Shedding New Light Onto the Ceiling and Floor? A Quantile Regression Approach to Compare EQ-5D and SF-6D Responses

Many studies now exist that compare different generic quality of life instruments. A particular issue is the existence of floor and ceiling effects, or the lack of sensitivity of a measure to changes in health status at the extremes of the distribution. Previous studies have used relatively simple methods to examine these effects. This has included the use of cross-section data that has ignored the effects of unobserved individual heterogeneity, and also the use of least squares regression methods based on the mean of the distribution. The aim of this paper is to determine the extent to which quantile regression applied to longitudinal data improves our understanding of the relationship between quality of life instruments. The paper uses data from a randomised trial of a community pharmacist medicines management service for patients with coronary heart disease. The study used EQ-5D and SF-36 (converted to SF6D values) instruments with both a baseline and follow up measurement for both study and control groups. Compared to ordinary least squares methods, a first difference model shows that there is a much lower association between the measures. This suggests that OLS methods may lead to biased estimates of the association, as a significant part can be attributed to unobservable patient characteristics. The results also show statistically significant ceiling and floor effects. For people whose health improves over time, the association between instruments is higher for those who experienced the largest increases in health, whilst for people who deteriorate, the association is higher amongst those with the smallest reduction in health status. The results have important policy implications for resource allocation decisions where these measures are used. The cost-effectiveness of an intervention may depend on the choice of instrument and thus it is important to document the magnitude of differences in cost-effectiveness that can be attributed to a particular instrument. The choice of instrument should also depend on the initial health status of the population being studied, and their expected changes in health status