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
Canopy-A was a randomized, parallel, quadruple-masked phase 3 trial evaluating canakinumab versus placebo as adjuvant therapy in adults with stages AJCC/UICC v. 8 II-IIIA and IIIB (T>5cm N2) completely resected non-small cell lung cancer.
| Feature | Canopy-A |
|---|---|
| Phase | Phase 3 |
| Status | Terminated |
| Condition | Non-Small Cell Lung Cancer |
| Population | Adults with stages AJCC/UICC v. 8 II-IIIA and IIIB (T>5cm N2) completely resected non-small cell lung cancer |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Quadruple |
| Primary purpose | Treatment |
| Enrollment | 1382 |
| Interventions | Canakinumab and placebo |
| Primary endpoint | Disease Free Survival (DFS) by Local Investigator |
| Primary endpoint type | Time-to-event |
| Primary hypothesis | Superiority |
| Lead sponsor | Novartis Pharmaceuticals |
| Sponsor type | Industry |
2. Clinical Question
The primary statistical question was whether adjuvant canakinumab improved disease-free survival compared with placebo in adults with completely resected non-small cell lung cancer meeting the registered stage criteria.
Population
Adults with stages AJCC/UICC v. 8 II-IIIA and IIIB (T>5cm N2) completely resected non-small cell lung cancer.
Intervention
Canakinumab.
Comparator
Placebo.
Primary question
Does canakinumab improve disease-free survival relative to placebo?
3. Trial Design
Trial timeline
Study start
The registered study start date was 2018-03-16.
Primary completion
The registered primary completion date was 2022-03-17.
Terminated
The ClinicalTrials.gov record lists the study status as terminated.
4. Endpoints
The registered primary endpoint was a time-to-event measure assessed by the local investigator.
| Endpoint | Time frame | Definition |
|---|---|---|
| Disease Free Survival (DFS) by Local Investigator | Up to approximately 4 years | DFS is the time from the date of randomization to the date of the first documented NSCLC disease recurrence as assessed by local investigator radiologically or death due to any cause. Disease recurrence included diagnoses of new primary lung malignancies. Clinical deterioration was not considered as a recurrence of disease. |
In cases of non-conclusive radiological evidence, a biopsy assessment was performed to confirm NSCLC recurrence. The median DFS was estimated using the Kaplan-Meier method. DFS was censored if no DFS event was observed before the analysis cut-off date or if a participant received subsequent anti-neoplastic therapy for NSCLC. The censoring date was the date of last assessment before the cut-off date or the NSCLC-related anti-neoplastic therapy date.
5. Statistical Methodology
Primary analysis population
The primary analysis used the Full Analysis Set (FAS), defined in the registry analysis as including all participants to whom study treatment was assigned by randomization. This is closely aligned with the intention-to-treat principle because treatment assignment, rather than treatment received or adherence, determines the primary efficacy comparison.
Stratified log-rank test
The reported hypothesis test was a stratified log-rank test. This is a standard method for comparing two time-to-event distributions while accounting for prespecified stratification factors. The registry identifies stratified analysis as a concept in the primary analysis.
Stratified Cox model
The hazard ratio for DFS and its 95% confidence interval were calculated from a stratified Cox model using the same stratification factors as those for the log-rank test.
For a hazard ratio below 1, the model estimates a lower instantaneous rate of the event in the canakinumab group than in the placebo group. A hazard ratio of exactly 1 would indicate equal estimated hazards under the model.
Superiority framework
The primary hypothesis type was superiority. The reported 95% confidence interval was two-sided. Thus, the statistical question was whether the observed time-to-event distributions provided evidence of a difference favoring either direction, with the trial's stated hypothesis concerning superiority of canakinumab.
Kaplan-Meier estimation
The registry states that median DFS was estimated using the Kaplan-Meier method. Kaplan-Meier estimation is appropriate for time-to-event outcomes because participants can be followed for different lengths of time and some observations may be censored before an event occurs.
The Kaplan-Meier estimator updates the estimated event-free probability at observed event times while retaining participants who are censored in the risk set up to their censoring time.
6. Primary Result: Disease-Free Survival
The primary endpoint was Disease Free Survival (DFS) by Local Investigator, assessed for up to approximately 4 years. The analysis used the Full Analysis Set and compared canakinumab with placebo using a stratified log-rank test.
