Mapping the Clinical Chronic Obstructive Pulmonary Disease Questionnaire onto Generic Preference-Based EQ-5D Values

Abstract

ObjectivesTo develop a model to predict EuroQol five-dimensional questionnaire (EQ-5D) values from clinical chronic obstructive pulmonary disease (COPD) questionnaire (CCQ) scores. MethodsWe used data from three clinical trials (the Randomized Clinical Trial on Effectiveness of Integrated COPD Management in Primary Care [RECODE], the Assessment Of Going Home Under Early Assisted Discharge [GO-AHEAD], and the Health Status Guided COPD Care [MARCH]). Data were randomly split into an estimation sample and a validation sample. The conceptual similarity between patient-reported CCQ and preference-based EQ-5D scores was assessed using correlation and principal-component analysis. Different types of models were estimated with increasing complexity. We selected the final models on the basis of mean absolute error and root mean square error when comparing predicted and observed values from the same population (internal validity) and from different trial populations (external validity). We also developed models for different country-specific EQ-5D value sets. ResultsThe principal-component analysis showed that the CCQ domains functional state and mental state are associated with four dimensions of the EQ-5D. The EQ-5D dimension pain/discomfort formed a separate construct on which no CCQ item loaded. The mean observed EQ-5D values were not significantly different from the mean predicted EQ-5D values in internal validation samples but did significantly differ in external validation samples. The models underestimated EQ-5D values in milder health states and overestimated them in more severe health states. The predictive ability of the models was similar across different EQ-5D value sets. ConclusionsThe models can predict mean EQ-5D values that are similar to observed mean values in a similar population. The overestimating/underestimating of the low/high EQ-5D values, however, limits its use in Markov models. Therefore, mapping should be used cautiously.