[gov2001-l] Standard Errors of Probit

List overview All Threads
Download

newer

older

[gov2001-l] Removing duplicate...

[gov2001-l] MatchIt plots

sheldon.bond＠gmail.com

15 Apr 2006 15 Apr '06

10:16 p.m.

Show replies by date

yohai＠fas.harvard.edu

15 Apr 15 Apr

11:15 p.m.

Hi Sheldon, Without seeing the code, it is a little hard to diagnose the problem. Are you getting reasonable estimates for the coefficients? Best, Ian On Sat, 15 Apr 2006, Sheldon Bond wrote:

...

Greetings, In trying to produce standard errors from my optimization process, I got a hessian matrix with nothing but zeros. I have tried multiple things and can't seem to find the error in my coding. Anyone have any possible solutions? Regards, Sheldon _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

nyiry81＠yahoo.com

16 Apr 16 Apr

10:24 a.m.

New subject: [gov2001-l] question

Hey Ian (and list) I dont really understand the model(s) ( how they are used) and what I should do in this problemset. I would like to ask if anyone could give me a quick overview of the main ideas/concepts. I know I might ask a lot now...but honestly I got totally lost during the past two lectures/material. At this point I cant really get started on the homework so I figured I could give it a try and ask. At least try to understand even if I dont get points for it... sorry to bug you all. Anna Ian Brett Yohai <yohai(a)fas.harvard.edu> wrote: Hi Sheldon, Without seeing the code, it is a little hard to diagnose the problem. Are you getting reasonable estimates for the coefficients? Best, Ian On Sat, 15 Apr 2006, Sheldon Bond wrote:

...

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l --------------------------------- New Yahoo! Messenger with Voice. Call regular phones from your PC and save big.

yohai＠fas.harvard.edu

10:45 a.m.

New subject: [gov2001-l] question

Hi Anna, The key is always to begin with the stochastic and systematic components of the model. From this, it will be possible to write down the likelihood function, and then the log likelihood function. Then, you can translate that log likelihood into an R function and find the maximum using optim, as we have been doing on the last few problem sets. Although the models change, the basic procedure is the same. Gary discussed the probit model, which is used for binary dependent variables, in the single equation models slides. More background about the model can also be found in the Zelig help files, available off Gary's webpage. Best, Ian On Sun, 16 Apr 2006, anna nyiry wrote:

...

nyiry81＠yahoo.com

12:53 p.m.

New subject: [gov2001-l] question

Thanks Ian. I have more (probably very basic) questions. I found the model formula in the lecture notes. but it requires the phi. Dan said it is the cdf. Could you help me where I could find cdf for the normal distribution? anna Ian Brett Yohai <yohai(a)fas.harvard.edu> wrote: Hi Anna, The key is always to begin with the stochastic and systematic components of the model. From this, it will be possible to write down the likelihood function, and then the log likelihood function. Then, you can translate that log likelihood into an R function and find the maximum using optim, as we have been doing on the last few problem sets. Although the models change, the basic procedure is the same. Gary discussed the probit model, which is used for binary dependent variables, in the single equation models slides. More background about the model can also be found in the Zelig help files, available off Gary's webpage. Best, Ian On Sun, 16 Apr 2006, anna nyiry wrote:

...

Hey Ian (and list) I dont really understand the model(s) ( how they are used) and what I should do in this problemset. I would like to ask if anyone could give me a quick overview of the main ideas/concepts. I know I might ask a lot now...but honestly I got totally lost during the past two lectures/material. At this point I cant really get started on the homework so I figured I could give it a try and ask. At least try to understand even if I dont get points for it... sorry to bug you all. Anna Ian Brett Yohai wrote: Hi Sheldon, Without seeing the code, it is a little hard to diagnose the problem. Are you getting reasonable estimates for the coefficients? Best, Ian On Sat, 15 Apr 2006, Sheldon Bond wrote:

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l --------------------------------- Love cheap thrills? Enjoy PC-to-Phone calls to 30+ countries for just 2?/min with Yahoo! Messenger with Voice.

dhopkins＠fas.harvard.edu

1:32 p.m.

