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
ANVIL is a randomized, parallel, open-label phase 3 treatment trial evaluating nivolumab after surgery and chemotherapy in patients with stage IB-IIIA non-small cell lung cancer. The registry lists 903 participants and two treatment arms, with disease-free survival and overall survival as the primary endpoints.
| Feature | ANVIL |
|---|---|
| Trial acronym | ANVIL |
| NCT identifier | NCT02595944 |
| Phase | Phase 3 |
| Population | Patients with Stage IB, Stage II, or Stage IIIA non-small cell lung cancer under AJCC v7 classifications |
| Design | Randomized, parallel, open-label |
| Primary purpose | Treatment |
| Enrollment | 903 |
| Number of arms | 2 |
| Lead sponsor | National Cancer Institute (NCI) |
| Sponsor type | NIH |
| Registry status | Active, not recruiting |
2. Clinical Question
The central statistical question is whether a strategy of nivolumab after surgery and chemotherapy produces different time-to-event outcomes from the comparator strategy of observation in patients with stage IB-IIIA non-small cell lung cancer.
Population
Patients with Stage IB Lung Non-Small Cell Carcinoma, Stage II Lung Non-Small Cell Cancer, or Stage IIIA Lung Non-Small Cell Cancer, classified using AJCC v7.
Intervention
Nivolumab after surgery and chemotherapy.
Comparator
Observation activity.
Primary question
How does randomized assignment affect disease-free survival and overall survival over follow-up assessed up to 10 years?
3. Trial Design
Nivolumab strategy
- Nivolumab
- Given in the setting described by the trial title as after surgery and chemotherapy
Observation strategy
- Observation activity
The registry does not provide a treatment-arm sample-size breakdown, randomization ratio, dosing schedule, treatment duration, or named randomization stratification factors in the information summarized here. Those details therefore cannot be used to characterize the design further without additional registry information.
4. Trial Population and Scope
The registry identifies three condition categories: Stage IB Lung Non-Small Cell Carcinoma AJCC v7, Stage II Lung Non-Small Cell Cancer AJCC v7, and Stage IIIA Lung Non-Small Cell Cancer AJCC v7.
| Population component | Registry description |
|---|---|
| Stage IB | Stage IB Lung Non-Small Cell Carcinoma AJCC v7 |
| Stage II | Stage II Lung Non-Small Cell Cancer AJCC v7 |
| Stage IIIA | Stage IIIA Lung Non-Small Cell Cancer AJCC v7 |
| Total enrollment | 903 |
From a statistical perspective, the inclusion of multiple disease stages is important because the overall treatment comparison is defined across the randomized population. If treatment effects differ substantially by stage, that issue would ordinarily be investigated through prespecified subgroup analyses or interaction testing. The registry description specifies that subset analyses are planned for all randomization stratification factors and known prognostic factors, including performance status, age, and gender.
5. Primary Endpoints
| Endpoint | Registry definition and time frame | Planned primary analysis |
|---|---|---|
| Disease-free survival (DFS) | Time from randomization to the earliest event defined as disease recurrence, any new lung cancer (even in the opposite lung), or death from any cause, assessed up to 10 years. | Kaplan-Meier estimation; Cox proportional-hazards models for treatment hazard ratios; stratified log-rank testing. |
| Overall survival (OS) | From randomization to death from any cause, assessed up to 10 years. | Kaplan-Meier estimation; Cox proportional-hazards models for treatment hazard ratios; stratified log-rank testing. |
Why DFS is a composite time-to-event endpoint
DFS is defined by the first occurrence of any one of several events: disease recurrence, a new lung cancer including one in the opposite lung, or death from any cause. Statistically, this means that the analysis does not wait for a single biological event. The first qualifying event determines the DFS event time.
This construction can be useful when the clinical objective is to capture whether a patient remains alive without one of several specified cancer-related or mortality events. It also means that the interpretation of a DFS hazard ratio applies to the composite endpoint as defined, rather than to recurrence alone.
