7 Nov
2004
7 Nov
'04
7:19 a.m.
Good question, Jason.
For #2: Optimze over s^2, but treat nu as fixed and optimize
iteratively by increasing nu in small increments, e.g. for nu =
1:1000.
----- Original Message -----
From: "Jason Morris Lakin" <jlakin(a)fas.harvard.edu>
To: <olau(a)fas.harvard.edu>
Sent: Sunday, November 07, 2004 11:30 AM
Subject: RE: #2
hey olivia. i don't really understand problem 2.
are you
asking us not to
maximize over v (as we do in prob 1 with sig2)? i was treating
v and s as
ancillary parameters like sig2 in prob 1. but then you can't
maximize
given different v's. so is just s an ancillary parameter?
should i
maximize across only 5 instead of 6 parameters?
thanks
j
7 Nov
7 Nov
7:24 a.m.
just to clarify one point: what olivia is suggesting here is an
optimization trick for likelihoods that optim has hard time with. if
optim were smarter we wouldn't have to do this and could treat nu
identically to sigma and beta. logically, we are treathing all the
parameters the same, but we're just giving the program a bit of a boost.
Gary
On Sun, 7 Nov 2004, Olivia Lau wrote:
Good question, Jason.
For #2: Optimze over s^2, but treat nu as fixed and optimize
iteratively by increasing nu in small increments, e.g. for nu =
1:1000.
----- Original Message -----
From: "Jason Morris Lakin" <jlakin(a)fas.harvard.edu>
To: <olau(a)fas.harvard.edu>
Sent: Sunday, November 07, 2004 11:30 AM
Subject: RE: #2
hey olivia. i don't really understand
problem 2. are you
asking us not to
maximize over v (as we do in prob 1 with sig2)? i was treating
v and s as
ancillary parameters like sig2 in prob 1. but then you can't
maximize
given different v's. so is just s an ancillary parameter?
should i
maximize across only 5 instead of 6 parameters?
thanks
j
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
7:36 a.m.
New subject: AW: [gov2001-l] Fw: #2
I understand what prob 2 asks for, but I am still a bit confused about why
exactly we want to maximize over nu. this may be a dumb question, but nu is
simply the df, right? and for a certain number of obs and a certain number
of coeff. we want to estimate, the df should be fixed. in our case we have
14 obs, and estimate 4 parameters so we are left with 10df for the model
right?
if we maximise the liklehood function for df:1,...1000 and find that the
highest value is at let's say df=250, what have we learned from this, if all
we really have are 10 df?
-----Ursprungliche Nachricht-----
Von: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]
Gesendet: Sunday, November 07, 2004 12:25 PM
An: gov2001-l(a)lists.fas.harvard.edu
Betreff: Re: [gov2001-l] Fw: #2
just to clarify one point: what olivia is suggesting here is an
optimization trick for likelihoods that optim has hard time with. if
optim were smarter we wouldn't have to do this and could treat nu
identically to sigma and beta. logically, we are treathing all the
parameters the same, but we're just giving the program a bit of a boost.
Gary
On Sun, 7 Nov 2004, Olivia Lau wrote:
Good question, Jason.
For #2: Optimze over s^2, but treat nu as fixed and optimize
iteratively by increasing nu in small increments, e.g. for nu =
1:1000.
----- Original Message -----
From: "Jason Morris Lakin" <jlakin(a)fas.harvard.edu>
To: <olau(a)fas.harvard.edu>
Sent: Sunday, November 07, 2004 11:30 AM
Subject: RE: #2
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
hey olivia. i don't really understand
problem 2. are you
asking us not to
maximize over v (as we do in prob 1 with sig2)? i was treating
v and s as
ancillary parameters like sig2 in prob 1. but then you can't
maximize
given different v's. so is just s an ancillary parameter?
should i
maximize across only 5 instead of 6 parameters?
thanks
j
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
9:48 a.m.
