[gov2001-l] multinomial logit questions

List overview All Threads
Download

newer

older

[gov2001-l] Random effects logit

[gov2001-l] subsetting syntax error

jvaynman＠fas.harvard.edu

26 Mar 2008 26 Mar '08

4:49 p.m.

We are trying to estimate a multinomial logit model with robust standard errors (the cases are clustered by country). Can anyone help with these questions... 1) We have a dependent variable with three factors. It seems that Zelig takes one as the base and estimates coefficients and SE for the other two. Is there any way to know which ones it is taking? Is it possible to change which one the base should be and which should be estimated? The Zelig mlogit function does not seem to take a "robust=" argument. Does this mean it cannot do robust SE or is there some other way to set it up? 2) We also tried to use the multinom function in the nnet package. This function automatically gives us the right coefficients for the two factors we want, but the SE's are wrong, probably because they are not the robust ones. Is there a way to estimate robust SE's in the multinom function, or perhaps in another package or function? We have not found other options, but perhaps someone else has. Thanks!

Show replies by date

ogawa＠fas.harvard.edu

26 Mar 26 Mar

5:20 p.m.

New subject: [gov2001-l] What exactly is due?

Hi, Perhaps I missed this, but what are we supposed to turn in on Monday? Thank you, Chika

jhainmueller＠gmail.com

5:31 p.m.

New subject: [gov2001-l] What exactly is due?

Yes you did miss this :-) (see below)

...

-----Original Message----- From: Jens Hainmueller [mailto:jhainm at fas.harvard.edu] Sent: Monday, March 17, 2008 6:51 PM To: 'gov2001-l at lists.fas.harvard.edu' Subject: Important Announcements Hi all, please carefully read the following important announcements. Let us know if you have any questions. Thanks! 1. There is no official section this Thursday. Instead we will offer a walk in clinic where students can get help on their replications. We will hold the clinic 5-7pm in CGIS S450, the room for the first section. 2. There is no class and section next week, because of the university holiday. 3. There is no homework for next week, because your TFs are very generous people and we want you to put all the energy into the replications over spring break. 4. The current homework (PS 4) is due this Thursday 5PM as usual. Please submit a hard copy as either during the clinic or stuff it into my mailbox on the first floor. If you already left town you can also submit your homework online to the drop-box. 5. For all non-extension school students: Replications are due at the beginning of class on March 31st. As Gary explained today, your replication "packet" must contain the following items: a. The original paper as a pdf b. a draft of your own paper with little text but with figures and tables, and a proposed table of contents for your paper, in a relatively polished form. c. all the data, code, and information necessary to replicate the results of your analysis and reproduce your tables and figures. 5.1. The replication packet should be submitted in the form of a pointer to a private website where you uploaded the files. All Harvard grad students have access to their own web-space so this should be feasible for most of you. If not, you should submit the packet in the form of a cdrom where you saved all the files (we need two copies, so two cdroms). On that day we will assign your packet to another student who will then replicate your work. As a reminder: Extension school students do not have to turn in a replication. 6. The assigned reading for this week is chapter 6 in UPM. Good luck with the replications and happy spring break! Best, Jens

...

-----Original Message----- From: gov2001-l-bounces at lists.fas.harvard.edu [mailto:gov2001-l- bounces at lists.fas.harvard.edu] On Behalf Of Melodie Chika Ogawa Sent: Wednesday, March 26, 2008 5:20 PM To: gov2001-l at lists.fas.harvard.edu Subject: [gov2001-l] What exactly is due? Hi, Perhaps I missed this, but what are we supposed to turn in on Monday? Thank you, Chika _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

jbischof＠fas.harvard.edu

8:38 p.m.

New subject: [gov2001-l] What exactly is due?

Hey Jens, I'm glad you sent this out, but now I'm fairly confused as to the nature of this assignment. By "replication," I thought it meant reproduce the tables from the papers and submit the code. However, if I'm reading this right and you want the additions/changes that we will be doing for the actual paper, then that is something quite different. It is highly unlikely that we are going to reproduce the methods/tables whole cloth in our own papers, and so our recoding of the original methodology has little bearing on the actual paper if we are going to modify the estimation technique as the justification for our work. So, to clarify, what we're submitting is NOT the code necessary to reproduce the original tables, but code for our improvements on the authors' methodology and the resulting tables, as well as an outline of the paper? If so, I'm glad there was clarification, but it means that I have a lot more work to do before I can turn this in. Thanks, Jon 2008/3/26, Jens Hainmueller <jhainmueller at gmail.com>:

...

Yes you did miss this :-) (see below)

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

-- Jon Bischof Graduate Student Department of Government Harvard University

jhainmueller＠gmail.com

8:56 p.m.