Why OS is different
OS has a simpler event definition: death from any cause. A patient who has not died remains event-free for OS until the last time at which survival is known, at which point the observation can be censored according to the prespecified analysis rules.
6. Planned Analysis
Disease-free survival
The registry states that DFS distributions will be estimated using the Kaplan-Meier method. Treatment hazard ratios will be estimated with Cox proportional-hazards models. The primary comparison will use a log-rank test stratified on the randomization stratification factors.
The registry specifies a one-sided type I error rate corresponding to the significance level associated with the population being analyzed: 1.5% for the overall population and 1% for the ≥ 50% population. The latter test will only be performed if the primary test in all-comers is not statistically significant.
Overall survival
OS follows the same broad analytical structure. Kaplan-Meier methods will estimate the OS distribution, Cox proportional-hazards models will estimate treatment hazard ratios, and the primary treatment comparison will use a stratified log-rank test.
The same one-sided type I error specifications are given for the overall population and the ≥ 50% population, with the latter analysis conditional on the result of the primary all-comers test.
| Statistical component | DFS | OS |
|---|---|---|
| Time-to-event estimator | Kaplan-Meier | Kaplan-Meier |
| Treatment effect model | Cox proportional-hazards model | Cox proportional-hazards model |
| Primary comparison | Stratified log-rank test | Stratified log-rank test |
| Overall-population one-sided type I error | 1.5% | 1.5% |
| ≥50% population one-sided type I error | 1% | 1% |
| Testing sequence for ≥50% population | Only if all-comers test is not statistically significant | Only if all-comers test is not statistically significant |
| Confidence intervals | Adjusted for the type I error rates associated with the endpoint | Adjusted for the type I error rates associated with the endpoint |
The registry also states that other group comparisons will use log-rank testing and Cox modeling. Point estimates of all endpoints will be accompanied by corresponding confidence intervals adjusted for the type I error rates associated with the endpoint.
7. Statistical Methodology
Kaplan-Meier estimation
The Kaplan-Meier method estimates the probability of remaining event-free over time while allowing patients to have different lengths of observable follow-up. This is particularly appropriate for DFS and OS because not every participant will necessarily experience the endpoint before the end of observation.
where di is the number of events at time ti and ni is the number of participants at risk immediately before that time.
For ANVIL, the Kaplan-Meier curve for DFS would represent the probability of remaining free of the registry-defined composite DFS event. The OS curve would represent the probability of remaining alive. These are related but distinct estimands because their event definitions differ.
Stratified log-rank testing
The primary comparisons are specified as stratified log-rank tests. Stratification allows the comparison to account for the randomization stratification factors rather than treating all participants as though those factors had no role in the trial design.
The registry does not identify the individual stratification factors in the information reported here. The appropriate interpretation is therefore that the analysis will respect the prespecified randomization strata without assigning names to factors that are not reported.
Cox proportional-hazards model
The Cox model provides a relative treatment-effect estimate in the form of a hazard ratio. In this trial, the registry specifies Cox proportional-hazards models for estimating treatment hazard ratios for both DFS and OS.
A hazard ratio is a model-based relative measure. It is not the same as a relative risk, an absolute risk difference, or the percentage of participants who benefit.
For example, if a future analysis reported a DFS hazard ratio of 0.70, the appropriate mathematical interpretation would be that the estimated instantaneous rate of the defined DFS event is approximately 30% lower in the treatment group under the fitted model. It would not mean that exactly 30% fewer participants experience recurrence or death.
Confidence intervals
The registry specifies that point estimates of all endpoints will be accompanied by confidence intervals adjusted for the type I error rates associated with the endpoint. This is important because the confidence interval is part of the inferential framework, not merely a descriptive range around an estimate.
