5:58 p.m.
On Sat, 12 Feb 2011, Allin Cottrell wrote:
On Sat, 12 Feb 2011, Birger Baksaas wrote:
> Coefficients from the Ordered Logit function are correct in gretl, but
> the signs seem to be mixed up.
I note that the gretl coefficients and cut-points (magnitude and
sign) agree with those given by Wooldridge...
In addition:
<script>
open pension.gdt
logit pctstck const choice finc25 wealth89 prftshr
foreign language=R --send-data
library(MASS)
y <- factor(pctstck)
plr <- polr(y ~ choice + finc25 + wealth89 + prftshr,
method = "logistic")
summary(plr)
end foreign
</script>
The coefficients (and cut points) given by gretl and R are
identical in both magnitude and sign.
Allin Cottrell
8:45 a.m.
New subject: Sign mix-up for ordered logit
Thanks for all your replies. I did miss something. The Gretl program and
documentation are correct. In the Gretl Guide it is specified that model
predictions are found by first subtracting the effect of the independent
variables from the intercepts (cut-points). I did not notice the
subtraction:
P (αj < y ≤ αj+1 | xi) = F (αj+1 − zi) − F (αj − zi)
This is how SPSS PLUM does it also, according to the book mentioned
below, and polr in the R MASS library.
So, minus times minus gives plus...
However, in the R VGAM library the effect of independent variables are
added (as we are used to with OLS).
A function call in VGAM:
mo <- vglm(profread ~ male, cumulative(link = logit, reverse = FALSE,
parallel = TRUE))
Similarly in SPSS the parts are added. SPSS is the program used in
"Logistic Regression Models for Ordinal Response Variables" by Ann A.
O'Connell.
Another thing to notice: Gretl uses the ascending alternative. The odds
are then accumulated over the lower-ordered categories. The opposite,
the descending method, gives parameters directly comparable with OLS.
However, this is easily given by reversing all the signs (for cut-points
and independent variables) from the ascending method. The magnitudes
stay the same.
Perhaps some of this could be emphasized in the user guide?
Thanks again. Have a nice week.
Birger
lø., 12.02.2011 kl. 12.58 -0500, skrev Allin Cottrell:
On Sat, 12 Feb 2011, Allin Cottrell wrote:
> On Sat, 12 Feb 2011, Birger Baksaas wrote:
>
> > Coefficients from the Ordered Logit function are correct in gretl, but
> > the signs seem to be mixed up.
>
> I note that the gretl coefficients and cut-points (magnitude and
> sign) agree with those given by Wooldridge...
In addition:
<script>
open pension.gdt
logit pctstck const choice finc25 wealth89 prftshr
foreign language=R --send-data
library(MASS)
y <- factor(pctstck)
plr <- polr(y ~ choice + finc25 + wealth89 + prftshr,
method = "logistic")
summary(plr)
end foreign
</script>
The coefficients (and cut points) given by gretl and R are
identical in both magnitude and sign.
Allin Cottrell
9:42 a.m.
New subject: Sign mix-up for ordered logit
Correction:
It should be SAS instead of SPSS in the lower part of my message. SAS is
the program mainly used in "Logistic Regression Models for Ordinal
Response Variables" by Ann A. O'Connell.
Instead of the sentence "Similarly in SAS the parts are added", it might
be more correct to write: "Similarly using SAS, the analyst should add
the logits for respective cut-point and effect of independent variables
to find cumulative odds and probabilities."
And of course: the probability for one category is found by subtracting
the cumulative probability for the category above, as is stated in the
Gretl user guide.
Birger
ma., 14.02.2011 kl. 09.45 +0100, skrev Birger Baksaas:
Thanks for all your replies. I did miss something. The Gretl program
and
documentation are correct. In the Gretl Guide it is specified that model
predictions are found by first subtracting the effect of the independent
variables from the intercepts (cut-points). I did not notice the
subtraction:
P (αj < y ≤ αj+1 | xi) = F (αj+1 − zi) − F (αj − zi)
This is how SPSS PLUM does it also, according to the book mentioned
below, and polr in the R MASS library.
So, minus times minus gives plus...
However, in the R VGAM library the effect of independent variables are
added (as we are used to with OLS).
A function call in VGAM:
mo <- vglm(profread ~ male, cumulative(link = logit, reverse = FALSE,
parallel = TRUE))
Similarly in SPSS the parts are added. SPSS is the program used in
"Logistic Regression Models for Ordinal Response Variables" by Ann A.
O'Connell.
Another thing to notice: Gretl uses the ascending alternative. The odds
are then accumulated over the lower-ordered categories. The opposite,
the descending method, gives parameters directly comparable with OLS.
However, this is easily given by reversing all the signs (for cut-points
and independent variables) from the ascending method. The magnitudes
stay the same.
Perhaps some of this could be emphasized in the user guide?
Thanks again. Have a nice week.
Birger
lø., 12.02.2011 kl. 12.58 -0500, skrev Allin Cottrell:
> On Sat, 12 Feb 2011, Allin Cottrell wrote:
>
> > On Sat, 12 Feb 2011, Birger Baksaas wrote:
> >
> > > Coefficients from the Ordered Logit function are correct in gretl, but
> > > the signs seem to be mixed up.
> >
> > I note that the gretl coefficients and cut-points (magnitude and
> > sign) agree with those given by Wooldridge...
>
> In addition:
>
> <script>
> open pension.gdt
> logit pctstck const choice finc25 wealth89 prftshr
> foreign language=R --send-data
> library(MASS)
> y <- factor(pctstck)
> plr <- polr(y ~ choice + finc25 + wealth89 + prftshr,
> method = "logistic")
> summary(plr)
> end foreign
> </script>
>
> The coefficients (and cut points) given by gretl and R are
> identical in both magnitude and sign.
>
> Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
5060
days inactive
5062
days old
2 comments
2 participants
participants (2)
-
Allin Cottrell -
Birger Baksaas