[Gretl-devel] remove unsignificant variables

[Allin Cottrell]

On Wed, 9 Jan 2008, Ignacio Diaz-Emparanza wrote:

> I am trying to write a script that, based on a regression with a 
> lot of variables and lags, drop unsignificant variables and 
> leave only the significant regressors...

Funny you should ask...  I've just finished adding documentation 
("help omit") and a GUI handle for the newish --auto option to the 
omit command.

ols y 0 xlist
omit --auto

sequentially omits the variable with the highest p-value, stopping 
when the highest (two-sided) p-value is less than 0.10.  You can 
adjust the cutoff p-value in this way:

omit --auto=0.05

The GUI model test item now has an option to do sequential 
omission, with a spinner for the p-value to be used.

This is in CVS and the Windows snapshot.

> I am having problems to identify the variable to drop in each 
> iteration because gretl sometimes detects "exact" collinearity 
> (but not perfect collinearity) and removes one regressor, so 
> that the regressors I write in the ols command are not the same 
> as gretl uses in the estimation, so it makes difficult to 
> associate the coefficients in $coeff with my list of 
> regressors...
> 
> I think changing the ols source may be some solutions to resolve 
> this problem. Gretl after the ols estimation could:
>  1- save the variable with highest t pvalue in a $ variable (for 
> example its ID number in a scalar)
>  2- save a list with the regressors used in the estimation (or a 
> row matrix with the ID numbers)

I quite like the idea of making available the list of regressors 
actually used, after estimation.  I'll think about how that might 
be done.

Allin.