New subject: [gov2001-l] What exactly is due?

Jon, Thanks for asking. For Monday we want you do produce the replication and the code. This is all that is required. Adding stuff about potential improvements that you have in mind or that you already pursued is optional. It may help the person that is writing the memo on your replication to give you a bit more targeted feedback. So you can include improvements if you want, but you do not have to do so at all. The assignment for Monday is "just" the replication. Hope this clarifies it. Best, Jens

...

-----Original Message----- From: gov2001-l-bounces at lists.fas.harvard.edu [mailto:gov2001-l- bounces at lists.fas.harvard.edu] On Behalf Of Jon Bischof Sent: Wednesday, March 26, 2008 8:38 PM To: gov2001-l at lists.fas.harvard.edu Subject: Re: [gov2001-l] What exactly is due? Hey Jens, I'm glad you sent this out, but now I'm fairly confused as to the nature of this assignment. By "replication," I thought it meant reproduce the tables from the papers and submit the code. However, if I'm reading this right and you want the additions/changes that we will be doing for the actual paper, then that is something quite different. It is highly unlikely that we are going to reproduce the methods/tables whole cloth in our own papers, and so our recoding of the original methodology has little bearing on the actual paper if we are going to modify the estimation technique as the justification for our work. So, to clarify, what we're submitting is NOT the code necessary to reproduce the original tables, but code for our improvements on the authors' methodology and the resulting tables, as well as an outline of the paper? If so, I'm glad there was clarification, but it means that I have a lot more work to do before I can turn this in. Thanks, Jon 2008/3/26, Jens Hainmueller <jhainmueller at gmail.com>:

Yes you did miss this :-) (see below) > -----Original Message----- > From: Jens Hainmueller [mailto:jhainm at fas.harvard.edu] > Sent: Monday, March 17, 2008 6:51 PM > To: 'gov2001-l at lists.fas.harvard.edu' > Subject: Important Announcements > > Hi all, > > please carefully read the following important announcements. Let

> know if you have any questions. Thanks! > > 1. There is no official section this Thursday. Instead we will

offer a

> walk in clinic where students can get help on their replications.

> will hold the clinic 5-7pm in CGIS S450, the room for the first > section. > > 2. There is no class and section next week, because of the

university

> holiday. > > 3. There is no homework for next week, because your TFs are very > generous people and we want you to put all the energy into the > replications over spring break. > > 4. The current homework (PS 4) is due this Thursday 5PM as usual. > Please submit a hard copy as either during the clinic or stuff it

into

> my mailbox on the first floor. If you already left town you can

also

> submit your homework online to the drop-box. > > 5. For all non-extension school students: Replications are due at

the

> beginning of class on March 31st. As Gary explained today, your > replication "packet" must contain the following items: > a. The original paper as a pdf > b. a draft of your own paper with little text but with figures and > tables, and a proposed table of contents for your paper, in a > relatively polished form. > c. all the data, code, and information necessary to replicate the > results of your analysis and reproduce your tables and figures. > > 5.1. The replication packet should be submitted in the form of a > pointer to a private website where you uploaded the files. All

Harvard

> grad students have access to their own web-space so this should be > feasible for most of you. If not, you should submit the packet in

the

> form of a cdrom where you saved all the files (we need two copies,

> two cdroms). On that day we will assign your packet to another

student

who will then replicate your work. As a reminder: Extension school students do not have to turn in a replication. 6. The assigned reading for this week is chapter 6 in UPM. Good luck with the replications and happy spring break! Best, Jens

> -----Original Message----- > From: gov2001-l-bounces at lists.fas.harvard.edu [mailto:gov2001-l- > bounces at lists.fas.harvard.edu] On Behalf Of Melodie Chika Ogawa > Sent: Wednesday, March 26, 2008 5:20 PM > To: gov2001-l at lists.fas.harvard.edu > Subject: [gov2001-l] What exactly is due? > > Hi, > > Perhaps I missed this, but what are we supposed to turn in on

Monday?

Thank you, Chika _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

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

-- Jon Bischof Graduate Student Department of Government Harvard University _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

king＠harvard.edu

9:14 p.m.

New subject: [gov2001-l] What exactly is due?

and just to emphasize the 2nd part of what Jens just said: the more you turn in, the more help you will get with your paper from the other student replicating your work. when you get out of grad school, an absolutely critical factor in your ultimate success will be maintaining a group of trusted peers who will give you feedback on your work before it goes public. Getting someone to spend the time necessary to give you real feedback is really hard and completely essential. So we're simulating this process as part of this assignment. Many of those who have taken this class in the past (including, incidentally, several of the authors of the articles being replicated) made their first real connections like this with others among their fellow class members. So then, as now, you want to finish as much as possible so you will learn the most. but now, of course, its likely that you won't be able to do it all right now, so do what you can. Gary On Wed, 26 Mar 2008, Jens Hainmueller wrote:

...

