Re: st: Panel with low within variance and -xtfmb-

["Rodrigo A. Alfaro" <raalfaroa@gmail.com>]

Late, but not too late. FEVD procedure is different than taking FE "by 
hand". You have a difficult problem, which is small within variation in your 
x's. Probably you should try to deal with the endogeneity using IV variables 
once you take an average of your data (sort of BE estimator). With some 
within variation, you could also work with the Hausman-Taylor estimator 
(xthtaylor). Good luck, RA.



----- Original Message ----- 
From: "Ivan Etzo" <ivanetzo@hotmail.com>
To: <statalist@hsphsun2.harvard.edu>; <nicola.baldini2@unibo.it>
Sent: Tuesday, October 09, 2007 5:10 PM
Subject: Re: st: Panel with low within variance and -xtfmb-



Thank you Nicola,

I read the paper and used the command -xtfevd-. I then runned a fixed effect
estimation using -regress- with -xi:- to generate groups dummies, which
allows for the estimation of time invariant variables,  but I found quite
different coefficient for the time invariant (while time variant are the
same). I'm now wondering which is the difference of using the two different
techinques..

Moreover, estimates of time variant variables from -xtfevd- are the same as
those obtained with -xtreg- ( fe opt). I understand that -xtreg- is a within
estimation which takes into account only variation across time (years)
within each individual, the point is that my time variant variables are
almost time invariant", that is their within variation is very small
compared to the between one.

Ivan

-------Original Message-------

From: nicola.baldini2@unibo.it
Date: 10/06/07 13:12:03
To: statalist@hsphsun2.harvard.edu
Cc: ivanetzo@hotmail.com
Subject: Re: st: Panel with low within variance and -xtfmb-

Package -xtfevd- from http://www.polsci.org/pluemper may help. It is a
module to estimate time-invariant and rarely changing variables in panel
data models with fixed effects.
Nicola

At 02.33 05/10/2007 -0400, "Ivan Etzo" wrote:

> Hi,
>
> I'd like to test a basic form of my model (panel T=3D7 N=3D380) with= one
regressor that is time invariant and two regressors that have low wi= thin
variance with the following percentage
> Y: 15%
> X1: 0.6%
> X2:0.6%
> X3:0
>
> hausman test rejects the RE model (using -xtreg-) so I cannot estima= te
the model with the time invariant regressor that for me it's very impo=
rtant.
> I tried the -xtfmb- which is an implementation of the Fama and MacBe= th
(1973) two step procedure, I report the very clear description of the =
command from the help
> -xtfmb- "is an implementation of the Fama and MacBeth (1973) two ste= p
procedure. The procedure is
>    as follows: In the fi= rst step, for each single time period a
cross-sectional regression is
> =     performed. Then, in the second step, the final coeffic= ient
estimates are obtained as the
>    average of the f= irst step coefficient estimates."
>
> The coefficient I obtain from this procedure are practically the sam= e
with the RE model.. I attached the results  below. I wonder whethe= r the
-xtfmb- produces inconsistent estimates as well.. any help would be= very
appreciated.
> Ivan
>
> Random-effects GLS regression = ;Number of obs      =3D 2660
> Group varia= ble (i): region Number of groups   =3D 380
>
> R-sq:  within =3D 0.0097 Obs per group: min =3D = 7
> between =3D 0.8153 avg =3D 7.0
> overall =3D 0.= 7953 max =3D 7
>
> Random effects u_i ~ Gaussian Wald chi2(3)  &nbs=
> p;    =3D 1665.29
> corr(u_i, X) =3D 0 (assumed= ) Prob > chi2        =3D = 0.0000
>
>
> lnmig | Coef.   Std. Err. = z    P>|z|     [95% Conf. Inte= rval]
>
> lndpop | .9560363   .033868= 9 28.23   0.000     .8896544 1.02= 2418
> lnopop | 1.023108   .0338689 30.21  = ; 0.000     .9567262 1.08949
> lndist | -.= 3231335   .0587596 -5.50   0.000  &nbs=
> p; -.4383002 -.2079668
> _cons | -21.17337    .= 799259 -26.49   0.000    -22.73989 -19= =2E60685
>
> sigma_u | .68680964
> sigma_e |&n=
> bsp;.27499318
> rho | .86183578   (fraction of varia= nce due to u_i)
>
>
>
> Fama-MacBeth (1973) Two-Step procedure&nbs=
> p;Number of obs     =3D 2660
>  Num.= time periods =3D 7
>  F(  3,   &nbs=
> p; 6)     =3D 11189.96
>  Prob > = F          =3D 0.0000  avg. R-squared    =3D 0.7962
> &=
> nbsp;
> |        &nb=
> sp;   Fama-MacBeth
> lnmig |     = ; Coef.   Std. Err.      t P>|= t|     [95% Conf. Interval]
>   = ;
> lndpop |   .9552167   .0101939  &=
> nbsp; 93.70 0.000     .9302731 .9801604
> = lnopop |   1.017995   .0061794   164.7= 4 0.000     1.002874 1.033115
> lndist&nbs=
> p;|  -.3232731   .0117134   -27.60 0.000&nb=
> sp;   -.3519349 -.2946114
> _cons |  -21.08686&=
> nbsp;  .1716352  -122.86 0.000    -21.50683=  -20.66688
*
*   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/