In addition to theoretical justifications, many statistical methods have been used for selecting covariates to include in algorithms mapping nonutility measures onto utilities. However, it is not clear whether using exploratory factor analysis (EFA) as one such method improves the predictive ability of these algorithms.
This question is addressed within the context of mapping a non–utility-based outcome, the core 23-item Women’s Health Questionnaire (WHQ-23), onto two utility instruments: five-level EuroQol five-dimensional questionnaire (EQ-5D-5L) and the six-dimensional health state short form (derived from short form 36 health survey) (SF-6D).
Data on all three outcomes were collected from 455 women from the Australian general population participating in a study assessing attitudes toward in vitro fertilization. Statistical methods for selecting covariates included stepwise regression (SW), including all covariates (Include all), multivariable fractional polynomial (MFP), and EFA. The predictive accuracy of 108 regression models was assessed using five criteria: mean absolute error, root mean squared error, correlation, distribution of predicted utilities, and proportion of predictions with absolute errors of less than 0.0.5. Validation of “primary” models was carried out on random samples of the in vitro fertilization study.
The best results for EQ-5D-5L and SF-6D predictions were obtained from models using SW, “Include all,” and MFP covariate-selection approaches. Root mean squared error (0.0762–0.1434) and mean absolute error (0.0590–0.0924) estimates for these models were within the range of published estimates. EFA was outperformed by other covariate-selection methods.
It is possible to predict valid utilities from the WHQ-23 using regression methods based on SW, “Include all,” and MFP covariate-selection techniques.