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
KEEPsAKE2 compared risankizumab, an interleukin-23 inhibitor, with placebo in adults with active psoriatic arthritis (PsA). Its primary endpoint was ACR20 response at Week 24, analyzed as a stratified risk difference with a missing-data strategy adapted to the COVID-19 pandemic, which overlapped the trial's final months.
| Feature | KEEPsAKE2 |
|---|---|
| Phase | Phase 3 |
| Population | Active psoriatic arthritis, including participants with a history of inadequate response or intolerance to biologic therapy |
| Design | Randomized, parallel-group, double-blind, placebo-controlled |
| Arms | Risankizumab 150 mg vs placebo |
| Enrollment | 444 |
| Primary endpoint | Percentage of participants with ACR20 response at Week 24 |
| Primary analysis | Cochran-Mantel-Haenszel risk difference, stratified; Full analysis set with NRI-C |
| Dates | Start 2019-03-07; primary completion 2020-06-22 |
| ClinicalTrials.gov | NCT03671148 |
| Sponsor | AbbVie (industry) |
2. Clinical Question
KEEPsAKE2 asked whether risankizumab increases the proportion of patients achieving at least a 20% improvement in PsA disease activity (ACR20) at Week 24, compared with placebo, in a population that included patients previously exposed to biologics.
Population
Adults with active PsA, including those with inadequate response or intolerance to one or more biologic therapies.
Intervention
Risankizumab 150 mg.
Comparator
Placebo.
Primary question
Is the Week 24 ACR20 response rate higher with risankizumab than with placebo (superiority)?
3. Trial Design
Risankizumab 150 mg
- IL-23 inhibitor, biological
- Double-blind
Placebo
- Matching placebo, biological
- Double-blind
4. Stratification and Analysis Populations
Randomization was stratified by three baseline factors, and every posted CMH and MMRM analysis adjusts for the same factors:
- Current use of csDMARD (0 vs ≥ 1)
- Number of prior biologic therapies (0 vs ≥ 1)
- Extent of psoriasis (≥ 3% vs < 3% body surface area) at baseline
The biologic-exposure stratum is especially important: patients who have already failed biologics typically respond less, so balancing them across arms and adjusting for them in the analysis protects both validity and precision.
| Population | Used for |
|---|---|
| Full analysis set (FAS) | Primary ACR20 and most secondary endpoints |
| FAS with baseline psoriasis BSA ≥ 3% | PASI 90 response |
| FAS with baseline LEI > 0 | Resolution of enthesitis |
| FAS with baseline LDI > 0 | Resolution of dactylitis |
| Safety population | Serious adverse events: 224 risankizumab, 219 placebo |
5. Endpoints
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Primary: Percentage of participants with ACR20 response at Week 24 | Response requires (1) ≥ 20% improvement in 68-tender joint count; (2) ≥ 20% improvement in 66-swollen joint count; and (3) ≥ 20% improvement in at least 3 of 5: physician global assessment of disease activity, patient global assessment of disease activity, patient assessment of pain, HAQ-DI, and hsCRP | Baseline and Week 24 |
ACR20 is a composite responder definition: a continuous, multi-domain disease state is converted into a yes/no outcome. This makes the endpoint easy to communicate but discards information about the degree of improvement, which is why ACR50, ACR70 and continuous measures are reported as secondary endpoints.
6. Primary Result: ACR20 at Week 24
Response rate difference (risankizumab − placebo)
95% CI: 15.9 to 33.0 (two-sided) · P < 0.001
Stratified Cochran-Mantel-Haenszel · Full analysis set · NRI-C
What the estimate means. After adjusting for the three stratification factors, the proportion of participants achieving ACR20 at Week 24 was an estimated 24.5 percentage points higher with risankizumab than with placebo. Because it is an absolute difference, it can be read directly: roughly one additional ACR20 responder for every four patients treated, relative to placebo.
What it does not mean. It is not a 24.5% relative increase in response, and it does not identify which patients responded because of risankizumab rather than spontaneous fluctuation or placebo effects. ACR20 is also a modest threshold; it says nothing about the depth of response.
Precision. The 95% CI of 15.9–33.0 points excludes zero; even the lower limit represents a difference of about 16 percentage points. The interval spans roughly 17 points, reflecting the binary outcome and a trial of 444 participants.
P-value vs effect size. P < 0.001 indicates the data are incompatible with equal response rates. It says nothing about whether a 24.5-point difference is clinically sufficient; that depends on the risk difference and its interval.
Cautions. The estimate depends on the NRI-C missing-data strategy: participants with missing data are generally counted as non-responders, except missing data due to COVID-19, which were handled by multiple imputation. Non-responder imputation tends to be conservative when dropout is more common on placebo, but it defines a composite estimand (response without dropout) rather than the effect had everyone remained on study. The registry does not report per-arm response rates.
