%@ Language=VBScript %> <% CheckState() CheckSub() %>
World Association of | February 2003 |
Jon Pinnell
MarketVision Research
Today's
rampant product proliferation makes it hard to imagine that there once was a
simpler time with few products and easier choices. However, the days of one
product for all buyers has almost universally passed. In an effort to appeal to
buyers and differentiate their products from competitors' products,
manufacturers and service providers have developed more and more products, each
different than existing products. This trend has existed for years in virtually
every industry including healthcare.
Markets
consist of products and buyers. Products differ because buyers differ. Buyers
potentially differ in many ways: their attitudes, their perceptions of brands,
their buying patterns, the needs their products should meet, their price
sensitivity, and many other ways.
Wendell
Smith, who in 1956 wrote 'Product Differentiation and Market Segmentation as
Alternative Marketing Strategies,' formally introduced the idea of segmentation.
Before this, econometric models explained most market structure, with the key
determination being product demand and substitution as a function of price.
Items such as product benefits, buyers' needs, marketing communication, and
branding were classified as non-price competition and frequently ignored as not
material. An interesting early work is Robinson's The Economics of Imperfect
Competition.
Before the
idea of segmentation was formally introduced, manufacturers started to realize
that buyers differ, but didn't have the structure or depth of understanding of
the specific differences, or why these differences mattered. Knowing that buyers
differ, manufacturers began launching new products based largely on either their
internal beliefs or internal capabilities.
Both
product differentiation and market segmentation produce markets that contain
varied products. However, market segmentation aligns the varied products with
the wants and needs of the buyers. Market segmentation reflects the evolution
from a product-driven organization to a consumer-driven organization.
Over the
past 50 years, market segmentation has been one of the most researched marketing
topics, and a standard for most companies. Despite the prevalence of the 'idea'
of segmentation, there is little agreement on the 'practice' of segmentation.
There are an astonishingly large number of specific approaches and techniques
used for market segmentation. Questions (and heated debates) continue about what
dimension should be used to identify segments as well as what techniques should
be used to distinguish segments.
In deciding
how to form segments, several questions must be answered:
What method should be used to segment the market?
What basis should be used to segment the market?
How many segments should be formed?
The method
of segmentation refers to when the segments are 'defined'. There are basically
two methods of segmentation: a priori and post hoc.
Segmentation
requires that respondents be grouped based on some set of variables that are
identified before data collection. In a priori segmentation not only are the
variables identified, but also the segment definitions. For example, consider a
study in which respondents are to be grouped based on age and number of
children. In this scheme, each respondent could be placed in the proper category
as soon as these pieces of information were known.
Post hoc
segmentation, on the other hand, requires only that the variables be identified.
The segments are then determined by using an analytical technique (often cluster
analysis)[1]
after all the data is collected. For example, if post hoc segments were to be
formed on psychographics, the psychographic measures for all respondents would
be cluster analyzed. Thus, the segments will not be known until after running
cluster analysis. This allows groups to be formed based on natural breaks in the
responses. This method is sometimes referred to as response based.
After
determining what method of segmentation to use, a second critical question that
must be answered is what characteristics should be used to identify the
segments. These characteristics are referred to as the basis of segmentation.
Common basis variables are discussed below:
Historically,
the most common methods of segmentation have been demographic in nature
women between 18 and 44 years, for example. These schemes have frequently been
motivated from a perspective of shopping or media behavior and often do not
distinguish brand behavior well between segments.
Geographic
segmentation calls for dividing the market into geographic regions. This allows
the company to operate in only a few regions, because of limited demand or
resources, or the company can tailor some element of the marketing mix by
region. Food products are frequently varied by region. For example, because
tastes differ, recipes are varied for different parts of the country. The
differing products are sold in the same container with the same advertising and
promotion in different geographic regions.
Attitudinal
variables are used based on the idea that opinions, traits, interests, and
lifestyles influence buying behavior. Put another way, people with similar
psychographic profiles should respond similarly to various marketing mixes. In
particular, attitudinal segmentation can be useful to guide positioning and
communication strategies.
Product
usage variables are also used to segment markets based on category usage,
product usage, or brand loyalty. It is a common practice to segment the market
into users and non-users and also by level of use though the practice of
focusing on heavy users has been criticized (Twedt).