A confidence interval communicates the statistical precision of an estimated treatment effect under the model and sampling framework. A relatively narrow interval indicates greater statistical precision than a wide interval, although precision does not by itself determine clinical importance.
One-sided type I error
The registry explicitly specifies one-sided type I error rates for the primary comparisons. For the overall population, the significance level is 1.5%; for the ≥50% population, it is 1%.
A one-sided test allocates the prespecified type I error to a particular direction of treatment effect. That choice must be interpreted together with the trial's prespecified hypothesis and testing sequence. It should not be converted into a two-sided threshold without changing the inferential framework.
8. Sequential Testing of the ≥50% Population
The registry describes a hierarchical testing feature: the primary test in the overall population is evaluated first, and the test in the ≥50% population is performed only if the primary test in all-comers is not statistically significant.
This structure is statistically different from conducting two unrelated hypothesis tests and examining whichever produces the more favorable result. The sequence is prespecified: the ≥50% population analysis is conditional on the result of the all-comers analysis.
The registry does not define the meaning of the ≥50% population in the information reported here. It should therefore be treated as a registry-specified analysis population without assigning an unreported clinical characteristic to it.
9. Subset Analyses
The registry states that subset analyses are planned for all randomization stratification factors and all known prognostic factors such as performance status, age, and gender.
| Planned subset category | Registry description | Statistical role |
|---|---|---|
| Randomization stratification factors | All factors used for randomization stratification | Examination of treatment effects within prespecified strata |
| Performance status | Identified as a known prognostic factor | Assessment of consistency across performance-status groups |
| Age | Identified as a known prognostic factor | Assessment of consistency across age groups |
| Gender | Identified as a known prognostic factor | Assessment of consistency across gender groups |
Subset estimates should be interpreted as estimates within smaller portions of the randomized population. A difference between two subgroup-specific hazard ratios is not, by itself, evidence that treatment effects truly differ between those subgroups. A formal claim of effect modification ordinarily requires an interaction test or another prespecified comparison of treatment effects.
10. Statistical Methods Explained
Why use Kaplan-Meier analysis for DFS and OS?
Both primary endpoints are defined by the time from randomization to an event, and follow-up can end before an event occurs. Kaplan-Meier estimation handles this right-censoring structure while using the observed event times to estimate the event-free distribution.
Why use a Cox model?
The Cox model provides a compact relative comparison between the randomized groups through the hazard ratio. It uses the ordering and timing of events while accommodating censored observations and can incorporate the stratified structure specified by the trial.
Why use a stratified log-rank test?
The log-rank test compares the event experience of the randomized groups over follow-up. Stratification allows the primary comparison to reflect the randomization stratification factors, which can improve alignment between the statistical analysis and the trial's design.
What does the DFS event definition mean statistically?
DFS is a composite endpoint. The event occurs at the earliest qualifying occurrence of disease recurrence, any new lung cancer including one in the opposite lung, or death from any cause. Therefore, the DFS hazard ratio describes the randomized comparison for this entire composite definition rather than for recurrence alone.
Why does the registry specify different one-sided error rates?
The registry specifies a one-sided type I error rate of 1.5% for the overall population and 1% for the ≥50% population. These prespecified thresholds form part of the trial's inferential design and are paired with a testing sequence in which the ≥50% population test is performed only if the all-comers test is not statistically significant.
Why are confidence intervals important even when a hypothesis test is specified?
The p-value addresses compatibility with the prespecified null hypothesis under the testing framework. The confidence interval provides additional information about the estimated magnitude and precision of the treatment effect. A statistically significant finding can still have an imprecise effect estimate, while a precise estimate can be clinically interpreted differently depending on its magnitude.
11. Planned Analysis of Disease-Free Survival
Event definition
Disease recurrence, any new lung cancer including one in the opposite lung, or death from any cause.
Time origin
Randomization.
Analysis horizon
Assessed up to 10 years.
Primary comparison
Stratified log-rank test with a one-sided type I error framework.