7. Secondary Endpoint Results
All posted secondary comparisons test superiority. Binary endpoints are summarized as CMH-adjusted response rate differences with NRI-C; continuous endpoints are summarized as MMRM least-squares mean differences.
Binary endpoints (CMH risk differences, percentage points)
| Endpoint | Population | Difference | 95% CI | P-value |
|---|---|---|---|---|
| ACR20 at Week 16 | FAS | 22.6 | 13.9 to 31.2 | <0.001 |
| ACR50 at Week 24 | FAS | 16.6 | 9.7 to 23.6 | <0.001 |
| ACR70 at Week 24 | FAS | 6 | 0.8 to 11.3 | 0.024 |
| Minimal Disease Activity (MDA) at Week 24 | FAS | 14 | 7.0 to 21.0 | <0.001 |
| PASI 90 at Week 24 | FAS, baseline BSA ≥ 3% | 44.3 | 33.9 to 54.6 | <0.001 |
| Resolution of enthesitis at Week 24 | FAS, baseline LEI > 0 | 13.8 | 3.5 to 24.2 | 0.009 |
| Resolution of dactylitis at Week 24 | FAS, baseline LDI > 0 | 38.8 | 22.9 to 54.8 | <0.001 |
The ACR20 → ACR50 → ACR70 gradient (24.5, 16.6, 6.0 points) is typical of responder hierarchies: stricter thresholds are met by fewer participants in both arms, so absolute differences shrink even when the relative effect is maintained. The ACR70 interval (0.8 to 11.3) nearly reaches zero, indicating that the evidence for a difference in high-level response is much weaker than for ACR20.
The skin endpoint (PASI 90) and dactylitis resolution show the largest differences, but they are estimated in subsets of the FAS (participants with psoriasis BSA ≥ 3% or dactylitis at baseline), so their intervals are wider and they apply to different populations than the primary endpoint.
Continuous endpoints (MMRM least-squares mean differences)
| Endpoint | Direction favoring risankizumab | LS mean difference | 95% CI | P-value |
|---|---|---|---|---|
| Change in HAQ-DI at Week 24 | Negative (less disability) | −0.16 | −0.26 to −0.07 | <0.001 |
| Change in SF-36 PCS at Week 24 | Positive (better physical health) | 3.86 | 2.41 to 5.31 | <0.001 |
| Change in FACIT-Fatigue at Week 24 | Positive (less fatigue) | 2.2 | 0.6 to 3.9 | 0.009 |
These analyses used observed data up to Week 24, excluding values after rescue medication or initiation of concomitant PsA medication that could meaningfully affect efficacy. That is a hypothetical-type strategy: it estimates the effect had those treatments not been started, relying on MMRM's missing-at-random assumption for the excluded data. It is therefore a different estimand from the NRI-C approach used for binary endpoints.
8. Safety: Serious Adverse Events
| Measure | Risankizumab 150 mg | Placebo |
|---|---|---|
| Participants with ≥1 serious adverse event | 9 / 224 | 12 / 219 |
Serious adverse events were uncommon and numerically similar between arms. With so few events, the trial has little power to detect differences in serious harms, and absence of a difference is not evidence of equal safety. No formal statistical comparison of SAEs is posted, and rare or delayed adverse events require larger or longer datasets to characterize.
9. Statistical Methodology
Cochran-Mantel-Haenszel risk difference
The CMH approach estimates a common treatment difference across strata defined by the stratification factors. Each stratum contributes its own risk difference, and these are combined with weights that reflect the stratum's size and balance between arms.
p̂R,k and p̂P,k are response proportions in stratum k for risankizumab and placebo.
Non-responder imputation with COVID-19 multiple imputation (NRI-C)
Under standard NRI, any participant without an evaluable response is classified as a non-responder. KEEPsAKE2 overlapped the onset of the COVID-19 pandemic, when visits were missed for reasons unrelated to treatment. Counting those as failures would bias toward the null and conflate pandemic disruption with lack of efficacy. NRI-C therefore used multiple imputation for missing data attributable to COVID-19 while retaining non-responder imputation for other missing data.
MMRM for continuous endpoints
Treatment, visit, treatment-by-visit interaction and stratification factors as fixed effects, baseline value as covariate; within-participant residuals correlated across visits.
Intention-to-treat
The full analysis set analyzes participants by randomized assignment. Endpoint-specific subsets (psoriasis BSA ≥ 3%, LEI > 0, LDI > 0) are defined by baseline characteristics and so remain randomized comparisons within those subsets.
10. Statistical Methods Explained
Why was a CMH test used rather than a simple chi-square test?