New subject: [gov2001-l] question

Hi Anna, You can find the formula for the normal CDF here: http://en.wikipedia.org/wiki/Normal_distribution although I am not sure how necessary knowing the formula is for the problem set. In terms of R functions, pnorm() evaluates the normal CDF. Best, Dan Quoting anna nyiry <nyiry81(a)yahoo.com>om>:

...

Hey Ian (and list) I dont really understand the model(s) ( how they are used) and what I

should do in this problemset. I would like to ask if anyone could give me a quick overview of the main ideas/concepts. I know I might ask a lot now...but honestly I got totally lost during the past two lectures/material. At this point I cant really get started on the homework so I figured I could give it a try and ask. At least try to understand even if I dont get points for it...

sorry to bug you all. Anna Ian Brett Yohai wrote: Hi Sheldon, Without seeing the code, it is a little hard to diagnose the problem. Are you getting reasonable estimates for the coefficients? Best, Ian On Sat, 15 Apr 2006, Sheldon Bond wrote:

big. _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l --------------------------------- Love cheap thrills? Enjoy PC-to-Phone calls to 30+ countries for just 2?/min with Yahoo! Messenger with Voice.

-- PhD Candidate in Government Tutor, Currier House Harvard University (617) 493-7058

ghumphr＠fas.harvard.edu

6:26 p.m.

New subject: [gov2001-l] question

Hello people, Be careful if reading other probability articles on Wikipedia; some of them contain prodigious unmarked errors. Geoff Quoting Dan Hopkins <dhopkins(a)fas.harvard.edu>du>:

...

Hey Ian (and list) I dont really understand the model(s) ( how they are used) and what I

should do in this problemset. I would like to ask if anyone could give me a quick overview of the main ideas/concepts. I know I might ask a lot

now...but

honestly I got totally lost during the past two lectures/material. At this point I cant really get started on the homework so I figured I could give

a try and ask. At least try to understand even if I dont get points for

it...

> > sorry to bug you all. > > Anna > > Ian Brett Yohai wrote: > Hi Sheldon, > > Without seeing the code, it is a little hard to diagnose the problem. Are > you getting reasonable estimates for the coefficients? > > Best, > Ian > > On Sat, 15 Apr 2006, Sheldon Bond wrote: > > > Greetings, > > In trying to produce standard errors from my optimization process, > > I got a hessian matrix with nothing but zeros. I have tried multiple > > things and can't seem to find the error in my coding. Anyone have any > > possible solutions? > > > > Regards, > > Sheldon > > > > _______________________________________________ > > 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 > > > > --------------------------------- > New Yahoo! Messenger with Voice. Call regular phones from your PC and

save

2?/min

with Yahoo! Messenger with Voice.

-- PhD Candidate in Government Tutor, Currier House Harvard University (617) 493-7058 _______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

king＠harvard.edu

5:15 a.m.

this almost surely means you have an error in the likelihood code (altho it can happen from input data that are wrong so check that too). its often something small, but you have to get it exactly right obviously. when this happens to me, i write over the math from scratch, as neatly as possible (and as obsessively as possible; try making all your equal signs line up vertically too!) and as compactly as possible. go through every step and make sure you totally understand it and its right. once you do that, then apply the same level of obesssiveness to the translation of the likelihood to the code. you'll find it. Gary On Sat, 15 Apr 2006, Sheldon Bond wrote:

...

ghumphr＠fas.harvard.edu

6:50 p.m.

Hi, In part d), should we use the medians of simulated samples or the medians of the sample data? Also, if you want us to use the medians of the simulated samples (which seems like a reasonable thing to do), should we use booleans representing differences of scale rather than dummies and interpolate linearly if the medians are computed with linear interpolation, or is there a more intelligent way of going about it? Also, I would like to hear more about use of the convex hull in Riemannian manifolds, which are not necessarily Euclidean spaces, or where covariates are known to be correlated, if there is time. Geoff Quoting Gary King <king(a)harvard.edu>du>:

...

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

yohai＠fas.harvard.edu

7:09 p.m.

Hi Geoff, You should use the medians of the sample data. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

...

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

7:19 p.m.

The fact that the independent variables are noncontinuous is not a concern for simulating standard errors? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>:

...

