10:21 a.m.
Hi all,
I just stumbled over the following challenge.
In the first example, I run a binary logit with a numeric (discrete)
encoded variable "dirnum".
In the second example, the dependent is string-valued series "dirstr"
with only two distinct values ("down", "up").
The logit command estimates a binary logit for the first example (as
expected) and a ordered logit for the 2nd example, however.
The reason might be that for "stringified" series, the minimal distinct
value is 1 and not zero.
Even though the magnitude of the coefficients are equals, the sign may
differ. Also, some statistics such as R^2 and the contigency table are
not printed for the ordered case.
Is this an expected behavior?
<hansl>
set verbose off
clear
open nysewk.gdt
series ret = ldiff(close)
series dirnum = ret >= 0 ? 1 : 0 # 0-based
series dirstr = ret >= 0 ? 2 : 1 # has to be 1-based for stringify() to
work
strings direction = defarray("down", "up")
stringify(dirstr, direction)
print -o --range=1:3
# Both only have two distinct values
eval values(dirnum)
eval values(dirstr)
logit dirnum const # binary logit
logit dirstr const # ordered logit
</hansl>
Best
Artur
6:02 p.m.
On Mon, 13 Nov 2023, Artur T. wrote:
Hi all,
I just stumbled over the following challenge.
In the first example, I run a binary logit with a numeric
(discrete) encoded variable "dirnum".
In the second example, the dependent is string-valued series
"dirstr" with only two distinct values ("down", "up").
The logit command estimates a binary logit for the first example
(as expected) and a ordered logit for the 2nd example, however.
The reason might be that for "stringified" series, the minimal
distinct value is 1 and not zero.
Even though the magnitude of the coefficients are equals, the sign
may differ. Also, some statistics such as R^2 and the contigency
table are not printed for the ordered case.
Is this an expected behavior?
Well, I think it might be expected, since the logit doc starts with
"If the dependent variable is a binary variable (all values are 0 or
1)...". But it's not intentional, nor (I think) desirable.
Now in git master, if a string-valued series has just two values,
and reduces to a straight dummy variable on subtracting 1 from its
numeric codes, we'll treat it as a binary case. Note that this is a
special dispensation for string-valued series, which much always
have a minimum numeric code of 1; we're not going to do this for
regular numeric series that have values 1 and 2 -- subtract 1,
please!
On R-squared: there's no reason why we shouldn't show a
pseudo-R-squared for ordered logit as well as binary, and that too
is now in git.
On signs of coefficients: If you're using a string-valued dependent
variable with two values you need to pay attention to this point. In
the "stringifying" of a series the first string will get attached to
whichever numeric value occurs first, but for binary logit you
presumably want the "on"/"true"/"yes" value to attach to the
higher
of the two numeric codes (which will turn into 1 on subtraction).
The changes mentioned above will be in the snapshots before long.
An example follows:
<hansl>
open greene19_1.gdt
# regular binary logit
logit GRADE 0 GPA TUCE PSI --p-values
series y = GRADE + 1
# invokes the "ordered" variant: now shows McFadden
logit y 0 GPA TUCE PSI --p-values
# note: the order matters in defarray()
stringify(y, defarray("down", "up"))
# invokes the binary variant
logit y 0 GPA TUCE PSI --p-values
</hansl>
Allin
7:55 a.m.
On 13/11/2023 19:02, Allin Cottrell wrote:
Well, I think it might be expected, since the logit doc starts with
"If
the dependent variable is a binary variable (all values are 0 or 1)...".
But it's not intentional, nor (I think) desirable.
Now in git master, if a string-valued series has just two values, and
reduces to a straight dummy variable on subtracting 1 from its numeric
codes, we'll treat it as a binary case. Note that this is a special
dispensation for string-valued series, which much always have a minimum
numeric code of 1; we're not going to do this for regular numeric series
that have values 1 and 2 -- subtract 1, please!
This is very nice, thanks, Allin! I
guess that by the same logic we
could enable printing the value labels when we apply multinomial logit
to a string-valued series. Eg
<hansl>
# Replicate example in Wooldridge, table 15.2
set verbose off
open keane.gdt
stringify(status, defarray("School", "Neither", "Work"))
smpl (year==87) --restrict
logit status 0 educ exper expersq black --multinomial
</hansl>
-------------------------------------------------------
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
-------------------------------------------------------
10:46 a.m.
On 14/11/2023 08:55, Riccardo (Jack) Lucchetti wrote:
This is very nice, thanks, Allin! I guess that by the same logic we
could enable printing the value labels when we apply multinomial logit
to a string-valued series.
Never mind, I pushed this to git master already. @Allin: it was a rather
simple change, but please review.
-------------------------------------------------------
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
-------------------------------------------------------
1:03 p.m.
On Tue, Nov 14, 2023 at 5:46 AM Riccardo (Jack) Lucchetti
<p002264(a)staff.univpm.it> wrote:
On 14/11/2023 08:55, Riccardo (Jack) Lucchetti wrote:
> This is very nice, thanks, Allin! I guess that by the same logic we
> could enable printing the value labels when we apply multinomial logit
> to a string-valued series.
Never mind, I pushed this to git master already. @Allin: it was a rather
simple change, but please review.
Looks good! Thanks, Jack.
Allin
717
days inactive
718
days old
4 comments
4 participants
participants (4)
-
Allin Cottrell -
Artur T. -
Cottrell, Allin
-
Riccardo (Jack) Lucchetti