6:10 p.m.
Folks,
This is an econometrics question rather than a Gretl question, so apologies in advance.
I've estimated an ordered probit on a panel data set with lagged regressors. Gretl
seems to automatically include the same number of lags of the dependent variable, and
I'm wondering why that is. So far I haven't been able to find a reference for this
- nothing in the user's guide, Lee's ebook, or Greene, AFAICS. If anyone could
help me out, I'd appreciate it.
Thanks!
PS
[cid:logoimg]
8:44 p.m.
On Tue, 6 Jan 2015, Summers, Peter wrote:
I've estimated an ordered probit on a panel data set with lagged
regressors. Gretl seems to automatically include the same number of
lags of the dependent variable, and I'm wondering why that is.
How are you specifying the model? I'm not able to replicate what you
describe. Here's a test script:
<hansl>
nulldata 100
set seed 12347
setobs 10 1.1 --stacked-time-series
series u = uniform()
series y = u > .75 ? 3 : u > .50 ? 2 : u > .25 ? 1 : 0
series x = normal()
probit y 0 x x(-1)
</hansl>
No lagged dependent variable gets included. I also tried specifying
the same thing via the GUI, with the same result.
Allin
9:50 p.m.
Hi Allin,
First, thanks for the fix on "modeltab" -- that's working fine now.
However the ordered probit thing is still there. I'm using a script with generic
notation "model4 <- probit y 0 x(-1 to -5)." I get the same result (lagged y
included) via the gui, but running your script works as expected (no lagged y). I'll
send you my data and a script that generates the weirdness off-list.
PS
-----Original Message-----
From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On
Behalf Of Allin Cottrell
Sent: Tuesday, January 06, 2015 3:44 PM
To: Gretl list
Subject: Re: [Gretl-users] ordered probit
On Tue, 6 Jan 2015, Summers, Peter wrote:
I've estimated an ordered probit on a panel data set with lagged
regressors. Gretl seems to automatically include the same number of
lags of the dependent variable, and I'm wondering why that is.
How are you specifying the model? I'm not able to replicate what you describe.
Here's a test script:
<hansl>
nulldata 100
set seed 12347
setobs 10 1.1 --stacked-time-series
series u = uniform()
series y = u > .75 ? 3 : u > .50 ? 2 : u > .25 ? 1 : 0 series x = normal() probit
y 0 x x(-1) </hansl>
No lagged dependent variable gets included. I also tried specifying the same thing via the
GUI, with the same result.
Allin
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
9:58 p.m.
On Tue, 6 Jan 2015, Summers, Peter wrote:
Hi Allin,
First, thanks for the fix on "modeltab" -- that's working fine now.
Good to hear it.
However the ordered probit thing is still there. I'm using a
script with generic notation "model4 <- probit y 0 x(-1 to -5)." I
get the same result (lagged y included) via the gui, but running
your script works as expected (no lagged y). I'll send you my data
and a script that generates the weirdness off-list.
OK, please do.
Allin
1 a.m.
On Tue, 6 Jan 2015, Summers, Peter wrote:
[T]he ordered probit thing is still there. I'm using a script
with
generic notation "model4 <- probit y 0 x(-1 to -5)." I get the
same result (lagged y included) via the gui, but running your
script works as expected (no lagged y). I'll send you my data and
a script that generates the weirdness off-list.
Thanks. After looking at Peter's data and script I can see what's
happening.
Ordered probit is misbehaving when the dependent variable is not
integer-valued. (Peter's dependent variable has values {0.5, 1, 1.5,
2.5, 3, 3.5, 4, 4.5, 5, 6, 6.5, 7.5, 12}.) We'll take a closer look
at this -- with a view to either handling correctly or rejecting
such data -- but in the meantime a workaround would be to multiply
the dependent variable by 2, in which case it should be treated
correctly.
Allin
8:47 a.m.
Am 07.01.2015 um 02:00 schrieb Allin Cottrell:
Thanks. After looking at Peter's data and script I can see what's
happening.
Ordered probit is misbehaving when the dependent variable is not
integer-valued. (Peter's dependent variable has values {0.5, 1, 1.5,
2.5, 3, 3.5, 4, 4.5, 5, 6, 6.5, 7.5, 12}.) We'll take a closer look
at this -- with a view to either handling correctly or rejecting
such data -- but in the meantime a workaround would be to multiply
the dependent variable by 2, in which case it should be treated
correctly.
