The construction of preferences for crux and sentinel product attributes

Abstract

Designers often attempt to find preferences that users have for products and their attributes. Applying theory from behavioural psychology, we demonstrate that product preferences are not ‘found’ in people, but rather constructed by people on an as-needed basis. The demonstration explores the relationship between crux product attributes, which are both important and difficult for people to assess, and sentinel attributes, which are easy to assess and have a perceived association with a crux attribute. A relationship between crux and sentinel attributes is proposed, supported by the results of a case study involving design of paper towels, where a discrete choice survey is analysed using a new technique called the full factorial marketplace. We generalise our approach to a constructed preferences design method that can be used to identify crux/sentinel relationships between product attributes.

Keywords:

• product design
• construction of preferences
• context effects
• discrete choice
• conjoint analysis
• stated-choice
• preference model
• inferences

Acknowledgements

This research was supported by a National Science Foundation Graduate Research Fellowship, the Donald C. Graham Chair Endowment, and the Rackham Graduate School Antilium Project at the University of Michigan. This support is gratefully acknowledged. An earlier version of this work was presented as a poster at the 2007 International Conference on Engineering Design in Paris, France (MacDonald 2007a). We also thank Elea Feit, Fred Feinberg, and W. Ross Morrow for their helpful advice; and Luth Research and Sawtooth Software for providing survey tools.

Notes

We also performed Tukey tests on Equations (17) and (19). The results remain significant (p<0.05) even after performing the more conservative test accounting for the distribution of the range.

The fit was tested by estimating parameters that the three versions shared (quilting, pattern, packaging, recycled paper content) with LL maximisations of combined and separate models. The combined model had a LL of −2200 and the separate models had LL of −652, −792, and −721 for Versions A, B, and C. A χ² distribution with eight degrees of freedom was used to determine that the separate models provided a better fit than the combined model (p<0.001).