![[ Visit Client Website ]](images/banner.gif)
To derive an algorithm to predict EQ-5D utility using the St George’s Respiratory Questionnaire (SGRQ) that can be used to estimate utility for Chronic Obstructive Pulmonary Disease (COPD) patients.
Purpose Methods Results The following OLS-based algorithm performed well: 1-Ud=0.0335+0.0017T+0.0001T2–0.0279G (1-Ud=predicted utility score, T=total SGRQ score, G=gender (0=female, 1=male)) with RMSE=0.1746 and R2=0.45. These values are comparable to those reported elsewhere: 0.084-0.2 for RMSE, 0.17-0.51 for R2 (Brazier et al, 2007). Some over-prediction of utility at lower health states and under-prediction at higher health states was apparent. The 2-part model (using OLS) performed only slightly better (RMSE: 0.1744). Conclusion The OLS regression model performed favourably compared with more complex statistical models. Although utility values may be derived using such an algorithm, some caution should be taken as the algorithm may not predict utility for the worst and best health states very accurately. Therefore, until a better predicting algorithm or other linkage is developed, direct collection of utility values within trials, via an instrument such as the EQ-5D is recommended.
Although cost-utility analysis alongside clinical trials is becoming increasingly common, not all trials incorporate an instrument (eg EQ-5D), from which utility values may be elicited. In such cases, an algorithm to estimate utility scores may be useful. It is however important that the algorithm is appropriately developed and validated if it is to be used to support resource allocation decisions.
The TORCH (Towards a Revolution in COPD Health) trial (reported in Calverley et al, 2007), collected SGRQ and EQ-5D at baseline and every 24 weeks over three years for approximately 4000 subjects. Data were split non-randomly into a ‘fitting’ sample (all non-USA, 66% of subjects) and ‘validation’ sample (all USA, 34% of subjects). EQ-5D responses were converted to index scores using a UK population survey. Regression equations were used to map SGRQ to EQ-5D index score using demographic variables and total, domain or item scores from SGRQ. Functional forms considered were: OLS, random effects, GLM (gamma, log link), GEE (gamma, log link, exchangeable correlation), Tobit, and a 2-part model (logistic regression plus OLS, GLM or GEE). Algorithms developed using the fitting sample were used to predict index scores in the validation sample. Predictive ability was measured using Root Mean Squared Error (RMSE).