Modeling Health State Utility Values in Ankylosing Spondylitis: Comparisons of Direct and Indirect Methods

Abstract

ObjectivesCost-effectiveness analyses of technologies for patients with ankylosing spondylitis frequently require estimates of health utilities as a function of the Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) and the Bath Ankylosing Spondylitis Functional Index (BASFI). MethodsLinear regression, bespoke mixture models, and generalized ordered probit models were used to model the EuroQol five-dimensional questionnaire as a function of BASDAI and BASFI. Data were drawn from a large UK cohort study (n = 516 with up to five observations) spanning the full range of disease severity. ResultsLinear regression was systematically biased. Three- and four-component mixture models and generalized probit models exhibit no such bias and improved fit to the data. The mean, median, mean error, and mean absolute error favored the mixture model approach. Root mean square error favored the generalized ordered probit model approach for the data as a whole. Model fit assessed using these same measures by disease severity quartiles tended to be best using the mixture models. The value of moving from good to poor health may differ substantially according to the chosen method. Simulated data from the mixture and probit models yield a very similar distribution to the original data set. ConclusionsThese results add to a body of evidence that the statistical model used to estimate health utilities matters. Linear models are not appropriate. The four-class bespoke mixture model approach provides the best performing method to estimate the EuroQol five-dimensional questionnaire values from BASDAI and BASFI.