4:42 p.m.
These people were from State Planning Agency. They told me that they
have about 60 series (which have different levels of collinearity) and
they use SPSS to do principal components analysis to create regional
development indices (Turkey has 81 provinces). I am not very familiar
with pca but I can call them and learn more.
I think you just need the eigenvectors. Put all the series into a matrix,
get the
eigenvectors, and peel off the first few, and put these into
series; these are pc. Estimate model, and perform the inverse
transformation on the coefficients to get back to the coefficients in the
original data rotation if desired.
>> 2)- Some students suggested (and all others agreed) that it
would be
>> very useful to have a predict command, which will provide predicted
>> values as well as slopes (given Xs) for various nonlinear models such
>> as polynomial regressions, logit, probit etc. I think this could be
>> nice to have as a command as well as a GUI entry next to the forecast
>> item. Maybe a small goodie to consider for the 2.0 release? They said
>> Stata has this.
>
> I don't see what the difference is between "predicted values"
> and what we offer already (in sample fitted values and
> out-of-sample forecasts). Can you expand on what you mean?
Now this is maybe I didn't know how to fully use gretl in this
context. The issue arised on 2 occasions:
(1) I had a polynomial regression and I was showing them to enter from
the GUI the command something like:
prediction = $coeff[1] + $coeff[2]*x + $coeff[3]*x^2
(2) I was showing an ordered logit example and I had long commands like:
pcut0 = 1 / (1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
pcut1= exp(-$coeff[1]-x*$coeff[2]) /
(1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
pcut2= exp(-$coeff[1]-x*$coeff[3]) /
(1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
...and they said Stata (supposedly) has a command where you enter x
and get the prediction and slope for different models :-P
Indeed Stata does and it is very, very slick. I think it has taken the
Stata team
a long time to get it to the point that it is now. That said,
it's not very hard to do in gretl on a model by model basis See chapter
16 of http://learneconometrics.com/gretl/using_gretl_for_POE4.pdf for some
simple examples. The ability to jump seamlessly between matrices and gretl
results makes it amazingly easy to use. In it there are a few examples of
computing marginal effects and even standard errors for the average
marginal effects. A more competent programmer could do something more
elegant and general, but it it's not very hard to do on a case by case
basis.
Although, as Talha says, a 'button' or function is nice since most users
don't appreciate the challenge of getting these like some of us do...:)
One of the things that I've noticed about Stata's development is that they
have expanded many of the commands to encompass almost every model and they
continue to add functionality to them with each new version. For instance,
predict is used to get in-sample model predictions, residuals, variances
(e.g., from arch and garch), and to generate both static and dynamic
forecasts. margins computes different types of marginal effects (discrete,
continuous, average, at specific points, at the averages, and so on
depending on the options used). In fact, these commands do so much it
makes them a little hard to use correctly, IMO--for instance, it took some
time to figure out exactly how and which marginal effects were being
computed for the qualitative choice models.
So, that presents something of a design choice. Should one use commands
that apply to all circumstances (even though they may work differently
under the hood, depending on model being estimated), or should estimation
routines have their own set of postestimation commands for tests, marginal
effects, and so on. I don't know the answer to this. Stata flirts with
both models: ubiquitous (margins, predict) and model specific (estat
postestimation commands).
Open source suggests perhaps more modularity than one gets with a
proprietary software like Stata (though nearly everything it does is
executed in .do or .ado add-ons). What we discussed in Torun was the idea
that Allin and Jack would work on the back-bone and that others would try
to develop the expertise with the bundle concept to add specific
functionality. So, the question is, is enhancing prediction or marginal
effects a back-bone issue or an add-on? (I'm not sure)
My idea of a back-bone issue would be the introduction of factor variables
(version 2.0). In it, variables are defined as being continuous or
discrete. They can be interacted in various combinations
(continuous-discrete, continutous-continuous and discrete-discrete very
easily within Stata's .do files). For us, it may solve an issue with the
handling of missing values (0 or NAN) and permits a very straightforward
way to interact variables and compute marginal effects in interaction
models. Just a thought....
Cheers,
Lee
--
“An expert is a person who has made all the mistakes that can be made
in a very narrow field.” - Niels Bohr (1885-1962)
--
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
--
Lee Adkins
Professor of Economics
lee.adkins(a)okstate.edu
learneconometrics.com
9:47 p.m.
New subject: Report -- caution! long!
On Wed, 29 Feb 2012, Lee Adkins wrote (in response to Talha):
> These people were from State Planning Agency. They told me that
they
> have about 60 series (which have different levels of collinearity) and
> they use SPSS to do principal components analysis to create regional
> development indices (Turkey has 81 provinces). I am not very familiar
> with pca but I can call them and learn more.
>
I think you just need the eigenvectors. Put all the series into a matrix,
get the eigenvectors, and peel off the first few, and put these into
series; these are pc.
You can do PCA manually if you like, but I guess Talha's
students were not too keen on that idea. But in fact gretl has
a perfectly good "pca" command (accessible via GUI and script)
that takes a list of series as input, and outputs either all
of the components or just those whose associated eigenvalues
are above the mean. By default the correlation matrix is used
as the basis for the PCs but there's an option to use the
covariance matrix instead.
