1:36 p.m.
Hi everone,
I am new to gretl, i want to use its MLE for estimation. I have written
small script just to check that how MLE works but i have following error
message (criterion is not met a warning ). I have used sample data set data
2-1.gdt. here is my scrpit.
genr pi=3.14
scalar a=2*pi
scalar mu=1
scalar sigma=1
series s=sigma^2
sereis sq=2*s
series s3=sigma^3
scalar sq3=2*s3
series c=vsat-mu
series sq1=c^2
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
deriv mu=-c/s
deriv sigma=-1/sigma + (sq1/s3)
end mle
i do not know where is the problem
and if I use this code then i have error message that "*failed to invert OPG
matrix GG'"*
genr pi=3.14
scalar a=2*pi
scalar mu=1
scalar sigma=1
series s=sigma^2
sereis sq=2*s
series s3=sigma^3
scalar sq3=2*s3
series c=vsat-mu
series sq1=c^2
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
params mu sigma
end mle
or if i add at last
end mle --hessian
error message will be "*numerical hessian is not computable*"
Can anyone tell me about this error or how to solve it??
Thanks in advance.
--
Haider
3:07 p.m.
On Mon, 10 May 2010, RAZIA HAIDER wrote:
I am new to gretl, i want to use its MLE for estimation. I have
written
small script just to check that how MLE works but i have following error
message (criterion is not met a warning ). I have used sample data set data
2-1.gdt. here is my scrpit.
genr pi=3.14
scalar a=2*pi
scalar mu=1
scalar sigma=1
series s=sigma^2
sereis sq=2*s
series s3=sigma^3
scalar sq3=2*s3
series c=vsat-mu
series sq1=c^2
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
deriv mu=-c/s
deriv sigma=-1/sigma + (sq1/s3)
end mle
Please post a script that actually runs. The above has some basic
errors and will not get as far as producing a numerical error from
the mle command.
It fails first at "sereis sq=2*s". That typo is easily fixed, but
it then fails at "scalar sq3=2*s3". This is a type error: s3 is a
series so the formula "2*s3" yields a series result, which cannot
be coerced to a scalar. However, since sq3 is not actually used
below this line can be deleted.
Then we hit "vsat", which is presumably a data series. To
replicate your results, we need to know where that is coming from.
All that said, your mle block is not set up correctly. You have
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
deriv mu=-c/s
deriv sigma=-1/sigma + (sq1/s3)
end mle
It looks as if you are trying to estimate mu and sigma, but mu is
given no role in the loglikelihood so this cannot work. In
addition "sq" is apparently a function of sigma but it is not
updated within the mle block. Before the "deriv" lines you need to
update any quantities in the "logl" expression that depend on mu
and sigma.
Also note that "pi" is a built-in constant in gretl; you don't
need to define it yourself, nor do you need to calculate log(a)
within the mle block; it would suffice to do something like
scalar a = -0.5*log(2*pi)
before the mle block, then express logl as "a + ...".
Allin Cottrell
3:31 p.m.
Hi,
Thanks for quick reply. First i realized my basic mistakes :), now i have
changed script, which is given below
scalar a = -0.5*log(2*pi)
scalar m=mean(vsat)
scalar mu=m/2
scalar sigma=m/mu
mle logl=a-log(sigma)-(sq1/sq)
series s=sigma^2
series sq=2*s
series s3=sigma^3
series c=vsat-mu
series sq1=c^2
params mu sigma
end mle --hessian
if I run this i have same error of "*The convergence criterion was not met"*
*But*
if i choose in "preferences" L-BFGS-B , it works properly and give me
results.
I do not know reason but it works now :), if you can explain me the reason
behind it, I will be thankful to you.
Ciao!
haider
On 10 May 2010 17:07, Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Mon, 10 May 2010, RAZIA HAIDER wrote:
> I am new to gretl, i want to use its MLE for estimation. I have written
> small script just to check that how MLE works but i have following error
> message (criterion is not met a warning ). I have used sample data set
data
> 2-1.gdt. here is my scrpit.
>
> genr pi=3.14
> scalar a=2*pi
> scalar mu=1
> scalar sigma=1
> series s=sigma^2
> sereis sq=2*s
> series s3=sigma^3
> scalar sq3=2*s3
> series c=vsat-mu
> series sq1=c^2
>
> mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
> deriv mu=-c/s
> deriv sigma=-1/sigma + (sq1/s3)
> end mle
Please post a script that actually runs. The above has some basic
errors and will not get as far as producing a numerical error from
the mle command.
It fails first at "sereis sq=2*s". That typo is easily fixed, but
it then fails at "scalar sq3=2*s3". This is a type error: s3 is a
series so the formula "2*s3" yields a series result, which cannot
be coerced to a scalar. However, since sq3 is not actually used
below this line can be deleted.
Then we hit "vsat", which is presumably a data series. To
replicate your results, we need to know where that is coming from.
