INVESTIGATING THE IMPACT OF PRIOR RANKING ON VAS RATINGS FOR EQ-5D HEALTH STATES

Purpose: Significant interest has been shown recently in the construction of cardinal scales of value for health derived from ordinal preference data (McCabe et al, Journal of Health Economics, 2006). Ordinal preferences can be determined using various methods, including the rank ordering of health states. Such rankings are often inferred from Visual Analogue Scale (VAS) ratings, rather than from directly observed rankings. This procedure assumes an identity between the rank order of a given health state inferred from VAS and that observed from a direct ranking task. The current study examines whether such assumption is valid or not and tests its impact in estimating values for EQ-5D health states.

Method: 205 university students were asked to rank and rate (using a 20-cm VAS scale) a set of 16 EQ-5D health states selected from the US Valuation Survey (Shaw et al, Medical Care, 2005). Half the participants performed the ranking task first (Group A); the remainder completed the rating task first (Group B). Each participant was rewarded with a £10 pound voucher (~$20 US) after completing the experiment. The inferred rank of each health state was derived by ranking the VAS ratings. Directly observed ranks were obtained from the ranking task in the experimental protocol. Spearman’s rank-order correlation was computed on a with-subject basis for each pair of rankings. A conditional logistic regression model was applied separately to the inferred and observed rankings to estimate values for 243 EQ-5D health states using 2 dummies for each of the 5 EQ-5D dimensions as independent variables.

Results: There were no statistically significant differences in the demographics of Groups A and B. VAS scores were averagely higher in Group A by 4 points. The mean correlation between observed and inferred rankings for Group A (r=0.93) was significantly higher (p<.02) than for Group B (r=0.91). Within each Group the coefficient estimates were higher when based on observed rankings. These were significantly different (p<.001) for the mobility, pain/discomfort and anxiety/depression dimensions.

Conclusion: Results show discrepancies between values computed using inferred and directly observed rankings. Prior exposure to ranking yields significantly different VAS ratings for health states. The current practice of modelling rank data derived from VAS ratings to generate an interval scale of values needs to be treated with some caution.