OBJECTIVES: The lack of preference-based utility data places great importance on the accuracy of mapping functions. The objective of this study is to assess the predictive accuracy of statistical models which address the unique properties of EQ-5D. METHODS: A large dataset from the Patient Reported Outcome Measures (PROMs) programme reporting EQ-5D and OHS values for patients who have undergone total hip replacement, during April 2010 and March 2011, was used to develop 6 mapping functions using different statistical methods: Ordinary Least Squares (OLS), standard Tobit, adjusted Tobit, two-part logit (TPL), response-mapping and censored least absolute deviation (CLAD). Three different model specifications were investigated, including the total OHS, individual item score and individual item responses. Each model specification was examined using goodness-of fit measures. The predictive accuracy of each model was analysed using the mean absolute error (MAE) and mean squared error (MSE). Model performance was compared in an internal and external validation. RESULTS: The OHS individual item response variables proved to give the best model fit and were therefore used across all models. The OLS and TPL models consistently demonstrated the highest predictive accuracy, providing the lowest MSE and the closest estimation of the mean EQ-5D. The response-mapping approach was the poorest predictor in estimating individual values; however it was able to predict the median with perfect precision. Models using Tobit and CLAD frameworks provided the poorest predictions. CONCLUSIONS: The OLS and TPL models proved to be the most accurate in predicting EQ-5D on an individual level, whilst the response-mapping model is recommended for predicting the median. Using inaccurate mapping functions such as the Tobit models developed in this study can have a substantial impact on CEA results and reimbursement decisions.