Distribution-Based MID

Distribution-Based MID for EQ-5D and PRO Scores

Trials that collect the EQ-5D, the SF-6D or a disease-specific questionnaire report changes in scores, and a reader has to judge whether a change of, say, 0.05 on a utility scale matters to patients. Distribution-based methods answer with a yardstick taken from the data, such as half a standard deviation, one standard error of measurement, a benchmark effect size or a reliable change criterion. These yardsticks describe precision and not importance, which is why methodologists and the US Food and Drug Administration (FDA) rank them behind anchor-based estimates, and a minimal important difference has no place in QALY arithmetic. Published EQ-5D and SF-6D figures and an illustrative EQ-5D-5L example are set against them later.

Why statistical yardsticks for important change exist

The Cochrane Handbook describes the minimal important difference (MID) as the smallest change in instrument score that patients perceive as important. The site's parent page uses the older name, minimal clinically important difference, and distribution-based methods are one family of approaches to estimating it. A direct estimate needs evidence of what patients regard as important, usually an external anchor such as a global rating of change.

Anchors are not always available. An instrument may be new, a trial may not have collected a global rating, or patients may know their treatment; the FDA's 2009 guidance on patient-reported outcome measures notes that patient ratings of change are less useful as anchors when patients are not blinded. Distribution-based methods fill the gap with statistics every dataset contains, namely the spread of scores and the reliability of the instrument.

The half standard deviation rule

The best-known estimate comes from a systematic review. Norman, Sloan and Wyrwich found 38 studies that had computed an MID for a health-related quality of life instrument and reported enough data to express it as an effect size, giving 62 estimates. For all but six studies the MID was close to half a standard deviation, with a mean of 0.495 (SD 0.155). The rule is written as follows.

$$MID_{0.5SD} = 0.5 \times SD_0$$

where $MID_{0.5SD}$ is the estimate in the instrument's own units and $SD_0$ is the standard deviation of scores in the sample at baseline.

The pattern held for generic and disease-specific instruments and did not depend on the number of response options. The authors linked it to evidence from psychology that people's ability to discriminate across many tasks is limited to about 1 part in 7, very close to half a standard deviation. Population-based estimation and brief follow-up gave smaller estimates, acute conditions larger ones, and the conclusion was framed for chronic diseases. The rule is therefore a summary of earlier MIDs re-expressed in standard deviation units. It matches the medium benchmark for Cohen's d, whose page covers those conventions and their caveats.

The standard error of measurement and the one-SEM criterion

The standard error of measurement (SEM) links the spread of scores to the reliability of the instrument. It estimates how far an observed score is likely to stray from a person's true score through measurement error, in the instrument's units. De Vet and colleagues give it as the formula below.

$$SEM = SD \sqrt{1 - R}$$

where $SEM$ is the standard error of measurement, $SD$ is the standard deviation of scores in the sample and $R$ is a reliability coefficient. Cronbach's alpha, a measure of internal consistency reliability, can be used, but de Vet and colleagues regard test-retest reliability in a stable population as more appropriate for change between two time points.

Wyrwich and colleagues proposed one SEM as a criterion for meaningful individual change. In 605 cardiac outpatients, one-SEM changes corresponded well to patient-derived MCID standards on the Chronic Heart Failure Questionnaire, and in 471 outpatients with chronic obstructive pulmonary disease one SEM closely approximated the standards for the Chronic Respiratory Disease Questionnaire. De Vet and colleagues summarise a later comparison in which one set of studies, on musculoskeletal disorders and using receiver operating characteristic methods, gave anchor-based values of 2.3 or 2.6 SEM. They suggest the gap may be partly explained by that set's definition of important change as a larger change on the anchor, and conclude that one SEM equalling the minimally important change is not a universal truth. By the formula, one SEM equals half a standard deviation when reliability is 0.75.

Effect-size benchmarks and the reliable change index

Two further tools express change in standardised units. The effect size used in responsiveness work divides mean change by the baseline standard deviation, and the standardised response mean divides it by the standard deviation of change scores. Benchmarks for standardised effects are covered on the Cohen's d page. The FDA's 2009 guidance notes that distribution-based methods can categorise changes as small, moderate and large.

The reliable change index of Jacobson and Truax asks whether one person's change exceeds what measurement error would plausibly produce. Blampied sets it out as follows.

$$RC = \frac{x_2 - x_1}{S_{diff}}, \quad S_{diff} = \sqrt{2 \, SEM^2}$$

where $RC$ is the reliable change index, $x_1$ and $x_2$ are the person's scores at the two measurements, $S_{diff}$ is the standard error of the difference between two scores and $SEM$ is the standard error of measurement. A value beyond 1.96 in either direction would arise from measurement error alone with probability of 5% or less. De Vet and colleagues express the same threshold in score units as the minimally detectable change.

$$MDC = 1.96 \times \sqrt{2} \times SEM$$

where $MDC$ is the minimally detectable change, 1.96 comes from the 95% confidence interval of no change and $\sqrt{2}$ reflects the two measurements in a change score.