In current CVS and snapshots I've enhanced the pca command
slightly: there's now a --quiet option if you don't care to
see all the eigen-analysis, and the --save option now takes an
optional integer parameter so you can specify exactly how many
components to save.
>>> 2)- Some students suggested (and all others agreed) that
it would be
>>> very useful to have a predict command, which will provide predicted
>>> values as well as slopes (given Xs) for various nonlinear models such
>>> as polynomial regressions, logit, probit etc. I think this could be
>>> nice to have as a command as well as a GUI entry next to the forecast
>>> item. Maybe a small goodie to consider for the 2.0 release? They said
>>> Stata has this.
>>
>> I don't see what the difference is between "predicted values"
>> and what we offer already (in sample fitted values and
>> out-of-sample forecasts). Can you expand on what you mean?
> Now this is maybe I didn't know how to fully use gretl in this
> context. The issue arised on 2 occasions:
> (1) I had a polynomial regression and I was showing them to enter from
> the GUI the command something like:
> prediction = $coeff[1] + $coeff[2]*x + $coeff[3]*x^2
Yes, in gretl that sort of thing is required if you want to
predict for some x-value that's not in the dataset. Or you
could do it by adding one observation to the dataset,
containing the x-values for which you want a prediction, then
asking for an out of sample forecast. I guess we could add a
variant of "fcast" (or maybe a new command-word is wanted),
for use after estimating a model, which takes a vector of
x-values as input and produces the prediction plus a standard
error. That woud be quite simple.
> (2) I was showing an ordered logit example and I had long
commands like:
> pcut0 = 1 / (1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
> pcut1= exp(-$coeff[1]-x*$coeff[2]) /
> (1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
> pcut2= exp(-$coeff[1]-x*$coeff[3]) /
> (1+exp(-$coeff[1]-x*$coeff[2])+exp(-$coeff[1]-x*$coeff[3]))
>
> ...and they said Stata (supposedly) has a command where you enter x
> and get the prediction and slope for different models :-P
Prediction for nonlinear limited dependent variable models are
trickier; as Lee says, there are various things one might to
see under this heading, and gretl's offering are limited. This
is a case where people can write gretl addons to do the job.
Open source suggests perhaps more modularity than one gets
with a proprietary software like Stata (though nearly
everything it does is executed in .do or .ado add-ons).
What we discussed in Torun was the idea that Allin and Jack
would work on the back-bone and that others would try to
develop the expertise with the bundle concept to add
specific functionality. So, the question is, is enhancing
prediction or marginal effects a back-bone issue or an
add-on? (I'm not sure)
Good question, but I tend to think it could be addressed by
addons (though the writers of such addons might reasonably
request some additional built-in functions to make their lives
easier).
My idea of a back-bone issue would be the introduction of
factor variables (version 2.0). In it, variables are
defined as being continuous or discrete. They can be
interacted in various combinations (continuous-discrete,
continutous-continuous and discrete-discrete very easily
within Stata's .do files).
Gretl does have the continuous/discrete distinction. But we
could make more use of it.
Allin
7:25 a.m.
New subject: Report -- caution! long!
Thanks very much for the recent additions and improvements.
>> Now this is maybe I didn't know how to fully use gretl in
this
>> context. The issue arised on 2 occasions:
>> (1) I had a polynomial regression and I was showing them to enter from
>> the GUI the command something like:
>> prediction = $coeff[1] + $coeff[2]*x + $coeff[3]*x^2
Yes, in gretl that sort of thing is required if you want to
predict for some x-value that's not in the dataset. Or you
could do it by adding one observation to the dataset,
containing the x-values for which you want a prediction, then
asking for an out of sample forecast. I guess we could add a
variant of "fcast" (or maybe a new command-word is wanted),
for use after estimating a model, which takes a vector of
x-values as input and produces the prediction plus a standard
error. That woud be quite simple.
Maybe I am wrong but having to add an observation to the data set to
compute a prediction sounds a bit awkward. I also understand that
these can be easily done with small scripts. On the other hand, if
that is so, implementing a predict command that can encompass a
reasonable number of model types should not be very hard. My 2 Turkish
Kurus (which are less than 1 Euro Cent) is that starting a predict
command (and a button) that will initially work for linear and
polynomial OLS models as well as bivariate logit and probit models
(and maybe 1-2 more most commonly used types). Over time, as needs
arise maybe more can be added.
As Prof. Adkins observes, most people don't appreciate the challenge
of figuring out things or scripts. This is especially true for
professionals who want to get things done fast and as easily as
possible. So one thing that we consider trivial can be a major thing
or a show stopper for those people.
Here is (more or less) the exact converstion that took place:
Talha: OK here we enter these commands to obtain the probabilities for
the 3 cuts. Notice the sum adds up to 100%.
Students: There has to be an easier way to do this.
Talha: You know, gretl is so easy that it is spoiling us.
Students: Stata has it.
Talha: ??? :-P
Cheers
--
“An expert is a person who has made all the mistakes that can be made
in a very narrow field.” - Niels Bohr (1885-1962)
--
4432
days inactive
4433
days old
2 comments
3 participants
participants (3)
-
Allin Cottrell -
Lee Adkins -
Talha Yalta