People
decide which product to buy in different ways. The influence that people
exercise over decisions, as well as their brand loyalty, variety seeking
behavior, information search, distribution channel used and promotional
sensitivities all impact the brand that they will purchase. While this focus is
more prevalent in business-to-business settings, it is perfectly applicable to
consumer buying as well.
The
firmographic variables of industry, revenue, and employee size are standard
bases for segmenting business-to-business markets. Firmographic segmentation is
based on the idea that companies can be targeted based simply on what industry
they are in or how large they are. The special needs of segmenting the
business-to-business market are discussed in Moriarty and Reibstein.
Rather than
relying on descriptive factors (like demographics, region, brand loyalty, or
product usage), needs based or benefit based segments are derived based on
causal factors relating to the purchase process. People buy products, and decide
between different brands, because of the unique benefits that one product can
provide over another and the relative importance of those benefits. Benefit
segmentation partitions the market into segments based on benefits sought. The
resulting segments include people that are seeking to realize some benefit(s)
with a particular product based.
Another
question that must be answered in any segmentation study is how many segments to
use. The decision on how many segments has considerations from both research
(statistical) and managerial viewpoints. The segments can always be made more
homogeneous by increasing the number of segments. This could continue until the
number of segments becomes unmanageable. One practical guideline, although not a
mandate, is that unless the market can be broken into at least three groups, the
advantages of segmenting are not cost effective over a general market approach;
and attempting to market to more than seven segments is not practical for most
companies.
There are
obviously many possible ways to segment a market. There is no standard right
answer and each situation can produce a different 'best' solution. What should
be obvious is that although segmentation research involves substantial
scientific processes, it also relies inescapably on judgment and the art of
segmentation. While not wanting to further fuel this specific debate, two
solutions that we have found particularly effective include using product needs
or buyers' motivations as the bases of segmentation.
The idea of
using buyers' needs as a basis of segmentation was introduced by Haley (1968).
It suggests the most effective way to divide the market to guide product design
is based on the unique needs that buyers are trying to meet with the product.
For example, with a managed care product, the needs might include monthly
premium price, co-pay price, access, network size, and availability of
preventive services with different groups of buyers placing varying importance
of the various needs. Similarly, for a surgical automation device, the needs
might include device price, cost per use, ease of use, efficacy, level of
automation, and breadth of capabilities. To the extent that potential buyers
have differing needs, segmentation and therefore multiple products makes
sense.
An
alternative basis of segmentation is attitudinal or motivational segmentation,
broadly referred to as psychometric. While less clearly useful for product
design, this approach has found success guiding product positioning and
marketing communications. Early discussions in this area of segmentation can be
found in Lazarfeld and in Dichter. Psychographics are discussed in general in
Gunter and Furnham as well as in Piirto.
By
specifically mentioning these two bases of segmentation we are not suggesting
that others are inappropriate. Rather, each segmentation exercise is unique, and
various solutions can each be appropriate for each specific situation. These two
approaches, though, seem to have the broadest appeal in a wide variety of
instances.
Before
determining if a firm should pursue a segmented marketing strategy, it has been
suggested the segments must meet six key criteria: identifiability,
substantiality, accessibility, stability, responsiveness, and actionability
(Frank, Massy, and Wind). These requirements are each discussed below.
The
individual segments must be different from each other. If all buyers in the
market behave basically the same, there is no need to segment the market.
Likewise, if segments are found in which the behaviors across groups are
basically indistinguishable, the segmentation scheme is not accomplishing what
it should be.
Markets for
non-differentiated, or commodity products would not benefit from a segmented
approach. A manufacturer of these products can more efficiently reach the entire
market with one marketing strategy than with several. There are, however, few
markets that are absolutely not segmentable. That is, almost every market has
identifiable groups that behave differently and by focusing on these
differences, companies can increase their marketing effectiveness.
The
segments must be large enough to sustain the additional cost incurred by
segmentation (like developing specialized communication messages). Segment size
is a relative question based on the size of the company, the size of the market,
and the buying power of each customer. Some firms can define each potential
customer as a separate segment and tailor their products and marketing mix
accordingly due to each buyer's individual buying power. A manufacturer of over
the counter pharmaceuticals, however, must segment the market on a more general
basis.