> know if you have any questions. Thanks! > > 1. There is no official section this Thursday. Instead we will

offer a

> walk in clinic where students can get help on their replications.

> will hold the clinic 5-7pm in CGIS S450, the room for the first > section. > > 2. There is no class and section next week, because of the

university

into

> my mailbox on the first floor. If you already left town you can

also

> submit your homework online to the drop-box. > > 5. For all non-extension school students: Replications are due at

the

Harvard

> grad students have access to their own web-space so this should be > feasible for most of you. If not, you should submit the packet in

the

> form of a cdrom where you saved all the files (we need two copies,

> two cdroms). On that day we will assign your packet to another

student

Monday?

Thank you, Chika _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

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

jhainmueller＠gmail.com

6:19 p.m.

Jane, vglm(VGAM) also a multinomial logit. Zelig by default uses the lowest value as reference category so recoding that is an easy way to change the baseline. As far as I am aware there is no canned way to do clustered SEs from multinomial logit in R. Unless you find a canned way to do this, the best you can probably do is either code up the clustered SEs n your own (the sandwich library has a canned function bread() which gives you the bread matrix but the meat matrix you have to code construct on your own, you can use the estfun() in sandwich to get the score values and with these you have to compute the outer products cluster by cluster (see the Stata (programming) manual on _robust for the equations of the MLE sandwich estimator). Another possibility is to estimate J-1 binary logits (with clustered SEs from svyglm(survey)), these will give you coefficients that are very close but not identical to the multinomial logit. It's fair to say that R is not very good at providing functions for clustered SEs for MLE models. The main reason for this that most statisticians agree that robust (or clustered robust) SEs don't make much sense for MLE models. They do make sense for linear models were the beta estimates remain consistent and unbiased even if not all the usual assumptions for the errors are satisfied so that "correcting" the variances for violations of non-independence is sensible. In non-linear models, however, the beta estimates are usually inconsistent if the errors are not homoscedastic, so you either need to model the dependence explicitly in the likelihood function or your betas will be inconsistent if the errors are indeed heteroskedastic. So in this case, "correcting" the variances using robust SEs is not very attractive, because you are effectively correcting the standard errors of inconsistent parameters. Stata is famous for providing robust SEs for most non-linear models, in R most people think this is a rather silly thing to do so canned functions are rare (the sandwich package being one exception and robcov as well). Here is a short article on the topic http://www.stat.berkeley.edu/~census/mlesan.pdf Maybe Gary can talk more about this issue in lecture at some point if people are interested. Jens

...

-----Original Message----- From: gov2001-l-bounces at lists.fas.harvard.edu [mailto:gov2001-l- bounces at lists.fas.harvard.edu] On Behalf Of Jane E. Vaynman Sent: Wednesday, March 26, 2008 4:49 PM To: gov2001-l at lists.fas.harvard.edu Subject: [gov2001-l] multinomial logit questions We are trying to estimate a multinomial logit model with robust standard errors (the cases are clustered by country). Can anyone help with these questions... 1) We have a dependent variable with three factors. It seems that Zelig takes one as the base and estimates coefficients and SE for the other two. Is there any way to know which ones it is taking? Is it possible to change which one the base should be and which should be estimated? The Zelig mlogit function does not seem to take a "robust=" argument. Does this mean it cannot do robust SE or is there some other way to set it up? 2) We also tried to use the multinom function in the nnet package. This function automatically gives us the right coefficients for the two factors we want, but the SE's are wrong, probably because they are not the robust ones. Is there a way to estimate robust SE's in the multinom function, or perhaps in another package or function? We have not found other options, but perhaps someone else has. Thanks! _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

jcastner＠fas.harvard.edu

27 Mar 27 Mar

8:38 a.m.

New subject: [gov2001-l] logicals

Does anyone know how, in R, to do the equivalent of the stata command: drop if country == "UK" & durest ==31 I would really apprechiate anyones help! Johannes

5893

days inactive

5894

days old

gov2001@lists.gking.harvard.edu

Manage subscription

7 comments

6 participants

tags (0)

participants (6)

jbischof＠fas.harvard.edu
jcastner＠fas.harvard.edu
jhainmueller＠gmail.com
jvaynman＠fas.harvard.edu
king＠harvard.edu
ogawa＠fas.harvard.edu