Overview of the Choice Modeling Platform
Choice modeling, pioneered by McFadden (1974), is a powerful analytic method used to estimate the probability of individuals making a particular choice from presented alternatives. Choice modeling is also called conjoint choice modeling, discrete choice analysis, and conditional logistic regression.
A choice experiment studies customer preferences for a set of product or process (in the case of a service) attributes. Respondents are presented sets of product attributes, called profiles. Each respondent is shown a small set of profiles, called a choice set, and asked to select the preference that he or she most prefers. Each respondent is usually presented with several choice sets. Use the Choice platform to analyze the results of a choice experiment.
Note: You can design your choice experiment using the Choice Design platform. See Choice Designs in the Design of Experiments Guide.
Because customers vary in how they value attributes, many market researchers view market segmentation as an important step in analyzing choice experiments. Otherwise, you risk designing a product or process that pleases the “average” customer, who does not actually exist, and ignoring the preferences of market segments that do exist.
For background on choice modeling, see Louviere et al. (2015), Train (2009), and Rossi et al. (2005).
The Choice Platform
The Choice Modeling platform uses a form of conditional logistic regression to estimate the probability that a configuration is preferred. Unlike simple logistic regression, choice modeling uses a linear model to model choices based on response attributes and not solely upon subject characteristics. In choice modeling, a respondent might choose between two cars that are described by a combination of ten attributes, such as price, passenger load, number of cup holders, color, GPS device, gas mileage, anti-theft system, removable-seats, number of safety features, and insurance cost.
The Choice platform allows respondents to not make a choice from among a set of profiles. The no choice option is treated as a product with a single attribute (“Select none of these”) that respondents are allowed to select. The parameter estimate for the No Choice attribute can then be interpreted in many ways, depending on the assumptions of the model. The Choice platform also enables you to obtain subject-level information, which can be useful in segmenting preference patterns.
You can obtain bias-corrected maximum likelihood estimators as described by Firth (1993). This method has been shown to produce better estimates and tests than MLEs without bias correction. In addition, bias-corrected MLEs improve separation problems that tend to occur in logistic-type models. See Heinze and Schemper (2002) for a discussion of the separation problem in logistic regression.
Note: The Choice platform is not appropriate to use for fitting models that involve ranking, scoring, or nested hierarchical choices. You can use PROC MDC in SAS/ETS for these analyses.
Choice Designs in Developing Products and Services
Although customer satisfaction surveys can disclose what is wrong with a product or service, they fail to identify consumer preferences with regard to specific product attributes. When engineers design a product, they routinely make hundreds or thousands of small design decisions. If customer testing is feasible and research participants (subjects) are available, you can use choice experiments to guide some design decisions.
Decreases in survey deployment, modeling, and prototyping costs facilitate the customer evaluation of many attributes and alternatives as a product is designed. Choice modeling can be used in Six Sigma programs to improve consumer products, or, more generally, to make the products that people want. Choice experiments obtain data on customer preferences, and choice modeling analysis reveals such preferences.
Segmentation
Market researchers sometimes want to analyze the preference structure for each subject separately in order to see whether there are groups of subjects that behave differently. However, there are usually not enough data to do this with ordinary estimates. If there are sufficient data, you can specify the subject identifier as a “By groups” in the Response Data or you could introduce a subject identifier as a subject-side model term. This approach, however, is costly if the number of subjects is large.
If there are not sufficient data to specify “By groups,” you can segment in JMP by clustering subjects using the Save Gradients by Subject option. The option creates a new data table containing the average Hessian-scaled gradient on each parameter for each subject. For an example, see Example of Segmentation. For more information about the gradient values, see Gradients.
In JMP Pro, you can request that the Choice platform use a Hierarchical Bayes approach in order to facilitate market segmentation. Bayesian modeling provides subject-specific estimates of model parameters (also called part-worths). These parameters i can be analyzed with hierarchical clustering or some other type of cluster analysis to reveal market segments.