Primary DFS hazard ratio
95% CI: 0.78–1.14 · P = 0.258
Analysis: Full Analysis Set; stratified log-rank test; stratified Cox model for the hazard ratio.
| Primary endpoint | Canakinumab vs placebo | 95% CI | P-value | Analysis |
|---|---|---|---|---|
| Disease Free Survival (DFS) by Local Investigator | HR 0.94 | 0.78–1.14 | 0.258 | Stratified log-rank test; stratified Cox model |
The hazard ratio of 0.94 means that the estimated instantaneous rate of a DFS event under the fitted model was 0.94 times the corresponding rate under placebo. Expressed descriptively, this is an estimated hazard ratio 6% below 1, but that relative interpretation should not be confused with a 6% absolute improvement in disease-free survival.
The estimate does not mean that 6% more participants were disease-free, that an individual participant's risk fell by 6%, or that the treatment prevented recurrence in a fixed proportion of patients.
The 95% confidence interval of 0.78–1.14 expresses uncertainty around the estimated hazard ratio. It includes 1, so the interval is compatible with a range of underlying hazard ratios that includes no difference between groups. The width of the interval also shows that the point estimate alone should not be treated as a highly precise description of the treatment effect.
The P-value of 0.258 is a measure of how compatible the observed data are with the statistical null hypothesis under the specified testing framework. It is not a measure of the size of the treatment effect, the probability that the null hypothesis is true, or the probability that the treatment has no clinical effect.
Because DFS is a censored time-to-event endpoint, interpretation also depends on the censoring rules and on the assumptions underlying the Cox model. In particular, a hazard ratio is a model-based relative measure and should not automatically be interpreted as a constant risk ratio at every time point.
7. Secondary DFS Analyses: PD-L1 Subgroups
The registry also reports DFS analyses in participants with a valid baseline measurement of PD-L1 expression. These analyses used the same general survival-analysis framework: a stratified log-rank test and a stratified Cox model using the same stratification factors as the log-rank test.
| PD-L1 subgroup | Hazard ratio | 95% CI | P-value |
|---|---|---|---|
| PD-L1 <1% | 1.09 | 0.76–1.58 | 0.676 |
| PD-L1 ≥1% and <49% | 0.60 | 0.34–1.05 | 0.036 |
| PD-L1 ≥50% | 1.33 | 0.73–2.43 | 0.823 |
How to read the subgroup estimates
The subgroup hazard ratios describe the canakinumab-versus-placebo DFS comparison within each specified PD-L1 category. They should not be interpreted as direct evidence that the treatment effect differs between categories merely because the point estimates differ.
The PD-L1 <1% analysis produced an HR of 1.09 with a 95% CI of 0.76–1.58. The estimate is above 1, but the confidence interval spans 1, indicating substantial uncertainty about the direction and magnitude of the underlying effect.
The PD-L1 ≥1% and <49% analysis produced an HR of 0.60 with a 95% CI of 0.34–1.05. The point estimate is below 1, while the confidence interval extends slightly above 1. The associated P-value of 0.036 is a hypothesis-test result within this subgroup analysis; it does not establish that the subgroup has a different treatment effect from the other PD-L1 categories.
The PD-L1 ≥50% analysis produced an HR of 1.33 with a 95% CI of 0.73–2.43. The relatively wide interval illustrates the additional uncertainty that can arise when an overall randomized comparison is partitioned into subgroups.
8. Secondary DFS Analyses: CD8 Subgroups
DFS was also analyzed according to baseline CD8 expression among participants with a valid baseline measurement. The registry reports two CD8 categories based on the median.
| CD8 subgroup | Hazard ratio | 95% CI | P-value |
|---|---|---|---|
| CD8 < median | 1.22 | 0.87–1.72 | 0.872 |
| CD8 ≥ median | 0.91 | 0.62–1.33 | 0.303 |
For participants with CD8 below the median, the estimated HR was 1.22, with a 95% CI of 0.87–1.72. For participants with CD8 at or above the median, the estimated HR was 0.91, with a 95% CI of 0.62–1.33.
The difference between the two point estimates should be treated as descriptive. A subgroup with an HR below 1 and another with an HR above 1 does not, by itself, demonstrate a treatment-by-subgroup interaction. Formal effect-modification claims require an appropriate comparison of treatment effects across subgroups.
