Hi Allin,
"We already report the likelihood ratio test relative to the
constant-only model for most (all? I'm not sure) ML estimators.
What would you do differently?"
The deviance compares the estimated model to the full model!
Here is the code of my function (overdispersion does not work at the
moment).
Since you already test on overdispersion it would be great if the
calculation of the deviance and
the scaled deviance could follow (SAS gives them both).
function scalar DV (series y "Endogeneous variable of the last estimated
model", string dist "with following distribution: Normal, Bernoulli,
Binomial, Poisson", string disp[null] "Overdispersion?")
genr yh = $yhat
scalar p = cols($coeff)
if dist = "Normal"
matrix yh2 = yh
matrix y2 = y
matrix ssr = y2.-yh2
matrix ssr2 = ssr.^2
scalar DV = sum(ssr2)
elif dist = "Bernoulli"
smpl y=0 --restrict
matrix yh2 = yh
matrix y2 = y
matrix l3 = -ln(1.-yh2)
matrix part1 = l3
scalar sum1 = 2*sum(part1)
smpl y=1 --restrict --replace
matrix yh2 = yh
matrix y2 = y
matrix l1 = ln(y2./yh2)
matrix part2 = y2.*l1
scalar sum2 = 2*sum(part2)
scalar DV =sum1+sum2
elif dist = "Binomial"
smpl y=0 --restrict
matrix yh2 = yh
matrix y2 = y
matrix l3 = -ln(1.-yh2)
matrix part1 = l3
scalar sum1 = 2*sum(part1)
smpl y=1 --restrict --replace
matrix yh2 = yh
matrix y2 = y
matrix l1 = ln(y2./yh2)
matrix part2 = y2.*l1
scalar sum2 = 2*sum(part2)
smpl y>1 --restrict --replace
matrix yh2 = yh
matrix y2 = y
matrix l1 = ln(y2./yh2)
matrix l2 = ln(1.-y2)
matrix l3 = ln(1.-yh2)
matrix l4 = (1.-y2)
matrix part3 = y2.*l1+l4.*(l2.-l3)
scalar sum3 = 2*sum(part3)
scalar DV =sum1+sum2+sum3
elif dist = "Poisson"
smpl y>0 --restrict
matrix yh2 = yh
matrix y2 = y
matrix l1 = ln(y2./yh2)
matrix l2 = (y2.*l1) - (y2.-yh2)
scalar part1 = 2*sum(l2)
smpl y=0 --restrict --replace
matrix yh2 = yh
scalar part2 = 2*sum(yh2)
scalar DV = part1 + part2
endif
printf "Deviance = %g\n", DV
return DV
end function
I guess there are better ways to do these operations.
Well, that's for now.
Pindar
-----Ursprüngliche Nachricht-----
Von: gretl-users-bounces(a)lists.wfu.edu
[mailto:gretl-users-bounces(a)lists.wfu.edu] Im Auftrag von Allin Cottrell
Gesendet: Sonntag, 9. Mai 2010 18:34
An: Gretl list
Betreff: Re: [Gretl-users] Bug in quantile reg output export routine and a
question in terms of DEVIANCE
On Sun, 9 May 2010, Pindar wrote:
> I guess there is a small bug: I can only copy the output of a
> quantile reg as plain text, all other formats crush.
You should be able to copy quantile regression output in all
formats when the specification gives a single value of tau.
LaTeX output is not supported when multiple values of tau are
given, but that non-support was not properly registered in the
GUI. That is now fixed in CVS (i.e. the LaTeX option will not
appear in the copy or save menus).
> More important is my question, whether it would be possible to
> add the deviance goodness of fit measure to all non linear
> models?
We already report the likelihood ratio test relative to the
constant-only model for most (all? I'm not sure) ML estimators.
What would you do differently?
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