New subject: AW: [gov2001-l] Fw: #2
nu, beta, and sigma are all parameters of the likelihood. they have
different substantive interpretations, but mathematically they each have
the identical logical status. when you max the likelihood function, you
search for the values of nu, beta, and sigma such that the log-likelihood
is at the maximum.
be careful not to get the method of optimization (which we have you trying
in 2 steps) confused with the big picture goal of the procedure (as in my
para above).
Gary
On Sun, 7 Nov 2004, Jens Hainmueller wrote:
I understand what prob 2 asks for, but I am still a
bit confused about why
exactly we want to maximize over nu. this may be a dumb question, but nu is
simply the df, right? and for a certain number of obs and a certain number
of coeff. we want to estimate, the df should be fixed. in our case we have
14 obs, and estimate 4 parameters so we are left with 10df for the model
right?
if we maximise the liklehood function for df:1,...1000 and find that the
highest value is at let's say df=250, what have we learned from this, if all
we really have are 10 df?
-----Ursprungliche Nachricht-----
Von: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]
Gesendet: Sunday, November 07, 2004 12:25 PM
An: gov2001-l(a)lists.fas.harvard.edu
Betreff: Re: [gov2001-l] Fw: #2
just to clarify one point: what olivia is suggesting here is an
optimization trick for likelihoods that optim has hard time with. if
optim were smarter we wouldn't have to do this and could treat nu
identically to sigma and beta. logically, we are treathing all the
parameters the same, but we're just giving the program a bit of a boost.
Gary
On Sun, 7 Nov 2004, Olivia Lau wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
Good question, Jason.
For #2: Optimze over s^2, but treat nu as fixed and optimize
iteratively by increasing nu in small increments, e.g. for nu =
1:1000.
----- Original Message -----
From: "Jason Morris Lakin" <jlakin(a)fas.harvard.edu>
To: <olau(a)fas.harvard.edu>
Sent: Sunday, November 07, 2004 11:30 AM
Subject: RE: #2
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
hey olivia. i don't really understand
problem 2. are you
asking us not to
maximize over v (as we do in prob 1 with sig2)? i was treating
v and s as
ancillary parameters like sig2 in prob 1. but then you can't
maximize
given different v's. so is just s an ancillary parameter?
should i
maximize across only 5 instead of 6 parameters?
thanks
j
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
9:56 a.m.
New subject: AW: [gov2001-l] Fw: #2
Once you give #2 a try, you'll see why we had you use a finite
number for nu...
----- Original Message -----
From: "Gary King" <king(a)harvard.edu>
To: <gov2001-l(a)lists.fas.harvard.edu>
Sent: Sunday, November 07, 2004 2:48 PM
Subject: Re: AW: [gov2001-l] Fw: #2
nu, beta, and sigma are all parameters of the likelihood.
they have
different substantive interpretations, but mathematically they
each have
the identical logical status. when you max the likelihood
function, you
search for the values of nu, beta, and sigma such that the
log-likelihood
is at the maximum.
be careful not to get the method of optimization (which we
have you trying
in 2 steps) confused with the big picture goal of the
procedure (as in my
para above).
Gary
On Sun, 7 Nov 2004, Jens Hainmueller wrote:
I understand what prob 2 asks for, but I am still
a bit
confused about why
exactly we want to maximize over nu. this may be a dumb
question, but nu is
simply the df, right? and for a certain number of obs and a
certain number
of coeff. we want to estimate, the df should be fixed. in our
case we have
14 obs, and estimate 4 parameters so we are left with 10df
for the model
right?
if we maximise the liklehood function for df:1,...1000 and
find that the
highest value is at let's say df=250, what have we learned
from this, if all
we really have are 10 df?
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
-----Ursprungliche Nachricht-----
Von: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]
Gesendet: Sunday, November 07, 2004 12:25 PM
An: gov2001-l(a)lists.fas.harvard.edu
Betreff: Re: [gov2001-l] Fw: #2
just to clarify one point: what olivia is suggesting here
is an
optimization trick for likelihoods that optim has hard time
with. if
optim were smarter we wouldn't have to do this and could
treat nu
identically to sigma and beta. logically, we are treathing
all the
parameters the same, but we're just giving the program a
bit of a boost.
