Minimal Clinically Important Difference

Minimal Clinically Important Difference in Anchor-Based Estimation, Responder Analysis, Trials and HTA

A trial can show a statistically significant gain on a quality of life questionnaire that patients would barely notice, or a modest mean gain that still matters to many of them. The minimal clinically important difference (MCID) is the yardstick used to tell these cases apart, and the preferred estimates link questionnaire scores to an external anchor such as a patient's own rating of change. The idea dates from 1989 and has gathered competing names, and the anchor-based methods are shown below with an illustrative mean-change estimate and ROC cut-off, followed by responder analysis, trial design and the limits of any single number.

From Jaeschke's 1989 definition to the MID

The term comes from Jaeschke, Singer and Guyatt, writing in Controlled Clinical Trials in 1989, when quality of life instruments were becoming primary outcomes in randomised trials. In the wording quoted by de Vet and colleagues from page 408, it is "the smallest difference in score in the domain of interest which patients perceive as beneficial and which would mandate, in the absence of troublesome side-effects and excessive cost, a change in patient management". Their method compared changes in questionnaire scores with patients' global ratings of change.

In three studies of instruments measuring dyspnoea, fatigue and emotional function in chronic heart and lung disease, the MCID corresponded to a mean change of about 0.5 per item on a seven-point scale, a value the Cochrane Handbook still cites for the Chronic Respiratory Questionnaire. The authors expected the figure to help interpret scores for individuals and groups and to help plan new trials.

MCID, MID, MIC and the meaningful score difference

The concept has gathered several names, and some mark real differences in meaning. Schünemann and Guyatt dropped the "clinically" to shift the focus from the clinical arena to patients' daily experience, and de Vet, Terwee and colleagues separate change within a person from differences between people.

Terwee and colleagues give an example. The difference between patients rating themselves "a little better" and "about the same" is an MID between groups, not a within-person MIC. Here MCID is used for the general concept, and the narrower terms where a source uses them.

Anchors and the global rating of change

An anchor is an external variable, not derived from the instrument being interpreted, for which a meaningful difference is already understood. Anchor-based estimation maps a meaningful step on the anchor onto the score scale of a patient-reported outcome or other clinical outcome assessment. Revicki, Hays, Cella and Sloan recommend patient-rated, clinician-rated and disease-specific anchors as the primary source of estimates.

The most common anchor is the global rating of change, or transition rating, which asks whether the patient is better or worse than before and by how much. Devji and colleagues note that patients find it hard to recall earlier health states, more so as time passes, and the FDA's 2023 draft adds present state bias, where the rating reflects status at follow-up more than the change experienced.

Three checks follow. The anchor should correlate with the change score. Guyatt and colleagues asked for at least a moderate correlation, Terwee and colleagues set a minimum of 0.30, and the credibility instrument of Devji and colleagues treats at least 0.5 as moderate to high. For a transition rating, the correlation with the change score should be appreciably greater than with the follow-up score, or the rating probably reflects current status only. The analyst must also choose which anchor category counts as minimal. De Vet and colleagues show that the estimate can differ by a factor of about 2.5 depending on whether the first category above no change or moderate to much improvement is chosen, and that the choice cannot rest on statistical grounds.

Revicki and colleagues treat the MID together with responsiveness, an instrument's ability to detect change, and de Vet and colleagues frame the link as one question, namely whether an instrument can detect changes as small as its minimal important change.

Mean change and ROC thresholds

Two estimators dominate the anchor-based literature. Both use the same anchor groups but summarise them in different ways. The first, from Jaeschke and colleagues, takes the mean change among patients rating themselves "a little better" as the estimate, written here as $MIC_{mean}$. Subtracting the mean change of patients reporting no change gives the between-group version.

$$MID_{between} = \bar{\Delta}_{LB} - \bar{\Delta}_{S}$$

where $\bar{\Delta}_{LB}$ is the mean change among patients rating themselves "a little better" and $\bar{\Delta}_{S}$ the mean change among those rating themselves "about the same".

The second, as described by de Vet and colleagues, treats the change score as a diagnostic test and the anchor as the reference standard. At each cut-off, sensitivity is the share of anchor-improved patients whose change reaches it and specificity the share of all other patients whose change falls below it. The ROC curve cut-off minimises the sum of the two misclassification proportions.

$$c^{*} = \arg\min_{c} \left[ (1 - Se_c) + (1 - Sp_c) \right]$$

where $c^{*}$ is the chosen threshold, $c$ a candidate cut-off, and $Se_c$ and $Sp_c$ the sensitivity and specificity at that cut-off. This is the same as maximising the Youden index, $Se_c + Sp_c - 1$.

