Indirect Utility Assessment

Indirect Utility Assessment with Preference-Based Questionnaires

In indirect utility assessment, patients describe their own health on a standard questionnaire such as the EQ-5D, SF-6D or HUI3, and a scoring model estimated earlier from valuations by members of the public turns each description into a utility on the scale where full health is 1 and dead is 0. No patient is asked to trade years of life or accept a risk of death, so the method can be built into almost any clinical trial or cohort study. The NICE reference case for technology appraisal is built around this route, with EQ-5D responses from patients scored by a UK value set as the preferred source of utilities. Applied to identical answers, two value sets can give QALY gains that differ by enough to move an ICER across a decision threshold.

Patients describe and the public values

The method splits measurement between two groups of people. NICE DSU Technical Support Document 8 explains that patients report their health on instruments sometimes called multi-attribute health status classification systems, and values from existing datasets for each instrument are then applied to each reported state, an approach it says is known in the literature as indirect methods. Arnold and colleagues, writing in the BMJ, add that the tariffs come from earlier exercises in which a general population sample valued health states with a trade-off method.

The NICE manual (PMG36, section 4.3.2) states that deriving a utility usually has two elements, measuring health-related quality of life in people in the relevant state and valuing it against preferences for other states, usually perfect health and death. Section 4.3.3 asks for the measurement to come from patients, or from their carer rather than healthcare professionals when patients cannot report, and section 4.3.4 asks for the valuation to come from a representative sample of the UK population using a choice-based method. Indirect assessment meets both requirements at once, because the patient completes only the descriptive part and the public preferences arrive already packaged in the value set.

The alternative is direct elicitation, in which the person valuing the state answers a time trade-off or standard gamble question. The general process of valuing health states is covered under health state valuation.

Scoring questionnaire answers with a value set

A preference-based questionnaire classifies health into a fixed number of dimensions, each with a fixed number of levels, so every completed form corresponds to one health state. NICE's interim methods statement on the EQ-5D-5L (PMG51, section 1.5) describes the value set as a mathematical model that converts EQ-5D questionnaire responses to utility values for use in economic models. The national EQ-5D-3L time trade-off value sets tabulated by Szende and colleagues are expressed largely as a constant and decrements for each dimension at each level above the first, and Dolan's UK model adds an intercept for any move away from full health, which gives the additive form below.

$$u(s) = 1 - c \cdot I(s) - \sum_{d=1}^{D} \beta_{d, l_d(s)}$$

where $u(s)$ is the utility of health state $s$, $c$ is a constant decrement, $I(s)$ equals 1 if any dimension is above the best level and 0 otherwise, $D$ is the number of dimensions, $l_d(s)$ is the level reported on dimension $d$, and $\beta_{d, l_d(s)}$ is the decrement for that level, which is 0 at the best level. Some value sets add interaction terms or use other functional forms, so the published algorithm is always the one applied.

The coefficients are estimated from valuations of a designed subset of states, since a descriptive system has far more states than any survey could value one by one. The UK EQ-5D-5L value set by Rowen and colleagues rests on 102 states valued by composite time trade-off.

Instruments and the values they produce

The choice of questionnaire changes the number that comes out, because instruments differ in content and in the valuation technique behind their value sets. NICE DSU Technical Support Document 11, published in 2011, listed the EQ-5D with five dimensions at three levels, giving 243 states valued by time trade-off, and the SF-6D with six dimensions including energy, valued by standard gamble. HUI3 has eight attributes and 972,000 states, valued by a visual analogue scale transformed into standard gamble values, and TSD 11 notes that the HUI instruments focus on impairments such as vision, hearing and dexterity.

These differences produce disagreement on the same patients. TSD 11 reports only poor to moderate agreement between measures, around 0.3 to 0.5 by the intraclass correlation coefficient, and adds that mean differences of little more than 0.05 between SF-6D, EQ-5D and HUI3 mask large differences in the distribution of scores. The EQ-5D appears to suffer from ceiling effects, with many respondents reporting no problems, while the SF-6D has a floor at 0.301 against minus 0.594 for the EQ-5D. TSD 11 also states that EQ-5D and SF-6D value sets from different countries have been shown to be significantly different, hence the NICE requirement for UK general population values.

