Estimating Utility Functions in the Presence of Response Error

This paper explores the nature and extent of response error when direct multiattribute utility assessment procedures are used as a basis for modeling preferences for risky multiattribute alternatives. The analysis is based on an experimental study of preferences for alternative air pollution control policies whose consequences were characterized by three value attributes: cost to consumers, level of pollution related illness, and level of pollution related mortality. The study generated the following findings: (i) direct assessments of preferences for outcomes were quite reliable and stable over a two-week time period; (ii) parameter estimates for additive utility functions fitted to direct utility assessments were both precise and stable over a two-week time period; (iii) statistically fitted additive utility models provided very accurate predictions of directly assessed preferences two weeks later (or earlier); (iv) ranking outcomes before assigning utilities to them resulted in high levels of serial correlation of errors in direct assessments; and (v) using a parameter estimation procedure that adjusted for serial correlation of errors had little effect on the accuracy of the model's predictions of preferences in a different time period.