The confidence intervals for both CD8 categories include 1. The P-values of 0.872 and 0.303 are hypothesis-test quantities and should not be read as estimates of effect size or as probabilities that the treatment works or does not work.
9. What the Hazard Ratio Means in This Trial
The hazard ratio is the central effect measure for all six posted DFS analyses. Understanding what it represents is therefore essential to interpreting Canopy-A.
A value of 0.94 means that the estimated instantaneous event rate in the canakinumab group was 94% of the corresponding estimated event rate in the placebo group under the fitted stratified Cox model.
What HR 0.94 does not mean
- It is not a 6% absolute increase in disease-free survival.
- It does not mean that 6% of participants avoided recurrence.
- It does not mean that every participant experienced a 6% reduction in event risk.
- It is not the same measure as a risk ratio at a particular time point.
- It does not provide a median DFS unless the corresponding survival distribution is separately reported.
Why the confidence interval matters
The primary 95% CI of 0.78–1.14 is more informative than the point estimate alone. It describes the uncertainty surrounding the estimated hazard ratio under the model and sampling framework. Because the interval contains 1, the data are compatible with both a lower and a higher hazard under canakinumab relative to placebo within the interval.
Why the P-value is not an effect-size measure
The primary P-value of 0.258 addresses the statistical test associated with the primary comparison. It does not quantify how clinically important the observed HR is. Effect magnitude is conveyed by the hazard ratio and its confidence interval; the P-value addresses statistical evidence under the specified null hypothesis.
10. Censoring and the Definition of DFS
Canopy-A's DFS definition makes censoring particularly important. Participants were censored when no DFS event had been observed before the analysis cut-off date or when they received subsequent anti-neoplastic therapy for NSCLC. The censoring date was based on the last assessment before the cut-off date or the date of NSCLC-related anti-neoplastic therapy.
Why censoring is needed
A participant who has not experienced recurrence or death by the end of available follow-up has not reported a complete event time. Survival analysis uses the participant's observed follow-up without treating the observation as if an event had occurred.
Why censoring is consequential
The statistical interpretation depends on how censoring relates to future event risk. Informative censoring can affect the validity of a time-to-event comparison.
Subsequent therapy
The registry's DFS definition specifies censoring when subsequent anti-neoplastic therapy for NSCLC was received, making treatment after randomization part of the endpoint's censoring rules.
Radiologic confirmation
When radiological evidence was non-conclusive, biopsy assessment was used to confirm NSCLC recurrence.
11. Statistical Methods Explained
Why was a stratified log-rank test used?
A log-rank test compares the experience of two groups over time rather than comparing only a single endpoint proportion. Stratification allows the comparison to account for the trial's stratified analysis framework. In Canopy-A, the registry specifically identifies the stratified log-rank test as the reported method for the primary DFS analysis and each posted subgroup analysis.
Why was a Cox model used in addition to the log-rank test?
The log-rank test supplies the hypothesis test, while the Cox model provides an interpretable relative effect estimate: the hazard ratio, together with its confidence interval. The registry states that the DFS HR and its 95% confidence interval were calculated from a stratified Cox model using the same stratification factors as the log-rank test.
What does HR 0.94 mean?
It means the estimated instantaneous DFS-event hazard in the canakinumab group was 0.94 times the corresponding estimated hazard in the placebo group under the fitted model. It does not mean that disease recurrence was reduced by 6% for every participant or that the absolute probability of remaining disease-free increased by 6 percentage points.
Why does the 95% confidence interval matter?
The interval communicates precision. The primary interval of 0.78–1.14 spans 1, so the point estimate of 0.94 should not be interpreted without acknowledging the uncertainty around it. A confidence interval is not a prediction interval for individual patients and does not describe the range of individual treatment effects.
Why is DFS analyzed with Kaplan-Meier methods?
DFS is a time-to-event endpoint, and participants can have different lengths of follow-up or be censored before an event is observed. Kaplan-Meier estimation accommodates this structure and provides an estimate of the probability of remaining event-free over time.
Why does the analysis population matter?
The primary analysis used the Full Analysis Set, including all participants to whom study treatment was assigned by randomization. Analyzing participants according to randomized assignment preserves the principal comparison created by randomization and avoids redefining the treatment groups based on later treatment exposure.