Randomization was stratified on three factors. The CMH test respects this stratification by comparing arms within strata and combining the results. A simple chi-square test would ignore the design and could be less efficient, particularly because prior biologic exposure strongly influences response.
Why report a risk difference rather than an odds ratio?
The response rate difference is on the absolute scale that clinicians and patients use: a 24.5-point increase in the proportion of responders. It also leads directly to a number-needed-to-treat interpretation, which odds ratios do not.
What is NRI-C, and why did COVID-19 matter?
Non-responder imputation treats missing outcomes as failures, a conservative choice when dropout reflects poor efficacy or tolerability. Missing visits caused by the pandemic do not carry that meaning, so multiple imputation was applied to them instead, avoiding an artificial dilution of the treatment effect.
Why do binary and continuous endpoints handle missing data differently?
Binary endpoints use NRI-C, in which dropout and rescue count against response (a composite strategy). Continuous endpoints use MMRM on observed data, excluding values after rescue, which targets a hypothetical estimand under missing-at-random assumptions. The two frameworks answer slightly different questions, so their results should not be combined as if they were on the same footing.
Why does ACR70 show a much smaller difference than ACR20?
ACR70 requires a 70% improvement, which few participants in either arm achieve. Absolute differences are bounded by how common the outcome is, so a smaller risk difference at a stricter threshold is expected and is not in itself evidence of a weaker relative effect.
How should a HAQ-DI difference of −0.16 be read?
It is the adjusted average difference in change from baseline in disability score between arms at Week 24. The 95% CI (−0.26 to −0.07) excludes zero, but whether −0.16 is clinically important depends on accepted minimal important differences for HAQ-DI, which the registry does not state.
11. Limitations and Interpretation Issues
- Missing-data assumptions: NRI-C and MMRM target different estimands, and both rely on assumptions (non-response for dropouts; missing at random for imputed or excluded data) that cannot be verified from the data.
- Pandemic disruption: the primary completion date of 2020-06-22 places the final assessments in the early COVID-19 period, a source of missing visits that required a special imputation rule.
- Per-arm rates not shown: the posted analyses report differences but not each arm's response rate, so relative effects cannot be judged from the record alone.
- Subset endpoints: PASI 90, enthesitis and dactylitis are analyzed in baseline-defined subsets and apply to those patients only.
- Multiplicity: the testing hierarchy is not described in the registry.
- Short placebo-controlled period: the randomized comparison is limited to 24 weeks; durability and long-term safety require longer follow-up.
- Safety precision: with 9 and 12 participants with SAEs, differences in serious harms cannot be reliably estimated.
12. Why This Trial Matters Statistically
KEEPsAKE2 is a clear example of a rheumatology responder trial analyzed on the absolute scale, with stratification carried into the analysis and a missing-data rule adapted to a real-world disruption.
| Concept | How it appears in KEEPsAKE2 |
|---|---|
| Cochran-Mantel-Haenszel test | Stratified risk differences for ACR20, ACR50, ACR70, MDA, PASI 90, enthesitis and dactylitis |
| Covariate adjustment | csDMARD use, prior biologics and psoriasis extent in every model |
| Multiple imputation / missing data | NRI-C for COVID-19-related missing data |
| MMRM | HAQ-DI, SF-36 PCS, FACIT-Fatigue change from baseline |
| Intention-to-treat | Full analysis set for efficacy |
| Confidence intervals | Narrow for ACR20; near zero at the lower limit for ACR70 |
| Estimands | Composite (NRI-C) vs hypothetical (MMRM excluding post-rescue data) |
13. Related Tutorials
Learn more about the methods used in this trial:
14. Related Calculators
15. Sources
- ClinicalTrials.gov: NCT03671148 — A Study Comparing Risankizumab to Placebo in Participants With Active Psoriatic Arthritis Including Those Who Have a History of Inadequate Response or Intolerance to Biologic Therapy(Ies).
- PubMed: PMID 34815219.
- PubMed: PMID 38739215.
- PubMed: PMID 39120849.
- PubMed: PMID 41028616.
- PubMed: PMID 41318485.
Continue exploring clinical trial statistics
Connect each trial's endpoints and methods to a deeper statistical tutorial, then to an appropriate calculator.
16. Summary
KEEPsAKE2 enrolled 444 adults with active PsA. The primary stratified CMH analysis showed an ACR20 response rate difference at Week 24 of 24.5 percentage points in favor of risankizumab (95% CI 15.9–33.0; P < 0.001), using NRI-C for missing data. Secondary responder, skin, enthesitis, dactylitis and patient-reported outcomes pointed in the same direction, with smaller and less precise differences at stricter response thresholds. A careful reading weighs the absolute risk differences, their confidence intervals, the estimand implied by each missing-data strategy, and the multiplicity of secondary comparisons.