This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
GEMINI III was a randomized, parallel-group, triple-masked phase 3 trial evaluating whether vedolizumab increased the proportion of participants in clinical remission at Week 6 compared with placebo in moderate to severe Crohn's disease, with the primary endpoint assessed in the subpopulation of participants who had previously failed a TNFα antagonist.
| Feature | GEMINI III |
|---|---|
| Phase | Phase 3 |
| Condition | Crohn's disease (moderate to severe) |
| Design | Randomized, parallel-group, triple-masked, placebo-controlled |
| Primary purpose | Treatment |
| Interventions | Vedolizumab (drug); placebo |
| Enrollment | 416 |
| Primary endpoint | Percentage of participants in clinical remission (CDAI ≤ 150) at Week 6 in the TNFα antagonist failure subpopulation |
| Primary hypothesis | Superiority |
| Status | Completed; results posted |
| Study dates | Start November 2010; primary completion February 2012 |
| Lead sponsor | Millennium Pharmaceuticals, Inc. (industry) |
| ClinicalTrials.gov | NCT01224171 |
2. Clinical Question
The central question was whether vedolizumab, compared with placebo, would increase the proportion of patients reaching clinical remission at Week 6 among people with moderate to severe Crohn's disease whose disease had already failed treatment with a TNFα antagonist. This is a clinically distinct and typically harder-to-treat group, and the trial was built around it as the primary analysis population.
Population
Participants with moderate to severe Crohn's disease; the primary analysis was restricted to those meeting the TNFα antagonist failure criterion.
Intervention
Vedolizumab, administered under triple masking.
Comparator
Matching placebo.
Primary question
Is the adjusted percentage in clinical remission at Week 6 higher with vedolizumab than with placebo in the TNFα antagonist failure subpopulation?
3. Trial Design
Vedolizumab
- Active drug: vedolizumab
- Masked to participants, care providers and investigators/assessors (triple masking)
- Primary assessment of clinical remission at Week 6
Placebo
- Matching placebo
- Same masking and assessment schedule
- Primary assessment of clinical remission at Week 6
The two safety denominators (209 and 207) sum to the 416 enrolled participants. The primary efficacy comparison, however, was made only in the TNFα antagonist failure subpopulation, and the registry fields summarized here do not state how many participants in each arm belonged to that subpopulation.
4. Endpoints
| Endpoint | Definition (registry) | Time frame |
|---|---|---|
| Primary: Percentage of participants in clinical remission in the TNFα antagonist failure subpopulation | Clinical remission defined as a Crohn's Disease Activity Index (CDAI) score ≤ 150 points. Participants who prematurely discontinued for any reason were considered as not achieving clinical remission. | Week 6 |
How the CDAI is built
The CDAI quantifies Crohn's disease symptoms using eight factors, each summed after adjustment with a weighting factor. The total ranges from 0 to approximately 600, with higher scores indicating greater disease activity. The registry lists the components as:
- Number of liquid or soft stools each day for 7 days
- Abdominal pain (graded 0–3 on severity) each day for 7 days
- General well-being, subjectively assessed from 0 (well) to 4 (terrible) each day for 7 days
- Presence of complications
- Taking Lomotil or opiates for diarrhea
- Presence of an abdominal mass (0 none, 2 questionable, 5 definite)
- Hematocrit of < 0.47 in men and < 0.42 in women
- Percentage deviation from standard weight
Because remission is defined as falling at or below a threshold on a continuous score, the endpoint is binary: each participant either is or is not in remission at Week 6. This dichotomization determines the analysis approach: a comparison of proportions rather than of mean scores.
5. Primary Endpoint Results
The registry reports one formal statistical analysis for the primary endpoint: a stratified Cochran-Mantel-Haenszel (CMH) comparison of vedolizumab and placebo in the TNFα antagonist failure intent-to-treat subpopulation, with the effect expressed as an adjusted risk difference.
Adjusted risk difference in Week 6 clinical remission
Percentage points (vedolizumab minus placebo) · 95% CI: −4.5 to 10.5 (two-sided)
CMH chi-square test P = 0.4332 · Superiority hypothesis at a 5% significance level
| Element | Registry-reported value |
|---|---|
| Outcome measure | Percentage of participants in clinical remission, TNFα antagonist failure subpopulation |
| Time frame | Week 6 |
| Analysis population | TNFα antagonist failure ITT subpopulation: all randomized participants who received any amount of blinded study drug and met the TNFα antagonist failure criterion |
| Statistical method | Cochran-Mantel-Haenszel |
| Effect measure | Risk difference (adjusted percent vedolizumab − adjusted percent placebo) |
| Estimate | 3 |
| 95% CI (two-sided) | −4.5 to 10.5 |
| P-value | 0.4332 |
| Stratification factors | Concomitant oral corticosteroids; concomitant immunomodulators (6-MP, azathioprine, or methotrexate) |
What the estimate means. After adjustment for the two stratification factors, the estimated percentage of participants in clinical remission at Week 6 was 3.0 percentage points higher with vedolizumab than with placebo in the TNFα antagonist failure subpopulation. This is an absolute difference in proportions, not a relative change: a 3.0 point difference does not mean remission was "3% more likely" in relative terms.
