Marketing Research Practice using
Please note the live questionnaires accessible from this page have only illustrative purpose. They have been stripped of explanation screens, help pop-up windows, conditionally formulated questions and other important aspects of field questionnaires. The purpose is to show, in a concise way, the core possibilities of the interviewing methods that rely on estimation methods based on RUT - Random Utility Theory and DCM - Discrete choice Modeling. All the questionnaires have been programed using the SSI Web software from Sawtooth Software, Inc.
| Demo questionnaires
|| Previews and links
(click to run the demo)
A product can often have optional features. Which of them are the most effective can be obtained from a CBC with attribute levels composed of selectable options. The advantage of this approach compared to the reduced MXD - Maximum Difference Scaling is the possibility to estimate interactions of the options in context to the products accepted by respondents.
A CBC task can be completed with an in-place calibration question. In contrast to a common calibration run with a fixed set of profiles independent from the conjoint exercise, the in-place calibration relies on the profiles acceptable for the respondent. The systematic bias due to the fixed calibration set is excluded.
The design allows for the standard pricing, promotion pricing and free-bees effects to be merged into a single simulation model.
To avoid the bias in estimation of part-worths due to impulsive behavior of customers in short-term promotional campaigns, an approach of using two CBC exercises, one without and the other with promotions applied, has been used.
Reduced MXD - Maximum Difference Scaling
The items tested in MaxDiff are often required to be viewed from a higher number of aspects leading to
different types of preferences. If item descriptions are complicated, as is common in medicinal,
pharmaceutical, financial, insurance, telecommunication and other products, a lot of interviewing time can
be saved, and fatigue from interviewing alleviate, by grouping several preference choices in a grid.
Consecutive questions in the same format are known to decrease reliability of responses. This can be rectified by ranking of the items selected in the MXD block. Calibration questions can be asked in course of the ranking as well.
Ranking of items from a fixed set is a common way of obtaining preferences but quantification of the
results is not as free of problems as one would wish. Bayesian estimation with restrictions derived from
the properties of permutohedron
allows to obtain estimates of multinomial utilities with low bias and reasonable scaling. Ranking of the
least preferred items can often be omitted if there is little interest in them as the latest choices
provide only a negligible amount of information.
Thanks to its simplicity the method can be used with items in card format in face-to face (P&P, CAPI) interviewing. The data can be seamlessly merged with data from other discrete choice based method(s). The following usage is common:
Numbers of items, options or products available on the current markets are usually very high. Individual customers are often inclined to prefer only a fraction of them. It is well known that a simple evaluation of items each being presented in an isolation does not give sufficiently reliable estimates. If just a screening of the market behavior is required, usage of the complete armory of a full-fledged DCM method for a single (cf. reduced MXD - Maximum Difference Scaling) or two-attribute (brand-price CBC) study is often superfluous and, unfortunately, a bit clumsy when the number of items is large.
With the mentioned caveats on mind the two approaches have been combined. The resulting method is much simpler and quicker than a conjoint approach, and gives estimates far better than the traditional one-by-one evaluation. The evaluation scale drift often observed during an interview is avoided.
The method was originally developed as a replacement of a self-standing evaluation of conjoint
attributes with many levels (a.k.a. self-explicated conjoint), and a simplification of calibration of a
relatively large number of profiles (cf. SCE - Sequential Choice
Exercise). It is feasible with a general survey interviewing software.
Data from a brand-price CBC with a large number of brands often bear a frustrating level of noise. This
is mostly due to the requirement to chose from among many items a respondent is only little, if at all,
interested in. A solution is to reduce the set so that it contains brands considered by the respondent as
possible choices, and of course, the brands of interest in the test.
As the amount of information obtained from a reduced set is lower, it has to be compensated with a larger
sample. This is especially true when a small number of brands dominate the market, and the study is aimed
at some new or less frequently bought ones.
The starting selection block of an interview is carried out as a SCE
- Sequential Choice Exercise. If the SCE choices are completed with questions about purchase intention a
market potential estimation is possible.
The producer of First snacks was considering an introduction of smaller packages containing 1 or 3
packets (servings) for a presumably significant segment of light users. The problem was to decide if this
would increase reach and/or share with or without replacement of some of the current 4 piece packages. The
major competing brands Second and Third were available only in packages of 4.
The 3 package sizes and 3 flavors for the First brand products gave 9 products compared to 3 products of each other brand. This would lead to a gross selection bias in the standard CBC design. A compromise solution was found in 5 item choice sets composed of 2 First products and at most 2 products of any of the competing brands in a choice set using a class-based design. This allowed to obtain data for estimation of availability and cannibalization effects among all the tested products. Another class-based design fully balancing brands in 9-item choice sets shown in the second part of the questionnaire is an example of an improper design due to different availability, substitution and cannibalization effects among the brands.
Most conjoint exercises deal with products whose properties are composed of attribute levels directly shown to respondents. The goal of a parametric study is to determine effects of some parameters on (acceptability of) the tested objectives. Parameter values may or may not be shown to respondents.This is typical for financial products such as accounts, transactions, loans, mortgages, installment sales, etc. The parameters can be interest rate, up-front payment percentage, length of a contracted period, values in various fee formulas, etc. Parameters may be, and often are, mutually interdependent and/or constrained. The shown values are computed using the tested parameters.
In this example, installment sales parameters, namely percentage of an upfront payment and of price
surcharge, are tested on an evoked set of cellular phone handsets. The set is determined as the three top
choices from a pool.
Similar to the previous example, profiles in a choice set are generated using varying term parameters of
the loan provision. The loan amount and maturity period are stated directly by a respondent. The ranges of
tested parameters (namely monthly fees and interest rates) are conditional to the stated amount and
maturity of the loan and to the presence of the free-of-charge early repayment.
The CBC task is followed by a block of calibration questions realized as a simple SCE - Sequential
Common Scale Discrete Choice Analysis (CSDCA)
The method has been developed to achieve a better discrimination power between objects based on their
aspects than is available from scale-based questions. Aspects for an object are ranked rather than
evaluated. Ranking is used also for objects in respect to their aspects.
The OBIMA design is based on the idea published by McCullough (2013) by replacing the proposed MaxDiff and Q-Sort methods with SCE exercises that are faster and easier to manage. Obtaining perception thresholds (anchoring) is inherent to the method without using any additional questions. An early experience shows an interviewing time is about 50% longer than using evaluation (scale based) questions but the results seem to better discriminate between objects.
Results from this particular survey are on DCM Blog Cz.
Ranked Grid Analysis (RGA)
method has been developed as a substantially faster and simpler variant of the above mentioned OBIMA
method. Ranking is carried out in a checkbox grid either in horizontal or vertical dimension. Minimal and
maximal number of ranks can be set. The rank data are analyzed by HB-MCMC
Concept or Package Test
or Package Test is a MaxDiff-based method for a special case when concepts are fixed and described as
a set of aspects. The objective is to estimate preferences between the product concepts and their aspects.
More detailed information on uses of DCM can be found on pages
- Discrete Choice Modeling in Marketing Research Practice