SF-12

SF-12 in Health Outcomes Measurement and Cost-Utility Analysis

Many trials and population surveys record health status with the SF-12 alone, which leaves a health economist who needs QALYs with summary scores that were never designed as utilities. Ware, Kosinski and Keller built the SF-12 by selecting twelve SF-36 questions that reproduce the long form's physical and mental summary scores in a questionnaire a third the length. Those two summaries describe health-related quality of life on a scale anchored on the US general population, with no anchor at full health and none at death, so they cannot be multiplied by time to give QALYs. Utilities come from SF-12 data by one of two routes, the six-dimension SF-6D classification that Brazier and Roberts derived from the twelve items and valued with the standard gamble, or a published mapping function that predicts EQ-5D values from SF-12 answers.

How the twelve items were chosen from the SF-36

The SF-12 was built by statistical reduction of a longer instrument, so the way the items were picked shapes what the short form can and cannot do. Ware, Kosinski and Keller used regression methods in the general US population, 2,333 people, to select twelve SF-36 items and score them so that they reproduced the SF-36 Physical Component Summary (PCS) and Mental Component Summary (MCS).

The items selected by multiple regression explained most of the variance in the long-form summaries. The authors reported multiple $R^2$ values of 0.911 for the PCS and 0.918 for the MCS. When the scoring algorithms built in the general population were cross-validated in the Medical Outcomes Study, they reached $R^2$ values of 0.905 against the SF-36 PCS and 0.938 against the SF-36 MCS. The design was therefore aimed at the two summaries, and the short form was never meant to reproduce every SF-36 scale with the same precision.

The developer, QualityMetric, describes the current SF-12v2 as a 12-item abridged version of the SF-36v2 that covers the same eight health domains with one or two questions per domain. The domains are physical functioning, role limitations due to physical health (role-physical), bodily pain, general health, vitality, social functioning, role limitations due to emotional problems (role-emotional) and mental health.

Physical and mental component summary scores

The SF-12 is usually reported as two numbers, PCS-12 and MCS-12, and the meaning of a change in either depends on how they are scaled. The summaries are norm-based. Hayes and colleagues, describing the SF-12v2 they used in a 2017 validation study, reported that each component is scored with a mean of 50 and that higher scores mean better health. Under the developer's standard US norm-based scoring, a respondent with a PCS-12 below 50 therefore reports worse physical health than the average of the US reference population.

Ware, Kosinski and Keller tested the summaries in their 1996 paper. Two-week test-retest reliability correlations in 232 members of the general US population were 0.89 for PCS-12 and 0.76 for MCS-12. They also compared the short form with the best SF-36 scale in validity tests published earlier for the long form. In 14 tests involving physical criteria, relative validity estimates for PCS-12 ranged from 0.43 to 0.93 with a median of 0.67, and in 6 tests involving mental criteria the estimates for MCS-12 ranged from 0.60 to 1.07 with a median of 0.97. The mental summary lost less ground against the long form than the physical summary did.

These results describe the SF-12 as a profile of group health. For an economic evaluation a difference in PCS-12 between trial arms says that one group reports better physical functioning than the other, but the PCS-12 scale has no exchange rate against length of life, so a summary score difference cannot be used as a QALY gain.

Version 2 and its recall periods

Two versions of the SF-12 are in use, and the version matters both for comparing studies and for later utility scoring. The second version, SF-12v2, is the one the developer currently supplies, as the abridged form of the SF-36v2.

QualityMetric lists four-week and one-week recall forms of the adult SF-12v2. Studies that used different recall periods are not strictly comparable, since a shorter recall can capture a different part of a fluctuating condition.

The response options also changed. In the second versions of the SF-36 and the SF-12, the role-physical and role-emotional items have five response categories, where the original SF-6D model expects two. Mukuria and colleagues, in their 2025 review of the SF-6D, note that SF-6D scoring of version 2 data requires the five categories for role-physical and role-emotional to be collapsed into two levels. Information recorded by the respondent is lost in that step.

SF-6D utilities derived from SF-12 responses

The SF-6D is the route by which SF-12 data reach a cost-utility analysis without any further questionnaire. The SF-6D is a health state classification with a value set, so every pattern of answers maps to a single health state utility value on a scale where full health scores 1.

