3:47 p.m.
Dear Gretl community,
When I try to perform a "White's test for heteroskedasticity" on OLS
models with only one term (x = b0 + b1y) Gretl shows me only the "White's
test for heteroskedasticity (squared only)". I believe that this could be a
little bit confusing and I suggest to introduce an alert when we try to
perform an "White's test for heteroskedasticity" on regressions (OLS) that
have only one explanatory variable.
I'd have attached a session file that shows this problem.
Best,
Henrique C. de Andrade
Doutorando em Economia Aplicada
Universidade Federal do Rio Grande do Sul
www.ufrgs.br/ppge
9:29 p.m.
On Wed, 17 Jun 2009, Henrique wrote:
When I try to perform a "White's test for
heteroskedasticity" on OLS models with only one term (x = b0 +
b1y) Gretl shows me only the "White's test for
heteroskedasticity (squared only)". I believe that this could be
a little bit confusing...
In CVS I have disabled the "(squares only)" item for White's test
if the model has only one regressor (besides the constant).
Allin Cottrell
9:35 p.m.
Hi,
I don't quite see why the current behavior would be confusing. But maybe
I didn't understand what you expected instead.
cheers,
sven
Henrique schrieb:
Dear Gretl community,
When I try to perform a "White's test for heteroskedasticity" on
OLS models with only one term (x = b0 + b1y) Gretl shows me only the
"White's test for heteroskedasticity (squared only)". I believe that
this could be a little bit confusing and I suggest to introduce an alert
when we try to perform an "White's test for heteroskedasticity" on
regressions (OLS) that have only one explanatory variable.
I'd have attached a session file that shows this problem.
Best,
Henrique C. de Andrade
Doutorando em Economia Aplicada
Universidade Federal do Rio Grande do Sul
www.ufrgs.br/ppge <http://www.ufrgs.br/ppge>
------------------------------------------------------------------------
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
3:18 a.m.
On Wed, 17 Jun 2009, Sven Schreiber wrote:
Henrique schrieb:
> When I try to perform a "White's test for
> heteroskedasticity" on OLS models with only one term (x = b0 +
> b1y) Gretl shows me only the "White's test for
> heteroskedasticity (squared only)".
I don't quite see why the current behavior would be confusing.
But maybe
I didn't understand what you expected instead.
My original thought on reading Henrique's message was this: if you
have estimated a model with only one non-constant regressor, then
cross-products are not applicable, so it might be confusing to
be faced with a choice of two versions of White's test (regular,
and "squares only"), as in the Tests menu in the model window.
My solution was to disable the "squares only" option in this case
since it is redundant.
But looking again at the bit of Henrique's message quoted above,
I'm not so sure: he says, "Gretl shows me _only_ the "White's test
for heteroskedasticity (squared only)" [emphasis added]. I thought
the 'problem' was that both variants of White's test were shown in
the menu.
Henrique?
Allin Cottrell
4:19 a.m.
On Wed, 17 Jun 2009, Sven Schreiber wrote:
> I don't quite see why the current behavior would be confusing. But maybe
> I didn't understand what you expected instead.
Let's suppose I've estimated a model like this:
EQUATION 1: Y = const + beta_1*X1 + beta_2*X2 + u
If we perform "White's test for heteroskedasticity" Gretl will
estimate the following equation:
EQUATION 2: uhat^2 = const + alfa_1*X1 + alfa_2*X2 + alfa_3*sq_X1
+ alfa_4*sq_X2 + alfa_5*X1*X2
And if we perform "White's test for heteroskedasticity (squared only)"
Gretl will estimate the
following equation:
EQUATION 3: uhat^2 = const + alfa_1*X1 + alfa_2*X2 + alfa_3*sq_X1
+ alfa_4*sq_X2
The last term of "equation 2", the cross-product, can only exist if we
have more than one
non-constant regressor, ok? So, if we have a model with just one
non-constant regressor,
like this:
EQUATION 4: Y = const + beta_1*X1 + u
In that case I think it's impossible to perform "White's test for
heteroskedasticity" because
in that test we need the cross-product term.
