Product Configuration: Evidence for Effectiveness
White Paper about TeXo - Part IIBy Rajan Sambandam & Pankaj Kumar
In the companion piece [Configuration: An Approach for the Times] the basics of configuration were explained. It is an effective approach that mimics the real world of customer driven product design to obtain simple yet deep understanding into consumer decision making, and its implications for the practicalities of new product design. In this piece we will look at an example from one study, the kinds of information that can be derived and the possibilities provided by advanced statistical analysis. The latter are particularly interesting, as such capabilities (utilities, simulation) have till now been the province of methods such as conjoint analysis.
The example in question is in the auto insurance industry. The topic is interesting for a few different reasons. All drivers need it and most adults choose their own provider. It can be customized for a driver and it has some complexity built into the process, especially with regard to differential pricing. The decision-making may not be straightforward with rules being used to arrive at an optimal product. It is often renewed every six months, providing an opportunity to re-visit the decision-making process with somewhat high frequency. And, of course, it is quite amenable for a configuration exercise.
The study was set-up as a task for choosing auto insurance for oneself. A basic product (largely hewing to state mandated minimums) was described followed by the configuration exercise where respondents were offered choices on six features. Each feature had three to four options including a base option and respondents could choose to stay with the base product or shift to one of the other offered options. Some options would increase the total price, while others would lead to a lower price. Given the customized pricing used in auto insurance, we kept the task realistic by asking respondents for their current expenditure and using that as a basis for building the price for the overall product. Respondents build their ideal product from the choices provided as they proceed through the exercise.
Of the 822 respondents in the study only 20% chose the base option in every feature. (Figure 1)
Almost half of respondents opt for some form of Accident Forgiveness option while about that proportion indicate they would prefer policy terms longer than 6 months. In both cases respondents are showing that they are willing to pay for such amenities, thus providing an auto insurance company with valuable input on pricing these kinds of innovative features. Profiling people by the choices they make also provides interesting information. This is clarified more when we run a segmentation analysis on the choices that people make when building the product. Using a Neural Network based segmentation method (called Self-Organizing Maps) we can identify segments with clearly distinguishing characteristics.
The primary information that comes from a configuration exercise is simple, intuitive and very useful. But we don’t have to stop there. Advanced econometric modeling can be applied to the data to draw out conjoint-like insights even though the design is not set up accordingly. While the problem is quite complex because of the design flexibility, it is possible to derive individual-level utilities or attractiveness scores for every option in every feature. Of course, this provides us the same level of flexibility on the back end that has been the hallmark of conjoint designs. In essence, we overcome the front-end design constraints of conjoint while availing ourselves of its back-end flexibility. In technical terms this is called having your cake and eating it too.
That is all great but where is the proof that the utilities calculated through this method are accurate? We use validation to show that this is the real deal. After the configuration exercise was completed we asked respondents to indicate their willingness to buy a few pre-specified products. If the individual utilities we calculated are accurate they should identify what is important for individual respondents. Using that information we should be able to predict the willingness to buy for each respondent and compare it with what they actually said in the survey. Doing this calculation for the example data, we are able to correctly predict the buy/no buy status of 81% of the respondents. Clearly, if the utilities are not properly calculated this kind of result would not be possible.
So what does it practically mean to have individual level utilities?
To read the rest of this article and have access to the entire TRC white paper library, please click here.