Response mapping for transforming SF36 data into utility scores

Presenter: Duncan Mortimer, Monash University & University of South Australia

Abstract

Background: Response mapping has recently been proposed as a means of improving the quality of transformations between descriptive and utility-based quality of life instruments. Response mapping assigns each individual to a response level on each item of a utility-based instrument (such as the EQ-5D); based on data from a descriptive instrument (such as the SF-36). Response mapping preserves the main design features of the utility-based instrument and directly addresses many of the criticisms that have previously been leveled at regression-based transformations, including: (i) predicted utility scores that lie outside the range of the utility-based instrument, and (ii) predicted utility scores that need not correspond to any health state on the descriptive system of the utility-based instrument.

While predicted scores obtained via response mapping may be more consistent with the theoretical properties of the utility-based instrument, there may be an associated loss of precision. Incorrect assignment to the response level of even a single item can translate into a substantial error if the scoring tariff of the utility-based instrument imposes large increments/decrements when moving between health states. Response mapping may therefore be much more suited to use in instruments that are able to detect small changes in utility. To test this conjecture, the present study applies response mapping to derive transformations from the SF-36 to the EQ-5D, HUI3 and AQoL.

Methods: Multinomial logistic regressions to estimate the probability of selecting each response level on each item of each utility-based instrument (EQ-5D, HUI3 & AQoL) from SF-36 scale scores. Monte Carlo simulation used to obtain predicted response levels for each respondent. Scoring tariffs then applied to predicted response levels to obtain predicted EQ-5D, HUI3 and AQoL utility scores. Mean absolute error (MAE) and mean squared error (MSE) evaluated for each quintile of each instrument’s utility scale. For each quintile, MAEs and MSEs for each instrument compared with the magnitude of utility increments/decrements between health states in that instrument’s descriptive system.

Data: Sample of 996 persons drawn from three strata and comprised of: (a) 396 non-institutionalised persons aged over 16 years and resident in Victoria, Australia; (b) 334 outpatients from two Group A public hospitals in metropolitan Melbourne; and (c) 266 admitted inpatients from three Group A public hospitals in metropolitan Melbourne. The study sample was randomly split into an estimation set (N=455) and a validation set (N=447) to allow out-of-sample validity testing.

Results: MAEs and MSEs between predicted and actual scores were closely correlated with the magnitude of utility increments/decrements between health states in each instrument’s descriptive system. As a consequence, MAEs and MSEs between predicted and actual HUI3 and AQoL utility scores were generally lower than for the EQ-5D.

Conclusions: Results suggest that the precision with which response mapping can predict utility scores depends on the characteristics of the utility instrument being targeted. For instruments with relatively small increments/decrements between health states such as the HUI3 and AQoL, response mapping can deliver predicted utility scores with acceptable theoretical properties and improved precision.