What it does not mean. The result is not evidence that vedolizumab has no effect in this population. A non-significant superiority test indicates that the observed data are compatible with no difference; it does not demonstrate equivalence. The trial did not test a non-inferiority or equivalence hypothesis, so no margin-based conclusion is available.
What the confidence interval says. The 95% CI of −4.5 to 10.5 percentage points spans zero. Values ranging from a modest disadvantage for vedolizumab to a benefit of roughly ten percentage points are all reasonably compatible with the data. The width of the interval reflects the precision available in the subpopulation at Week 6.
Why the p-value is not an effect size. P = 0.4332 describes how surprising a stratum-adjusted difference at least this large would be if the true difference were zero. It does not quantify how large the benefit is or how likely vedolizumab is to work. The estimate and interval carry that information.
Cautions. Missing Week 6 assessments due to premature discontinuation were counted as non-remission, a conservative rule that applies equally to both arms but can attenuate differences. The endpoint was assessed at a single early time point; the registry record summarized here does not report later time points for this measure.
6. Serious Adverse Events
The registry reports serious adverse events (SAEs) by arm as the number of participants affected over the number at risk.
| Safety measure | Vedolizumab | Placebo |
|---|---|---|
| Participants with ≥1 serious adverse event (affected / at risk) | 13 / 209 | 16 / 207 |
The counts of participants with serious adverse events were similar in the two arms, with slightly fewer in the vedolizumab arm. No formal statistical comparison of SAE rates is reported, and SAE tables in registries are descriptive: the trial was not powered to detect differences in individual or aggregate safety events, and absence of a difference in a short study of this size does not rule out less common harms.
7. Statistical Methodology
Cochran-Mantel-Haenszel test
The primary endpoint was tested using the CMH chi-square test at a 5% significance level. The CMH test compares a binary outcome between two groups while conditioning on strata. Within each stratum (here defined by concomitant corticosteroid use and concomitant immunomodulator use) a 2×2 table of treatment by remission is formed, and evidence of association is pooled across strata.
where ak is the number of vedolizumab-treated participants in remission in stratum k, and E(ak) and Var(ak) are its expectation and variance under no treatment effect, given the stratum margins. Under the null hypothesis, QCMH approximately follows a chi-square distribution with 1 degree of freedom.
Adjusted risk difference
The registry defines the effect measure as the adjusted percent in remission with vedolizumab minus the adjusted percent with placebo, adjusted for the stratification factors. A stratum-adjusted risk difference is typically a weighted average of stratum-specific differences, with weights (for example, CMH-type weights) that give more influence to larger, more informative strata.
where p̂V,k and p̂P,k are remission proportions for vedolizumab and placebo in stratum k, and wk is the stratum weight.
Analysis population
The primary analysis used a modified intent-to-treat subpopulation: all randomized participants who received any amount of blinded study drug and met the TNFα antagonist failure criterion. Two features matter. First, requiring receipt of study drug is a common modification of strict ITT that is usually minimally consequential when masking is maintained. Second, restricting to a baseline-defined subgroup preserves randomization within that subgroup, because TNFα antagonist failure status was determined before treatment assignment could influence it.
Handling of missing data
Participants who prematurely discontinued for any reason were classified as not achieving clinical remission (a non-responder imputation). This converts missing outcomes into failures, which avoids excluding participants and aligns the estimand with "remission while remaining on study through Week 6."
Hypothesis framework
The registry lists the hypothesis as superiority, tested two-sided at the 5% level. With a single registered primary endpoint and a single reported primary analysis, no multiplicity adjustment across primary endpoints was required for this comparison.
8. Statistical Methods Explained
Why was a Cochran-Mantel-Haenszel test used rather than a simple chi-square test?
Randomization was balanced within levels of concomitant corticosteroid and immunomodulator use, and those therapies can affect remission on their own. The CMH test compares arms within each stratum and then pools, so the treatment comparison is not distorted by any chance imbalance in background therapy and the analysis mirrors the design. An unstratified chi-square test ignores this structure and is typically somewhat less efficient.
What does a risk difference of 3.0 mean in practical terms?
