Assessing Attribute Significance in Conjoint Analysis: Nonparametric Tests and Empirical Validation

Statistical testing of attribute significance is not possible in conjoint studies that use nonmetric algorithms to analyze respondent ranks of multiattribute product profiles. Procedures that test attribute significance at an aggregate (e.g., segment) level but maintain individual differences in preferences can be used (1) to confirm differences among benefits sought by hypothesized segments of respondents, (2) to eliminate insignificant attributes, reducing the time and cost of conjoint choice simulations, and (3) to design subsequent conjoint studies for the same product class. The author presents two tests of attribute significance in conjoint analysis. One is appropriate when consumer preferences for attribute levels can be ordered a priori and the other can be used when such ordering is not permissible. Each test permits different levels of an attribute to appear in different numbers of product profiles. The proposed tests assess attribute significance across multiple respondents with idiosyncratic preferences. Because they use rank order data, the testing procedures are not limited to a specific scaling algorithm. A Monté Carlo simulation indicates that eliminating insignificant attributes does not affect share-of-choices predictions for new product concepts if the number of insignificant attributes is not very large. Otherwise, the usual tradeoff between parsimony and predictive accuracy is necessary.