Two-For-One Conjoint: Applying Cross Category Learning to Better Predict Customer Behavior
By Professor Eric T. Bradlow (@ebradlow) (GBK Collective and the Wharton School)
Conjoint analysis is a preferred approach used across industries to optimize decisions around product design, features and pricing to predict and optimize market success. Conjoint is popular and effective because it allows you to model the tradeoff decisions your target customers make in the real world in an easy-to-do set of hypothetical choice tasks.
To date, companies have conducted separate conjoint analyses by product category. For example, if you wish to know consumer preferences for yogurt features, you run a conjoint that uses choices between yogurts.
In my latest research paper with my Wharton Marketing Professor colleague John McCoy and Wharton Marketing Ph.D. Student Rachele Ciulli, we propose a new conjoint methodology whereby brands can apply learnings across two (or more) product categories with similar attributes (e.g. yogurt and ice cream) to successfully predict customer behavior, and/or to compare choices between a product in a focal category and a non-focal category.
Why cross-category learning is essential
Conjoint has been around for over 50 years. Literally, the first marketing paper that included conjoint analysis was published by Green and Rao, 1971. In all that time, I find it quite surprising that no research has “borrowed utility estimates (partworths)” across product categories with shared attributes and levels. That is, can practitioners leverage relevant data from past conjoints performed on related products and services to inform or predict the outcomes for an adjacent product or category - perhaps without even needing to execute a new conjoint study? Clearly this is an untapped area of opportunity for brands!
“By pairing conjoint studies for two product categories with similar attributes, we show that brands are able to predict customer preferences with high levels of accuracy.”
As a Bayesian statistician and Vice Dean of Analytics at Wharton, much of my career has focused on understanding customer behavior and improving statistical estimates through “mixing data” or what is commonly called “data fusion”. An area that has long been on my “conjoint wish list” is to do a formal study that demonstrates the impact that Bayesian cross-category conjoint learning can make in understanding customer preferences.
Together with my Wharton colleagues, we’ve done just that. By pairing conjoint studies for two product categories together with similar attributes, we show that brands are able to predict customer preferences with high levels of accuracy, while also unlocking new insights to inform decision making. Let’s take a closer look at our research.
Bayesian cross-category learning (who is learning what?)
At its core, conjoint is a decompositional approach. Products or services are made up of attributes and levels. Attributes are discrete features or benefits. For example, an attribute for a conjoint study in the automotive industry could be Color, and then the levels for that attribute could include White, Black, Yellow, Gray, Brown, etc.
In conjoint analysis, one of the key outcomes are the “partworths,” which are the relative utility values that people place on various levels of attributes. For example, how much value (or utility) does a particular customer or group of customers place on a White versus Yellow car?
In our study, we choose three product category pairs (ice cream and yogurt; hiking jackets and sleeping bags; and TVs and computer monitors) where the categories in a pair have overlapping attributes. For example, ice cream and yogurt share the attributes of relative sweetness, flavors, calories, sugar content, fat content, and package size. Hiking jackets and sleeping bags share the attributes of color, material, weight, and insulation (warmth).
TVs and computer monitors share the attributes of screen size, resolution and energy efficiency. Next, we ran conjoints for each of these pairings separately to see how the partworths changed across these related product categories.
The results? We saw very little difference in the partworths across the paired categories, which means you can use one set of results to predict the other. For example, people who tended to value sweeter ice cream also preferred sweeter yogurt. While this may sound intuitive, there has been no research - until now - that demonstrates the quantifiable relationships between shared attributes across related categories.
In fact, we saw an out-of-sample prediction accuracy of more than 90% across categories with little or no data from the focal category. So, that means a practitioner who has conducted a conjoint in a related category could leverage many or even most of the part worths from that prior study to predict the partworths for an adjacent product - without even needing to conduct another study.
Mixed-category conjoint pairing
Another novel area our study explored is what we call mixed-category conjoint pairing. We also showed (besides the two independent single-category conjoint studies described above) survey respondents two product categories together (i.e., hiking boots and sleeping bags) and had them choose which of the two products they liked more.
Not only are these results highly predictive of customer preferences – it’s literally two-for-one. With respondents giving choice preferences across two categories simultaneously, companies can predict both categories' demand with one conjoint task.
Applying cross-category learning to decision making
Now let’s bring it up to a 30,000-foot view. Imagine you're a marketing leader who is tasked with conducting conjoint studies across multiple product categories to apply to decision-making.
Everything I’ve just described with cross-category learning can be done using standard software. You can run a conjoint study for one product and literally take the computer output and apply it to a second product category.
Through data fusion, brands can use both sets of conjoint studies to infer people's preferences for both sets of products. In other words, by fusing the data together, you actually get more precise estimates of the partworths.
Another key benefit of data fusion is that brands can more accurately predict partworths and the related interaction effects with a given product category. Say for example you show you a pairing where you have several brands of ice cream and yogurts. From a mathematical perspective, this allows you to estimate the interaction terms. In other words, does the actual product category (ice cream vs. yogurt), interact with the brand name, the price, or the sweetness?
And that’s the beauty of our approach with cross-category learning. Category interaction terms, which are normally not available in a conjoint study, are now directly estimable from the software. For example, is Sweetness the same in both product categories? Is Price? To the degree that's not true, you'll see positive-interaction terms or negative-interaction terms, and this is a form of what we call in the field of marketing, context effects. In other words, the context of the category influences your choice. If it doesn't, those interaction terms will be zero.
This research is extremely practical - meaning any company can run a conjoint study with mixed product categories. And the beauty is it opens up new insights on customer preferences to a wide range of decisions.