Purpose: Combining individual health state utilities (single states, SS) into joint states (JS) is necessary for accurate decision modeling. We previously showed our linear index model surpasses other theoretical models for estimating JS from SS related to prostate cancer. Some subjects give logically inconsistent ratings in utility surveys, however, which may affect data-based predictions. This research measures the impact of excluding inconsistent ratings on aggregate linear index prediction values, and identifies socio-demographic features linked with inconsistent utility ratings.
Method: Men completed a utility elicitation survey after prostate biopsy (n=279), during which they rated SS and JS related to prostate cancer treatment using time tradeoff. Ratings were tested for logical inconsistency: a JS rating should not be higher than either composing SS. An alternate definition to accommodate measurement error considered ratings inconsistent only if the JS exceeded the population average for the SS by 1+ SD. Aggregate mean SS utilities with and without inconsistent responses were entered into the linear index model to predict JS values. Univariate and multivariate regression showed associations between socio-demographic features and rating inconsistently.
Result: Linear index prediction values are lower by 4 to 7 points on a 0-100 utility scale for JS of Impotence with each of Asymptomatic Localized Disease, living Post-Prostatectomy, and Incontinence when inconsistent responses for these health states are removed. Excluding ratings by people who give 1 SD rating inconsistencies drops prediction values slightly less, by 3 to 4 points for each of the three CS. Univariate regression analyses indicate associations between rating inconsistency and being married, anxious, living at home, and taking more time with each rating. Large errors are made by African-Americans, people with lower education, worse current health, and lower income. In multivariate analysis, marriage and anxiety significantly predict inconsistency by both definitions. Lower education and rating time were also significant for giving any inconsistent rating, while worse current health was significant for rating inconsistently by the 1 SD criterion.
Conclusion: Excluding inconsistent responses lowers JS utilities predicted using our linear index model. Clinically important differences have not been established for utilities. Whether excluding inconsistent responses will change a decision depends on preference sensitivity of the decision. Associations with inconsistent ratings indicate that elicitation methods should account for patient emotions, marital status, and education.
Candidate for the Lee B. Lusted Student Prize Competition