The
segments must also be reachable. To the extent that a segment exists, but cannot
be educated about a product's potential benefits by advertising, point of
purchase, or another method, it is of very little use to the company and of no
more value than the unsegmented market.
Stability
over time is also a reasonable concern for segmentation studies. The question
here is not so much that one person or one company stays in the same segment all
the time people and companies change, but that over time, the same basic
segments exist, in about the same size, with the same basic profiles.
The two key
criteria in our work are identifiability and responsiveness.
Identifiability
requires that each segment be unique from other segments. That is, that there is
heterogeneity between segments and homogeneity within segments. Without meeting
this requirement there is no need to or benefit of segmenting a market.
Responsiveness
requires that the heterogeneity in buyers be related to the buyers' product
choice processes.
Additional
segmentation background is discussed from a strategic perspective in Myers, and
from a more theoretical and technical perspective in Wedel and
Much of the
early work in segmentation was based on clustering or grouping consumers based
on similar traits. Many different techniques have been used to form market
segments, but cluster is the most common. Most clustering algorithms can be
classified as one of two standard types: hierarchical or partitioning.
Hierarchical
methods often start with each observation as a unique segment and joins segments
together until all observations are joined into the same cluster. The segments
are joined by joining 'similar' elements or by joining elements such that the
resulting clusters have the smallest possible within-group sums of squares. With
these kinds of clustering, once two units are joined together, they cannot be
pulled apart. That is, a true hierarchical relationship is assumed. (Some
hierarchical methods start with one large cluster and split out clusters until
each observation is its own cluster.)
An
alternative to the hierarchical structure is the family of partitioning or
divisive methods. The partitioning clustering methods are more varied, but they
work on similar principals. These clustering methods are iterative methods. The
typical process is to search the data set to find seed points. The method used
to find the original seeds is one of the major differences among these methods,
and generally rather critical (Milligan). After the original seeds are found,
another pass is made through the data set assigning each case to one of the
seeds based on a calculated distance between each observation and all seeds. The
observation is assigned to the nearest seed. After this pass the group means are
re-calculated and these become the new seeds. Several passes through the data
set are made to reclassify cases based on the updated seeds. A convergence
criterion is specified to stop the iterations when the cluster solution becomes
stable.
Although
these partitioning methods were originally designed to be efficient clustering
algorithms for very large data sets, they have been widely accepted in marketing
research applications, especially for segmentation. Punj and Stewart report that
'Ward's Minimum Variance, average linkage, and several variants of the iterative
partitioning method tend to outperform all other methods. K-means (a standard
partitioning method) appears to outperform both Wards and average linkage if a
non-random starting point is specified.' They go on to say that the 'K-means
procedure appears to be more robust than any of the hierarchical methods with
respect to the presence of outliers... '.
Rather than
clustering data for each person, much recent segmentation has been focused on
disaggregating a total market into segments based on consumer responses or
behaviors. Recent examples include Allenby, Aurora and Ginter; Lenk, Desarbo,
Green and White; Desarbo et al; and Meyer et al. This interest in disaggregation
is appropriate for a wide variety of marketing applications, including market
opportunity assessment, product design, portfolio optimization, positioning,
advertising copy development, media placement, direct marketing, satisfaction
assessment, pricing, and channel management. The balance of this paper explores
unique ways to disaggregate buyers in three specific applications for entities
in healthcare industries: satisfaction research, product design research, and
prospect targeting.
While the
need to recognize the differences between buyers is real and growing in nearly
every industry, several trends heighten the need to understand the variety of
buyers in the healthcare industries specifically. These include:
mergers
and acquisitions;
direct
to consumer marketing including CRM;
deregulation;
information
available to the user, buyer, patient especially via the Internet;
increasing
cost pressure and price awareness/sensitivity.
Given
these new forces and changes, the need to understand customers and their unique
needs and characteristics is stronger than ever, allowing healthcare marketers
to:
retain
current customers;
develop
new products and services;
acquire
new customers.
Each is
discussed below.
In the
healthcare arena, satisfaction research is conducted among a wide variety of
audiences (patients, users, providers) for a wide variety of purposes. The
purposes might include accreditation, compensation, regulatory information, or
process improvement. It is the last item that is of relevance here when the
research is conducted to guide process improvements with the intent of
increasing outcomes such as overall satisfaction, brand loyalty or financial
performance. Some of the interactions between these items are illustrated in figure
1.
