StatsPlease
Validation Blog Start free
← Back to Blog

Interpretation · ~8 min read

The MCID Was Built to Judge One Patient. Using It to Judge a Trial's Average Effect Is a Different Claim

Statistics for clinical researchers and surgical trainees

Published

In short

The minimal clinically important difference was defined and validated to answer whether one patient's change would matter to them. Applying that same threshold to a trial's group average, or to classify individual patients as "responders" against a group-derived number, is a more contested move than it looks. It has a documented failure mode, and it is also a widely used practice with genuine tradeoffs on both sides. The difference between doing this correctly and doing it carelessly comes down to which threshold, derived which way, gets applied to which question.

What the original definition actually validated

Jaeschke, Singer, and Guyatt's 1989 definition of the MCID is specific about what it measures: the smallest change a patient would perceive as beneficial, small enough that continuing a treatment is still worth it absent excessive cost or side effects.1 That is a claim about one person's experience, anchored to that person's own report of whether they felt better. A 2025 perspective on MCID in Alzheimer's disease trials states the implication of this directly: the MCID is not an alternative trial outcome, and it does not apply to group measures at all. It exists to determine whether one patient has crossed a threshold, not whether a group's average did.2

A separate 2025 critique catalogues specific ways this gets misapplied when a threshold built from individual data gets pointed at a group's average treatment effect instead: it ignores the role of time in progressive diseases, it does not transfer cleanly from the open-label study where it was often derived to the blinded trial where it is applied, it becomes unreliable in small trials, and it gets murky when patients in the placebo arm also cross the threshold at meaningful rates.3

Where "responder" classification actually fails

The sharper, more specific problem shows up in responder analysis, where a group-level MCID or MIC threshold gets used to sort individual patients into "responder" and "non-responder" categories. A 2021 paper found this produces systematically over-optimistic conclusions: patients who have not meaningfully changed get classified as responders, because the threshold used to draw that line was never built to separate real individual change from measurement noise in the first place. That is a different question, with a different answer: the MDC, not the MIC.4

The recommended fix is to use individual-level statistics purpose-built for that comparison, such as the reliable change index, rather than repurposing a group-derived threshold to do a job it was not validated for.

The one-sentence version: a threshold that was validated to answer "is this change big enough to matter?" cannot also answer "is this change bigger than my instrument's measurement error?" Those are two questions, and a responder rate built on the first one silently assumes an answer to the second.

The other side: why responder analysis persists anyway

None of this makes responder analysis simply wrong, and treating it that way would be unfair to a genuinely useful, widely endorsed tool. A widely cited 2007 paper in Trials lays out the honest tradeoff: dichotomizing a continuous outcome into responders and non-responders costs real statistical efficiency and does not fully solve the problem it is meant to solve, but it also does something a bare group-mean comparison does not. It translates a trial's result into a number clinicians and patients can actually reason about, "most patients on this treatment felt meaningfully better" rather than a mean difference on a scale nobody outside the field has intuition for.5

That trade-off is why responder analysis persists in regulated settings rather than being quietly dropped: it does a communicative job a group mean cannot do on its own.5 If you are planning one for a submission, check the current patient-reported outcome guidance from the relevant regulator directly, and prespecify your responder definition rather than choosing it after seeing the data.

How this actually gets done correctly

The two uses do not have to conflict, and well-designed trials already keep them separate rather than picking one. The pattern to copy, which you will find spelled out in the more careful published statistical analysis plans (heart failure trials are a good place to look for worked examples), is this: the group-level threshold used to judge whether the trial's average effect was clinically meaningful is derived by a distribution-based method specifically for that comparison, while a separate, anchor-based threshold is used for individual-patient responder classification. The plan says outright that these are different numbers answering different questions, and that the individual-level threshold is expected to be larger than the group-level one. That is the model: two thresholds, two purposes, neither one substituting for the other.

Two questions, two thresholds. A conceptual summary of the distinction drawn in this article, not a computed result.
  Group-level claim Individual-level claim
The question Was the trial's average effect large enough to matter? Did this patient change enough to matter?
Unit being judged A mean difference between arms One patient's change score
Typical derivation Distribution-based, chosen for that comparison Anchor-based, validated against patient report
Must clear measurement error? No, the mean averages error down Yes, which is the MDC question, not the MIC question
Relative size The smaller of the two Expected to be larger
What it licenses you to write "The average effect exceeded the threshold" "This proportion of patients improved meaningfully"

What to check in your own manuscript

If you are reporting that a group's average change exceeded a published MCID, that is a defensible descriptive claim about the average, not a claim about how many individual patients benefited, and it should not be phrased as one.

If you are reporting a responder rate, the threshold behind it should come from individual-level validation, not a group-derived number borrowed because it was the only threshold available for that outcome measure. Ideally it should account for measurement error the way a reliable change index does, which a raw MIC or MCID threshold does not.

