Re: [Gretl-users] GRETL - User Suggestions - Minus Signs and Binary Logit Function options
On Fri, 5 May 2017, Rao G.V. wrote:
Hi Riccardo
I got your details from the GRETL user community site. A great friendly
product from a user perspective!
Thank you. I'm taking the liberty to forward my reply to the gretl user
list, so that our exchange becomes public and possibly useful to other
people too.
I have used a number of functions but mainly the logit regression and
the Bayesian function. All good.
However I noted that when the output for the logit function is saved in
word and copied to an excel spreadsheet the signs are not proper minus
signs. See below.
GRETL OUTPUT
Coefficient Std. Error z p-value
V3D_A1_6 −175812 50541.4 −3.479 0.0005 ***
compared to
-175812
-3.479
Which means one needs to correct the output if further analytical work
needs to be performed. Now, in an earlier version of GRETL the signs
were in the correct format !............ so may be a suggestion to
correct this. The sign has remained unchanged for quite some time now.
This has come up several times. If you copy-n-paste from gretl to a text
processing application we use a character that is the typographically
appropriate glyph for the minus operator, which is unfortunately different
from the minus sign that mathematical apps use. However, if you
copy-n-paste directly from gretl to a spreadsheet (without the
intermediate word processing passage) everything should be ok.
Another suggestion would be to add a stepwise function (backward or
forward) for sequential addition/deletion of variables to determine the
final logit model (example in the binary logit) with the existing
p-values (say 0.1) as the criteria of adding (or removing) significant
variables. This is similar to functions/options available in SPSS,
Statgraphics, Stata and other packages. This would greatly assist a
user identify the significant variables and model quicker.
Admittedly this will add to the processing time and memory to the final
output. However it will definitely be a big improvement and if it is
added to the binary logit function as an option in the menu like other
software packages (compared to a separate function package) even
better!. Currently this is a very manual repetitive process of removing
each non-significant variable from the model, especially if the list of
regressors is large and therefore time consuming.
I noted that there is a function package available for sequential
variables, however this is only for OLS not other functions such as
binary logit. Also the function seems to require a list 1 of initial
regressors and then the candidates. What if all regressors are
candidates to begin with......Also there does not appear to be p-value
criteria which is a standard feature in most packages.
Actually, the existing function package could be extended to cover a wider
class of models than OLS. Let me think about it.
I hope my input is useful
It surely is. Thank you for your feedback!
-------------------------------------------------------
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
-------------------------------------------------------