[讨论][转帖]multinomial regression with sample selection bias

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

Date: Thu, 8 Sep 2005 15:28:48 -0400 From: Helen Labun Jordan <hjordan@UVM.EDU> Subject: Mlogit with sample selection I'm trying to run a multinomial regression with sample selection bias. The context is that I have four clusters and want to predict the factors that assign survey respondents to one of the clusters. . . the problem is that the 4th is simply "undifferentiated", so I would not expect a coherent set of significant factors for assignment to this cluster. I have a probit equation written to predict whether or not a response can be classified. . . but then I can't seem to hold those results over to the multinomial (mlogit). Help? Thank you, Helen Jordan (hjordan@uvm.edu) -- Helen Labun Jordan Center for Rural Studies 207 Morrill Hall Burlington, VT 05405 Office Phone: 802 656-0254

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏0 回帖

关键词：Multinomial regression Selection regressio Election regression Sample Multinomial Bias Selection

Dear Users,

I am currently using NLOGIT to estimate a random parameter logit. Does anyone know whether the p-values attached to the standard deviation = estimates adjust for the fact that this is a positive quantity? Or, = should we not worry about it? Many thanks best wiji

[此贴子已经被作者于2005-9-18 2:02:12编辑过]

Date: Thu, 8 Sep 2005 07:56:35 +1000 From: David <garym@ITLS.USYD.EDU.AU> Subject: RP\SP RP\SP is very straightforward if you use nested logit set up with 2 branches , one for RP and one for SP. You can set IV or scale of one branch to 1 and allow other to be free and you get the differences in variance or scale. You can simply treat RP and SP as separate observations that are pooled. But note that we find that the differences in sacle are often between alts across SP and RP and not between SP and RP (since scale is simply picking up differences in the unobserved effects content). We discuss all of this plus how to do it in Nlogit in our new book David Hensher, John Rose and Bill Greene (2005) Applied Choice Analysis : a primer, Cambridge uni press, Cambridge. David

1. How to get the SP-RP scale parameter for logit model calibrate with RP & SP data? (2)

----------------------------------------------------------------------

Date: Wed, 7 Sep 2005 09:47:55 +0800 From: Susanna Tam <susanna.tam@POLYU.EDU.HK> Subject: How to get the SP-RP scale parameter for logit model calibrate with RP & SP data?

Anybody know how to run a multinomial logit model using RP and SP data? Did the SP-RP scale parameter run out automatically, or I need to write some commands to get it? Anyone can help? Many thanks.

------------------------------

Date: Tue, 6 Sep 2005 22:11:27 -0400 From: Fred Feinberg <feinf@UMICH.EDU> Subject: Re: How to get the SP-RP scale parameter for logit model calibrate with RP & SP data?

Susanna Tam wrote:

Anybody know how to run a multinomial logit model using RP and SP data? Did the SP-RP scale parameter run out automatically, or I need to write some commands to get it? Anyone can help? Many thanks.

We do this all the time in LIMDEP, but not in a very elegant way. Because the 'coefficient' of the Gumbel error in standard logit is fixed to one, what you really want to do is say that, say, the RP error coefficient is 1 (standard logit), and the SP coefficient is some number you wish to estimate. A simple, brute force way to do this is to write a batch job to divide all your SP data by a variety of values, and run a logit model on your SP data and RP data together. Because this is a discrete choice model, dividing all your data by some value is the same as multiplying your error by that value. Monitor the LLs until you get the optimum divisor, and that's your scaling factor. You can then do a likelihood ratio test against the logit run with no scaling to assess significance.

My co-author Linda Salisbury (who did this in her dissertation) and I have done this hundreds of times, and it's only failed to converge maybe twice. Logit runs very fast, so running a batch job is no problem. You can usually estimate the scale parameter and run all tests in a few minutes, and the likelihood ratio test is preferred econometrically to looking at asymptotic t-like values from the logit output itself.

There is probably a simpler way to do this using an HEV model, but I don't see how to get around manual scaling.

Fred

=====

Prof. Fred Feinberg Stephen M. Ross School of Business University of Michigan feinf@umich.edu

[此贴子已经被作者于2005-9-18 2:03:43编辑过]

Hello list,

I'm trying to run a panel model, and test for fixed/random effects, and I get the error: Could not invert VC matrix for Hausman test. I'm assuming this is because the VC matrix is singular, or close to it? I've printed it, looked at it, and don't see anything obvious. Is multicollinearity the likely culprit, or is there some other obvious thing I should be looking at?

Many thanks in advance for advice,

Darla Munroe

Hello,

Singularity is probably the problem. One possibility is to rescale your variables. I ran into a similar situation when analyzing experimental data. I rescaled my variables (e.g. lab dollars to actual dollar earnings) and was able to run the test.

Good luck,

Chris McIntosh

We get the "Could not invert VC matrix for Hausman test" warning in perhaps one-third of our state-by-year panel analyses. Our reading of page R11-28 of the version 8.0 reference guide makes us think this isn'treally a problem. Computation of the Hausman statistic involves taking the inverse of the DIFFERENCE of two VC matrices (fixed effects minus random effects). If those VC matrices are extremely similar, this couldlead to inversion problems. So Greene writes "when the differencematrix is not PD, you should use zero for the Hausman statistic". Note that LIMDEP shows a zero statistic in this instance (failing to reject random effects), along with the warning above.

Bill Ponicki Prevention Research Center Berkeley, CA, USA

[此贴子已经被作者于2005-9-18 2:00:41编辑过]

I HAVE SOME QUESTIONS AS WHAT YOU HAD EXPLAINED,  BUT ALSO PUZZLED