home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!uvaarpa!darwin.sura.net!paladin.american.edu!auvm!MHS.UNC.EDU!CROSS.CPC
- X-Delivery-Notice: SMTP MAIL FROM does not correspond to sender.
- Message-ID: <9301261944.AA04963@hermes.oit.unc.edu>
- Newsgroups: bit.listserv.stat-l
- Date: Tue, 26 Jan 1993 14:33:00 LCL
- Sender: STATISTICAL CONSULTING <STAT-L@MCGILL1.BITNET>
- From: CROSS <CROSS.CPC@MHS.UNC.EDU>
- Subject: summary of multinomial logit
- Lines: 218
-
- On Tuesday, 1/19 I posted the following question to STAT-L:
-
- ************************************************************************
- Users of multinomial logits:
- What software do you use and why? Do any programs also calculate
- probabilities for you? Our center is looking for an alternative to PROC
- MLOGIT which runs under version 5 of SAS. Will post responses to the
- list. Thanks!
-
- Catherine Cross
- Carolina Population Center
- CB#8120, University Square East
- University of North Carolina
- Chapel Hill, NC 27516-3997
- ************************************************************************
-
- I received responses from 9 multinomial logit users over the past week.
- SAS procedures were mentioned most frequently but I was surprised at
- the wide variety of programs listed. The following programs were
- recommended: SAS(PROC CATMOD-4 mentions; PROC LOGISTIC-2 mentions;
- PROC PHREG), GLIM(2 mentions), LIMDEP(2 mentions), NOMINAL, SPIDA,
- STATA, HOTZTRAN, LOGIT procedure under TSP, LOGIT procedure in QUANT
- module in MARKOV (running under GAUSS), and LOGIT module of SYSTAT.
- See responses below for details. Thanks to everyone who replied!
- ************************************************************************
-
- Response #1
-
- Awhile back (pre SAS vers 6) I too was on a software hunt for the
- multinomial logit (unordered categorical outcome) problem. PROC CATMOD
- worked if you knew the right tricks to use -- but the outcome was not
- easily decipherable -- estimates were in a consistent but not readily
- apparent order -- once you figured out how SAS ordered the coefficients
- you could consistently decode the output. I recently tried to run my
- CATMOD tricks in the new version 6, but it seems they were specific to
- version 5 -- I have tried to figure out how to translate the version 5
- tricks into version 6, but have not yet succeeded.
-
- I was on this software hunt about the time MLOGIT was released. It
- seemed to me that MLOGIT was great for ordered categorical outcomes,
- which reflected the needs of the econometricians (as evidenced by the
- references in the documentation). For unordered MLOGIT wasnt all that
- useful.
-
- The easiest software to use, in my opinion, was the LOGIT procedure
- in TSP (surprising, given that TSP is the domain of econometricians).
- If you can get your data in a rectangular array, you simply use a few
- TSP commands (use the @FIT options for fitted probabilities for the
- observations in your data array -- if you want probabilities for
- hypothetical values, it may require some work) and you have your model
- estimately relatively quickly and with well labeled output.
-
- At the time, the most sophisticated software for the problem was
- called NOMINAL (obtained from Univ Laval in Quebec). This was the
- software used by Dubin and Pasternak in their very readable article in
- AJE entitled "Risk Assessment for Case-Control Subgroups by
- Polychotomous Logistic Regression", Am J Epi, 123(6): 1101-1117 (1986).
- Since I was working in a biostatistical context this was a very useful
- reference for me, and I tried to use the software. However, I never
- got it to run successfully (mostly due to the limitation of PCs at the
- time -- should not be a problem now that we are in the 486 age).
-
- Of course a "real" biostatistician would recommend GLIM -- it does it
- all if you know what you want and how to program GLIM.
-
- Hosmer and Lemeshow, in their book Applied Logistic Regression, give a
- good explication of multinomial (aka polychotomous, aka polytomous),
- but no software recommendations. I have other references if you need
- them, but that's about it for software. Hope this is of at least some
- help.
-
- Ken Petronis
- The Urban Institute
- Washington, DC
- "UI4300::PETRONIS"@ui.urban.org
- ************************************************************************
-
- Response #2
-
- You can use the PHREG procedure in version 6 of the SAS System to fit
- multinomial logit models. If you send me your postal address, I will
- be glad to send you more information. [And he did!]
-
- Warren F. Kuhfeld
- Statistical R & D
- SAS Institute Inc.