Hi Geoff, You should use the medians of the sample data. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Hi, In part d), should we use the medians of simulated samples or the medians

of the

sample data? Also, if you want us to use the medians of the simulated samples (which

seems

like a reasonable thing to do), should we use booleans representing

differences

of scale rather than dummies and interpolate linearly if the medians are computed with linear interpolation, or is there a more intelligent way of

going

about it? Also, I would like to hear more about use of the convex hull in Riemannian manifolds, which are not necessarily Euclidean spaces, or where covariates

are

known to be correlated, if there is time. Geoff Quoting Gary King <king(a)harvard.edu>du>: > > this almost surely means you have an error in the likelihood code (altho > it can happen from input data that are wrong so check that too). its > often something small, but you have to get it exactly right obviously. > when this happens to me, i write over the math from scratch, as neatly as > possible (and as obsessively as possible; try making all your equal signs > line up vertically too!) and as compactly as possible. go through every > step and make sure you totally understand it and its right. once you do > that, then apply the same level of obesssiveness to the translation of

the

likelihood to the code. you'll find it. Gary On Sat, 15 Apr 2006, Sheldon Bond wrote:

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

7:22 p.m.

nevermind Quoting ghumphr(a)fas.harvard.edu:

...

The fact that the independent variables are noncontinuous is not a concern for simulating standard errors? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>:

Hi Geoff, You should use the medians of the sample data. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Hi, In part d), should we use the medians of simulated samples or the medians

of the

sample data? Also, if you want us to use the medians of the simulated samples (which

seems

like a reasonable thing to do), should we use booleans representing

differences

of scale rather than dummies and interpolate linearly if the medians are computed with linear interpolation, or is there a more intelligent way of

going > about it? > > Also, I would like to hear more about use of the convex hull in

Riemannian

> manifolds, which are not necessarily Euclidean spaces, or where

covariates

are > known to be correlated, if there is time. > > Geoff > > > Quoting Gary King <king(a)harvard.edu>du>: > > > > > this almost surely means you have an error in the likelihood code

(altho

> > it can happen from input data that are wrong so check that too). its > > often something small, but you have to get it exactly right obviously. > > when this happens to me, i write over the math from scratch, as neatly

> > possible (and as obsessively as possible; try making all your equal

signs

> > line up vertically too!) and as compactly as possible. go through

every

> > step and make sure you totally understand it and its right. once you

> that, then apply the same level of obesssiveness to the translation of

the > > likelihood to the code. you'll find it. > > Gary > > > > On Sat, 15 Apr 2006, Sheldon Bond wrote: > > > > > Greetings, > > > In trying to produce standard errors from my optimization process, > > > I got a hessian matrix with nothing but zeros. I have tried multiple > > > things and can't seem to find the error in my coding. Anyone have

any

> possible solutions? > > Regards, > Sheldon > > _______________________________________________ > 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

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

7:24 p.m.

How large should we make the confidence intervals? 95%? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>:

...

Hi Geoff, You should use the medians of the sample data. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Hi, In part d), should we use the medians of simulated samples or the medians

of the

sample data? Also, if you want us to use the medians of the simulated samples (which

seems

like a reasonable thing to do), should we use booleans representing

differences

of scale rather than dummies and interpolate linearly if the medians are computed with linear interpolation, or is there a more intelligent way of

going

about it? Also, I would like to hear more about use of the convex hull in Riemannian manifolds, which are not necessarily Euclidean spaces, or where covariates

are

the

likelihood to the code. you'll find it. Gary On Sat, 15 Apr 2006, Sheldon Bond wrote:

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

yohai＠fas.harvard.edu

7:27 p.m.

Yes, 95% confidence intervals. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

...

How large should we make the confidence intervals? 95%? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>:

Hi Geoff, You should use the medians of the sample data. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

Hi, In part d), should we use the medians of simulated samples or the medians

of the

sample data? Also, if you want us to use the medians of the simulated samples (which

seems

like a reasonable thing to do), should we use booleans representing

differences

of scale rather than dummies and interpolate linearly if the medians are computed with linear interpolation, or is there a more intelligent way of

going

about it? Also, I would like to hear more about use of the convex hull in Riemannian manifolds, which are not necessarily Euclidean spaces, or where covariates

are

the

likelihood to the code. you'll find it. Gary On Sat, 15 Apr 2006, Sheldon Bond wrote: > Greetings, > In trying to produce standard errors from my optimization process, > I got a hessian matrix with nothing but zeros. I have tried multiple > things and can't seem to find the error in my coding. Anyone have any > possible solutions? > > Regards, > Sheldon > > _______________________________________________ > 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

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

17 Apr 17 Apr

12:01 a.m.