Researchers
frequently conduct statistical analysis on satisfaction data to help prioritize
areas for process improvement. For example, satisfaction with specific process
characteristics, such as ease of scheduling an appointment, length of time
waiting to see a doctor, accuracy of insurance filings, and access to a
specialist physician, might be analyzed to determine which one has the strongest
relationship with overall satisfaction. It is not uncommon for researchers to
use linear regression, logistic regression, structural equation modeling, or
simple correlations to explain these relationships. Sometimes referred to as key
driver analysis, this approach develops a 'weight' that describes either the
strength (as with correlation analysis) or the magnitude (as with regression
analysis) of the relationship between each process area and overall
satisfaction.
The weights
developed through any of the techniques just mentioned are identical for all
respondents. This implies that each consumer has the same motivations with the
same relative influence assigned to each motivation. This implicit assumption of
homogeneous motivations is clearly a false over-simplification in nearly every
instance.
Latent
Class Analysis is a variation on regression analysis[2]
that simultaneously finds segments based on respondents' answers and conducts
regression for each identified segment. In this application, the segments are
derived based on the similarity of each segment's weights, or relative
motivations in explaining overall satisfaction or other outcome measures. The
procedure of conducting a Latent Class Analysis is discussed more completely in
Pinnell (2001).
Consider
the following example with synthetic data (into which we added error to each
respondents' responses). Suppose there are two populations in the data with
known and different importance structures (weights) as shown in table
1.
After
conducting a simple linear regression by regressing X1, X2, and X3 onto the
overall rating, we observe the parameter estimates (ί1, ί2, and ί3) shown in table
2.
The model
appears to fit well as the Adjusted R2 is a respectable .75 and all three
parameter estimates are significant (p < 0.05). From a modeling standpoint,
such regression results would please many analysts.
However, we
see that the model suggests that all attributes are, for all practical purposes,
of equal importance, and suggests that the third attribute is the most
important. A researcher conducting this aggregate level analysis would be
unaware that two distinct populations exist and might suggest that the third
attribute be an area of focus, even though it is the least important to the
largest percentage of the population.
After
fitting a two-class latent class regression model to these data, the researcher
obtains the parameter estimates shown in table
3.
This
researcher has located two distinct and identifiable groups in the data. The
first group places the highest degree of importance on X1, while the second
group places the most importance on X3. The researcher would now correctly
conclude that, for a large percentage of the population, efforts to improve the
organization's performance on X1 would yield the most influence on overall
satisfaction. Similarly, efforts to improve organizational performance on X3
would impact overall satisfaction for a smaller, but readily identifiable,
percentage of the customer population.
The
following demonstrates Latent Class Analysis with actual survey data. The data
are from approximately 1,500 interviews conducted among members of a large
health plan. Please note that the data have been disguised.
Table
4 shows the relative importance weights predicting overall satisfaction. The
items are broken into three tiers of importance with five items in each tier,
from most important to least important. (+++ = most important, + = least
important). The most important attributes dealt broadly with medical advice,
coverage information, and convenience of seeing your PCP.
Next, we
conduct a latent class analysis to explore if there are segments of plan members
with differences in their key drivers. Table 5 shows the results of the LCA with
the five most important key drivers for each segment indicated (with a +) - see table
5.
Notice
that 14 of the 20 key drivers for the four segments are not among the top five
drivers in the aggregate model. Unique items of importance for each segment
include medical information via phone, access to specialists, availability of
preventative care, panel size, and choice of doctors none of which were
among the five most important overall.
For the
purpose of identifying areas for improvement, focusing on key drivers, which can
become key differentiators, the segment level prioritization should be
considered first.
To guide
product design efforts, researchers have used a number of traditional techniques
and methods. Researchers frequently use direct methods of questioning such as
scalar importance questions for product design research. A researcher who wants
to understand the importance of product features when designing a managed care
service plan, for example, might ask benefits managers:
Using a
scale from 1 to 9 where 1 means 'not at all important' and 9 means 'extremely
important,' how important is:
the
number of PCPs in the plan?
the
number of Specialists in the plan?
the
office visit co-pay?
the
pharmacy coverage?
the
out of network coverage?
the
monthly premium?
