On Fri, 24 Aug 2007, Franck Nadaud wrote:

> I am currently working on AIDS models to be estimated as sytems.
> So I use the command :
> 
> system
> ...
> ...
> end system
> 
> with restrictions inside. However those restrictions imply a 
> singular covariance matrix. In other pacakges (SAS for example) 
> estimation is done by deleting one equation, and the missing 
> coefficient is recovered by the restrictions.
> 
> I would like to know how GRETL handles such cases, and more 
> specifically whith which methods.

Gretl first estimates the unrestricted system, then estimates the 
restricted system.  The method for the second step is to form 
augmented versions of the X and y matrices using R, R' and q, as 
in the representation of the restriction as Rb = q.  Depending on 
whether an F-test or a likelihood ratio test is appropriate, we 
either save the original coefficient vector and covariance matrix 
(F-test) or the original log-likelihood.

The procedure is found in plugin/sysest.c in the gretl source: the 
main driver is the function system_estimate().  See William 
Greene, Econometric Analysis, 4e, section 7.3, "The Restricted 
Least Squares Estimator".  Note 7 on page 281 deals with the 
question of possible singularity of X'X, which doesn't matter so 
long as the restriction-augmented matrix is of full rank.

Allin.

> I would like to know if it is possible in system estimation to 
> recover all the coefficients and standards errors, because i 
> have to compute the AIDS elasticities.

Right now, it's not.  But I should add that.  The question is, how 
should one define the matrix of coefficients for an arbitrary set 
of equations?  Two possibilities occur to me:

(a) Create a matrix with a column for each equation, and a number 
of rows equal to the maximum number of coefficients in any 
equation.  For each column (equation) enter the coefficients in 
the order they're printed, starting from the top, and pad out the 
rest of the column with zeros if required.

(b) More complicated: create an n x g zero matrix, where n is the 
total number of distinct variables appearing on the right hand 
side of *any* equation and g is the number of equations.  Fill in 
the appropriate values.

I'm inclined towards (a) but I'd welcome comments.  Just to 
clarify, consider this example:

open greene13_1.gdt
system name=Grunfeld
  equation I_GE 0 F_GE C_GE
  equation I_WE 0 F_WE C_WE
end system
estimate Grunfeld method=sur

Under method (a) the coefficient matrix would be 3 x 2, with no 
zero entries, while under method (b) it would be 5 x 2, with two 
blocks of zeros, since the variables on the right-hand sides of 
the two equations differ.

Allin.

Reply via email to