Gary
On Sun, 7 Nov 2004, Olivia Lau wrote:
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
Good question, Jason.
For #2: Optimze over s^2, but treat nu as fixed and
optimize
iteratively by increasing nu in small increments, e.g.
for nu =
1:1000.
----- Original Message -----
From: "Jason Morris Lakin" <jlakin(a)fas.harvard.edu>
To: <olau(a)fas.harvard.edu>
Sent: Sunday, November 07, 2004 11:30 AM
Subject: RE: #2
> hey olivia. i don't really understand problem 2. are
> you
> asking us not to
> maximize over v (as we do in prob 1 with sig2)? i was
> treating
> v and s as
> ancillary parameters like sig2 in prob 1. but then you
> can't
> maximize
> given different v's. so is just s an ancillary
> parameter?
> should i
> maximize across only 5 instead of 6 parameters?
>
> thanks
> j
>
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
8 Nov
8 Nov
6:24 a.m.
New subject: [gov2001-l] Another question re: #2
Every time I run my function for #2, I get this message:
Error in optim(start, llike.t, method = "BFGS", control = list(fnscale =
-1), : non-finite finite-difference value [4]
I can't figure out exactly what's going wrong in the function (or even
what that means...). I've pasted the code below...anyone have any ideas?
Thanks,
Matt
# (a) Log-likelihood function ##this definitely works for given values of
the parameters
llike.t <- function(par,X,Y,nu){
xbeta <- t(X)%*%par[1:nrow(X)]
gam <- par[nrow(X) + 1]
nu <- nu
sum(log(gamma((nu+1)/2))-log(gamma(nu/2))-1/2*log(nu)-1/2*gam+(nu+1)/2*log(1
+ ((Y-xbeta)^2)/(nu*exp(gam))))
}
#(b) Numerical Optimization
start <- c(lm(Y ~ t(X) - 1)$coef, 1)
results <- matrix(0, nrow=98, ncol=6)
t.optim <- function(){
for (i in 3:100){
nu <- i
MLE <- optim(start, llike.t, method="BFGS", control=list(fnscale =
-1), hessian = TRUE, X=X, Y=Y, nu=nu)
point.estimate <- c(MLE$par[1:4], exp(MLE$par[5]))
results[i-2,] <- c(point.estimate, MLE$value)
}
}
6:50 a.m.
New subject: [gov2001-l] Another question re: #2
The non-finite thing means that you have a different number of
starting values than you have parameters. I would check the
number of columns in X vs the number of starting values. If you
have k columns in X, you should have k+1 starting values.
Also, in your log likelihood function, I think you need to
exp(gamma) -- I'm not sure one this because I didn't decipher
your math fully -- but the idea is that all the paremeters
should be optimized over an unbounded space. So for s and
sigma, you have to parameterize them so that the parameter
you're optimizing isn't bounded below at zero.
Hope this helps!
Olivia
----- Original Message -----
From: "Matthew Ryan Smith" <smith21(a)fas.harvard.edu>
To: <gov2001-l(a)lists.fas.harvard.edu>
Sent: Monday, November 08, 2004 11:24 AM
Subject: [gov2001-l] Another question re: #2
Every time I run my function for #2, I get this
message:
Error in optim(start, llike.t, method = "BFGS", control =
list(fnscale =
-1), : non-finite finite-difference value [4]
I can't figure out exactly what's going wrong in the function
(or even
what that means...). I've pasted the code below...anyone have
any ideas?
Thanks,
Matt
# (a) Log-likelihood function ##this definitely works for
given values of
the parameters
llike.t <- function(par,X,Y,nu){
xbeta <- t(X)%*%par[1:nrow(X)]
gam <- par[nrow(X) + 1]
nu <- nu
sum(log(gamma((nu+1)/2))-log(gamma(nu/2))-1/2*log(nu)-1/2*gam+(nu+1)/2*log(1
+ ((Y-xbeta)^2)/(nu*exp(gam))))
}
#(b) Numerical Optimization
start <- c(lm(Y ~ t(X) - 1)$coef, 1)
results <- matrix(0, nrow=98, ncol=6)
t.optim <- function(){
for (i in 3:100){
nu <- i
MLE <- optim(start, llike.t, method="BFGS",
control=list(fnscale =
-1), hessian = TRUE, X=X, Y=Y, nu=nu)
point.estimate <- c(MLE$par[1:4], exp(MLE$par[5]))
results[i-2,] <- c(point.estimate, MLE$value)
}
}
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
9 Nov
9 Nov
12:41 p.m.