For DFS, Kaplan-Meier estimation provides the event-free distribution, while the Cox model supplies the treatment hazard ratio. The combination of these methods is informative because the Kaplan-Meier analysis describes the estimated event-free experience over time, whereas the hazard ratio summarizes the relative instantaneous event rate under the fitted model.
The registry's definition also makes clear that death is an event rather than a censoring event for DFS. This is an important distinction: a participant who dies before documented recurrence has nevertheless experienced the DFS endpoint.
Any future DFS hazard ratio should be interpreted specifically for the composite event defined by the registry. It should not automatically be described as a hazard ratio for recurrence alone. Likewise, a future Kaplan-Meier DFS estimate would represent freedom from the complete composite endpoint, not simply freedom from radiographic recurrence.
12. Planned Analysis of Overall Survival
Event definition
Death from any cause.
Time origin
Randomization.
Analysis horizon
Assessed up to 10 years.
Primary comparison
Stratified log-rank test with Cox proportional-hazards modeling for the treatment hazard ratio.
OS is a time-to-death endpoint and therefore fits naturally into the same Kaplan-Meier, log-rank, and Cox framework specified for DFS. The key difference is the event definition: OS does not combine recurrence with mortality. The event is death from any cause.
Because the trial remains active, not recruiting and has a listed primary completion date of 2026-12-31, the registry's 10-year endpoint horizon is broader than the currently available registry status information.
A future OS hazard ratio below 1 would indicate a lower estimated instantaneous rate of death in the nivolumab group under the fitted Cox model. It would not represent a fixed percentage reduction in each individual's probability of death, nor would it by itself describe absolute survival at a particular time point.
13. Multiplicity and Endpoint Hierarchy
The registry identifies two primary endpoints, DFS and OS, and specifies type I error rates for the overall population and the ≥50% population. It also specifies that the ≥50% population test is performed only if the primary test in all-comers is not statistically significant.
| Component | Role in the registered analysis | Statistical implication |
|---|---|---|
| Disease-free survival | Primary endpoint | Primary time-to-event comparison |
| Overall survival | Primary endpoint | Primary time-to-event comparison |
| Overall population | Primary population for the specified testing sequence | One-sided type I error rate of 1.5% |
| ≥50% population | Conditional population analysis | One-sided type I error rate of 1%; tested only if the all-comers primary test is not statistically significant |
| Subset analyses | Exploratory/consistency assessment | Planned for stratification and prognostic factors |
The sequential structure is important because it defines how the registry intends the two population-level analyses to relate to one another. It is not equivalent to independently testing every possible population and endpoint at the same nominal threshold.
14. Confidence Intervals and Type I Error
The registry states that point estimates of all endpoints will be accompanied by confidence intervals adjusted for the type I error rates associated with the endpoint. This connects the interval estimates to the same inferential framework used for the hypothesis tests.
For a hazard ratio, the null value is 1.00. A confidence interval entirely below 1.00 would be consistent with a lower estimated event hazard in the nivolumab group, while an interval crossing 1.00 would indicate that the data are compatible with both directions of the treatment contrast under the specified interval framework.
The direction of interpretation depends on the treatment contrast used in the analysis. The registry describes Cox modeling for treatment hazard ratios but does not provide a posted numerical estimate.
15. What a Future Hazard Ratio Would — and Would Not — Mean
If a future DFS hazard ratio were reported as 0.75, that would mean the fitted model estimated the instantaneous rate of the defined DFS event to be approximately 25% lower in the treatment group than in the comparison group over the analyzed follow-up.
It would not mean that 25% of participants avoided recurrence, that each participant had exactly a 25% lower probability of an event, or that the absolute difference in DFS at every time point was 25 percentage points.
The confidence interval would quantify uncertainty around the estimated hazard ratio under the prespecified statistical framework. Its width would provide information about precision, while its location relative to the null value of 1.00 would inform the corresponding hypothesis-test interpretation.