It means that, on the adjusted scale, about 3 more participants per 100 would be expected to be in remission at Week 6 with vedolizumab than with placebo in the TNFα antagonist failure subpopulation. Because the interval runs from −4.5 to 10.5, the data also accommodate 4.5 fewer or 10.5 more per 100.
Why does the confidence interval crossing zero matter?
For a risk difference, zero is the value of no effect. A two-sided 95% CI that contains zero corresponds to a two-sided test that does not reject the null hypothesis at the 5% level, which is consistent with the reported P = 0.4332. The interval and the p-value tell the same story from different angles, but only the interval conveys the plausible size of the effect.
Why is non-responder imputation considered conservative?
Participants who leave the study early are counted as treatment failures regardless of why they left. If discontinuation is related to lack of efficacy, this is realistic; if it happens for unrelated reasons, it reduces the observed remission percentage in both arms and tends to shrink the difference toward zero. Either way, it does not credit either arm with outcomes that were never observed.
Why is the primary analysis restricted to a subpopulation of the enrolled participants?
The clinical question concerned patients with prior TNFα antagonist failure. Because that status is a baseline characteristic, comparing the randomized arms within it remains a randomized comparison. The trade-off is that the effective sample size is smaller than the full enrollment of 416, which widens the confidence interval.
Is a non-significant result the same as showing that the treatments are equivalent?
No. Equivalence would require a prespecified margin and a confidence interval lying entirely within it. Here the interval extends to 10.5 percentage points on the favorable side, so a clinically meaningful benefit is not excluded, and a small disadvantage is not excluded either.
9. Limitations and Interpretation Issues
- Single early time point: the primary endpoint was measured at Week 6. The registered primary endpoint does not address whether remission rates diverge with longer exposure.
- Subpopulation size not stated in the fields summarized here: without arm-level counts for the TNFα antagonist failure subpopulation, readers cannot directly judge how much of the 416-participant enrollment contributed to the primary analysis.
- Arm-level percentages: the adjusted difference is reported, but the fields summarized here do not include the adjusted percentages in each arm, which limits interpretation of the difference relative to the placebo response.
- Dichotomized endpoint: converting the continuous CDAI into a remission threshold of ≤ 150 discards information about the magnitude of symptom change and reduces statistical power relative to an analysis of the continuous score.
- Composite, partly subjective score: the CDAI includes patient-reported items such as abdominal pain and general well-being, which can have high placebo responses; triple masking helps protect against assessment bias.
- Non-responder imputation: conservative handling of discontinuation can attenuate between-arm differences.
- Safety descriptive only: SAE counts are small and not formally compared; the trial was not designed to detect differences in uncommon adverse events.
10. Why This Trial Matters Statistically
GEMINI III is a useful teaching case because it shows a well-designed randomized trial whose primary analysis did not reach statistical significance, and because its analysis links stratified randomization directly to a stratified test and an adjusted absolute effect measure.
| Concept | How it appears in GEMINI III |
|---|---|
| Stratified analysis | CMH test stratified by concomitant corticosteroids and immunomodulators |
| Risk difference | Adjusted absolute difference in remission of 3.0 percentage points |
| Confidence interval | 95% CI −4.5 to 10.5, spanning the null value of 0 |
| P-value | 0.4332 under a two-sided superiority test at the 5% level |
| Intention-to-treat | Modified ITT subpopulation defined by a baseline characteristic |
| Missing data | Premature discontinuation counted as non-remission |
| Binary endpoints | Continuous CDAI dichotomized at ≤ 150 points |
| Absence of evidence | A non-significant superiority result is not evidence of equivalence |
11. Related Tutorials
Learn more about the methods used in this trial:
12. Related Calculators
13. Sources
- ClinicalTrials.gov: NCT01224171 – Study of Vedolizumab in Patients With Moderate to Severe Crohn's Disease.
- PubMed: PMID 30615117
- PubMed: PMID 30203005
- PubMed: PMID 29857145
- PubMed: PMID 26893500
- PubMed: PMID 25996351
Explore more trial statistics
Connect the methods behind each trial to step-by-step tutorials and calculators.
14. Record Summary
GEMINI III compared vedolizumab with placebo in moderate to severe Crohn's disease, with a primary endpoint of Week 6 clinical remission in the TNFα antagonist failure subpopulation. The stratified CMH analysis produced an adjusted risk difference of 3.0 percentage points (95% CI −4.5 to 10.5; P = 0.4332), so superiority was not demonstrated at the 5% level. Serious adverse events occurred in 13 of 209 vedolizumab-treated and 16 of 207 placebo-treated participants. The most informative reading of the trial combines the absolute effect estimate, the width of its confidence interval, and the design choices (early time point, subpopulation analysis, non-responder imputation) that shape what the result can and cannot show.