Why should subgroup HRs not be compared by eye?
Separate subgroup estimates have their own sampling variability. An HR of 0.60 in one subgroup and 1.09 in another does not automatically establish that the treatment effect differs between those populations. A formal interaction or heterogeneity analysis is required for that type of conclusion.
12. Secondary Endpoint Results Summary
The posted statistical analyses include five secondary DFS subgroup analyses: three based on PD-L1 expression and two based on CD8 expression.
| Endpoint subgroup | HR | 95% CI | P-value | Analysis population |
|---|---|---|---|---|
| PD-L1 <1% | 1.09 | 0.76–1.58 | 0.676 | Randomized participants with valid baseline PD-L1 measurement |
| PD-L1 ≥1% and <49% | 0.60 | 0.34–1.05 | 0.036 | Randomized participants with valid baseline PD-L1 measurement |
| PD-L1 ≥50% | 1.33 | 0.73–2.43 | 0.823 | Randomized participants with valid baseline PD-L1 measurement |
| CD8 < median | 1.22 | 0.87–1.72 | 0.872 | Randomized participants with valid baseline CD8 measurement |
| CD8 ≥ median | 0.91 | 0.62–1.33 | 0.303 | Randomized participants with valid baseline CD8 measurement |
All five secondary analyses used the stratified log-rank test, with the hazard ratio and 95% confidence interval obtained from a stratified Cox model. The subgroup analyses therefore share the same basic statistical architecture as the primary DFS analysis while applying it to smaller analysis populations defined by baseline biomarker measurements.
13. Safety
The registry reports serious adverse events by treatment arm for the on-treatment period. These counts provide a separate safety dimension that should not be conflated with the efficacy analysis.
| Arm | Serious adverse events, on-treatment |
|---|---|
| Canakinumab | 141/692 affected/at risk |
| Placebo | 146/689 affected/at risk |
The serious-adverse-event results are reported using affected participants over participants at risk during the on-treatment period. These safety data are therefore distinct from the primary efficacy analysis, which used the Full Analysis Set based on randomized treatment assignment.
The denominators also illustrate why efficacy and safety populations should not automatically be treated as interchangeable. The primary DFS analysis was based on participants assigned by randomization, whereas the serious-adverse-event summary is explicitly reported for the on-treatment period.
14. Intention-to-Treat Analysis and Randomization
Randomization is central to the causal interpretation of the primary DFS comparison. Before treatment exposure and subsequent follow-up occur, random allocation establishes the treatment groups according to the trial design. The primary analysis then retains participants according to their randomized assignment.
For a time-to-event endpoint, intention-to-treat analysis does not eliminate censoring or other statistical considerations. Rather, it preserves the randomized comparison while the survival-analysis methods address differing follow-up and incomplete event observation.
15. Interpreting the Primary Result in Context
The primary DFS estimate should be read as a complete statistical statement rather than as a single number. The estimate was HR 0.94, the 95% confidence interval was 0.78–1.14, and the two-sided P-value was 0.258.
| Component | Primary result | What it tells the reader |
|---|---|---|
| Effect estimate | HR 0.94 | Relative estimated event hazard under the stratified Cox model |
| Precision | 95% CI 0.78–1.14 | Uncertainty around the estimated hazard ratio |
| Hypothesis test | P = 0.258 | Evidence from the specified statistical test against its null hypothesis |
| Endpoint | DFS | Recurrence or death according to the registered definition |
| Population | Full Analysis Set | All participants assigned treatment by randomization |
This presentation prevents a common statistical mistake: treating the P-value as the main result and the effect estimate as secondary. The hazard ratio describes the estimated magnitude and direction of the treatment comparison; the confidence interval describes uncertainty; and the P-value addresses the corresponding hypothesis test.
16. Primary vs Subgroup Inference
The overall primary DFS analysis and the biomarker-defined subgroup analyses answer different questions.
Primary question
What is the randomized comparison of DFS between canakinumab and placebo in the Full Analysis Set?
Subgroup question
What is the DFS comparison within participants meeting a particular PD-L1 or CD8 baseline category?
Precision
Subgroup confidence intervals can be wider because each subgroup contains fewer participants than the full randomized population.
Effect modification
Demonstrating different treatment effects across subgroups requires a formal comparison between subgroup-specific treatment effects.
