Reducing Respondent Burden in Ranking Tasks: Hierarchical Bayesian Analysis of Pairwise Comparisons with Covariates

• Activity Number: 192
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and Marketing
• Author(s): Well Howell*+
• Companies: Harris Interactive
• Keywords: hierarchical Bayesian analysis ; Bradley-Terry Model ; pairwise comparisons

Abstract

Hierarchical Bayesian (HB) analysis of pairwise comparison data with respondent covariates as suggested by Allenby and Ginter reduces the number of comparisons needed per respondent. Marketing practitioners seek respondent level information about which product attributes are more important in stimulating a first purchase or which features make a product more useful. The key issue is how attributes are ranked on some utility scale. However, self-reported utility measurements are notoriously unreliable. Pairs of attributes can be compared instead, but this quickly becomes an overwhelming task, as an 11-item list would produce 55 possible pairs. This task can be fractionalized so that each respondent sees only a portion of all pairs. The data are then analyzed with an HB model that includes respondent-level covariates in the prior. This paper shows that useful results can be produced on a test dataset with reduced respondent burden when covariates are included in the model, and that these results appear more accurate than those produced by alternative HB models or an aggregate-level analysis.