Brazier and Roberts set out to derive a preference-based measure of health from the SF-12 for use in economic evaluation and to compare it with the original SF-36 preference-based index. They reduced the SF-12 items to a six-dimension classification that defines 7,500 health states. A sample of 241 of those states was valued by 611 members of the UK general population using the standard gamble, and regression modelling of those valuations gave values for every state in the classification. The SF-6D dimensions, as Mukuria and colleagues describe them, are physical functioning, role limitations, social functioning, pain, mental health and vitality.

Brazier and Roberts reported in 2004 that the coefficients of the SF-12 model were similar to those of the SF-36 version, with a similar fit. The model over-predicted the value of the poorest health states, and the authors recommended a final model in which some inconsistent estimates were merged. Mukuria and colleagues list the same two problems for the first version of the SF-6D in general, non-monotonicity, where some better states receive lower values than worse ones, and systematic prediction errors such as overestimation of the values of poor health states.

The SF-6D therefore gives an SF-12 study a route to QALYs, with known features that affect results. Over-predicting poor states raises utilities at the severe end, which narrows the gap between severe and mild states and can shrink the QALY gain from a treatment that moves people out of poor health. Mukuria and colleagues describe the SF-6Dv2 as a response to criticisms of the first version, and report that recent SF-6D valuations use online discrete choice experiments where the original used standard gamble interviews.

Mapping SF-12 responses to EQ-5D utilities

A health economist working to a reference case that names the EQ-5D may need EQ-5D values when only the SF-12 was collected. A crosswalk or mapping study estimates the statistical relationship between the two instruments in a dataset that contains both, and then applies that relationship to the study that collected only the SF-12.

Gray, Rivero-Arias and Clarke published one such mapping in 2006. Instead of predicting the EQ-5D utility directly, they used multinomial logit regression to estimate the probability of each EQ-5D response level from SF-12 item responses and summary scores, then used Monte Carlo simulation to generate predicted EQ-5D responses that were converted into utility scores. They compared this response mapping with direct mapping to utility scores by ordinary least squares. Their equations were estimated on 12,967 adult survey responses from the 2000 US Medical Expenditure Panel Survey, and out-of-sample testing used 13,304 adults from the 1996 Health Survey for England.

Response mapping keeps the EQ-5D descriptive system in the prediction, so a different value set can be applied to the predicted responses later. A mapping still predicts an average relationship, and individual predictions carry error that a probabilistic analysis should reflect.

Illustrative worked example of QALYs from SF-6D scores in a trial

The numbers in this example are illustrative and do not come from any study. A trial records the SF-12 at baseline, six months and twelve months, and each response is scored with the SF-6D. Mean SF-6D utilities in the treatment arm are 0.62, 0.68 and 0.70, and in the control arm 0.62, 0.64 and 0.65. The treatment adds GBP 1,300 to costs over the year.

QALYs over the year are the area under each utility curve, found by the trapezium rule.

$$Q = \sum_{t=1}^{T} \frac{u_{t-1} + u_t}{2} \times \Delta t$$

where $Q$ is the QALYs accrued over follow-up, $u_t$ is the mean utility at assessment $t$, $u_{t-1}$ is the mean utility at the previous assessment, $\Delta t$ is the time between the two assessments in years and $T$ is the number of intervals.

1. Treatment arm. The first interval gives (0.62 + 0.68) / 2 = 0.65, and 0.65 × 0.5 = 0.325. The second gives (0.68 + 0.70) / 2 = 0.69, and 0.69 × 0.5 = 0.345. The total is 0.325 + 0.345 = 0.67 QALYs.

2. Control arm. The first interval gives (0.62 + 0.64) / 2 = 0.63, and 0.63 × 0.5 = 0.315. The second gives (0.64 + 0.65) / 2 = 0.645, and 0.645 × 0.5 = 0.3225. The total is 0.315 + 0.3225 = 0.6375 QALYs.

3. Incremental QALYs and cost per QALY. The QALY gain is 0.67 − 0.6375 = 0.0325, and the incremental cost-effectiveness ratio is 1,300 / 0.0325 = 40,000, so the treatment costs GBP 40,000 per QALY gained on SF-6D values.

4. A different utility source. Suppose that EQ-5D values mapped from the same SF-12 answers gave a QALY gain of 0.05. The ratio would then be 1,300 / 0.05 = 26,000, or GBP 26,000 per QALY.

