What we have learned from 20 years of Conjoint Analysis

Conjoint analysis as a commercial method has been available for more than 20 years (Green, Wind and Jain 1972). Ten years ago, 1987, at a Sawtooth conference, I raised the question of how conjoint could work so well, given obvious differences between the task and market choice. I proposed that conjoint works primarily because the simplification in a conjoint task mirrors the simplification in the marketplace. That is, the complexity of the marketplace encourages people to make choices based on relatively few attributes, effectively selecting the attributes that they will attend to and value in a given choice. In the same way, the complexity of a full-profile conjoint task also encourages respondents to attend to a subset of the attributes. Thus conjoint works by simulating the attribute selections process that occurs in actual choices. Much has changed in the last decade. The biggest change is the expanded arsenal of methods that enable us to measure people's values (Green and Srinivasan 1990). There are sophisticated self-explicated methods that break down choice into values for levels of attributes, weights for different attributes, and perceptions of alternatives on those attributes. ACA in its new version permits different kinds of self-explicated assessments and allows various scales to be used. Full-profile can be built with more general choice designs (Kuhfeld, Tobias and Garratt 1994). Perhaps the biggest change has been the availability of easy-to-use choice-based conjoint systems, which measure tradeoffs with responses to hypothetical choice sets. The question for the next millennium is not whether conjoint works, but defining the contexts in which different methods are appropriate. One fact is irrefutable: different methods, and even the same method in different contexts, give different results. Consider the following three examples: 1. On high technology products such as computers, ACA's price partworth for knowledgeable respondents needs to be doubled to accurately predict subsequent choice-based conjoint (Pinnell 1994). That is, choices among brands are best predicted if the raw ACA utilities are doubly weighted before being entered into a choice simulator. 2. The relative value of price over brand doubles from the early choice tasks compared to later ones (Johnson and Orme 1996). 3. In contrast to ACA and full-profile conjoint, choices reflect 20%-30% greater emphasis on the most important attributes and put 30%-40% more weight on the least preferred levels of each attribute. 2 These results have been profoundly disquieting to me, and should be to you. …