This
approach is able to include a large number of attributes and can be applied in a
large variety of settings. These ratings can be used (via cluster analysis) to
identify market segments. These resulting segments should relate to behavior, as
these are the characteristics on which buyers will choose between products.
However, if we were to learn that there is a segment for which the monthly
premium is the most important plan characteristic, we don't learn what price
they would find acceptable, just that price is the most important. Similarly,
another segment might be most concerned about network size, but we don't know if
they require 100 or 400 PCPs. Also, researchers using this approach often find
that everything is important. There is nothing forcing respondents to trade-off
one feature for another. When shopping, consumers rarely have the opportunity to
have more of every feature that is desirable and less of every feature that is
undesirable. Rather, they must make trade-offs between product features or
characteristics. To illustrate, the ideal product might be 100% safe and 100%
efficacious, but that ideal might not be possible. Rather, a trade-off
(realizing that this is not the way the topics would typically be presented)
might be required between the following alternatives.
Alternative
1: Safe in 99.9% and Efficacious in 70% of patients, and
Alternative
2: Safe in 99.1% and Efficacious in 85% of patients.
A
physician could be presented with these two alternatives and asked which one he
would be most likely to prescribe. This question represents a discrete choice
modeling task. Discrete choice is a consumer research methodology[3]
that is related to conjoint analysis. It allows researchers to identify
individual's preferences for unique elements of a product or service.
In discrete
choice studies, respondents are presented with several configured products
(combinations of product features) and asked which one they would purchase. This
exercise, or choice task, is repeated several times, each time carefully varying
the products shown to the respondent. The following are data from a discrete
choice study conducted among benefit managers to guide managed care product
design. The attributes in the study are shown in figure
2a and 2b.
Using a
discrete choice methodology we can learn how important specific product features
or service characteristics are to respondents. These preferences are expressed
through utilities. Table
6a and 6b
shows the derived average utilities from the respondents in the study.
Improving
from a $10 prescription co-pay to a $3.50 prescription co-pay produces a
relatively small gain in utility (24 utiles), but improving from a $20 office
co-pay to a $5 office co-pay produces nearly twice the change in utility (46
utiles). So in total, there are 70 utiles of improvement that can be made from
lowering transaction co-pays from the highest studied level to the lowest
studied level. However, that 70 utile difference is less than the 93 utile
difference between a $75 monthly premium and $200 monthly premium, suggesting
that this firm would benefit from lowering monthly premiums, even if co-pays are
increased.
For the
purposes of our discussion related to segmentation, the utilities are
interesting but the differences in utilities between respondents are even more
interesting. Historically, discrete choice studies have been conducted by
developing utilities only at the aggregate, or maybe subgroup, level. Recent
advances in statistical computing have made another set of techniques available
for discrete choice modeling. Hierarchical Bayesian (HB) methods[4]
now allow the estimation of individual level parameters from choice data.
Effectively, Bayesian methods allow respondents to share data amongst themselves
so that the researcher is able to produce utilities for each individual. The
effectiveness of HB with discrete choice studies is discussed in Pinnell (2000).
Conducting
the analysis using HB and cluster analyzing those individual level parameters,
we find four clear segments of the population, each approximately equal in size.
Each segment has different key buying factors. The utilities of the key segments
are summarized in the figure
3.
Segment 2
is most driven by transaction costs (office and prescription co-pays) but is
much less sensitive to monthly premium. Segment 4, on the other hand, is very
sensitive to monthly premium and is less concerned with transaction costs.
Segment 3 is driven by network coverage, and the Segment 1 is most concerned
with brand and access.
Recall from
our aggregate model, it would appear that lowering monthly premiums would make
the product more appealing even if co-pays are increased. It appears that this
is truly the case in Segment 4, however with each change that lowers premiums
and increases co-pays the product becomes less appealing to Segment 2. The
finding from this study is that both Segments 2 and 4 should be targeted with
specific products identical except for their fee structure provided that
offering both is economically feasible.
With this
approach, the company can also determine the costs and benefits of increasing
network coverage and further investments in the brand name and image. Only by
considering segments of buyers could these realizations be made.