The same questionnaire responses lead to two ratios that differ by GBP 14,000 per QALY, purely through the choice of utility route. The SF-6D compresses differences at the severe end of the scale, so a treatment whose benefit lies in moving people out of poor health can look less cost-effective on SF-6D values than on mapped EQ-5D values. A sensitivity analysis that reports both routes shows the decision maker how much the result depends on that choice.

How the SF-12 differs from the SF-36, SF-6D and EQ-5D

The SF-12 sits among several related instruments, and confusing them leads to errors in both data handling and reporting. The SF-36 is the parent, the SF-6D and SF-6Dv2 are scoring systems, and the EQ-5D is a rival instrument.

The SF-12 is a patient-reported outcome measure and a generic instrument, since its questions apply across diseases. The SF-12 is a non-preference-based measure in its own right because PCS-12 and MCS-12 are standardised against population norms and not valued by the public. The preference weights belong to the SF-6D, which is a separate scoring system applied to the same answers.

Limitations and common misreadings

Several errors recur when SF-12 data are used in economic evaluation. Some come from treating the summary scores as more than they are, and others from the scoring rules.

• Summary scores used as utilities. PCS-12 and MCS-12 are centred on 50 for the US population and have no anchor at death, so neither score, nor any rescaling of either to a 0 to 1 range, is a utility.
• Correlated physical and mental health. Lau and colleagues tested the SF-12v2 in 6,126 Singapore residents in 2021 with confirmatory factor analysis and found a two-factor structure in which physical and mental health were correlated. Because the traditional scoring method treats the two components as uncorrelated, they advised that algorithm-derived PCS and MCS scores be interpreted with caution.
• Precision at the individual level. The two-week test-retest correlation of 0.76 for MCS-12 reported by Ware, Kosinski and Keller leaves room for noise in individual change scores, and the instrument was designed to reproduce summaries for groups.
• Version mixing. Pooling SF-12 and SF-12v2 data, or scoring version 2 data with the original SF-6D model without collapsing the role items, mixes response formats.
• Compression at the severe end. The SF-6D from the SF-12 over-predicts the value of the poorest states, which reduces the measured gain for treatments that help the most severely ill.
• Mapping treated as observation. Predicted EQ-5D values from a mapping such as that of Gray, Rivero-Arias and Clarke carry prediction error and depend on the population in which the mapping was estimated.

Sources

• Brazier JE, Roberts J. The estimation of a preference-based measure of health from the SF-12. Medical Care. 2004;42(9):851-859. doi:10.1097/01.mlr.0000135827.18610.0d
• Gray AM, Rivero-Arias O, Clarke PM. Estimating the association between SF-12 responses and EQ-5D utility values by response mapping. Medical Decision Making. 2006;26(1):18-29. doi:10.1177/0272989X05284108
• Hayes CJ, Bhandari NR, Kathe N, Payakachat N. Reliability and validity of the Medical Outcomes Study Short Form-12 Version 2 (SF-12v2) in adults with non-cancer pain. Healthcare. 2017;5(2):22. doi:10.3390/healthcare5020022
• Lau JH, Abdin E, Vaingankar JA, Shafie S, Sambasivam R, Shahwan S, et al. Confirmatory factor analysis and measurement invariance of the English, Mandarin, and Malay versions of the SF-12v2 within a representative sample of the multi-ethnic Singapore population. Health and Quality of Life Outcomes. 2021;19:80. doi:10.1186/s12955-021-01709-9
• Mukuria C, Rowen D, Mulhern B, McDool E, Kharroubi S, Bjorner JB, Brazier JE. The Short Form 6 Dimensions (SF-6D): development and evolution. Applied Health Economics and Health Policy. 2025;23(1):19-33. doi:10.1007/s40258-024-00919-8
• QualityMetric. SF-12v2 Health Survey. Developer product page, accessed 17 October 2026. https://qualitymetric.com/health-surveys/the-sf-12v2-health-survey
• Ware J Jr, Kosinski M, Keller SD. A 12-Item Short-Form Health Survey: construction of scales and preliminary tests of reliability and validity. Medical Care. 1996;34(3):220-233. doi:10.1097/00005650-199603000-00003