[Limdep Nlogit List] Best-worst estimation

[William Greene]

Walt: NLOGIT allows rank data and it allows ties. For your case, 
the best is ranked 1, the worst is ranked 3 and the rest are tied
for second.  The full set of options for the random parameters
logit model is provided for this model, so you can handle repeated
measures (stated choice experiments) as well. Sawtooth does not
actually provide the coefficients for each person; that is a
misconception. They provide the posterior expectation of the parameter
vector conditioned on the data for each person. (E.g., two people
with the same data who make the same decisions have the same conditional
mean even though the actual realizations might be different because
of the random unobserved part.)  NLOGIT provides the same using 
classical methods, but the logic is the same.  It goes beyond the
"hierarchical Bayes" estimator in Sawtooth in that using NLOGIT, you
can build an upper level model for the coefficients with covariates,
as well as heteroscedasticity and autocorrelation.  Note, you need
NLOGIT for this capability. It is not supported in LIMDEP.
/Bill Greene



----- Original Message -----
From: "Data Analytics Corp." <dataanalytics at earthlink.net>
To: limdep at limdep.itls.usyd.edu.au
Sent: Wednesday, August 27, 2008 6:49:53 AM GMT -05:00 US/Canada Eastern
Subject: [Limdep Nlogit List] Best-worst estimation

Hi

A potential client asked for a form of analysis done in market research 
called best-worst or maximum-difference (maxdiff) analysis.  Basically, 
consumers are presented with a choice set containing a series of product 
attributes (say, four at a time or a "quad") and then are asked to 
select the one from the choice set they prefer the most and the one they 
prefer the least; the other items in the choice set are not ranked.  For 
a quad, the other two non-selected attributes just fall in the middle 
with no specific ranking (maybe we can asume they are ties?).  As an 
example, consumers may be asked to rank four marketing messages in a 
choice set.  They select the message they like the most and the one they 
like the least.  This is repeated several times with the messages 
rotating following an experimental design just as in a regular discrete 
choice problem.  So, in essence, this is a discrete choice problem with 
an incomplete ranking.  Does anyone know how to handle this problem in 
Limdep?  Sawtooth Software has a program to do this using hierarchical 
Bayes estimation to give utilities (coefficients) for each individual, 
which is what I actually need for other work for this problem.  So, 
whatever I do in Limdep, I'd like to be able to get the utilities 
(coefficients) at the individual level.

Thanks,

Walt

-- 
________________________

Walter R. Paczkowski, Ph.D.
Data Analytics Corp.
44 Hamilton Lane
Plainsboro, NJ 08536
________________________
(V) 609-936-8999
(F) 609-936-3733
dataanalytics at earthlink.net
www.dataanalyticscorp.com

_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au
http://limdep.itls.usyd.edu.au