I've had luck with using Moore-Penrose matrix inverses for finding some of the standard errors, but I can't find anything useful on generalized Cholesky decomposition (and the relevant R packages appear broken)--any clues? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>:

...

Yes, 95% confidence intervals. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

How large should we make the confidence intervals? 95%? Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>: > Hi Geoff, > > You should use the medians of the sample data. > > Best, > Ian > > On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote: > > > Hi, > > > > In part d), should we use the medians of simulated samples or the

medians

> of the > > sample data? > > > > Also, if you want us to use the medians of the simulated samples (which > seems > > like a reasonable thing to do), should we use booleans representing > differences > > of scale rather than dummies and interpolate linearly if the medians

are

> > computed with linear interpolation, or is there a more intelligent way

> going > > about it? > > > > Also, I would like to hear more about use of the convex hull in

Riemannian

> > manifolds, which are not necessarily Euclidean spaces, or where

covariates

> are > > known to be correlated, if there is time. > > > > Geoff > > > > > > Quoting Gary King <king(a)harvard.edu>du>: > > > > > > > > this almost surely means you have an error in the likelihood code

(altho

> > > it can happen from input data that are wrong so check that too). its > > > often something small, but you have to get it exactly right

obviously.

> > > when this happens to me, i write over the math from scratch, as

neatly as

> > > possible (and as obsessively as possible; try making all your equal

signs

> > > line up vertically too!) and as compactly as possible. go through

every

> > > step and make sure you totally understand it and its right. once you

> > > that, then apply the same level of obesssiveness to the translation

> the > > > likelihood to the code. you'll find it. > > > Gary > > > > > > On Sat, 15 Apr 2006, Sheldon Bond wrote: > > > > > > > Greetings, > > > > In trying to produce standard errors from my optimization

process,

> > > > I got a hessian matrix with nothing but zeros. I have tried

multiple

> > > > things and can't seem to find the error in my coding. Anyone have

any

> > possible solutions? > > > > Regards, > > Sheldon > > > > _______________________________________________ > > 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 > _______________________________________________ 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

ghumphr＠fas.harvard.edu

12:07 a.m.

nevermind- found it Quoting ghumphr(a)fas.harvard.edu:

...

Yes, 95% confidence intervals. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

medians > > of the > > > sample data? > > > > > > Also, if you want us to use the medians of the simulated samples

(which

> seems > > like a reasonable thing to do), should we use booleans representing > differences > > of scale rather than dummies and interpolate linearly if the medians

are > > > computed with linear interpolation, or is there a more intelligent

way

> going > > about it? > > > > Also, I would like to hear more about use of the convex hull in

Riemannian

> > manifolds, which are not necessarily Euclidean spaces, or where

covariates

(altho > > > > it can happen from input data that are wrong so check that too).

its

> > > often something small, but you have to get it exactly right

obviously.

> > > when this happens to me, i write over the math from scratch, as

neatly as

> > > possible (and as obsessively as possible; try making all your equal

signs

> > > line up vertically too!) and as compactly as possible. go through

every > > > > step and make sure you totally understand it and its right. once

you

> > > that, then apply the same level of obesssiveness to the translation

process,

> > > > I got a hessian matrix with nothing but zeros. I have tried

multiple > > > > > things and can't seem to find the error in my coding. Anyone

have

any

> > > possible solutions? > > > > > > Regards, > > > Sheldon > > > > > > _______________________________________________ > > > 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 > > > > > > > _______________________________________________ > 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

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

1:05 a.m.