Was that variable marked as discrete? I vaguely remember a discussion
about the possibility to tell gretl that even non-integer data could
come from a discrete variable. So from a user's point of view I'd say
that (ordered) probit should work when the variable is "officially"
discrete. If it's not marked as such *and* is not obviously discrete
such as not being integer-valued I think (fwiw) it's ok if gretl refuses
to do ordered probit on that.
cheers,
sven
12:17 p.m.
Hi Sven,
Yes, it was marked as discrete. I agree with you that this should be necessary for proper
ordered probit behavior.
PS
Sent from my iPad
On Jan 7, 2015, at 3:50 AM, Sven Schreiber <svetosch(a)gmx.net>
wrote:
> Am 07.01.2015 um 02:00 schrieb Allin Cottrell:
>
>
> Thanks. After looking at Peter's data and script I can see what's
> happening.
>
> Ordered probit is misbehaving when the dependent variable is not
> integer-valued. (Peter's dependent variable has values {0.5, 1, 1.5,
> 2.5, 3, 3.5, 4, 4.5, 5, 6, 6.5, 7.5, 12}.) We'll take a closer look
> at this -- with a view to either handling correctly or rejecting
> such data -- but in the meantime a workaround would be to multiply
> the dependent variable by 2, in which case it should be treated
> correctly.
>
Was that variable marked as discrete? I vaguely remember a discussion
about the possibility to tell gretl that even non-integer data could
come from a discrete variable. So from a user's point of view I'd say
that (ordered) probit should work when the variable is "officially"
discrete. If it's not marked as such *and* is not obviously discrete
such as not being integer-valued I think (fwiw) it's ok if gretl refuses
to do ordered probit on that.
cheers,
sven
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
3:03 p.m.
On Wed, 7 Jan 2015, Summers, Peter wrote:
Hi Sven,
Yes, it was marked as discrete. I agree with you that this should be
necessary for proper ordered probit behavior.
I have no idea of the actual problem, but if your dependent variable is
measured on a non-arbitrary scale, then you probably need interval
regression, rather than ordered probit.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
6:09 p.m.
On Tue, 6 Jan 2015, Summers, Peter wrote:
[T]he ordered probit thing is still there. I'm using a script
with
generic notation "model4 <- probit y 0 x(-1 to -5)." I get the same
result (lagged y included) via the gui, but running your script
works as expected (no lagged y). I'll send you my data and a script
that generates the weirdness off-list.
That should now be fixed in CVS and snapshots.
The spurious "extra regressors" you were seeing were not actually
lags, they were dummies for some of the levels of the dependent
variable; they somehow got into the mix due to failure of the
algorithm that was supposed to normalize the dependent variable (that
is, to transform it such that it comprises consecutive integer values
starting at zero). I've replaced that algorithm with a smarter one.
It should now be safe to pass a series with non-integral values as the
dependent variable in ordered probit, provided it has been sucessfully
marked as discrete.
(I'm still not sure exactly what triggered the problem in your test
case: it turned out that y having non-integral values was not a
sufficient condition. But anyway the new algorithm should handle all
cases.)
Allin
6:40 p.m.
On Wed, 7 Jan 2015, Allin Cottrell wrote:
It should now be safe to pass a series with non-integral values as
the
dependent variable in ordered probit, provided it has been sucessfully
marked as discrete.
Excuse me, I may be missing something, but I fail to see the logic in
this. In an ordered probit model, the support of the dependent variable is
supposed to be a sequence of increasing numbers, which indicate increasing
"degress of intensity" of a certain unobserved variable, whose conditional
mean is what we're trying to estimate. Of course they could be any
sequence, as long as it's increasing, but I would guess that common sense
dictates they should be increasing _integers_, since it's a purely
conventional way of saying labelling different degrees of intensity across
observations. IMHO allowing the dependent variable to be non-integer could
easily lead to failure to spot an incorrect application of ordered probit
(eg, when you're using the wrong variable from your dataset) and gives you
nothing in return.
As I said before, if the dependent variable is truly quantitative (as
opposed as being a conventional coding for an unobserved continuous latent
variable) and for example 12.5 is just a way of saying "somewhere between
12 and 13" or whatever, the right tool for the job is interval regression,
not ordered probit.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
7:03 p.m.
On Wed, 7 Jan 2015, Riccardo (Jack) Lucchetti wrote:
On Wed, 7 Jan 2015, Allin Cottrell wrote:
> It should now be safe to pass a series with non-integral values as
> the dependent variable in ordered probit, provided it has been
> sucessfully marked as discrete.
Excuse me, I may be missing something, but I fail to see the logic
in this. In an ordered probit model, the support of the dependent
variable is supposed to be a sequence of increasing numbers, which
indicate increasing "degress of intensity" of a certain unobserved
variable, whose conditional mean is what we're trying to estimate.