Finally, we
present a technique to help target customers. These targets can be identified
from data mining conducted on an existing database, such as sales data or
transactions, or from survey data. We will consider another example using
discrete choice, this time for a medical device. The purpose is to identify
those doctors who are likely to be early adopters of this device. The results
have been disguised. While this example is presented as identifying early
adopters, the technique is broadly appropriate for targeting and classification,
such as identifying likely to churn customers, those with specific medical
outcomes, those with unusually high utilization, those likely to respond to a
specific offer, or those physicians with more satisfied patients.
Using a
family of techniques broadly referred to as classification and regression trees
(CART), or sometimes as classification trees or decision trees (which we will
use), we can identify the characteristics of early adopters in our example. The
benefit of decision trees is that they can be easily communicated in words
(rather than formulas) and graphically (through a tree structure). Decision
trees produce a series of splits (often dichotomous) of a database, with each
split identifying the single split that best discriminates the segments so that
the two resulting groups are most different on the dependent variable. Decision
Trees are discussed[5]
in Breiman and in Loh and Vanichsetakul.
Figure 4
shows the results of the decision tree conducted on the adoption likelihood of a
specific medical device - see figure
4. Here we see that the overall adoption likelihood is 38% for the total
population. We also see that by simply dividing the database into two segments,
we identify two strong segments of doctors with different adoption propensity
(as shown in the first branch of the tree). And this first cut is based on only
one variable. As we can continue to 'grow' the decision tree we find six unique
and easily identifiable segments that range in their adoption propensity from
12% to 62%, over a 5 X improvement from highest to lowest. Targeting these early
adopters will provide the largest benefit relative to cost for launching this
product.
Marketers
frequently follow a development plan, as illustrated in figure
5, that shows a progression from 'Understanding the Market and Market
Opportunities,' to 'Developing Products for the Market,' and 'Tracking the
Performance of the Products in the Market.'
For each
of the three phases in this cycle, we have identified a common type of marketing
research or analysis and shown that segmenting and targeting the market can lead
to a deeper understanding and more efficient product development and marketing
effort.
The three
techniques highlighted: latent class analysis, discrete choice with HB analysis,
and decision trees, were each illustrated with a specific application. Each
could easily be used for other applications. To summarize, latent class analysis
will simultaneously conduct a regression analysis and identify segments of
people that have different regression results; hierarchical Bayesian methods
will estimate regression results for each individual even when there are not
enough observations to do so with traditional methods; and decision trees can
identify the characteristics that best distinguish between specific groups of
individuals.
The methods
are again summarized in the table
7 and 7b.
The
discussion is motivated by an understanding that buyers differ, and that many
potentially conflicting targets are involved with healthcare product development
and marketing. However, by systematically understanding the differences between
end users, influencers, and buyers, product and service providers are able to
better design products for specific market segments and more effectively target
those segments.
This paper
presents three techniques that we have found particularly useful for specific
applications. We anticipate that the debates that rage about the 'how' to
segment will continue. We have presented three techniques for three applications
that we have found useful and have hopefully illustrated the 'why' to segment
and the resulting benefits.
References
Allenby,
Greg, Neeraj Arora and James Ginter (1995). Incorporating Prior Knowledge into
the Analysis of Conjoint Studies. JMR, 32, 152-162.
Breiman,
Leo (1984). Classification and Regression Trees.
Desarbo,Wayne
et al (1997). Representing Heterogeneity in Consumer Response Models. Marketing
Letters, 8 (July) 335-348.
Dichter,
Ernest (1958). Typology. Motivational Publications, 3, 3.
Frank,
Ron; William Massy and Yoram Wind (1972). Market Segmentation,
Gunter,
B. and A. Furnham (1992). Consumer Profiles: An Introduction to Psychographics.
Haley,
Russell (1968). Benefit Segmentation: A Decision-Oriented Research Tool. JM,
Vol. 32, 30-35.
Lazarfeld,
Paul (1935). The Art of Asking Why. National Marketing Review, 1, 26-38.
Lenk,
Peter, Wayne Desarbo, Paul Green, and Martin Young (1996). Hierarchical Bayes
Conjoint Analysis: Recovery of Partworth Heterogeneity from Reduced Experimental
Designs. Marketing Science, 15, 173-191.
