8:15 p.m.
Sorry to keep harping on this, but the "add" and "omit" have
been rather messy for a long time, and I'm struggling to
straighten them out, at least partially.
Part of the difficulty is that one could think of "omit" in
either of two ways:
1) It's shorthand for a special case of "restrict", with the
side effect (which one may or may not want) of estimating the
restricted model and making it the "last model".
2) It's a shorthand means of re-estimating the last model
with a smaller specification, with the side effect of
reporting a test for the implied restriction.
Lots of scripts in the gretl package use omit in the second
way, which makes me a bit reluctant to switch the default
behavior to NOT producing a new model.
In CVS I had introduced a --test-only switch but on further
thought that did not seem very clean, since we already had a
--wald switch which suppresses the "produce a new model"
behavior. So here's what I now have in CVS, and on which I'd
value comments:
1) For "omit", if you don't want a new model, use the --wald
flag. This will now produce the F-form of the Wald test by
default, but if you prefer the chi-square version, use the new
--chi-square option. Note that if you have any scripts which
use the --quiet option to "omit" to suppress the new model,
switch to --wald and you should be OK. (As it should, --quiet
now only affects what gets printed, not the substance of what
gets done.)
2) "add" is a bit different, but I've introduced an --lm
option that does an LM test using an auxiliary regression.
That way you don't get a new model, but right now the --lm
option is only for OLS. Otherwise you do get the augmented
model, even if you choose not to print it. I think this is OK,
since "add" (apart from the new LM option) is a sort of
synthetic command anyway: in effect, estimate the bigger model
and do an "omit" on it.
The documentation for add and omit is updated in CVS. If we're
OK with this I'll put a entry into the backward
incompatibility log.
Oh, one other thing: the latest version is not yet in the
snapshots for Windows and OS X. I thought I'd wait to hear
some reactions before doing that.
Allin
3:52 p.m.
Am 25.07.2011 22:15, schrieb Allin Cottrell:
Lots of scripts in the gretl package use omit in the second
way, which makes me a bit reluctant to switch the default
behavior to NOT producing a new model.
agreed
In CVS I had introduced a --test-only switch but on further
thought that did not seem very clean, since we already had a
--wald switch which suppresses the "produce a new model"
behavior. So here's what I now have in CVS, and on which I'd
value comments:
1) For "omit", if you don't want a new model, use the --wald
flag. This will now produce the F-form of the Wald test by
default, but if you prefer the chi-square version, use the new
--chi-square option. Note that if you have any scripts which
use the --quiet option to "omit" to suppress the new model,
switch to --wald and you should be OK. (As it should, --quiet
now only affects what gets printed, not the substance of what
gets done.)
But what if you DO want a new model, but you want to use the chi-square
test?
What about a switch "--test=wald" or "--test=chi-sq" to choose the
test
variant, irrespective of whether a new model is produced or not? (And
maybe even "--test=lm", as per your remark below.) And then an
additional switch to indicate the wish to keep the old model (e.g.
"--keep" or whatever".
2) "add" is a bit different, but I've introduced an --lm
option that does an LM test using an auxiliary regression.
That way you don't get a new model, but right now the --lm
option is only for OLS. Otherwise you do get the augmented
model, even if you choose not to print it. I think this is OK,
since "add" (apart from the new LM option) is a sort of
synthetic command anyway: in effect, estimate the bigger model
and do an "omit" on it.
I never really understood the justification for 'add', and this seems to
confirm my prejudice. Why not just ditch it on the way to gretl 2.0? I
know it's radical, but...
cheers,
sven
4:14 p.m.
On Tue, 26 Jul 2011, Sven Schreiber wrote:
Am 25.07.2011 22:15, schrieb Allin Cottrell:
> In CVS I had introduced a --test-only switch but on further
> thought that did not seem very clean, since we already had a
> --wald switch which suppresses the "produce a new model"
> behavior. So here's what I now have in CVS, and on which I'd
> value comments:
>
> 1) For "omit", if you don't want a new model, use the --wald
> flag. This will now produce the F-form of the Wald test by
> default, but if you prefer the chi-square version, use the new
> --chi-square option. Note that if you have any scripts which
> use the --quiet option to "omit" to suppress the new model,
> switch to --wald and you should be OK. (As it should, --quiet
> now only affects what gets printed, not the substance of what
> gets done.)
But what if you DO want a new model, but you want to use the chi-square
test?
I didn't make that clear, but the --chi-square switch works
whether or not we're producing a new model.
But here's another point I came across: a while back we
introduced options --both and --instrument for tsls models:
the default is to omit the specified variables as regressors
only, leaving the instrument set unchanged, but these flags
can be used to drop variables from both slots, or only in
their role as instruments.
I can't really remember why we did this, but the effect is
a bit weird, since (AFAIK) we can't really do a meaningful
test with either of those options in place -- although they
are OK if considered just as means of specifying a new model.
Right now we do a test if --both is given, but it's just the
test for dropping the variables as regressors.
> 2) "add" is a bit different, but I've introduced an
--lm
> option that does an LM test using an auxiliary regression.
> That way you don't get a new model, but right now the --lm
> option is only for OLS. Otherwise you do get the augmented
> model, even if you choose not to print it. I think this is OK,
> since "add" (apart from the new LM option) is a sort of
> synthetic command anyway: in effect, estimate the bigger model
> and do an "omit" on it.
I never really understood the justification for 'add', and this seems to
confirm my prejudice. Why not just ditch it on the way to gretl 2.0? I
know it's radical, but...
That bears thinking about, certainly.
Allin
6:31 p.m.
One more follow-up:
The --wald option to "omit" (with the effect of not replacing the
orignal model) is definitely a misnomer. The help text is OK, in a
sense: it glosses this option as "Wald test _only_". But in almost
all cases of both omit and add, what we do is in fact a Wald test,
whether or not we're saving the modified model.
Only in the (new) case of "add --lm" are we doing something that's
truly not a Wald test. The other semi-exception is the case where
the original model is plain OLS (no --robust option) and we're
saving the modified model, and the --chi-square option is not given:
in that case we calculate the F-test based on the sums of squared
residuals from the restricted and unrestricted models, which is
in fact numerically equivalent to a "true" Wald test calculated
using the covariance matrix of the unrestricted model.
So: to complete the clean-up I'm now proposing to rename the --wald
option as --test-only. This also allows us to add an --lr option to
do a likelihood ratio test instead of a Wald test, where applicable.
I'm thinking one should be able to do:
omit X1 --test-only --lr
while it would make no sense to say
omit X1 --wald --lr
That is:
(a) Everything is a Wald test unless you give --lm (for "add" only)
or --lr (not implemented yet, but might be worth it).
(b) The original model is always replaced as "last model" unless you
say --test-only.
The renaming of --wald is not yet in CVS, but I'll put it there
barring any strong objections. The recommended backward fix would
be:
* Wherever you have omit --wald, replace with omit --test-only.
* If perchance you have used omit --quiet in the undocumented sense
of not saving the modified model, replace with --test-only.
Allin
4943
days inactive
4944
days old
3 comments
2 participants
participants (2)
-
Allin Cottrell -
Sven Schreiber