Re: [Gretl-users] issues with the DF/EG tests
On 05/09/2012 03:16 AM, Daniel Ventosa S. wrote:
slightly modifies the critical values. Curiously enough, including
the
trend in the cointegrated equation also modifies the critical values of
the DF test on the residuals (i.e., according to my simulation, both
sets of simulated CV are different to the EG critical values).
Could you provide more details? So you are running a E-G test in gretl,
specifying a linear trend component, and the critical values used in
gretl are different from what your simulation shows? Are you actually
simulating (close-to-) asymptotic critical values or are you implicitly
running a small-sample simulation?
2.- I agree with Riccardo (Jack) that if the outcome of the test
changes
when a few observations are removed, then the test is not quite
reliable. But that was not my argument. If you have 120 obs, why should
we only use 90 obs. to draw inference? Sven said that getting the same
sample sizes for all the variables would make possible a comparison
between the ADF tests. I would rather say that the main point of doing
ADF tests to the observed variables is to be sure that the variables are
I(1), and not to compare the tests. It's always better to draw
inference using all the available information.
I think you are missing the problem that comes with the lag choice. If
the lag length were known in advance, there wouldn't be any problem and
everybody agrees that all observations should be (and are) used. But the
argument is statistically flawed IMHO if you use one sample to determine
your lag length in a data-based way, and then fix that lag choice and
jump to a different sample.
(OTOH, it is well known that even fixing the sample still has the
pre-testing problem which is not really accounted for in the unit-root
test that follows. So this is not an argument that gretl's way is the
one and only, but it is not dominated by your suggestion.)
3.- Allin said the test cannot be considered as rocket science and I
agree. That said, it remains an option in Gretl, next to the Johansen
test. The latter allows the user to include/remove deterministic
elements in the cointegrating vector. Why would that not be possible
with the EG?
Actually there are prominent econometricians who would like to remove
some of the options from the Johansen test (not in gretl in particular,
but in general) and in that sense make it more similar to the E-G-setup
as it is in gretl right now. So the argument cuts both ways, to some
extent. OTOH the Johansen framework is richer because you can specify
different deterministics for the I(0) and the I(1) parts ("directions")
in a way that just isn't possible in the simple E-G world. And this
difference is important here.
And again: Some of the setups lead to different test properties
(asymptotics), so the implementation is not as simple as just adding the
options in the dialog window.
cheers,
sven