Loh,
Wei-Yin and N. Vanichsetakul (2001). Tree Structured Classification and
Regression,
Meyers,
Robert et al (1997). Dynamic Influences on Individual Choice Behavior. Marketing
Letters, 8, 349-360.
Milligan,
G.W. (1980). An Examination of the Effect of Six Types of Error Perturbation on
Fifteen Clustering Algorithms. Psychometrika, 45, 325-342.
Moriarty,
Rowland and David Reibstein (1982). Benefit Segmentation: An Industrial
Application, Report No. 82-110 (November),
Myers,
James (1996). Segmentation and Positioning for Strategic Marketing Decisions.
Piirto,
Rebecca (1991). Beyond Mind Games: The Marketing Power of Psychographics.
Pinnell,
Jon (2001). All Customers Are Not Created Equal: Using Latent Class Analysis to
Identify Individual Differences. Quirks Marketing Research Review (December).
Pinnell,
Jon (2000). Customized Choice Designs: Incorporating Prior Knowledge and Utility
Balance in Discrete Choice Experiments. Presented at Sawtooth Software
Conference.
Punj,
Girish and David Stewart (1983). Cluster Analysis in Marketing Research: Review
and Suggestions for Application. JMR, 20, 134-48.
Robinson,
J. (1938). The Economics of Imperfect Competition.
Smith,
Wendell (1956). Product Differentiation and Market Segmentation as Alternative
Marketing Strategies. JM, Vol. 21, 3 8.
Twedt,
Dik
Wedel,
Michel and Wagner
[1]
Standard
cluster analysis references include: Hartigan, John (1975), Clustering
Algorithms.
[2]
The
interested reader is referred to Goodman, Leo (1978). Analyzing
Qualitative/Categorical Data.
[3]
Discrete
choice modeling is discussed in detail in Louviere, Jordan and George
Woodworth (1983), 'Design and Analysis of Simulated Consumer Choice or
Allocation Experiements: An Approach Based on Aggregate Data', Journal of
Marketing Research, Vol.20 (November), 350-67; Huber, Joel and Jon Pinnell
(1995), 'Consistent Differences between Experiemental Choice and
ratings-Based Trade-offs, Presentation at INFORMS Marketing Science
Conference, Sydney, New South Wales, Austrailia; Pinell Jon (1994), 'Multistage
Conjoint to Measure Price Sensitivity', Presented at AMA Advanced Research
Techniques Forum; and Pinnell, Jon and Sherry Englert (1997), 'The Number of
Choice Alternatives in Discrete Choice Modeling', Presented at the Sawtooth
Software Conference.
[4]
The
interested reader is referred to Carlin, Bradley P. and Thomas
A. Louis (1996), Bayes and Empirical Bayes Methods for Data Analysis,
London: Chapman & Hall; Gamerman, Dani (1997), Markov Chain Monte Carlo:
Stochastic Simulation for Bayesian Inference, London: Chapman &
Hall; Gelman, Andrew, John B. Carlin, Hal S. Stern and Donald B. Rubin
(1995), Bayesian Data Analysis, London: Chapman & Hall; Gilks, W.R., S.
Richardson and D.J. Spiegelhalter (1996), Markov Chain Monte Carlo in
Practice, London: Chapman & Hall; Johnson, Richard M. (1999), 'The Joys
and Sorrows of Implementing HB Methods for Conjoint Analysis,' Presented at
Bayesian Applications and Methods in Marketing Conference, The Ohio State
University; and Zellner, Arnold (1971), An Introduction to Bayesian
Inference in Econometrics, New York: Wiley.
[5]
The
interested reader is referred to Kass, G.V. (1980), 'An Exploratory
Technique for Investigating Large Quantities of Categorical Data,"
Applied Statistician, 29, No. 2, 119-127. Original expositions on AID
include Sonquist, J.N. and J.A. Morgan (1964), The Detection of Interaction
Effects, Monograph No. 35. Survey Research Center, Institute for Social
Research, University of Michigan; and Morgan, J.A. and J.N. Sonquist (1963),
'Problems in the Analysis of Survey Data: and a proposal,' Journal of the
American Statistical Association, 58, 415-434.
http://www.warc.com | © ESOMAR. All rights reserved |