Outlier Detection for a Hierarchical Bayes Model for Patient Preferences Elicited Through Discrete Choice Experiments

• Activity Number: 198
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Anna Liza M. Antonio* and Catherine Crespi and Robert E. Weiss and Christopher Saigal
• Companies: UCLA and UCLA and UCLA and UCLA
• Keywords: Best-worst choice experiments ; conjoint analysis ; diagnostics ; conditional predictive ordinate ; missing data ; random-effects

Abstract

To provide patient-centered care, it is important to identify and characterize patients with outlying preferences. In our study, patients participated in a discrete choice experiment designed to elicit their preferences among health states that might occur subsequent to prostate cancer treatment. Perfect health being ideal but improbable, the health states implicitly force patients to make tradeoffs between treatment side effects (e.g., better or worse sexual or urinary functioning) and other treatment attributes (surgical or non-surgical, family/doctor support). In our design, patients were presented with sets of four health states and asked to choose their most and least preferred, reducing respondent burden but leaving some health states unranked. We use a hierarchical Bayes random-effects rank-ordered multinomial logit model which accounts for missing ranks by marginalizing over all possible permutations of unranked profiles to estimate patient-specific regression coefficients. The conditional predictive ordinate is used to identify patients with outlying preferences with respect to the population and outlying choice sets that deviate from estimated patient preferences.