Terwee and colleagues point out that patients who all reported minimal improvement should, in theory, have a mean change above the threshold for minimal improvement, so the mean change method tends to overstate the MIC. They prefer the ROC method or a predictive modelling method, note that the ROC estimate is biased when the share of improved patients is not 50%, and recommend predictive modelling as the best option because it is more precise and can be corrected for that bias. De Vet and colleagues add a stricter 95% limit cut-off, defined as the mean change of the patients the anchor classes as not importantly changed, plus 1.645 standard deviations of their change scores.

Worked example with one anchor and three thresholds

All figures are illustrative. A validation cohort of 200 patients completes a disease-specific questionnaire scored 0 to 100 (higher is better) at baseline and 12 weeks, with a five-category global rating of change. The change score correlates 0.52 with the anchor, passing both the 0.30 and 0.5 criteria.

1. Mean change. The 50 patients in the "a little better" group give $MIC_{mean}$ = 8.0 points, with no reference group.

2. Between-group estimate. The reference group is the 70 patients rating themselves "about the same".

$$MID_{between} = 8.0 - 2.0 = 6.0$$

where 8.0 and 2.0 are the mean changes of the "a little better" and "about the same" groups.

3. ROC cut-off. The 90 patients rating themselves better are classed as improved; the reference group is all 110 others, rating themselves about the same or worse.

At 6 points, 72 of the 90 improved patients reach the cut-off, so sensitivity is 72 / 90 = 0.800, and 94 of the 110 others fall below it, so specificity is 94 / 110 = 0.855. The misclassification sum is 0.200 + 0.145 = 0.345, the smallest in the table, so the ROC threshold is 6 points.

4. 95% limit cut-off. The reference group is the 70 patients rating themselves "about the same", taken here as the not importantly changed group. If their change scores have a standard deviation of 5.0 the cut-off is as follows.

$$c_{95} = 2.0 + 1.645 \times 5.0 = 10.2$$

where $c_{95}$ is the 95% limit cut-off and 1.645 is the one-sided 95% point of the normal distribution.

The same data support thresholds of 6.0, 8.0 and 10.2 points. The mean change estimate exceeds the ROC value, as Terwee and colleagues predict, and the improved share is 45%, not 50%, so a predictive modelling estimate would be a useful check. A report should name the estimator it used.

Because no single anchor or estimator settles the value, Revicki and colleagues recommend triangulating methods, with anchor-based estimates primary and distribution-based MID methods in support.

Responder analysis, from a threshold to a proportion

A threshold turns a continuous score into a yes or no outcome for each patient. The FDA's 2009 guidance derives the evidence for a responder definition from anchor-based methods and suggests that cumulative distribution displays of change in each arm may be preferable to a single categorical definition. The 2023 draft asks sponsors who dichotomise to prespecify and justify one threshold and to explore treatment effects over a range of thresholds.

Guyatt and colleagues described two steps. The first is to establish the smallest change patients consider important on average, and the second is to estimate the proportion of patients who achieve it. Snapinn and Jiang list the costs. Under their normal model a responder analysis needed about 60% more participants than a comparison of means even at the most favourable threshold, and with a large enough sample any small difference in responder rates becomes statistically significant.

Continuing the illustration, suppose a trial finds a mean change of 2 points with control and 6 with treatment, a common standard deviation of 12 and normally distributed changes. The 4-point difference in means falls short of the between-group MID of 6.0, the estimate suited to a difference between arms, yet more patients cross the within-person ROC threshold, which is also 6 points.

$$P_C = 1 - \Phi\left(\frac{6 - 2}{12}\right) = 1 - \Phi(0.333) = 0.369$$

where $P_C$ is the share of control patients reaching the threshold and $\Phi$ is the standard normal cumulative distribution function. With treatment the mean equals the threshold, so $P_T = 0.500$, and the responder difference is $P_T - P_C = 0.500 - 0.369 = 0.131$, so about 13 percentage points more treated than control patients reach it.

MCIDs in trial design, sample size and HTA

Trial designers use an important difference as the target difference in a sample size calculation. The DELTA2 guidance says the target difference for a definitive trial should be important to at least one key stakeholder group, need not be the minimum important value if a larger difference is realistic, and should come from anchor or opinion-seeking methods, not the distribution method. For a continuous outcome and two equal arms, the usual normal approximation is as follows.

$$n = \frac{2 (z_{1-\alpha/2} + z_{1-\beta})^2 \sigma^2}{\delta^2}$$

where $n$ is the number per arm, $z_{1-\alpha/2}$ and $z_{1-\beta}$ are standard normal quantiles for the two-sided significance level and the statistical power, $\sigma$ is the outcome standard deviation and $\delta$ the target difference. Taking the between-group MID of 6.0 as $\delta$, since $\delta$ is a difference between arms, with a standard deviation of 12, 5% significance and 90% power, the calculation is as follows.

$$n = \frac{2 \times (1.96 + 1.2816)^2 \times 12^2}{6^2} = \frac{2 \times 10.508 \times 144}{36} = 84.1$$

where 1.96 and 1.2816 are the two quantiles, so 85 patients per arm are needed. A 4-point target gives 189.1, or 190 per arm, so the chosen difference drives trial size and cost.