I'm not having any luck with generalized Cholesky decomposition (imaginary standard errors). Has anyone gotten it to work? Quoting ghumphr(a)fas.harvard.edu:

...

nevermind- found it Quoting ghumphr(a)fas.harvard.edu:

Yes, 95% confidence intervals. Best, Ian On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote:

medians > > of the > > > sample data? > > > > > > Also, if you want us to use the medians of the simulated samples

(which > > > seems > > > > like a reasonable thing to do), should we use booleans representing > > > differences > > > > of scale rather than dummies and interpolate linearly if the

medians

are > > > computed with linear interpolation, or is there a more intelligent

way

> going > > about it? > > > > Also, I would like to hear more about use of the convex hull in

Riemannian

> > manifolds, which are not necessarily Euclidean spaces, or where

covariates

(altho > > > > it can happen from input data that are wrong so check that too).

its > > > > > often something small, but you have to get it exactly right > obviously. > > > > > when this happens to me, i write over the math from scratch, as > neatly as > > > > > possible (and as obsessively as possible; try making all your

equal

> signs > > > > > line up vertically too!) and as compactly as possible. go

through

every > > > > step and make sure you totally understand it and its right. once

you > do > > > > > that, then apply the same level of obesssiveness to the

translation

process,

> > > > I got a hessian matrix with nothing but zeros. I have tried

multiple > > > > > things and can't seem to find the error in my coding. Anyone

have

any

> > > > possible solutions? > > > > > > > > Regards, > > > > Sheldon > > > > > > > > _______________________________________________ > > > > 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 > > > > > > > > > > > > > _______________________________________________ > > 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 > _______________________________________________ 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

ghumphr＠fas.harvard.edu

1:54 a.m.

The "it" in this case is http://gking.harvard.edu/files/sechol.s (a few changes are require to avoid floating point rounding errors; see the Gauss version). Quoting ghumphr(a)fas.harvard.edu:

...

I'm not having any luck with generalized Cholesky decomposition (imaginary standard errors). Has anyone gotten it to work? Quoting ghumphr(a)fas.harvard.edu:

nevermind- found it Quoting ghumphr(a)fas.harvard.edu: > I've had luck with using Moore-Penrose matrix inverses for finding some

> the > standard errors, but I can't find anything useful on generalized Cholesky > decomposition (and the relevant R packages appear broken)--any clues? > > Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>: > > > Yes, 95% confidence intervals. > > > > Best, > > Ian > > > > On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote: > > > > > How large should we make the confidence intervals? 95%? > > > > > > Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>: > > > > > > > Hi Geoff, > > > > > > > > You should use the medians of the sample data. > > > > > > > > Best, > > > > Ian > > > > > > > > On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote: > > > > > > > > > Hi, > > > > > > > > > > In part d), should we use the medians of simulated samples or the > > medians > > > > of the > > > > > sample data? > > > > > > > > > > Also, if you want us to use the medians of the simulated samples > (which > > > > seems > > > > > like a reasonable thing to do), should we use booleans

representing

> > > differences > > > > of scale rather than dummies and interpolate linearly if the

medians > > are > > > > > computed with linear interpolation, or is there a more

intelligent

> way > > of > > > > going > > > > > about it? > > > > > > > > > > Also, I would like to hear more about use of the convex hull in > > Riemannian > > > > > manifolds, which are not necessarily Euclidean spaces, or where > > covariates > > > > are > > > > > known to be correlated, if there is time. > > > > > > > > > > Geoff > > > > > > > > > > > > > > > Quoting Gary King <king(a)harvard.edu>du>: > > > > > > > > > > > > > > > > > this almost surely means you have an error in the likelihood

code

> > (altho > > > > > > it can happen from input data that are wrong so check that

too).

equal

> signs > > > > > line up vertically too!) and as compactly as possible. go

through > > every > > > > > > step and make sure you totally understand it and its right.

once

you > do > > > > > that, then apply the same level of obesssiveness to the

translation

of > > the > > > > likelihood to the code. you'll find it. > > > > Gary > > > > > > > > On Sat, 15 Apr 2006, Sheldon Bond wrote: > > > > > > > > > Greetings, > > > > > In trying to produce standard errors from my optimization process, > > > > > I got a hessian matrix with nothing but zeros. I have tried multiple > > > > > things and can't seem to find the error in my coding. Anyone

have

any > > > > > possible solutions? > > > > > > > > > > Regards, > > > > > Sheldon > > > > > > > > > > _______________________________________________ > > > > > 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 > > > > > > > > > > > > > > > > > > > _______________________________________________ > > > 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 > > > > > > > _______________________________________________ > 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

_______________________________________________ gov2001-l mailing list gov2001-l(a)lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

ghumphr＠fas.harvard.edu

23 Apr 23 Apr

6:53 a.m.

