I am estimating a fixed effects model with the command xtreg
and every time I estimate, exactly the SAME model, I get
different estimation results: different coefficients,
and different levels of significance. Variables change the
level of significance from being non sig to being significant at 10%
to being significant at 5%!
I don't see how or why this may happen.
This is the command I am writting:
xtreg ltheil2ie lunemp lreedu lrprim lrsup lelectricity lelectricitysq ldepindex
lshare2 if (year>1997),fe;
(all variables are in logs)
The following three results are an example of how different my
results can be:
(1)
Fixed-effects (within) regression Number of obs = 157
Group variable (i): id2 Number of groups = 28
R-sq: within = 0.4867 Obs per group: min = 4
between = 0.3483 avg = 5.6
overall = 0.4355 max = 6
F(9,120) = 12.64
corr(u_i, Xb) = -0.0689 Prob > F = 0.0000
------------------------------------------------------------------------------
ltheil2ie | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lunemp | .2229948 .0249649 8.93 0.000 .173566 .2724236
lreedu | .1180611 .0428117 2.76 0.007 .0332969 .2028253
lrprim | -1.649794 .5064185 -3.26 0.001 -2.652467 -.6471206
lrsec | .2077582 .1722232 1.21 0.230 -.1332318 .5487482
lrsup | -.0384691 .088997 -0.43 0.666 -.2146771 .1377388
lelectricity | -.0027886 .0652681 -0.04 0.966 -.1320149 .1264376
lelectrici~q | .0155034 .0636444 0.24 0.808 -.110508 .1415149
lshare2 | -.2545261 .0878008 -2.90 0.004 -.4283656 -.0806867
ldepindex | .5007121 .2549344 1.96 0.052 -.0040403 1.005465
_cons | -2.365994 .9358431 -2.53 0.013 -4.218898 -.51309
-------------+----------------------------------------------------------------
sigma_u | .07689923
sigma_e | .10595896
rho | .34499518 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(27, 120) = 2.15 Prob > F = 0.0026
(2) Results of exactly the SAME model, just run at another time:
Fixed-effects (within) regression Number of obs = 157
Group variable (i): id2 Number of groups = 28
R-sq: within = 0.4629 Obs per group: min = 4
between = 0.3493 avg = 5.6
overall = 0.4298 max = 6
F(9,120) = 11.49
corr(u_i, Xb) = -0.0264 Prob > F = 0.0000
------------------------------------------------------------------------------
ltheil2ie | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lunemp | .2215847 .0258759 8.56 0.000 .1703523 .272817
lreedu | .0925941 .0434583 2.13 0.035 .0065496 .1786386
lrprim | -1.507966 .5257904 -2.87 0.005 -2.548994 -.4669373
lrsec | .2356492 .1899493 1.24 0.217 -.1404372 .6117357
lrsup | -.1277904 .0892154 -1.43 0.155 -.3044306 .0488498
lelectricity | -.0177035 .0670554 -0.26 0.792 -.1504686 .1150616
lelectrici~q | .0415298 .0638548 0.65 0.517 -.0848982 .1679579
lshare2 | -.2643193 .0881324 -3.00 0.003 -.4388153 -.0898234
ldepindex | .3461164 .2478695 1.40 0.165 -.1446479 .8368808
_cons | -1.976496 .9175012 -2.15 0.033 -3.793084 -.1599073
-------------+----------------------------------------------------------------
sigma_u | .07472599
sigma_e | .10806872
rho | .3234682 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(27, 120) = 1.90 Prob > F = 0.0103
(3) Results of exactly the SAME model, just run at another time:
Fixed-effects (within) regression Number of obs = 157
Group variable (i): id2 Number of groups = 28
R-sq: within = 0.5041 Obs per group: min = 4
between = 0.3077 avg = 5.6
overall = 0.4283 max = 6
F(9,120) = 13.56
corr(u_i, Xb) = -0.0703 Prob > F = 0.0000
------------------------------------------------------------------------------
ltheil2ie | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lunemp | .2167134 .0231063 9.38 0.000 .1709645 .2624622
lreedu | .0971098 .041291 2.35 0.020 .0153566 .178863
lrprim | -1.409938 .4957602 -2.84 0.005 -2.391509 -.4283672
lrsec | .2751781 .163676 1.68 0.095 -.0488889 .5992451
lrsup | -.0548135 .0821648 -0.67 0.506 -.2174942 .1078671
lelectricity | -.0651719 .0649134 -1.00 0.317 -.1936959 .063352
lelectrici~q | .0817531 .0609845 1.34 0.183 -.038992 .2024982
lshare2 | -.260746 .078065 -3.34 0.001 -.4153093 -.1061827
ldepindex | .6116727 .2411518 2.54 0.012 .134209 1.089136
_cons | -2.727725 .8835034 -3.09 0.003 -4.477 -.9784494
-------------+----------------------------------------------------------------
sigma_u | .08130385
sigma_e | .10147583
rho | .39096577 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(27, 120) = 2.75 Prob > F = 0.0001
As you can see, the coefficients are different every time, the F value too.
In particular the variable "lrsec" that is not sig the first and second time I
run the regression, it turns out significant at 10% the third time.
Similarly, the variable "ldepindex" has a coefficient of 0.5 an 10% level of
significance the first time I run the regression, while the coefficient is 0.346
the second time I estimate and is not significant, and the third time I run the
regression has a coefficient of 0.61 and is significant at 5%!
I really don't know what may be going on, since it is a quite standard model
that I am using. Do I need to set any "start" value or sth like that so that
every time I run the fixed effects model I get the same results?
Please, if someone has a hint on this I would really appreciate it.
Sincerely yours,
Maria.
-----------------------------------------------------------------
Santos, Maria Emma
Vanderbilt University
Email: [email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