Health technology assessment bodies apply their own rules. IQWiG's General Methods (version 8.0, December 2025) note that published MIDs for one instrument can range widely and that an empirically determined MID cannot at present be selected on quality criteria. IQWiG therefore uses prespecified responder analyses whose criterion is at least 15% of the scale range; otherwise it uses the standardised mean difference (Hedges' g), where the whole confidence interval must lie above an irrelevance threshold of 0.2, or a post hoc responder analysis at exactly 15%. On the 0 to 100 scale of the example, a 6-point responder analysis would as a rule not be used. In meta-analysis, the Cochrane Handbook allows results in MID units but warns that a pooled difference below the MID may be read as trivial when a substantial proportion of patients achieved an important benefit.

Why QALYs use the whole difference

The NICE manual (PMG36) gives each health state a utility and multiplies the time spent in each state by that utility, so every utility difference counts towards the QALY gain and no importance threshold enters a cost-utility analysis; the distribution-based MID page works through an example. The threshold still matters indirectly. A model that gives responders and non-responders different health state utility values inherits the responder definition, so moving from the 6-point ROC threshold to the 10.2-point 95% limit changes the responder shares and the QALYs.

Context dependence and the gap between individual and group thresholds

An MCID belongs to an instrument used in a particular population and setting. Revicki and colleagues state that the MID is not an immutable characteristic of an instrument and may vary by population and context. De Vet and colleagues show that it depends on the anchor, the definition of minimal importance, the baseline score and the direction of change, and IQWiG adds disease severity and type of intervention. The FDA's 2023 draft asks sponsors to check whether one value holds across baseline scores and for both improvement and deterioration.

The number also applies at a particular level. Terwee and colleagues treat the MIC as the mean of individual thresholds, so a responder classification can give the right share of responders for a group while misclassifying many individuals, and measurement error adds more. The DELTA2 guidance notes that the difference important to an individual may differ from the value at population level.

Two errors are common in reports. One treats a mean difference below the MCID as proof of no benefit; the other treats an MCID as a fixed property that transfers across instruments, populations and directions of change. Reporting the anchor, its correlation, the estimator and a range of values, with the full distribution of change in each arm, avoids both.

Sources

• Cook JA, Julious SA, Sones W, Hampson LV, Hewitt C, Berlin JA, et al. DELTA2 guidance on choosing the target difference and undertaking and reporting the sample size calculation for a randomised controlled trial. BMJ. 2018;363:k3750.
• 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.
• de Vet HC, Ostelo RW, Terwee CB, van der Roer N, Knol DL, Beckerman H, Boers M, Bouter LM. Minimally important change determined by a visual method integrating an anchor-based and a distribution-based approach. Quality of Life Research. 2007;16(1):131-142.
• Devji T, Carrasco-Labra A, Qasim A, Phillips M, Johnston BC, Devasenapathy N, et al. Evaluating the credibility of anchor based estimates of minimal important differences for patient reported outcomes: instrument development and reliability study. BMJ. 2020;369:m1714.
• 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. Sections II.A.2, III.B and III.C.
• Guyatt GH, Osoba D, Wu AW, Wyrwich KW, Norman GR; Clinical Significance Consensus Meeting Group. Methods to explain the clinical significance of health status measures. Mayo Clinic Proceedings. 2002;77(4):371-383.
• Institute for Quality and Efficiency in Health Care (IQWiG). General Methods. Version 8.0 of 19 December 2025 (English translation). Cologne: IQWiG; 2026. Section 10.3.3.
• Jaeschke R, Singer J, Guyatt GH. Measurement of health status: ascertaining the minimal clinically important difference. Controlled Clinical Trials. 1989;10(4):407-415.
• 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.
• 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, Guyatt GH. Commentary: goodbye M(C)ID! Hello MID, where do you come from? Health Services Research. 2005;40(2):593-597.
• 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.
• Snapinn SM, Jiang Q. Responder analyses and the assessment of a clinically relevant treatment effect. Trials. 2007;8:31.
• Terwee CB, Peipert JD, Chapman R, Lai JS, Terluin B, Cella D, Griffiths P, Mokkink LB. Minimal important change (MIC): a conceptual clarification and systematic review of MIC estimates of PROMIS measures. Quality of Life Research. 2021;30(10):2729-2754.