| This site has the April 2006 version of Proscal5.0 for SPlus which is currently in
beta test mode. Changes from Proscal4.5 include the following:
- an ability to evaluate liking ratings under the assumption of
dependent sampling
- an improved initial estimation procedure for liking ratings under
independent sampling
- a new optimization procedure which updates coordinates and
measurement constant parameters at the same time instead of separately.
- calculation of CAIC and BIC information criterion statistics
Dependent sampling has the provocative property of being able to
uniquely estimate a multidimensional solution with just one ideal object
using only liking rating data. Dependent sampling must, of course, characterize the data as well as the model. Tests of whether
data fit a dependent or independent sampling model may be made using the CAIC statistic.
Another advantage of dependent sampling is that the variances on the
real and ideal objects are uniquely estimated, even when the only data
available are liking ratings. (With independent sampling, an
arbitrary constant may be added/subtracted from the real/ideal object
variances without changing the likelihood, as long as all variances are
positive.)
Dependent sampling assumes that when a subject evaluates real
objects, a single sample is made from the subject's ideal distribution
and the resulting values are used in making comparisons with values
drawn from the distributions of all of the other real objects.
Independent sampling assumes that a different ideal sample is drawn for
each comparison to a real object. When each row of a data matrix
represents a different subject, it seems plausible that dependent
sampling is the more appropriate model.
The ability to evaluate data with only one ideal object is valuable
when a market is homogeneous or when it is suspected that different market segments perceive products
differently. This is a common occurrence. Consumers who are
diet conscious, for example, frequently perceive product properties
differently from consumers who are not diet conscious. Traditional
analyses, based upon the simultaneous analysis of multiple segments,
assume that all segments perceive products equally and that differences
in segment preferences are due only to differences in the estimation of
ideal products.
A liking rating data set for one ideal with dependent sampling is
provided in the SPlus Data zip file under the name deplikratd.dat.
The data were simulated using a reversed nine point integer scale where
a one means a product is very much liked and a nine indicates it is very
much not liked. Parametric target files for coordinates and variances are named
deplikratc.dat and deplikratv.dat. The data are for 12 real
objects and one ideal object. An isotropic model with one variance
for all objects on all dimensions should be estimated. No
measurement model estimates need be requested.
The manuals for the SPlus and command line versions of Proscal have also been updated..
For complex liking rating data sets, ones in which individual
respondents may be using very different transformations, it should be
remembered that better results may be obtained if the liking ratings are
first transformed to and analyzed as binary choices. Binary
choices may be analyzed under the assumption of independent or dependent
sampling.
Feedback on Proscal5.0 is desired. Please send your comments to
mackay@proscal.com.
|
