3 TREATMENT PREFERENCES DERIVED USING ADAPTIVE BEST-WORST CONJOINT (ABC) ANALYSIS

Friday, October 19, 2012
The Atrium (Hyatt Regency)
Poster Board # 3
Decision Psychology and Shared Decision Making (DEC)

Ely Dahan, PhD1, Sylvia Lambrechts, MPH, MA1, Robert M. Kaplan, PhD2, Catherine M. Crespi, PhD1, Elizabeth Garcia, BS1 and Christopher S. Saigal, MD, MPH1, (1)UCLA, Los Angeles, CA, (2)University of California Los Angeles, Los Angeles, CA

Purpose: We apply a newly developed preference measurement method, Adaptive Best-worst Conjoint (ABC), to improve questioning efficiency for patient treatment preferences. With this approach, utility functions with 7-10 parameters are estimated instantaneously at the individual level with as few as 12-15 tasks completed in 10 minutes.

Method: Conjoint analysis respondents choose the best and worst of four treatment alternatives (attribute bundles). The method adaptively presents the next four treatment alternatives, and after 12-15 tasks a utility function is estimated and an individualized report is printed. This report summarizes the patient's treatment preferences and priorities, and serves to enhance the doctor-patient discussion of possible treatments. Adaptive Best-worst Conjoint with four options-at-a-time identifies five of the six possible paired comparisons (Best > option B, Best > option C, Best > Worst, option B > Worst, option C > Worst; only B is not compared to C). So ABC is 66% more efficient than traditional choice-based conjoint even without adaptive questioning. This inherent efficiency advantage of best-worst questioning is further enhanced through adaptive questioning based on transitivity of preference. That is, we assume that if full-profile A is preferred to full-profile B, and if B > E, then A is also > E, even though we never directly compared A to E. Such transitivity may resolve even more paired comparisons than direct questioning. For example, with 16-full-profiles, there are 16 x 15 / 2 = 120 possible paired comparisons, over 50% of which are resolved through transitivity.

Results: The presentation highlights three key results: •Internal consistency:  The estimated utility functions explain 88%, on average, of the variance in treatment scores. •Estimation and prediction: a comparison of linear regression, LINMAP, and Hierarchical Bayes shows that the abc method has high predictive accuracy (e.g., 68% first choice-out-of-four hit rate for holdout questions).  Also, 80%-90% of paired comparisons in holdouts are consistent with the estimated utility function. •Respondent and physician reactionsto this system have been favorable as compared with a control group.

Conclusion: Adaptive Best-worst Conjoint analysis compares favorably with Ratings Scale and Time Tradeoff as a way of measuring patient preferences.  This presentation will lay out the Excel-based method as a direct takeaway from the SMDM conference.