Indirect utilities in the NICE reference case

PMG36 section 4.3.5 states that different methods of measuring health-related quality of life produce different utility values, so results from different methods or instruments cannot always be compared. NICE therefore prefers one instrument. PMG51 gives new wording for PMG36 section 4.3.6, making the EQ-5D-5L the preferred measure in adults, scored with the UK value set by Rowen and colleagues in reference-case analyses.

When the preferred data are missing, NICE still expects an indirect route. PMG51 section 1.10 asks for EQ-5D-3L descriptive data to be mapped to utilities based on the UK 5L value set using the mapping model developed by the Decision Support Unit and the Economic Evaluation Policy Research Unit. Section 1.11 asks that any model for mapping non-EQ-5D data be fully described, justified and tested in sensitivity analyses, and section 1.13 rules out 3L-based utilities in sensitivity or scenario analyses. A mapping function adds a statistical prediction step between the questionnaire and the utility.

In the revised figure 4.1 set out in PMG51, another generic and then a condition-specific preference-based measure remain the first alternatives when evidence shows the EQ-5D is not appropriate, ahead of vignettes and direct valuation of own health. PMG36 section 4.3.10 asks for that evidence to show that key dimensions of health are missing, supported by poor performance on tests of construct validity and responsiveness.

Indirect utilities in trial-based analyses

A short questionnaire can be completed at every trial visit. TSD 8 recommends a sensitivity analysis using any other preference-based measure collected in the trial, such as the SF-6D or HUI, and, where the analysis rests on a single trial, using each arm separately with baseline differences adjusted where necessary. Manca, Hawkins and Sculpher note that baseline mean utility is invariably imbalanced between arms and likely to be highly correlated with QALYs over follow-up, and show that widely used methods for differential QALYs give incorrect results when it is imbalanced. They show that multiple regression controlling for baseline utility can give appropriate estimates.

Some patients cannot complete the questionnaire. TSD 8 notes that agreement between proxies and patients tends to be good for observable dimensions such as mobility and self-care, but weaker or poor for pain and depression. For children, PMG36 section 4.3.14 states that NICE does not recommend specific measures and that not all paediatric instruments have a UK value set.

Worked example with one trial and two value sets

All numbers in this example are illustrative, and both value sets are invented for teaching. A 12-month trial records a five-dimension, three-level questionnaire at baseline, 6 months and 12 months. States list the level on mobility, self-care, usual activities, pain and anxiety or depression in that order, with 1 meaning no problems.

Value set A uses a constant of 0.05 and decrements at levels 2 and 3 of 0.06 and 0.25 for mobility, 0.08 and 0.18 for self-care, 0.04 and 0.10 for usual activities, 0.10 and 0.32 for pain, and 0.07 and 0.20 for anxiety or depression. Value set B is identical except for usual activities, at 0.10 and 0.20, and anxiety or depression, at 0.15 and 0.30.

1. Record the responses. Every patient on the new treatment reports 2-1-2-3-2 at baseline and 1-1-1-2-1 at 6 and 12 months. Every patient on usual care reports 2-1-2-3-1 at baseline and 2-1-2-2-1 at 6 and 12 months.

2. Score with value set A. The new treatment arm scores 1 − 0.05 − 0.06 − 0.04 − 0.32 − 0.07 = 0.46 at baseline and 1 − 0.05 − 0.10 = 0.85 afterwards. Usual care scores 1 − 0.05 − 0.06 − 0.04 − 0.32 = 0.53 at baseline and 1 − 0.05 − 0.06 − 0.04 − 0.10 = 0.75 afterwards.

3. Convert to QALYs. Joining the measurements with straight lines, the new treatment gives (0.46 + 0.85) / 2 × 0.5 = 0.3275 for the first half year and 0.85 × 0.5 = 0.425 for the second, a total of 0.7525 QALYs. Usual care gives (0.53 + 0.75) / 2 × 0.5 = 0.32 and 0.75 × 0.5 = 0.375, a total of 0.695. The unadjusted gain is 0.7525 − 0.695 = 0.0575 QALYs.

4. Note the baseline imbalance. The new treatment arm started 0.53 − 0.46 = 0.07 lower, and baseline carries a weight of 0.5 / 2 = 0.25 years in the first interval, so the imbalance alone removes 0.07 × 0.25 = 0.0175 QALYs from the gain. A real analysis would adjust by regression on baseline utility.

