Re: [Gretl-users] Nonlinear switching regression vs Loop constructs
On Tue, 1 Feb 2011, GREg Ory wrote:
First of all I would like to express most sincere thanks to
everybody
who answered on the matter. All gretl users must have seen this very
fine result of scripting and it�s rare to receive fast and quality
help and get exactly what you were looking for when dealing with
econometric packages. Allin and Jack � I may become your fan.
[...]
A question: why estimate �y � GDV� and not �y� as a dependent
variable. I understand that it�s equivalent in the linear regression
just want to know the reason for this.
The coefficient on the variable constructed by Jack's GDV()
function is constrained to be 1.0 in the model you're estimating.
So the obvious way of implementing this is to subtract it from the
"real" dependent variable.
The problem with the script is - NLS isn't estimated if FE = 1
because series u = $unit doesn't work with the current
periodicity. Some pointers on how to solve this would be nice.
Apparently your dataset is not recognized as a panel. See the
"setobs" command, or the menu item /Data/Dataset structure in the
GUI.
Another problem I encountered is my little knowledge of the NLS
estimation process � I can�t figure out how Jack uses linear
combinations, nelem function and covarince matrix to get the NLS
results.
The lincomb() (linear combination) function is just a way of doing
the matrix mutiplication X*b, in the case where X is not actually
a matrix but a list of series.
The nelem() function just returns the number of members in a list
of series.
In gretl's nls you need to construct a formula for the dependent
variable -- in terms of various parameters and right-hand-side
variables -- then tell gretl what are the parameters you want to
adjust to minimize the sum of squared residuals.
The model you wanted to estimate is basically linear (hence the
"lincomb" part), with an element of nonlinearity introduced by the
logistic term (hence the need for nls).
Allin Cottrell