MEASURING STATED PREFERENCES TO IDENTITY ATTRIBUTE IMPORTANCE: IS IT A CASE OF 1 ATTRIBUTE WITH N LEVELS OR N ATTRIBUTES WITH ONLY 1 LEVEL

Monday, October 24, 2011
Grand Ballroom AB (Hyatt Regency Chicago)
Poster Board # 30
(ESP) Applied Health Economics, Services, and Policy Research

John F.P. Bridges, PhD and Gisselle Gallego, PhD, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD

Purpose: Increasingly researchers are using stated preference methods to estimate preferences in health care using N attributes with 2 or more levels, often reporting on attribute importance (the value that one places on the attributes presented) ignoring the impact of level importance (the value associated with the differences in the levels of a given attribute). 

Method: We present two case studies examining attribute importance focused on liver cancer control. The first used discrete choice experiment (DCE) to examine 11 possible liver cancer control strategies where respondents chose among paired subsets of the 11 strategies based on an orthogonal array. The second used best-worst scaling (BWS) (Case 1) to explore the likely future impact of 11 emerging liver cancer control technologies using a balance incomplete block design (BIBD) that generated sets of 5 technologies, with respondents identifying the best and worst in each set. Both studies were analyzed assuming random utility (RUT) theory via a conditional logistic regression. The first simply regressed choice on the attributes present in the possible cards, but the second required an assumption of sequential best worst and the use of effects coding to estimate the marginal value of the 11 technologies to avoid the dummy variable trap.

Result: 120 experts in liver cancer completed both surveys, a response rate of 37%. Respondents includes hepatologists (40%), oncologists (22%), radiologists (13%), surgeons (18%) and other experts (19%) involved in hepatocellular carcinoma (63%), hepatitis (n=16%), transplantation (13%) and metastatic liver cancer (8%). From the DCE, the highest valued strategy was monitoring ask risk populations (p<0.001). From the BWS molecular targeted therapy was most valued (p<0.001). 

Conclusion: We demonstrate two methods for the estimation of attribute importance within a RUT framework.  The application of DCE requires an assumption that there are N attributes of only one level each, while BWS case 1 assumes that there are only 1 attribute, but with N levels. While we demonstrate that both the DCE and BWS can be used to estimate attribute importance, these competing methods make different assumptions, and hence, results are on different scales. More research is needed to directly compare the results of these two methods of estimating level importance on a single research question and common attributes/levels.