Naming which kind of threshold you used, and why it matches the question you are actually asking, closes the gap this article is about in one sentence.

This is also the part of a manuscript where a chatbot is least useful and most confident. Ask ChatGPT for "the MCID" of a given instrument and it will hand you a single number with no indication of whether it was anchor-based or distribution-based, whether it was validated on individual change or on between-group difference, or which of those two questions you were about to attach it to. The threshold is a literature-sourced judgement you have to defend in your methods section. The change scores you compare it against should be computed from your data, not generated.

Frequently asked questions

Does the MCID validly apply to a clinical trial's group average treatment effect?

This is genuinely contested. The MCID was originally defined and validated to judge whether an individual patient's change mattered to them; recent methodological critiques argue it was not built to be applied directly to a group's average difference in a trial.

Why can responder analysis produce misleading results?

When a group-level threshold, derived from a different validation process, is used to classify individual patients as responders, it can systematically misclassify patients who have not meaningfully changed as responders, because the threshold wasn't built to separate real change from measurement noise at the individual level.

Why is responder analysis still used if it costs statistical efficiency?

Because it translates a trial result into something clinicians and patients can reason about, which a mean difference on an unfamiliar scale does not. Dichotomising a continuous outcome does cost real statistical efficiency, so the trade-off is deliberate rather than free. If you are planning one for a regulatory submission, check the current patient-reported outcome guidance directly and prespecify the responder definition.

Try it yourself

Separate the two claims: in SPSS or StatsPlease

Take any trial-shaped dataset with a baseline score, a follow-up score, and an allocation column. The exercise is to produce the group-level claim and the individual-level distribution as two separate outputs, so that neither one gets quietly used to support the other.

In SPSS

  1. Build the change score: Transform → Compute Variable, target change, expression follow_up - baseline.
  2. Group-level claim: Analyze → Compare Means → Independent-Samples T Test. Put change in Test Variable(s) and arm in Grouping Variable. Read the mean difference and its 95% confidence interval; compare that interval, not just the point estimate, against your group-level threshold.
  3. Individual-level distribution: Analyze → Descriptive Statistics → Explore, with change as Dependent and arm as Factor. Under Plots tick Histogram. This is the spread the group mean hides.
  4. Only if you have an individually validated threshold: Transform → Recode into Different Variables to flag change above it, then Analyze → Descriptive Statistics → Crosstabs against arm for the responder rate. SPSS will not warn you that the threshold you typed in was derived for a different question.

In StatsPlease

  1. Upload the CSV.
  2. Open the Analysis tab.
  3. For the group-level claim, set change as the Outcome and arm as the Group, then press the Group Comparison preset button. There is no separate run step: pressing the preset is what runs the analysis. The engine screens normality and routes to the appropriate test itself, so you pick the preset, not the test.
  4. For the individual-level view, set the baseline and follow-up columns as the paired variables and press the Before vs After preset button. Again, pressing it runs the analysis.
  5. Read both on the Results tab: the mean difference with its confidence interval and effect size for the group claim, and the distribution of within-patient change for the individual claim. AMA by default, or APA if you switch styles.

Compare: both tools return the same mean difference and the same confidence interval, because both run the same test on the same numbers. The difference is what you are handed alongside it. Neither tool can tell you whether the threshold you brought to the table was validated for the question you are asking; that judgement stays yours, which is exactly why it belongs in your methods section rather than in a chatbot's answer.

You might also read

References

  1. Jaeschke R, Singer J, Guyatt GH. Measurement of health status. Ascertaining the minimal clinically important difference. Controlled Clinical Trials. 1989;10(4):407–415. https://doi.org/10.1016/0197-2456(89)90005-6
  2. Cummings J. Perspective: minimal clinically important difference (MCID) and Alzheimer's disease clinical trials. Alzheimer's & Dementia: Translational Research & Clinical Interventions. 2025;11:e70059. https://doi.org/10.1002/trc2.70059
  3. Ganju J. Why "minimal clinically important difference" for interpreting the magnitude of the treatment effect is not useful. Pharmaceutical Statistics. 2025;24:e70015. https://doi.org/10.1002/pst.70015
  4. Hays RD, Peipert JD. Between-group minimally important change versus individual treatment responders. Quality of Life Research. 2021;30(10):2765–2772. https://doi.org/10.1007/s11136-021-02897-z
  5. Snapinn SM, Jiang Q. Responder analyses and the assessment of a clinically relevant treatment effect. Trials. 2007;8:31. https://doi.org/10.1186/1745-6215-8-31

A responder claim is only as good as the threshold behind it.

StatsPlease reports both the group-level effect size and the individual-level distribution of change from your own uploaded data, so a responder claim in your manuscript can be built on a threshold actually validated for that question. Set your Outcome and Group and press Group Comparison, or set your paired columns and press Before vs After. Pressing the preset is what runs the analysis; the numbers are computed from your data, not generated.

Try StatsPlease free