17. Important Limitations and Interpretation Issues
- Time-to-event interpretation: DFS is not a simple binary endpoint. Follow-up time, censoring, and event timing all contribute to the analysis.
- Censoring rules: Participants could be censored because of the analysis cut-off or subsequent NSCLC-related anti-neoplastic therapy according to the registered definition.
- Hazard-ratio interpretation: A Cox HR is a relative, model-based measure and should not be treated as an absolute risk difference or individual-level treatment effect.
- Confidence interval: The primary 95% CI of 0.78–1.14 spans 1, so the point estimate alone does not capture the uncertainty in the treatment comparison.
- Subgroup analyses: PD-L1 and CD8 analyses apply the survival methodology to biomarker-defined populations and should be interpreted with their greater statistical uncertainty.
- Subgroup comparisons: Differences between subgroup-specific HRs cannot establish treatment-effect heterogeneity without a formal interaction or equivalent comparison.
- Safety population: The reported serious-adverse-event counts use an on-treatment affected/at-risk framework, whereas primary efficacy uses the Full Analysis Set.
- Registry scope: The ClinicalTrials.gov record reports the posted analyses summarized here, but it does not provide every possible descriptive statistic that could be generated from individual participant-level data.
18. Why This Trial Matters Statistically
Canopy-A is a useful teaching example because its primary and secondary analyses illustrate several foundational principles of modern survival analysis within one randomized phase 3 study.
| Concept | How it appears in Canopy-A |
|---|---|
| Randomization | Participants were randomly assigned to 2 parallel treatment arms. |
| Blinding | The trial was quadruple masked. |
| Intention-to-treat principle | The primary analysis used the Full Analysis Set containing participants assigned treatment by randomization. |
| Time-to-event endpoint | DFS was defined from randomization to recurrence or death. |
| Kaplan-Meier estimation | The registry states that median DFS was estimated using Kaplan-Meier methodology. |
| Stratified log-rank test | The primary and posted subgroup analyses used a stratified log-rank test. |
| Hazard ratio | DFS treatment effects were reported as hazard ratios. |
| Confidence interval | Each posted HR was accompanied by a two-sided 95% confidence interval. |
| Stratified Cox model | The HR and confidence interval were obtained from a stratified Cox model using the same stratification factors as the log-rank test. |
| Subgroup analysis | DFS was examined within PD-L1 and CD8 categories. |
| Safety analysis | Serious adverse events were reported separately by arm for the on-treatment period. |
19. A Practical Framework for Reading the Canopy-A Results
A statistically disciplined reading of the primary result can be reduced to five questions.
1. What was measured?
Disease-free survival, a time-to-event endpoint incorporating documented NSCLC recurrence or death from any cause.
2. Who was analyzed?
The Full Analysis Set, including all participants to whom study treatment was assigned by randomization.
3. What was the effect estimate?
HR 0.94 for canakinumab versus placebo.
4. How precise was it?
The 95% confidence interval was 0.78–1.14.
5. What did the hypothesis test show?
The two-sided P-value was 0.258.
What should not be inferred?
The HR and P-value do not by themselves provide an absolute DFS difference, an individual treatment benefit, or a probability that one hypothesis is true.
20. Sources
- ClinicalTrials.gov: NCT03447769 — Canopy-A.
- PubMed: PubMed record for PMID 37788412.
Continue with the Clinical Biostats methods library
Explore the statistical concepts underlying randomized time-to-event analyses, including hazard ratios, confidence intervals, Kaplan-Meier estimation, stratified analysis, and related clinical-trial methods.
21. Record Summary
Canopy-A provides a focused example of survival-analysis methodology in a randomized phase 3 trial. The primary DFS analysis used the Full Analysis Set, a stratified log-rank test, and a stratified Cox model, producing a hazard ratio of 0.94 with a 95% CI of 0.78–1.14 and a two-sided P-value of 0.258. The registry also reports five biomarker-defined DFS analyses, demonstrating why subgroup-specific estimates must be interpreted with attention to confidence intervals and the distinction between subgroup description and formal treatment-effect heterogeneity.
The most statistically informative way to read the record is to consider the endpoint definition, analysis population, hazard ratio, confidence interval, hypothesis test, and censoring framework together. That approach preserves the distinction between an estimated treatment effect, its statistical uncertainty, and the broader clinical meaning of a time-to-event result.