Of course they could be any sequence, as long as it's increasing,
but I would guess that common sense dictates they should be
increasing _integers_ [...]
I see your point, but I seem to remember some discussion long ago when
some people (Sven?) argued that a "degree of intensity" discrete
variable might in practice be coded with round-ish fractions (e.g.
multiples of 0.5). Also, FWIW, Stata allows this in ordered probit.
(Plus there's a constraint implied by the condition that the dependent
variable be "successfully marked as discrete" -- you can't define any
old junk as a discrete variable.)
Maybe Peter could comment on the dependent variable in his case?
Allin
7:06 p.m.
Here's some background:
I'm working on a model predicting periods of financial crises; the dependent variable
is usually coded 0/1 (no crisis/crisis) so probit/logit is fine. However Romer & Romer
(2014) have a new integer-valued classification scheme that ranges from 0 (no crisis) to
15 (extreme crisis). They present a chronology based on semi-annual observations, but my
data are annual. The half-integer values in my dependent variable arose because I averaged
R&R's classification over the year. So a country that goes from 0 in the first
half to 1 in the second gets a 0.5 from me.
I've never used interval regression, but a brief read of the user's guide re
'intreg' makes me think that my situation really is closer to ordered probit.
Also, note that the guide (p. 288) seems to specifically allow for the change Allin just
implemented: "In order to apply these models in gretl, the dependent variable must
either take on only nonnegative integer values, or be explicitly marked as discrete. (In
case the variable has non-integer values, it will be recoded internally.)"
PS
-----Original Message-----
From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On
Behalf Of Riccardo (Jack) Lucchetti
Sent: Wednesday, January 07, 2015 1:40 PM
To: Gretl list
Subject: Re: [Gretl-users] ordered probit
On Wed, 7 Jan 2015, Allin Cottrell wrote:
It should now be safe to pass a series with non-integral values as
the
dependent variable in ordered probit, provided it has been sucessfully
marked as discrete.
Excuse me, I may be missing something, but I fail to see the logic in this. In an ordered
probit model, the support of the dependent variable is supposed to be a sequence of
increasing numbers, which indicate increasing "degress of intensity" of a
certain unobserved variable, whose conditional mean is what we're trying to estimate.
Of course they could be any sequence, as long as it's increasing, but I would guess
that common sense dictates they should be increasing _integers_, since it's a purely
conventional way of saying labelling different degrees of intensity across observations.
IMHO allowing the dependent variable to be non-integer could easily lead to failure to
spot an incorrect application of ordered probit (eg, when you're using the wrong
variable from your dataset) and gives you nothing in return.
As I said before, if the dependent variable is truly quantitative (as opposed as being a
conventional coding for an unobserved continuous latent
variable) and for example 12.5 is just a way of saying "somewhere between
12 and 13" or whatever, the right tool for the job is interval regression, not
ordered probit.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
6:42 p.m.
That works great now, Allin. Thanks heaps!
PS
-----Original Message-----
From: gretl-users-bounces(a)lists.wfu.edu [mailto:gretl-users-bounces@lists.wfu.edu] On
Behalf Of Allin Cottrell
Sent: Wednesday, January 07, 2015 1:09 PM
To: Gretl list
Subject: Re: [Gretl-users] ordered probit
On Tue, 6 Jan 2015, Summers, Peter wrote:
[T]he ordered probit thing is still there. I'm using a script
with
generic notation "model4 <- probit y 0 x(-1 to -5)." I get the same
result (lagged y included) via the gui, but running your script works
as expected (no lagged y). I'll send you my data and a script that
generates the weirdness off-list.
That should now be fixed in CVS and snapshots.
The spurious "extra regressors" you were seeing were not actually lags, they
were dummies for some of the levels of the dependent variable; they somehow got into the
mix due to failure of the algorithm that was supposed to normalize the dependent variable
(that is, to transform it such that it comprises consecutive integer values starting at
zero). I've replaced that algorithm with a smarter one.
It should now be safe to pass a series with non-integral values as the dependent variable
in ordered probit, provided it has been sucessfully marked as discrete.
(I'm still not sure exactly what triggered the problem in your test
case: it turned out that y having non-integral values was not a sufficient condition. But
anyway the new algorithm should handle all
cases.)
Allin
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
3637
days inactive
3638
days old
12 comments
4 participants
participants (4)
-
Allin Cottrell -
Riccardo (Jack) Lucchetti -
Summers, Peter -
Sven Schreiber