All that said, your mle block is not set up correctly. You have
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
deriv mu=-c/s
deriv sigma=-1/sigma + (sq1/s3)
end mle
It looks as if you are trying to estimate mu and sigma, but mu is
given no role in the loglikelihood so this cannot work. In
addition "sq" is apparently a function of sigma but it is not
updated within the mle block. Before the "deriv" lines you need to
update any quantities in the "logl" expression that depend on mu
and sigma.
Also note that "pi" is a built-in constant in gretl; you don't
need to define it yourself, nor do you need to calculate log(a)
within the mle block; it would suffice to do something like
scalar a = -0.5*log(2*pi)
before the mle block, then express logl as "a + ...".
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Razia Haider
Ph.D. student
Dipartimento di Ingegneria dell'Informazione
Universita' di Modena e Reggio Emilia
Via Vignolese 905, 41100 Modena, Italy
e-mail: razia.haider(a)unimo.it
tel: +39 059 2056314
4:09 p.m.
On Mon, 10 May 2010, RAZIA HAIDER wrote:
Thanks for quick reply. First i realized my basic mistakes :), now i
have
changed script, which is given below
scalar a = -0.5*log(2*pi)
scalar m=mean(vsat)
scalar mu=m/2
scalar sigma=m/mu
mle logl=a-log(sigma)-(sq1/sq)
series s=sigma^2
series sq=2*s
series s3=sigma^3
series c=vsat-mu
series sq1=c^2
params mu sigma
end mle --hessian
if I run this i have same error of "*The convergence criterion
was not met"* *But* if i choose in "preferences" L-BFGS-B , it
works properly and give me results.
I suspect you are initializing sigma very far away from the ML
estimate, so BFGS is having numerical trouble. L-BFGS-B will
sometimes work better under these conditions. (If "vsat" refers to
verbal SAT scores, the standard deviation will be in the hundreds,
far from the value of 2.0 that you set as a starting point.)
Another point is that estimating the variance parameter jointly
with the location parameter can, in general, be a difficult
numerical problem. It is common to concentrate the variance
parameter out of the loglikelihood to yield a more stable
calculation.
Allin Cottrell
4:15 p.m.
On Mon, 10 May 2010, Allin Cottrell wrote:
On Mon, 10 May 2010, RAZIA HAIDER wrote:
> Thanks for quick reply. First i realized my basic mistakes :), now i have
> changed script, which is given below...
By the way, if you want to use analytical derivatives, which may
give better convergence, here's a version that does that:
<script>
open data7-9
series y = vsat
scalar a = -0.5 * log(sqrt(2*pi))
# your initialization here...
scalar mu = 400
scalar sigma = 100
mle logl = a - log(sigma) - (y-mu)^2 / (2*sigma^2)
deriv mu = (y-mu) / (2*sigma^2)
deriv sigma = -1/sigma + (y-mu)^2 / (sigma^3)
end mle
# check results
scalar n = $nobs
scalar sigma_hat = sqrt(var(y) * (n-1)/n)
printf "mean = %g, sigma_hat = %g\n", mean(y), sigma_hat
</script>
Allin Cottrell
10:25 a.m.
Hi,
Thanks; I have changed initialization values of parameters and now its
working fine , also i run your script its also gving perfect results.
Thank you very much
Bye
Haider
On 10 May 2010 18:15, Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Mon, 10 May 2010, Allin Cottrell wrote:
> On Mon, 10 May 2010, RAZIA HAIDER wrote:
>
> > Thanks for quick reply. First i realized my basic mistakes :), now i
have
> > changed script, which is given below...
By the way, if you want to use analytical derivatives, which may
give better convergence, here's a version that does that:
<script>
open data7-9
series y = vsat
scalar a = -0.5 * log(sqrt(2*pi))
# your initialization here...
scalar mu = 400
scalar sigma = 100
mle logl = a - log(sigma) - (y-mu)^2 / (2*sigma^2)
deriv mu = (y-mu) / (2*sigma^2)
deriv sigma = -1/sigma + (y-mu)^2 / (sigma^3)
end mle
# check results
scalar n = $nobs
scalar sigma_hat = sqrt(var(y) * (n-1)/n)
printf "mean = %g, sigma_hat = %g\n", mean(y), sigma_hat
</script>
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Razia Haider
Ph.D. student
Dipartimento di Ingegneria dell'Informazione
Universita' di Modena e Reggio Emilia
Via Vignolese 905, 41100 Modena, Italy
e-mail: razia.haider(a)unimo.it
tel: +39 059 2056314
3:55 p.m.