- Cary, NC 27513-2414
- saswfk@unx.sas.com
- ************************************************************************
-
- Response #3
-
- SAS and SPIDA come to my mind! In SAS, you can use PROC LOGISTIC and
- CATMOD to fit multinomial logit models. You can specify different LINK
- functions in PROC LOGISTIC to fit Proportional ODDS (cumulative logit
- models) and Adjacent category (Continuation ratio models), for ORDINAL
- (response) data. You can also use CATMOD to fit (among other forms of
- logits) a POLYTOMOUS LOGIT models. I'm refering to SAS 6.07!
-
- SPIDA is a relatively new software and can be used to fit (some of the
- ??) models mentioned above. I have not used SPIDA much, but have heard
- from other users that its good.
-
- Personally, I prefer to use SAS to fit multinomial models because I
- find SAS to be excellent for data management. (I don't want to switch
- between softwares for data management and analyses).
-
- Cande V Ananth
- Department of Biostatistics, SPH
- University of North Carolina at Chapel Hill
- ananth@UNC.bitnet
- ************************************************************************
-
- Response #4
-
- PROC CATMOD in SAS does multinomial logit models quite nicely. It will
- also output the predicted probabilities for each of C-1 logistic
- equations for dependent variables having C categories. It can also
- handle continuous independent variables via the DIRECT statement.
-
- I believe I ran the same model in MLOGIT and CATMOD once some time ago
- and compared the coefficients, log-liklihoods etc. If memory serves
- correctly, the results were the same in both.
-
- Hope this helps.
-
- Kenneth A. Hardy
- Assoc. Dir. for Information Systems
- Inst. for Research in Social Science
- Manning Hall CB#3355
- University of North Carolina
- Chapel Hill, N.C. 27599-3355
- KENHARDY@UNCVM1.BITNET
- KHARDY.IRSS@MHS.UNC.EDU
- KENHARDY@UNCVM1.OIT.UNC.EDU
- ************************************************************************
-
- Response #5
-
- I've used LIMDEP fairly successfully, and I believe that version 6 of
- SAS has this available as well -- I don't have my manuals handy, but I
- remember seeing it in SAS/STAT
-
- George Jakubson
- CORNELL UNIVERSITY
- AK5J@CORNELLA.CIT.CORNELL.EDU
- ************************************************************************
-
- Response #6
-
- I haven't done multinomial logit analysis myself, but I understand that
- GLIM is an excellent program for carrying out such analyses.
-
- Clive Payne at Oxford University (Nuffield College) is a GLIM
- enthusiast, and uses it for loglinear and logit analyses - and is
- usually willing to help with advice. His e-mail address is:
- nuff@oxford.ac.uk.
-
- Stephen Gourlay
- Senior Lecturer
- Kingston University
- bs_s467@neptune.king.ac.uk
- ************************************************************************
-
- Response #7
-
- I generally use Limdep 6.0 on the PC. UNC has Limdep on the mainframe,
- too. I like the PC version better than the mainframe. I also use
- Stata, which I am getting to like more and more as I get familiar with
- it. It is written for both Unix and DOS platforms. I have used both
- and find them to be very easy to use. The Intercooled DOS version
- (written for 386 based machines with at least three megabytes of
- memory) is unbelievably fast computation-wise. The 386 version of
- Limdep is very fast, too.
-
- David Guilkey is quite familiar with both, you could probably talk to
- him about it. Also available at UNC is Hotztran, which is not such a
- friendly program. In any case, Guilkey knows about as much about
- limited dependent variable computation as anyone on the planet, so you
- could get the best info from him.
-
- Cheers,
-
- George McCarthy
- MCCARTHY@LEVY.BARD.EDU
- ************************************************************************
-
- Response #8
-
- In addition to SAS PROC CATMOD and PROC LOGISTIC you can try the LOGIT
- procedure in the QUANT module in MARKOV (running under GAUSS). Use the
- utility program ATOG.EXE to create a GAUSS datafile from an ASCII file
- (see Chapter 17 in the GAUSS System and Graphics Manual). MARKOV is
- very easy to use. GAUSS is not. Get in touch with Ken Hardy at IRSS
- to get access to the program. It should be accessible on the social
- sciences network server. The program is not supported by the IRSS
- staff, so you'll have to learn it yourself. However, last year there
- was a statistics group comprised of faculty, staff, and grad students
- at IRSS which partly was intended as a GAUSS user group. It was
- organized by Walt Davis, a sociology graduate student.
-
- Good luck,
- Magnus
-
- MAGNUS STENBECK
- mastfoh%tellus.sos.se%sos@mail.swip.net
- ************************************************************************
-
- Response #9
-
- The LOGIT module of SYSTAT also fits multinomial logit models for a
- categorical response variable that is not ordinal. Agresti's book
- Categorical Data Analysis (1990 edition) gives an overview of extant
- computer programs for the different types of models.
-
- Jacqui Cater
- IEGH71U@tjuvm.tju.edu
- ************************************************************************
-