Effect of level of disaggregation on conjoint cross validations: Some comparative findings

Early formulations of conjoint models focused on part-worth estimation at the individual level. As the methodology's popularity grew so did industry demands for increasingly larger numbers of attributes and levels. In response to these demands, new approaches, based on partial or full data aggregation (such as clusterwise/latent class conjoint and choice-based conjoint), have appeared. This paper suggests that pooled-data models will often be successful in predicting market shares when researchers employ monotonic attributes. In these cases more of a good attribute (or less of a bad attribute) is always more preferred. In the more realistic case, in which some of the attributes may be nonmonotonic, we find that data aggregation does not predict holdout sample preferences as well as individual part-worth models.