AN ITEM RESPONSE THEORY APPROACH FOR PREDICTING THE EQ-5D RESPONSES FROM THE SF-12
Candidate for the Lee B. Lusted Student Prize Competition
Methods: We used data from the Medical Expenditure Panel Survey (MEPS) 2000-2003 to randomly divide observations into estimation and validation samples. Using the estimation sample, we applied a bi-factor IRT model that assumes the EQ-5D and the SF-12 instruments purport to measure the same construct at the item (question) level. The bi-factor model permits each item to tap the primary dimension (e.g. overall well-being) and one-subdomain (e.g. physical well-being). In the validation sample, predictions for each item of the EQ-5D were calculated using the estimates obtained from the bi-factor model assuming that only items from the SF-12 were available. We compared the prediction performance of the bi-factor model to those obtained from ordered logistic regression models, in which each EQ-5D item is modelled separately. We further discussed the connections between these approaches. We then transformed predicted item responses into the EQ-5D preference index employing US preference weights. We evaluated these predictions using mean average error (MAE) and root mean square error (RMSE).
Results: Item response predictions from the bi-factor model are highly accurate and better than those obtained by ordered logistic models. The percent of correctly predicted answers ranged from 89% to 75%, with lower level of agreement for items related to pain and mental health. The bi-factor model was also superior at predicting the EQ-5D preference index after applying preference weights, with improvements in both MAE and RMSE.
Conclusions: Several methods have been proposed for predicting the EQ-5D from the SF-12, as economic evaluations may be hindered when data on preferences are not available. The underlying assumption is that the EQ-5D and SF-12 instruments are related. In our work, we make this assumption explicit and use modern psychometric methods to jointly calibrate both instruments. We showed that the bi-factor IRT model improves prediction and provides further insight on the differences between the instruments. IRT is a general methodology that has great potential in economic evaluations as in many cases technologies and treatments are evaluated using different metrics which need to be harmonized before they can be compared.