5. Score the same responses with value set B. The new treatment arm scores 1 − 0.05 − 0.06 − 0.10 − 0.32 − 0.15 = 0.32 at baseline and 0.85 afterwards. Usual care scores 1 − 0.05 − 0.06 − 0.10 − 0.32 = 0.47 at baseline and 1 − 0.05 − 0.06 − 0.10 − 0.10 = 0.69 afterwards. The first half year gives (0.32 + 0.85) / 2 × 0.5 = 0.2925 for the new treatment and (0.47 + 0.69) / 2 × 0.5 = 0.29 for usual care. Adding the second half years of 0.425 and 0.69 × 0.5 = 0.345 gives totals of 0.7175 and 0.635, a gain of 0.7175 − 0.635 = 0.0825 QALYs.

6. Compare the decisions. With an incremental cost of GBP 2,000 per patient, the ICER is 2,000 / 0.0575 = GBP 34,783 per QALY under value set A and 2,000 / 0.0825 = GBP 24,242 per QALY under value set B. The gain is 0.025 QALYs, or about 43%, larger under B, since 0.025 / 0.0575 = 0.435.

7. Interpretation. The answers did not change, yet the ICER moved from the top of the GBP 25,000 to GBP 35,000 per QALY range used in PMG36 section 4.7.15 to below it. The valuation sample's preferences decide how much the gains in usual activities and mood are worth, which is why NICE fixes one instrument and one value set for the reference case.

How indirect and direct values compare

Arnold and colleagues reviewed studies in which the same patients gave direct utilities and completed a generic questionnaire scored with existing tariffs, and concluded that indirect methods give consistently lower utilities. Since death is fixed at zero, lower utilities leave more room for gains in non-fatal conditions, so they argue indirect methods favour moving resources away from treatments that prevent or delay death.

Limits and common misreadings

The convenience of the method makes several errors easy to commit. Most overlook how much the utility depends on the instrument and value set as well as on the answers.

1. Treating a questionnaire profile as a utility. A state such as 2-1-2-3-1 has no value until a value set is applied.
2. Mixing instruments across model states. PMG36 section 4.3.5 warns that results from different instruments cannot always be compared.
3. Applying another country's value set. Value sets differ significantly between countries, and NICE asks for UK general population values.
4. Ignoring mapping uncertainty. Mapped utilities are statistical predictions, and PMG51 asks for sensitivity analyses on the mapping model.
5. Ignoring baseline imbalance. A lower starting utility in one arm shrinks the unadjusted QALY gain, as step 4 shows.

Sources

• Arnold D, Girling A, Stevens A, Lilford R. Comparison of direct and indirect methods of estimating health state utilities for resource allocation: review and empirical analysis. BMJ. 2009;339:b2688. https://doi.org/10.1136/bmj.b2688
• Brazier JE, Longworth L. NICE DSU Technical Support Document 8: An introduction to the measurement and valuation of health for NICE submissions. Decision Support Unit; August 2011. https://doi.org/10.15131/shef.data.33738301
• Brazier JE, Rowen D. NICE DSU Technical Support Document 11: Alternatives to EQ-5D for generating health state utility values. Decision Support Unit; March 2011. https://doi.org/10.15131/shef.data.33744721
• Dolan P. Modeling valuations for EuroQol health states. Medical Care. 1997;35(11):1095-1108. https://doi.org/10.1097/00005650-199711000-00002
• Manca A, Hawkins N, Sculpher MJ. Estimating mean QALYs in trial-based cost-effectiveness analysis: the importance of controlling for baseline utility. Health Economics. 2005;14(5):487-496. https://doi.org/10.1002/hec.944
• National Institute for Health and Care Excellence. Interim methods statement: implementing the EQ-5D-5L value set (PMG51). Published 27 August 2026. Sections 1.5, 1.10, 1.11 and 1.13, appendix A and changes to figure 4.1. https://www.nice.org.uk/process/pmg51
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). Published 31 January 2022, last updated 13 October 2026. Sections 4.3.2 to 4.3.5, 4.3.10, 4.3.14 and 4.7.15. https://www.nice.org.uk/process/pmg36
• Rowen D, Mukuria C, Bray N, et al. A United Kingdom value set for the EQ-5D-5L. Value in Health. 2026;29(5):858-869. https://doi.org/10.1016/j.jval.2026.03.008
• Szende A, Janssen B, Cabases J, eds. Self-Reported Population Health: An International Perspective based on EQ-5D. Springer; 2014. Chapter 2, table 2.2. https://doi.org/10.1007/978-94-007-7596-1