> CBC - Choice Based Conjoint 
CBC - Choice Based Conjoint
> CBC - Choice Based Conjoint
Metric-based conjoint suffers from inherent drawbacks:
CBC - Choice-Based Conjoint, as a straightforward application of DCM - Discrete Choice Modeling, avoids the problems pertinent to metric methods. Currently, CBC is the most often used method of conjoint study.
Conjoint of any type is sensitive to the design efficiency. In general, metric-based conjoint is more sensitive to (deficient) orthogonality among the profiles than a DCM based conjoint. As CBC is relatively tolerant in this respect, other aspects should be accented. Referring to Joel Huber (1982), the issues to be considered in a CBC design can be summarized as follows.
In practice, the actual experiment has to be modified so that no effect causing a bias would prevail. A pragmatic experimental design balance must be achieved.
Users differ in their preferences and, therefore, in their consideration sets. To ensure that every respondent is presented the stimuli in their consideration scope, CBS - Choice Based Sampling can be used.
The standard orthogonal design of a conjoint exercise assumes all combinations of attribute levels are allowed. This requirement cannot be met in the design of realistic profiles. In order to remove or at least substantially suppress the detrimental influence of level combination prohibitions, and to have full control over them, a method of hierarchical (nested) design based on product classes of profiles can be used. It is used typically in assigning prices to products with various combinations of benefit attribute levels.
The method of incomplete profiles is inherent to ACA - Adaptive Conjoint Analysis. Incomplete profiles usually lead to equalization of importance of all attributes except for two or three dominant ones. The importance of an attribute is given by respondent's ability to discriminate between attributes in an incomplete profile. Due to the decreased number of attributes in the profiles, the same importance is given to a higher number of attributes. The attribute importance gets close to the result of determining the importance of each attribute separately.
In many practical cases some attributes cannot be omitted. If some attributes are always shown, their importance will decrease while the omitted attributes will get higher importance. An attribute that is more rarely displayed will attract more attention when it is displayed and its seeming importance will increase.
Incomplete profiles in a branded study should be avoided completely. Respondents would implicitly substitute the missing attributes with the levels common to the brand. This will sway not only the level part-worths of the omitted attributes but also of the other attributes.
The standard orthogonal design of a conjoint exercise assumes that all combinations of attribute levels are allowed. This requirement cannot be met in the design of realistic profiles. In order to remove or at least substantially suppress the detrimental influence of level combination prohibitions, and to have full control over them, a method of hierarchical (nested) design based on product classes of profiles can be used. It is mostly used in assigning prices to products with various combinations of benefit attribute levels such as brand-inherent levels.
An alternative approach is to prepare the most likely concepts to be used in the market. Preferences of these concepts can be evaluated in a MaxDiff block or a block analogous to product-price conjoint. Preference of individual levels of attributes can be assessed in separate simple tasks such as SCE, Gabor-Granger combined with direct evaluations on a proper Likert scale.
Problems with nominal attributes differing in the number of levels are common. The number of times each level of an attribute will be shown is inversely proportional to the attribute cardinality (the number of levels). Therefore, level part-worths of higher cardinality attributes will be estimated with higher error. A great help in this respect is the use of soft constraints between levels of a particular attribute (or a combination of attributes) constructed as tasks involving only levels of the attribute (or the combined attribute). MaxDiff, SCE, or simply an in-place sorting can be used. The use of soft constraints will also allow to dramatically decrease the number of CBC tasks.
Different cardinality of ordinal attributes is much less problematic thanks to possibility to constrain the part-worths.
Most CBC exercises deal with products whose properties are composed of attribute levels directly shown to respondents. However, visible properties of some other products can be computed from some values revealed by the customer. This is typical for banking products (accounts, transactions, loans, mortgages), installments sales methods with variable upfront payments, deduction and discount sales methods dependent on a purchased quantity or a total value of the purchase, etc. The parameters can be an amount of goods or services, interest rate, up-front payment percentage, length of a contracted period, values in various fee or discount formulas, etc. The parameters can be, and often are, mutually interdependent and/or constrained or bound. The common trait in these problems is the aim to determine influences of the parameters rather than the actually shown attribute values. The parameter values may or may not be shown to respondents.
