Improving Conjoint Analysis with Incorporating Unvertainty

This study provides evidence that the out-of-sample predictive performance of conjoint analysis can be improved by measuring and modeling the uncertainty of preference statements. Preferences are measured in terms of rating scores for products, while uncertainty is considered as an indicator of the stability of preferences. Uncertainty is measured in six different ways and for each of these measures a Hierarchical Bayes model was developed. The models were empirically tested and the results indicate that including uncertainty measures leads to an improvement in out-of-sample predictive performance and the precision of the predictions. The best performance was found for a combination of two implicit uncertainty measures, location and test-retest, using a weighted regression model. The estimated weights suggest that preferences with a low score, relative to all of the scores from an individual, should have a smaller weight or be given a higher impact. In addition, we found some evidence that preferences change over time. While there appears to be evolution, we found that valuable information could be extracted from the initial data when the low preference scores were allowed to have a higher impact relative to the other preference scores, supporting the idea that preference about products that a consumer does not like is more informative and more stable over time.