriate?
+
Is it safe to call the last line of code a time fixed-effect estimation?
+
What is the difference between vce(cluster id) VS. cluster(id)? If both applicable, which one is better?


Thanks,
Abdullah

[Carlo Lazzaro]

Abdullah:
welcome to this forum.
-

F test that all u_i=0: F(28, 373) = 0.98

answers your first question;
--i.time- is often called time-fixed effect (although you cannot retrieve it from -xtreg postestimation-);
- I usually go vce(cluster id), when needed; I'm not familiar with the other notation that probably dates back to some Stata releases ago.

[Abdullah Helal]

Thank you so much Carlo.