Hi again ,
whatever i run it was in MLE window.
but when i try to run script in script command i have following problems:
open data2-1
scalar a = -0.5*log(2*pi)
scalar m=mean(vsat)
scalar mu=m/2
scalar sigma=m/mu
mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
params mu sigma
end mle --hessian
with this code error is "*Failed to compute numerical Hessian"*
**
And if i use derivatives it says "*NLS: The supplied derivatives seem to be
incorrect"*
**
open data2-1
scalar a = -0.5*log(2*pi)
scalar m=mean(vsat)
scalar mu=m/2
scalar sigma=m/mu
mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
deriv mu=(vsat-mu)/sigma^2
deriv sigma=-1/sigma+((vsat-mu)^2/sigma^3)
end mle
Derivatives seems to me fine but i do not know may be something wrong!!!
I think i am doing too much basic mistakes but i am not getting those :(,
Thanks in advance.
Ciao
Haider
On 10 May 2010 17:07, Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Mon, 10 May 2010, RAZIA HAIDER wrote:
> I am new to gretl, i want to use its MLE for estimation. I have written
> small script just to check that how MLE works but i have following error
> message (criterion is not met a warning ). I have used sample data set
data
> 2-1.gdt. here is my scrpit.
>
> genr pi=3.14
> scalar a=2*pi
> scalar mu=1
> scalar sigma=1
> series s=sigma^2
> sereis sq=2*s
> series s3=sigma^3
> scalar sq3=2*s3
> series c=vsat-mu
> series sq1=c^2
>
> mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
> deriv mu=-c/s
> deriv sigma=-1/sigma + (sq1/s3)
> end mle
Please post a script that actually runs. The above has some basic
errors and will not get as far as producing a numerical error from
the mle command.
It fails first at "sereis sq=2*s". That typo is easily fixed, but
it then fails at "scalar sq3=2*s3". This is a type error: s3 is a
series so the formula "2*s3" yields a series result, which cannot
be coerced to a scalar. However, since sq3 is not actually used
below this line can be deleted.
Then we hit "vsat", which is presumably a data series. To
replicate your results, we need to know where that is coming from.
All that said, your mle block is not set up correctly. You have
mle logl=-0.5*log(a)-log(sigma)-(sq1/sq)
deriv mu=-c/s
deriv sigma=-1/sigma + (sq1/s3)
end mle
It looks as if you are trying to estimate mu and sigma, but mu is
given no role in the loglikelihood so this cannot work. In
addition "sq" is apparently a function of sigma but it is not
updated within the mle block. Before the "deriv" lines you need to
update any quantities in the "logl" expression that depend on mu
and sigma.
Also note that "pi" is a built-in constant in gretl; you don't
need to define it yourself, nor do you need to calculate log(a)
within the mle block; it would suffice to do something like
scalar a = -0.5*log(2*pi)
before the mle block, then express logl as "a + ...".
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Razia Haider
Ph.D. student
Dipartimento di Ingegneria dell'Informazione
Universita' di Modena e Reggio Emilia
Via Vignolese 905, 41100 Modena, Italy
e-mail: razia.haider(a)unimo.it
tel: +39 059 2056314
8:22 p.m.
On Mon, 10 May 2010, RAZIA HAIDER wrote:
Hi again ,
whatever i run it was in MLE window.
but when i try to run script in script command i have following problems:
open data2-1
scalar a = -0.5*log(2*pi)
scalar m=mean(vsat)
scalar mu=m/2
scalar sigma=m/mu
mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
params mu sigma
end mle --hessian
with this code error is "Failed to compute numerical Hessian"
The line
mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
should read
mle logl=a-log(sigma)-((vsat-mu)^2/2*(sigma^2)).
Moreover, initialsing mu and sigma to sensible values is not a bad idea.
Finally, the --verbose switch is your friend :-)
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
10:23 a.m.
Thanks, I have changed initializing parameters now its works well .
Once again Thank You,
Cioa!
Haider
On 10 May 2010 22:22, Riccardo (Jack) Lucchetti <r.lucchetti(a)univpm.it>wrote:
On Mon, 10 May 2010, RAZIA HAIDER wrote:
Hi again ,
>
> whatever i run it was in MLE window.
>
> but when i try to run script in script command i have following problems:
>
> open data2-1
> scalar a = -0.5*log(2*pi)
> scalar m=mean(vsat)
> scalar mu=m/2
> scalar sigma=m/mu
>
> mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
> params mu sigma
> end mle --hessian
>
> with this code error is "Failed to compute numerical Hessian"
>
The line
mle logl=a-log(sigma)-((vsat-mu)^2/2*sigma^2)
should read
mle logl=a-log(sigma)-((vsat-mu)^2/2*(sigma^2)).
Moreover, initialsing mu and sigma to sensible values is not a bad idea.
Finally, the --verbose switch is your friend :-)
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Razia Haider
Ph.D. student
Dipartimento di Ingegneria dell'Informazione
Universita' di Modena e Reggio Emilia
Via Vignolese 905, 41100 Modena, Italy
e-mail: razia.haider(a)unimo.it
tel: +39 059 2056314
5325
days inactive
5326
days old
8 comments
3 participants
participants (3)
-
Allin Cottrell -
RAZIA HAIDER -
Riccardo (Jack) Lucchetti