Purpose: We sought to create a robust mapping algorithm for predicting EQ-5D utility scores from descriptive quality of life responses in the Seattle Angina Questionnaire (SAQ) for use in patients with coronary artery disease.
Methods: Using data from the Alberta Provincial Project for Outcome Assessment in Coronary Heart Disease (APPROACH) database, we examined the relationship between scores in each of the five domains of the SAQ (physical limitation, anginal stability, anginal frequency, treatment satisfaction, and disease perception) and the EQ-5D utility score. We divided the cohort into an 80% derivation set and a 20% validation set. Potential regression models that were examined were a simple linear regression model, a polynomial regression model and a Tobit model. To account for the skewed distribution of the EQ-5D scores and the presence of heteroscedasticity, we applied Bayesian extensions to each of the models by specifying the variance of the error term as a function of the SAQ components. Model performance in the validation data set was assessed by the amount of variance explained (R2) by the model and the mean of the absolute difference between the predicted and actual EQ-5D scores.
Results: Our cohort consisted of 1992 patients with complete data. We found that the simple linear regression model had the best predictive ability, with a R2 of 0.38 and a mean absolute error of 0.088 (Table). In contrast, the Tobit model performed poorly, only explaining 29% of the variance in the validation set. The Bayesian extensions with specification of the variance term did not improve overall model performance for any of the models.
Conclusions: We found that EQ-5D scores can only be predicted with moderate accuracy from the SAQ using a simple linear regression. This limited performance is likely the consequence of mapping the disease-specific SAQ onto a generic utility measure. Our study suggests that further research is necessary to identify a subset of either the SAQ or EQ-5D that have greater overlap so as to improve predictive ability. Table:
Linear Regression | Linear Regression with specified variance | Polynomial | Polynomial with specified variance | Tobit | Tobit with specified variance | |
| mean (2.5% to 97.5%) | ||||||
| R2 | 0.38 (0.36 to 0.39 | 0.38 (0.36 to 0.39) | 0.37 (0.35 to 0.38) | 0.37 (0.35 to 0.38) | 0.29 (0.26 to 0.32) | 0.28 (0.25 to 0.30) |
| mean-error | 0.088 (0.087 to 0.090 | 0.088 (0.087 to 0.090) | 0.090 (0.088 to 0.091) | 0.090 (0.088 to 0.091) | 0.089 (0.088 to 0.091) | 0.090 (0.088 to 0.091) |
Candidate for the Lee B. Lusted Student Prize Competition