New subject: [gov2001-l] Another question re: #2
Hi Olivia,
I'm using the MatchIt software to do some nonparametric preprosessing of my
data set.
Once I have the matched sample, is there a way to export that sample in
stata format?
I've used the "matched.data" command to extract it and save it in R using
save(mydata, file="mydata.dta")
but when i try to open it in stata the format is not recognized.
Thanks very much for the advise,
Cecilia
1:36 p.m.
New subject: [gov2001-l] Another question re: #2
library(foreign)
write.dta(...)
You will be unable to analyze the data in Stata because the whole point of
matching is to generate an average treatment effect, or counterfactual
estimate. Zelig does this pretty much seamlessly but Stata (even with
Clarify) needs a ton of programming.
Try demo(match) in Zelig to see what I'm talking about...
----- Original Message -----
From: "Cecilia Vidal" <cecilia_vidal(a)harvard.edu>
To: <gov2001-l(a)lists.fas.harvard.edu>
Sent: Tuesday, November 09, 2004 5:41 PM
Subject: RE: [gov2001-l] Another question re: #2
Hi Olivia,
I'm using the MatchIt software to do some nonparametric preprosessing of
my
data set.
Once I have the matched sample, is there a way to export that sample in
stata format?
I've used the "matched.data" command to extract it and save it in R using
save(mydata, file="mydata.dta")
but when i try to open it in stata the format is not recognized.
Thanks very much for the advise,
Cecilia
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
4:16 p.m.
New subject: [gov2001-l] replication question
Hey Gary & Olivia,
For our replication, the data we obtained from the original author was the
original, uncleaned world bank survey data. (we were not able to obtain
the finished, cleaned data from the author)
Now as we are going through and cleaning the data, we're (tentatively)
finding that the original author has dropped more observations from the
original data then we have. It does not appear like we will end up with
the same cleaned set of data as the author (the author may have dropped
several additional observations out of a sample of 20,000, for reasons we
do not know).
If we can show that the observations that we preserved in our cleaning of
the very original data set is reasonable, is it ok if none of our
replication results are the same as the authors? (even the most basic
tables of means and standard deviations of variables will be different)
thanks
yang
6:50 p.m.
New subject: [gov2001-l] replication question
Try to push it forward further. contact the author if necessary.
if you can't replicate the authors' results, then anything you do to
improve on the author's results might be viewed by reviewers as due to the
failure to replicate rather than your improvements. so better to nail
down this issue as closely as possible.
Gary
On Tue, 9 Nov 2004, Yang Lin wrote:
Hey Gary & Olivia,
For our replication, the data we obtained from the original author was the
original, uncleaned world bank survey data. (we were not able to obtain
the finished, cleaned data from the author)
Now as we are going through and cleaning the data, we're (tentatively)
finding that the original author has dropped more observations from the
original data then we have. It does not appear like we will end up with
the same cleaned set of data as the author (the author may have dropped
several additional observations out of a sample of 20,000, for reasons we
do not know).
If we can show that the observations that we preserved in our cleaning of
the very original data set is reasonable, is it ok if none of our
replication results are the same as the authors? (even the most basic
tables of means and standard deviations of variables will be different)
thanks
yang
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
7131
days inactive
7133
days old
gov2001@lists.gking.harvard.edu
10 comments
6 participants
participants (6)
-
cecilia_vidal＠harvard.edu -
jens_hainmueller＠ksg05.harvard.edu -
king＠harvard.edu -
olau＠fas.harvard.edu -
smith21＠fas.harvard.edu -
yanglin＠fas.harvard.edu