The input values the shown values are derived from can be entered directly or indirectly by a respondent. An example of direct input is the desired amount and maturity period of a bank loan, as in the bank loan demo questionnaire. An indirect input can be based on an evoked set of products that enter the actual CBC, as in an installments sales demo questionnaire.
Alternatives that do not have common attributes are called specific. Their use has been first introduced in transportation problems where fuel prices, parking fees, transport fares, bike paths, etc. have profound influence on choice of a transportation means such as car, public transport (with sub-alternatives, such as train or bus), motorbike, bike or walk. Evidently, each of the alternatives has different attributes. In mass product marketing, similar problems are encountered as well. They are most often related to the quality of the products and the extent of services related to them. Again, the product classes approach proved to be adequate.
There are two basic types of choice set arrangement.
It is generally recommended to use one profile in a choice set as a constant alternative and estimate its utility. While the wording "None of these" is used most often, a constant alternative can have any other meaning such as a status quo ("I would stay with my current product"), a switch to another product from the category, abandoning usage of products from the category, etc.
| A constant alternative is regarded as a way to reveal segments that |
|
The most common "None of these" alternative should serve as a choice option when none of the offered products meets the respondent's needs or expectations. It might mean either a choice of another product based on a new belief acquired during the CBC exercise (acquisition), a choice of another product based on previous experience (switchback), a continuation of the current usage (retention, loyalty), or their mixture. However, we have often observed choice of the none alternative when a choice of a regular profile could be expected.
In our view, the none alternative often serves as a pure escape from a rational answer. Analyses of many studies show that the constant alternative has usually greater variability than levels of any of the attributes. The frequency of "None" choices is increasing with the number of choice tasks made. We have never found a significant correlation between the none utility and any other measure of willingness to buy or purchase intention stated in the interview. In non-compensatory modeling, the none choice appears even if a profile with all attribute levels above the acceptability thresholds is present. The reason may be the fatigue from a long interview, little interest in the object of the study, indecisiveness in selecting from profiles being hard to evaluate or distinguish between, or simply annoyance from a repetitive questioning. We usually leave out the none utility from further considerations.
If no item in the choice set is acceptable, and a "None" alternative or any other way out is unavailable, the respondent is forced to pick something up. The profile might be chosen for several reasons.
As all the above types of choices violate the preference axioms, the utilities estimated from them are biased.
Some authors recommend use of the dual-response none design. With no constant alternative present, some item must be chosen and only then assigned as acceptable or not. Unfortunately, the problem of forced choice remains.
There can be more constant alternatives in a choice set offered simultaneously. E.g. with hired services and the main question "If these were the only services available from your operator, which one would you chose?" there might be constant alternatives "I would switch to another operator" and "I would stop using the service completely". The alternatives might be placed right in the choice task, or when the main "None" choice has been made, or outside the CBC in a separate calibration procedure. Each of the levels can be characterized by their parameters.
The constant alternative is usually omitted from choice-based studies such as MXD - Maximum Difference Scaling, SCE - Sequential Choice Exercise or CEA - Cross Effect Analysis where it might cause excessive decrease of the collected information, or go against the principle of the method, or its goal. The same may be true for electable benefits (freebies, options), necessity products (power, gasoline, staple foods), etc.
Methodology selection is strongly dependent on the category of tested products. The tested category should be as
narrow as possible.
The most common is a price study. If the tested products are taken as completely independent, each product can be considered a class with prices on several levels. These prices are then mapped to a price attribute. If the price attribute is composed of discrete values, the part-worths can be estimated as part-wise linear. If the attribute is considered continuous, some linearization function must be used.
The product-price design can be extended by effects of packaging, discounts, gifts, promotions, etc., as in the live questionnaire example of coffees.
