(The capabilities described in this section are only available if you own the CBC Advanced Design Module.)
In spite of their advantages, CBC questions can overwhelm some respondents, particularly if there are many attributes. The base CBC system can measure up to ten attributes (shown in full profile), but more than about eight attributes may overwhelm respondents in some situations.
Some researchers have proposed "partial-profile" choice experiments as a way to estimate preferences for a large set of attributes. With partial-profile designs, each choice task includes a subset of the attributes (typically around five). Across all tasks and respondents, a much larger list of attributes is evaluated. The CBC Advanced Design Module permits up to 100 attributes.
Partial-profile choice designs have now been regularly used in CBC studies for nearly two decades. A growing body of research suggests that they can be valuable for some situations. We haven't yet formed a definite opinion about their performance versus other methods like Adaptive CBC (ACBC) and ACA for dealing with large numbers of attributes. We hope that including partial-profile in the Advanced Design Module encourages further research and experimentation.
With partial-profile designs, we assume respondents can evaluate the product concepts holding all attributes not represented constant. If respondents cannot maintain this ceteris paribus mind set, the resulting data may be incorrect. Therefore, when asking the choice question, we suggest including language such as, "Please assume that all features not shown are alike in all other ways," or "Please assume that these toasters differ only on the features shown." This may help respondents answer appropriately, but it is still no guarantee.
The use of the "None" concept in partial-profile CBC studies is problematic. The None weight varies significantly depending on how many attributes are displayed in the partial-profile task.
Analysis methods for partial-profile include logit, Latent Class, and HB. The success of the methods (particularly for Latent Class and HB) hinge upon how much information is available from each respondent on each attribute. We caution that individual-level estimation may not be stable for partial-profile designs if the information available from each respondent relative to the number of parameters to be estimated is low. Given enough information per respondent relative to the number of parameters to be estimated, Hierarchical Bayes estimation may provide useful individual-level utilities for use in market simulations where the results are summarized across respondents. Still, choice data are not as rich in terms of statistical information content as ratings-based data. The individual-level estimates for large partial-profile designs may contain a significant amount of noise and counter-intuitive relationships (reversals). If the goal of the research is to estimate stable utilities for individual-level profiling and predictions, the partial-profile approach alone may be insufficient. For sparse partial-profile designs, it may be helpful in CBC/HB software to use a lower prior variance assumption (such as 0.5) and a higher degrees of freedom for prior covariance matrix (such as a value equal to 1/4 or 1/2 of your sample size) to avoid potential overfitting.
Many split-sample partial profile / full profile methodological studies have shown that the parameters derived from either experiment are quite similar, including Price. Some recent methodological experiments presented at the Sawtooth Software conference suggested that the parameters were quite similar, but that the importance (slope) of Price was understated using partial-profile. Therefore, we suggest users should be aware of this potential outcome and proceed with caution when using large partial-profile designs in CBC for pricing research.
Specifying Partial-Profile Designs
Choose Compose | Write Questionnaire... and edit a CBC question within the CBC exercise to bring up the Conjoint Settings | CBC Settings... to bring up the CBC Exercise Settings dialog. From that dialog, click the Design Tab and then click the Partial-Profile CBC Design option (revealed by clicking the Show Advanced Settings button).
There are a number of control parameters governing partial-profile designs. For purposes of illustration, let's assume there are 12 total attributes in the study. The researcher wants to display 5 attributes per task with attributes 1 and 12 to appear in every choice task.
Number of Attributes to Show: In this example, 5 attributes are displayed in each choice task, so we specify a 5. Recent research suggests that between 2 to 4 attributes may be optimal to use in partial-profile studies. The length of the level text has a bearing on the appropriate number, along with the familiarity and interest respondents have for the product category.
Rotate Attributes into Concepts Starting with: In this example, attribute 1 appears in every choice task. Therefore, attributes rotate in and out of the choice tasks starting with attribute 2. (If you want all attributes to rotate into the tasks, always specify 1). Note that, all else equal, attributes displayed in every task are measured with greater precision than those rotated into the tasks.
and Ending with Attribute: In this example, attribute 12 appears in every choice task. Therefore, attributes rotate into choice tasks ending with attribute 11. (If you want all attributes to rotate into the tasks, specify the last attribute number.)
If the Attribute Randomization drop-down is enabled, the attributes appear in random order (held constant within respondent) within the concept. If No Randomization is selected (default), the attributes appear in their natural order.
Design Strategies for Partial-Profile Designs
With partial-profile designs, the design selection has two stages. The first stage involves choosing the subset of attributes displayed in the choice task. To formulate the first task, a subset of attributes is randomly chosen. For all subsequent tasks, the two-way joint frequency table of attribute presentation within choice tasks is examined. Attributes are chosen so that the off-diagonal cells in this table remain approximately balanced. This strategy is identical for the two methods.
The second stage of the design selection involves deciding which levels of the selected attributes are displayed for each concept.
For partial-profile designs, we generally suggest either the Balanced Overlap or the Complete Enumeration strategy (unless you have so many attributes that the design generation process is prohibitively slow). If the time to generate the design is too long, then you might consider 1) generating fewer total designs (e.g. 20 instead of 300) or 2) the Shortcut design strategy. In either case, use the Test Design design procedure (estimated standard errors via logit and relative D-efficiency test) to determine the effect of one strategy versus the other.
The purely Random design method should only be used if the measurement of interactions (through aggregate analysis) is the primary goal, the attribute list is relatively small and the sample size is quite large. However, partial-profile designs make it much more difficult to measure interactions between attributes, when these attributes may only occur together within a product concept a minority of the time.
Notes for Partial-Profile Designs
Estimation of main effects and especially interaction terms for attributes rotated into choice tasks is significantly less precise than with manageable full-profile CBC designs.
We strongly suggest you test your design prior to fielding to ensure you can estimate reasonably stable part-worths. The Test Design module provided with CBC may be used. The efficiency of part-worths for attributes rotated into the design is naturally lower than with a full-profile design (where the attribute is always present), since comparatively less information about that attribute is available.
We suggest using the Test Design procedure (simulated respondent data, logit estimation report, and relative D-efficiency) to assess the quality of different partial-profile designs. (But, remember that the most D-efficient design isn't necessarily the design that will lead to the most precise utility estimates. As more attributes are shown to respondents, the D-efficiency increases, but the difficulty of the task and the resulting respondent error increases.)
Fixed partial-profile tasks (or entire designs) may be specified.
Partial-profile designs are more robust in the face of level prohibitions than full profile designs. For example, consider a study with 10 attributes, where 4 attributes are shown at a time in product concepts. Further consider that there are prohibitions between the levels of attributes 1 and 2. When attributes 1 and 2 appear together within the same choice task, the prohibitions introduce correlations (and resulting lower precision) within the design matrix. However, for the many tasks in which attributes 1 and 2 appear separate from one another, the prohibitions have no negative effect on the design efficiency. These tasks in which attributes 1 and 2 appear separately provide uncompromised information (assuming no bias due to partial-profile displays) for stabilizing the parameters for attribute 1 and 2.