Precision is not importance

Every method above is built from variability and reliability, so none contains information about what patients value. De Vet and colleagues call this the major disadvantage of distribution-based methods, since they do not, in themselves, indicate the importance of an observed change. They argue that half a standard deviation and the SEM correspond more to minimally detectable than to minimally important change.

The estimates also move with the sample. A more heterogeneous population has a larger standard deviation and so a larger half-SD estimate, although patients' perception of change is unaltered. Revicki and colleagues note that an MID is not an immutable characteristic of an instrument and may vary by population and context. Two misreadings follow. One is quoting half a standard deviation as a fixed property of an instrument, and the other is applying an individual-level threshold such as the minimally detectable change to a difference in trial means.

Methodological and regulatory guidance putting anchor-based estimates first

Published recommendations agree on the order of evidence. Revicki, Hays, Cella and Sloan recommend triangulating several approaches, with anchor-based methods using patient-rated, clinician-rated and disease-specific variables as the primary estimates, and distribution-based methods in support or where anchor-based estimates are unavailable.

The FDA's December 2009 guidance on patient-reported outcome measures to support labelling claims says the empirical evidence for any responder definition is derived using anchor-based methods, and that distribution-based methods are supportive and not appropriate as the sole basis for one. The fourth guidance in the newer patient-focused drug development (PFDD) series, issued in draft in April 2023 and listed as a draft at the time of writing, in October 2026, uses the term meaningful score difference for the difference between two clinical outcome assessment scores that patients would view as meaningful. It states that distribution-based methods do not directly consider the patient voice and are insufficient as the sole basis for that difference, though they give helpful information about measurement variability, and it recommends multiple anchors. The draft says that, when final, the PFDD series will replace the 2009 guidance.

Published MID estimates for the EQ-5D and SF-6D

Utility instruments have their own MID literature, although the figures below are anchor-based and instrument-defined estimates, given here as comparators for distribution-based values and not as examples of them. Walters and Brazier used question 2 of the SF-36, a rating of general health compared with the last assessment, as the anchor in eight longitudinal studies covering 11 patient groups, treating patients who rated themselves somewhat better or somewhat worse as having changed by the MID.

The mean MID was 0.041 for the SF-6D (range 0.011 to 0.097) and 0.074 for the EQ-5D-3L (range -0.011 to 0.140). Standardised response means, a distribution-based check, averaged 0.39 for the SF-6D, mainly small to moderate by Cohen's criteria, which the authors read as supporting the MID results, and 0.24 for the EQ-5D-3L. The authors conclude that the two MIDs are not equal and differ in absolute values, the EQ-5D-3L figure being almost double, so an MID for one utility instrument cannot be carried over to another. The two are proportionally equivalent only relative to each scale's range, since the EQ-5D-3L scale (-0.59 to 1.00) spans about twice the range of the SF-6D (0.29 to 1.00).

For the EQ-5D-5L, McClure and colleagues used an approach they call instrument-defined, neither anchor-based nor distribution-based, which is a simulation of single-level transitions in the descriptive system. Across scoring algorithms from six countries the estimates were generally between 0.037 and 0.069, with 0.063 for the England algorithm, or 0.037 when the maximum-valued scoring parameters were excluded; the authors attribute the differences to population preferences, valuation techniques and modelling strategies.

Worked example with five yardsticks for one EQ-5D-5L dataset

All figures in this example are illustrative. A trial population has a baseline EQ-5D-5L index standard deviation of 0.20 and a test-retest reliability of 0.80. An anchor-based analysis finds that patients rating themselves "a little better" improved by 0.05 on average.

1. Half a standard deviation.

$$MID_{0.5SD} = 0.5 \times 0.20 = 0.10$$

where 0.20 is the baseline standard deviation of index scores.

2. Cohen's small benchmark.

$$ES_{0.2} = 0.2 \times 0.20 = 0.04$$

where $ES_{0.2}$ is the change equal to an effect size of 0.2 against the baseline standard deviation.

3. Standard error of measurement.

$$SEM = 0.20 \times \sqrt{1 - 0.80} = 0.20 \times 0.447 = 0.089$$

where 0.80 is the test-retest reliability.