Many consumables are marketed as lines. As each brand uses a certain pricing strategy, the prices of the items in
a line are usually mutually dependent. We have developed a method of a two-stage brand-price CBC where prices of
the products in a line are changed in parallel. The choice sets are composed of the complete lines of several
brands. The respondent first selects the brand, and then the most attractive brand variant. As yet (September
2022) the method has not been applied in practice.
In contrast to consumables, a choice of a durable product requires high involvement of the decision maker. The
CBC excise may contain only a very limited number of choice tasks with a small number of product profiles. The
profiles are often composed of more than 6 attributes, which is considered the top estimable limit for a
self-contained CBC. To achieve satisfactory results, additional query blocks are required. This was the main
motivation for the introduction of CSDCA - Common Scale Discrete Choice Analysis and CBCT - Choice-based Concept
Test. As a rule of thumb, CSDCA is suitable either as an introductory screening or if preferences between levels
of different attributes are not obtainable. CBCT was developed for the case when the final idea of what the
profiles should look like exists and also pricing is known, at least in rough outline.
The problem is similar to durables. A difference can be seen in the perception of the pricing of the individual components that make up the product. While the sensitivity to the price of individual components of durables is clearly detectable and estimable with reasonable values, the total price has the dominant role in services. The presence of an unimportant paid partial service is well accepted if the total price is acceptable and the composite service comprises all required partial services. This phenomenon can be captured in the measure we call perceptance. The estimated sensitivity to price of a partial service is often very vague.
Profiles of bundles and packages (such as travel, vacancy, etc.) are typical for many levels stating "not present". This leads to the numerical problem of s.c. sparse matrix. CSDCA - Common Scale Discrete Choice Analysis with the block of motivators estimation is useful in tackling the problem. Packages are often offered with optional features. CBCT - Choice-based Concept Test with selectable options inside the profiles can be used.
Choice tasks are an inefficient way of obtaining quantitative information on preferences. Respondents evaluate multiple concepts in a choice task, but only tell us which one of them they prefer most. It is not known how strong that preference is relative to the other product concepts.
Showing more product concepts in a choice-set increases the information content obtained from the task. As humans are quite efficient at processing information about products or their concepts, three to six relatively complicated concepts per task can be shown around. However, many more simple concepts can be shown. This is typical for CPG (Consumer Packed Goods) brand-price studies with concepts arranged as if on a shelf.
Motivations leading to a choice are a mixture of habits, experience, expectations, and randomness. It is believed that the choices in CBC can represent the behavioral patterns of purchasers. E.g., in a nationwide study of bottled beers, the respondents who stated either (1) frequent visits to two different outlets for a similar purchase purpose or (2) two dissimilar purchase purposes in the same or a different outlet, were asked to do an additional CBC task. The selection of products and the prices were set according to the outlet type. While the filters were an application of CBS - Choice Based Sampling rather than a behavioral segmentation, we computed 4 and 8-segment solutions. Percentages of the respondents who were assigned to the identical behavioral segments (table rows) from the two CBC exercises (table columns) are in the table below.
| Segmentation solution (a) | (1) An identical purpose of purchase stated; Different point of sale |
(2) A different purpose of purchase stated; Point of sale not distinguished |
|---|---|---|
| 4-Segment | 67.0 % | 79.1 % |
| 8-Segment | 66.0 % | 68.0 % |
| (a) | Latent class analysis program (Sawtooth Software, Inc.) was used for LCA - Latent Class Analysis |
The behavioral patterns of respondents are strongly persistent even when respondents state a different purpose for their purchases. The belief in CBC as a robust experimental instrument for behavioral studies is clearly supported.
CBC is the workhorse in most of our studies based on DCM - Discrete Choice Modeling. Examples are scattered over several pages of this web.
The live demonstration questionnaires in the English language are accessible on page DCM Demo Web Questionnaires. Some Czech examples are available on the DCM blog CZ.