Often our choice modeling (including conjoint analysis) studies support the development of new products, new services or new product features.
By using market simulators, clients can learn what percentage of respondents want the new product (service, feature), how price sensitive customers might be and from which competitors they might steal share.
In addition to these typical kinds of things, some clients also what to know how to target the customers most likely to be interested in the new product.
Profiling Respondents by Preferred Product/Service
One way to characterize customers likely to buy the new product would be to profile them in terms of their demographics, attitudes or behaviors as collected in the survey (creating Target Customer Profiles).
The conjoint simulator gives us the probability of each respondent choosing any given product in any given competitive marketplace. For any competitive scenario, therefore, we can identify the likely buyers of the new product.
We can then profile those likely buyers: what percent are female, how much do they agree with each of 10 items in an attitudinal battery, how old are they, how often do they purchase products in the category, etc. Here’s an excerpt from such a target profile (with disguised variable names):
For example, the Target Customer Profile for the product/service simulated in the above example is 24% more likely to be female than the total sample (index of 1.24) and is 31% more likely to use the product category once or more per day than the total sample (index of 1.31).
We can even build this profiling capability into an Excel simulator so that the profiles change dynamically depending on the specific competitive scenario.
For example, who are my new product’s customers when we price it at $43; who are my customers when I price it at $37 but a competitor has launched a similar product at a lower price? And so on.
Sometimes, however, the characteristics of customers interact. For example, my product may have a slightly higher appeal to males than to females, and a slightly higher appeal to respondents under the age of 45, but if the interaction of age and gender matter, I might find that certain combinations of age and gender have much greater affinity for the new product.
Decision Trees to Find Interactions among Respondent Characteristics
When the characteristics of customers interact with respect to preference for specific products/services, we sometimes get better results using a machine learning classification tool called a decision tree.
Again, we capitalize on the fact that the conjoint simulator gives us the probability of each respondent choosing a given product in a particular competitive marketplace.
Selecting a competitive scenario, we can identify the respondents most likely to prefer our new product. For this analysis we use whichever of a couple of different tree models (a classification and regression tree or a conditional inference tree) fit the data better.
Below is a tree from a recent study of a service available for a consumer durable, with labels changed to disguise the product and the demographic, attitudinal and behavioral predictors of preference. The dependent variable was the preference share for a specific version of the client’s brand that included a new feature.
From the total sample of 1,580 respondents, the tree searches for the single demographic, attitudinal or behavioral variable that best splits the sample into groups with higher and lower share.
In this disguised example it was whether the respondent was over age 64 or not (among respondents aged 65+, the new service has a predicted share of 16.3%, whereas the service has a predicted share of 61.1% among respondents under 65.
For each of the resulting subgroups, the analysis next looks for variables it can use to further split into higher-and lower share groups and at each step it selects the most discriminating split. This progression keeps running until it can make no more splits and the tree stops growing. In this case 4 sample splits produce 5 segments with shares ranging from under 11% to over 67%
The analysis thus produces targetable segments of customers, based on their propensity to buy the client’s new product.
As you’d imagine, this kind of repeated splitting of the sample works better as sample size increases: I recommend having at least 1,000 respondents to run this kind of analysis, though I have seen the method work with smaller samples when we have strong differentiation in preference shares based on the demographic/attitudinal/behavioral variables.
Though the examples above rely on simulators from Choice-Based Conjoint (CBC) studies, we can (and have) run the same kind of analyses for shares derived from MaxDiff, and Menu-Based Choice (MBC) studies.
Sawtooth Software’s market simulators currently don’t do the advanced kinds of displays and analysis I’ve described here. Without too much effort, the first type of analysis (reporting relative targeting indices for product alternatives) can be added to an Excel-based market simulator.
Decision tree analysis could be implemented using any number of commercial or open-source tools in R. Our analytics practice, Sawtooth Analytics, is available to assist you should you need help implementing either of these solutions.