4. Individual detectable change. The standard error of the difference is $\sqrt{2} \times 0.089 = 0.126$, which gives the detectable change below.

$$MDC = 1.96 \times 0.126 = 0.247$$

where $MDC$ is the minimally detectable change at 95% confidence. For a patient whose index rises from 0.60 to 0.75, the reliable change index is as follows.

$$RC = \frac{0.15}{0.126} = 1.19$$

where 0.15 is the patient's change and 0.126 the standard error of the difference. The value is below 1.96, so this change is not reliable at the 5% level.

5. Compare the yardsticks.

The figures span about a sixfold range (0.04 to 0.247), and the half-SD estimate is twice the anchor-based one. An analyst who adopted 0.10 would call a 0.06 improvement unimportant, although patients reporting themselves a little better changed by less. The anchor-based figure describes change within patients, so applying it to a difference between trial arms is a further judgement that should be stated.

What an MID does and does not do in cost-effectiveness analysis

A QALY uses the whole utility difference, however small. The NICE manual (PMG36) gives each health state a utility and multiplies the time spent in each state by that utility; no importance threshold enters. Suppose, illustratively, that a treatment with an incremental cost of £1,200 raises mean utility by 0.03 for two years (undiscounted).

$$ICER = \frac{1200}{0.03 \times 2} = \frac{1200}{0.06} = 20000$$

where $ICER$ is the incremental cost-effectiveness ratio in pounds per QALY gained, 0.03 is the mean utility gain per year and 2 is the number of years. The gain is below every group-level yardstick in the worked example, yet the result is £20,000 per QALY. Setting sub-MID differences to zero would discard health gain from a cost-utility analysis, and a gain above the MID can still be poor value if the cost is high.

MIDs matter elsewhere. They help readers judge patient-reported results reported alongside an evaluation, and they set responder definitions. A model that gives responders and non-responders different health state utility values inherits the threshold, so choosing a distribution-based or anchor-based estimate changes the responder proportion and the QALYs. They also feed the sample size calculation for trials designed to detect a minimal important change in quality of life.

Distribution-based MID and nearby measures

Several terms in this area describe different quantities on the same scale. The table separates them by what each rests on.

Sources

• Blampied NM. Reliable change and the reliable change index: still useful after all these years? The Cognitive Behaviour Therapist. 2022;15:e50.
• Cohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. Lawrence Erlbaum Associates; 1988. Section 2.2.3.
• de Vet HC, Terwee CB, Ostelo RW, Beckerman H, Knol DL, Bouter LM. Minimal changes in health status questionnaires: distinction between minimally detectable change and minimally important change. Health and Quality of Life Outcomes. 2006;4:54.
• Food and Drug Administration. Guidance for Industry. Patient-Reported Outcome Measures: Use in Medical Product Development to Support Labeling Claims. Silver Spring, MD: FDA; December 2009. Section IV.E.
• Food and Drug Administration. Patient-Focused Drug Development: Incorporating Clinical Outcome Assessments Into Endpoints for Regulatory Decision-Making. Draft guidance (PFDD Guidance 4). Silver Spring, MD: FDA; April 2023. Section III.B.1.
• Jacobson NS, Truax P. Clinical significance: a statistical approach to defining meaningful change in psychotherapy research. Journal of Consulting and Clinical Psychology. 1991;59(1):12-19.
• McClure NS, Sayah FA, Xie F, Luo N, Johnson JA. Instrument-defined estimates of the minimally important difference for EQ-5D-5L index scores. Value in Health. 2017;20(4):644-650.
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). London: NICE; 2022, last updated 13 October 2026. Section 4.3.
• Norman GR, Sloan JA, Wyrwich KW. Interpretation of changes in health-related quality of life: the remarkable universality of half a standard deviation. Medical Care. 2003;41(5):582-592.
• Revicki D, Hays RD, Cella D, Sloan J. Recommended methods for determining responsiveness and minimally important differences for patient-reported outcomes. Journal of Clinical Epidemiology. 2008;61(2):102-109.
• Schünemann HJ, Vist GE, Higgins JPT, Santesso N, Deeks JJ, Glasziou P, Akl EA, Guyatt GH. Chapter 15: Interpreting results and drawing conclusions. In: Cochrane Handbook for Systematic Reviews of Interventions version 6.5. Cochrane; 2024. Section 15.5.
• Walters SJ, Brazier JE. Comparison of the minimally important difference for two health state utility measures: EQ-5D and SF-6D. Quality of Life Research. 2005;14(6):1523-1532.
• Wyrwich KW, Nienaber NA, Tierney WM, Wolinsky FD. Linking clinical relevance and statistical significance in evaluating intra-individual changes in health-related quality of life. Medical Care. 1999;37(5):469-478.
• Wyrwich KW, Tierney WM, Wolinsky FD. Further evidence supporting an SEM-based criterion for identifying meaningful intra-individual changes in health-related quality of life. Journal of Clinical Epidemiology. 1999;52(9):861-873.