Hello list, After discussing the matter with Professor King, I found that I was applying the generalized Cholesky decomposition incorrectly. The proper way to apply it for finding pseudo-standard errors is, diag(sechol(diag(diag(pseudoinverse(-hessian))))) Cholesky decomposition is analagous to taking the square root of a matrix. I am still not entirely clear on how much information may be gained from Hessians which are not invertible or positive definite, but this technique is certainly interesting. It would of course be nice to coerce R into making use of the machine epsilon in order to avoid certain types of floating-point errors which could be encountered when using generalized Cholesky decomposition for other purposes. In cases where the machine epsilon is required but not known, another small constant might be used instead. Quoting ghumphr(a)fas.harvard.edu:

...

The "it" in this case is http://gking.harvard.edu/files/sechol.s (a few changes are require to avoid floating point rounding errors; see the Gauss version). Quoting ghumphr(a)fas.harvard.edu:

I'm not having any luck with generalized Cholesky decomposition (imaginary standard errors). Has anyone gotten it to work? Quoting ghumphr(a)fas.harvard.edu:

nevermind- found it Quoting ghumphr(a)fas.harvard.edu: > I've had luck with using Moore-Penrose matrix inverses for finding some

of > > the > > standard errors, but I can't find anything useful on generalized

Cholesky

> > decomposition (and the relevant R packages appear broken)--any clues? > > > > Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>: > > > > > Yes, 95% confidence intervals. > > > > > > Best, > > > Ian > > > > > > On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote: > > > > > > > How large should we make the confidence intervals? 95%? > > > > > > > > Quoting Ian Brett Yohai <yohai(a)fas.harvard.edu>du>: > > > > > > > > > Hi Geoff, > > > > > > > > > > You should use the medians of the sample data. > > > > > > > > > > Best, > > > > > Ian > > > > > > > > > > On Sun, 16 Apr 2006 ghumphr(a)fas.harvard.edu wrote: > > > > > > > > > > > Hi, > > > > > > > > > > > > In part d), should we use the medians of simulated samples or

the

> > > medians > > > > > of the > > > > > > sample data? > > > > > > > > > > > > Also, if you want us to use the medians of the simulated

samples

> (which > > > > seems > > > > > like a reasonable thing to do), should we use booleans

representing

> > > differences > > > > of scale rather than dummies and interpolate linearly if the

medians > > are > > > > > computed with linear interpolation, or is there a more

intelligent

code

> > (altho > > > > > > it can happen from input data that are wrong so check that

too). > > its > > > > > > > often something small, but you have to get it exactly right > > > obviously. > > > > > > > when this happens to me, i write over the math from scratch,

> neatly as > > > > > possible (and as obsessively as possible; try making all your

equal

> signs > > > > > line up vertically too!) and as compactly as possible. go

through > > every > > > > > > step and make sure you totally understand it and its right.

once > > you > > > do > > > > > > > that, then apply the same level of obesssiveness to the > translation > > > of > > > > > the > > > > > > > likelihood to the code. you'll find it. > > > > > > > Gary > > > > > > > > > > > > > > On Sat, 15 Apr 2006, Sheldon Bond wrote: > > > > > > > > > > > > > > > Greetings, > > > > > > > > In trying to produce standard errors from my

optimization

> > > process, > > > > > > > > I got a hessian matrix with nothing but zeros. I have

tried

> > > multiple > > > > > > > > things and can't seem to find the error in my coding.

Anyone

have > any > > > > > > possible solutions? > > > > > > > > > > > > Regards, > > > > > > Sheldon > > > > > > > > > > > > _______________________________________________ > > > > > > 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 > > > > > > > > > > > > > > > > > > > > > > > > > _______________________________________________ > > > > 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 > > > > > > > > > > > > > _______________________________________________ > > 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 > _______________________________________________ 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

6598

days inactive

6605

days old

gov2001@lists.gking.harvard.edu

Manage subscription

18 comments

6 participants

tags (0)

participants (6)

dhopkins＠fas.harvard.edu
ghumphr＠fas.harvard.edu
king＠harvard.edu
nyiry81＠yahoo.com
sheldon.bond＠gmail.com
yohai＠fas.harvard.edu