On Wed, 18 Jun 2009, Allin wrote
My original thought on reading Henrique's message was this: if you
have estimated a model with only one non-constant regressor, then
cross-products are not applicable, so it might be confusing to
be faced with a choice of two versions of White's test (regular,
and "squares only"), as in the Tests menu in the model window.
Yes, this is what I thought.
My solution was to disable the "squares only" option in
this case
since it is redundant.
I disagree. I suppose that the best solution is to disable "White's
test for heteroskedasticity"
(the regular version), because in the "one non-constant regressor"
case we just can perform
the "White's test for heteroskedasticity (squared only)".
But looking again at the bit of Henrique's message quoted above,
I'm not so sure: he says, "Gretl shows me _only_ the "White's test
for heteroskedasticity (squared only)" [emphasis added].
In the Tests menu the two versions of White's test appears. But when I
perform them Gretl
shows the same information to me. It is something like this:
I'd estimated "EQUATION 4" (presented above) and now I want to test
for heteroskedasticity using
White's test. I now that the only possible version of the test that I
can use is the "White's test for
heteroskedasticity (squared only)", because I have only non-constant
regressor, ok?
But if I try to perform the two versions of the White's test anyway,
Gretl prints only one in the
model window. The problem is that Gretl prints "White's test for
heteroskedasticity" (in the title)
and shows the results of the "White's test for heteroskedasticity
(squared only)".
I hope I could be clearer now :)
12:36 p.m.
Henrique schrieb:
So, if we have a model with just one
> non-constant regressor,
> like this:
> EQUATION 4: Y = const + beta_1*X1 + u
> In that case I think it's impossible to perform
"White's test for
> heteroskedasticity" because
> in that test we need the cross-product term.
Well I don't think there is a universal convention about the default way
of this test, with or without cross terms. Put differently, I think it's
perfectly legitimate to refer to the test w/o cross terms as "the White
test" (without the additional clarifier) as well.
Of course, to make things clearer, one could add the label "with cross
terms" to that version of the test, just as the other version carries
the label "squares only".
Allin's solution to make gretl even clever about this is already almost
a luxury IMHO and not strictly necessary, but I certainly don't have
anything against it :-)
Now even if I sound like Jack, if a user has a model with just one
regressor and is confused that "with cross terms" and "squares only"
coincide in that special case, I wonder if he should be doing that kind
of research. But I guess gretl should be user-friendlier than the way I
just sounded, so no offense please!
cheers,
sven
1:12 p.m.
On Thu, 18 Jun 2009, Sven Schreiber wrote:
Allin's solution to make gretl even clever about this is already
almost
a luxury IMHO and not strictly necessary, but I certainly don't have
anything against it :-)
Yes, the recent modification is in the "bending over backwards"
category.
I can't accept Henrique's point that what we show for a model with
a single non-constant regressor is not "White's test" (with no
qualification needed). In the simple regression case we do
include all k available cross-products, where k = 0.
Allin.
1:55 p.m.
On Thu, 18 Jun 2009, Sven Schreiber wrote:
Now even if I sound like Jack, if a user has a model with just one
regressor and is confused that "with cross terms" and "squares only"
coincide in that special case, I wonder if he should be doing that kind
of research. But I guess gretl should be user-friendlier than the way I
just sounded, so no offense please!
:-)
As you may imagine, I agree.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
6:58 p.m.
No problem. It was just a suggestion.
Best,
Henrique
On Thu, Jun 18, 2009 at 10:55 AM, Riccardo (Jack)
Lucchetti<r.lucchetti(a)univpm.it> wrote:
On Thu, 18 Jun 2009, Sven Schreiber wrote:
> Now even if I sound like Jack, if a user has a model with just one
> regressor and is confused that "with cross terms" and "squares
only"
> coincide in that special case, I wonder if he should be doing that kind
> of research. But I guess gretl should be user-friendlier than the way I
> just sounded, so no offense please!
:-)
As you may imagine, I agree.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Henrique C. de Andrade
Doutorando em Economia Aplicada
Universidade Federal do Rio Grande do Sul
www.ufrgs.br/ppge
5670
days inactive
5671
days old
8 comments
4 participants
participants (4)
-
Allin Cottrell -
Henrique -
Riccardo (Jack) Lucchetti -
Sven Schreiber