Abstract
Caution should be taken when using EQ-5D-3L responses mapped from the SF-12 to describe patient outcomes or when undertaking economic evaluation, due to the underestimation of disability associated with mapped values. The findings from this study could be used to adjust expected EQ-5D-3L preference weights when estimated from SF-12 item responses when combining data from studies that use either instrument.
Introduction
In health research, comparing outcomes across different studies is often complicated by the use of different instruments to measure Health-Related Quality of Life (HRQL). Two of the most common tools are the EQ-5D, a five-question survey, and the SF-12, a 12-item survey. To pool data from studies using different instruments, researchers often employ "mapping algorithms"—mathematical models that translate SF-12 responses into EQ-5D scores.
However, most existing algorithms were developed using data from relatively healthy general populations. This study addresses a critical gap: determining whether these mapping shortcuts remain accurate for populations with significant disabilities, such as those recovering from major trauma. Inaccurate mapping in these cohorts could lead to flawed economic evaluations and a fundamental misunderstanding of the true burden of injury.
Methods
The researchers evaluated a mapping algorithm developed by Gray et al., which utilizes multinomial logit regression to predict EQ-5D responses from SF-12 data. Unlike simpler models that only predict a final utility score, this approach attempts to predict the specific level of disability for each of the five EQ-5D dimensions: Mobility, Self-care, Usual activities, Pain/discomfort, and Anxiety/depression.
To test accuracy, predicted scores were compared against actual EQ-5D scores reported by the same patients using several statistical methods:
- Kappa & PABAK statistics: To measure the level of categorical agreement between predicted and actual responses.
- Bland-Altman plots: To visualize bias and the "limits of agreement" across the spectrum of health states.
- UK Value Sets: To convert categorical responses into "preference weights" (utility scores ranging from -0.56 to 1.00).
Datasets
The study utilized a large dataset from the Injury-VIBES project, drawing from two major Australian registries: the Victorian State Trauma Registry (VSTR) and the Victorian Orthopaedic Trauma Outcomes Registry (VOTOR). The final analysis included 10,166 adult trauma patients who completed both surveys at 6 and 12 months post-injury, representing one of the largest validation cohorts for this type of mapping in a high-disability population.
Results
The mapping algorithm consistently exhibited an "optimism bias," making patients appear healthier than they actually were by under-reporting disability in four out of the five dimensions. Actual agreement on individual items ranged from 65% for Pain/discomfort to 86% for Self-care.
At 6 months post-injury, the mapped preference weights were significantly higher (healthier) than actual reports, with a mean difference of -0.024 (p < 0.001). For the "Usual Activities" dimension, the algorithm underestimated disability for nearly 25% of participants. While the average difference in scores was relatively small, the Bland-Altman limits of agreement were wide, suggesting that for any individual patient, the mapped score could differ substantially from their actual experience.
Discussion and Limitations
The systematic underestimation of disability suggests that using mapped EQ-5D values in cost-effectiveness analyses might make interventions appear less valuable than they truly are, as the potential for improvement is artificially narrowed.
Several limitations were noted: the study focused on hospitalized trauma patients, meaning results might not generalize to minor injuries treated in primary care. Additionally, patients with severe traumatic brain injuries or dementia were excluded because the SF-12 cannot be administered by proxy. Future work should focus on developing adjustment factors to correct the inherent healthy-bias of current algorithms when applied to injury populations, ensuring more accurate economic and clinical assessments.
Summary updated