Hi again.

after investigating further - because the results were quite unsatisfactory - I found out that:

1) there is no memory problem with creating the unit dummies if done via 'dummify' (took millisecs for 865 dummies).
    Using 'genr unitdum' even did not succeed when shrinking down the dataset to only include 30 units. I guess there is some bug.

2) As theory tells us the FE and LSDV estimates and standard errors are the same:

                  FE OECD                 LSDV OECD 
const              -37.89**            

l_rgdppc1           1.388**          1.388**
                      (0.1428)             (0.1428)

l_pop1              1.105**            1.105**
                      (0.4690)             (0.4690)

l_rgdppc2           1.412**          1.412**
                      (0.1586)             (0.1586)

l_pop2              2.030**            2.030**
                      (0.3567)             (0.3567)

rta_rw             0.2072**         0.2072**
                     (0.04146)            (0.04146)

EU                 0.3190**         0.3190**
                     (0.03560)            (0.03560)

3) It does make a huge difference if one applies the mentioned formula of  $V[\widehat{u_{ijt}}]$  to the two models:

FE forecast -> not jet available in GREL

own calculation for se                        1.55001760177788    ->  what I meant with unsatisfactory

LSDV forecast
 obs    l_exp1to2          forecast            se                    95%-interval

    1:01     23.264466    22.893993     0.382118    22.144998 - 23.642989

own calculation for                            0.382118174334948

Fortunately I now can use the LSDV estimator. But the estimation process for LSDV takes really long compared to FE.


Hi Allin.

the LSDV method in my dataset does not work because GRETL stops in the 'genr unitdum' command with 'out of memory'.
Then I have to restart GERTL in order to work further.
Ok. thanks so far. I gonna use the mentioned formula for calculation of confidence intervals in FE model.

Have a nice weekend

On Tue. 9 Aug 2011. Leon Unger wrote:
I want to calculate $V[\widehat{u_{ijt}}]$ for *Fixed Effect *models.
In a first step I reproduce the forecast error variance of a pooled OLS 
Using the following formula

$V[\widehat{u_{ijt}}] = V[{\epsilon}] 

which provided the same values as GRETL produces with $fcerr.
Can I apply the same calculation after the Fixed Effect?
That would, I think, ignore the uncertainty due to estimation
of the fixed effects. One option would be to estimate the
fixed effects model manually, by adding the unit dummies to
the dataset and applying OLS (gretl estimates the FE model
by de-meaning the data).

Allin Cottrell
Gretl-users mailing list