EViews.com

Hello again,

scrolling through the User Guide, I was not able to validate whether eviews 6 uses the LSDV-estimator or the within-estimator to compute the fixed effects.

As I have quite a large number of entities (~350), I a am litte bit concerned that a LSDV would drastically reduce the degree's of freedom & the efficiency of the model.

Is there an easy way to use the within-estimator ("demeaned OLS algorithm") with eviews in a one- and a two-way model? Is it then still possible to combine it with random effects in the other dimension?

Thanks again for your help.

Cheers
Jörn

I'm not certain what distinction you are trying to make between LSDV and the within-estimators. Using common terminology, they are the same; the within estimator demeaning is a computational trick for estimating the fixed effect / dummy variable model without forming the full moment matrix (it's equivalent to performing the partitioned inverse). In both cases, you will lose your degrees-of-freedom.

EViews allows for mixing fixed in one dimension with random effects in a second dimension. There are some restrictions on when this can be done (I'll have to look up the conditions), but it is certainly supported for the standard balanced cases.

Hello,

as I understand it, the specification between LSDV and within-estimator is different. In LSDV you are including k+n regressors (the k X's, the n-1 binary variables, and the intercept) whereas you do the computational trick (demeaning the variables) in the other case.

In the end the estimators are the same, but in the LSDV case you are loosing more degrees of freedom as you introduce more regressors (i.e. n-1 variables for company or year dummies). Consequently, you would run up against a problem of having only a few degrees of freedom when you have a lot of entities (besides this being a tedious exercise..).

But as I understand your reply, eviews uses the computational trick (so you do not have to make the demeaning transformation manually). I was just a little bit confused because in the estimation output window it says "Cross-section fixed (dummy variables)" when I include fixed effects.

Did I understand you right?

Thanks
Jörn

Hello,

as I understand it, the specification between LSDV and within-estimator is different. In LSDV you are including k+n regressors (the k X's, the n-1 binary variables, and the intercept) whereas you do the computational trick (demeaning the variables) in the other case.

In the end the estimators are the same, but in the LSDV case you are loosing more degrees of freedom as you introduce more regressors (i.e. n-1 variables for company or year dummies). Consequently, you would run up against a problem of having only a few degrees of freedom when you have a lot of entities (besides this being a tedious exercise..).

But as I understand your reply, eviews uses the computational trick (so you do not have to make the demeaning transformation manually). I was just a little bit confused because in the estimation output window it says "Cross-section fixed (dummy variables)" when I include fixed effects.

Did I understand you right?

Thanks
Jörn

When you demean data you lose the same number of degrees of freedom. It's just that bad software might not remind you about this. As you point out they are the same estimators; the computational technique for getting to the answer doesn't affect the statistical properties.