Hazard ratios are relative measures. Kaplan-Meier estimates at clinically relevant time points would provide an absolute description of the proportion remaining event-free or alive. These measures answer different questions and should be interpreted together when results become available.
16. Proportional-Hazards Assumption
The registry specifies Cox proportional-hazards models. The usual interpretation of a Cox hazard ratio therefore relies on the proportional-hazards framework: the relative hazard represented by the model is treated as sufficiently stable over the relevant follow-up for the single hazard ratio to be an informative summary.
If hazards were strongly non-proportional, a single hazard ratio could conceal important changes in the relative treatment effect over time. In that situation, the Kaplan-Meier curves and time-specific survival estimates would become particularly important for understanding the shape and timing of treatment effects.
The registry information summarized here does not report a proportional-hazards diagnostic or any time-varying treatment-effect analysis.
17. Censoring and Follow-up
Both primary endpoints are time-to-event outcomes and are assessed up to 10 years. In this type of analysis, participants who have not experienced the relevant endpoint by their last known follow-up generally contribute information until that time and are then treated as censored according to the prespecified analysis rules.
Censoring is not equivalent to an event. A censored participant does not provide evidence that the endpoint did or did not subsequently occur; rather, the participant contributes information for the period during which the outcome is observed.
The registry description provided here specifies the endpoint definitions and statistical models but does not give detailed censoring rules, missing-data procedures, or imputation methods.
18. Limitations and Interpretation Issues
- No posted efficacy estimates: the registry does not currently provide numerical results or formal posted statistical analyses for DFS or OS.
- Incomplete design detail: the summarized registry information does not identify the individual randomization stratification factors, randomization ratio, treatment dosing schedule, or arm-specific enrollment.
- Composite DFS endpoint: DFS combines disease recurrence, any new lung cancer including one in the opposite lung, and death from any cause. A DFS effect therefore should not automatically be interpreted as a recurrence-only effect.
- Hazard-ratio interpretation: Cox hazard ratios are model-based relative measures and should not be treated as absolute risk differences or individual-level probabilities.
- Proportional hazards: a single Cox hazard ratio is most informative when the proportional-hazards framework is reasonable over the relevant follow-up.
- Subset analyses: subgroup estimates generally involve fewer participants than the overall randomized population and can therefore be less precise. Differences between subgroup estimates require formal interaction assessment before being interpreted as treatment-effect modification.
- Sequential testing: the registry specifies that the ≥50% population test is conditional on the result of the all-comers primary test, so the two analyses should not be interpreted as unrelated tests.
- Registry maturity: the study is listed as active, not recruiting, with a primary completion date of 2026-12-31. The absence of posted numerical results means the statistical evidence cannot yet be summarized through observed hazard ratios, confidence intervals, or p-values from the registry.
19. Why This Trial Matters Statistically
ANVIL provides a useful teaching example of how a randomized phase 3 oncology trial can be structured around two clinically distinct time-to-event endpoints and a prespecified inferential sequence.
| Concept | How it appears in ANVIL |
|---|---|
| Randomization | Participants are randomized to two parallel arms. |
| Open-label design | The registry specifies no masking. |
| Kaplan-Meier estimation | Used to estimate DFS and OS distributions. |
| Hazard ratio | Cox models are used to estimate treatment hazard ratios for DFS and OS. |
| Stratified log-rank testing | Primary comparisons are specified using log-rank tests stratified on randomization stratification factors. |
| Type I error | One-sided significance levels of 1.5% and 1% are specified for different populations. |
| Sequential testing | The ≥50% population test is performed only if the all-comers primary test is not statistically significant. |
| Composite endpoint | DFS combines recurrence, new lung cancer, and death from any cause. |
| Subset analysis | Analyses are planned for randomization factors and prognostic factors such as performance status, age, and gender. |
| Long follow-up | DFS and OS are assessed up to 10 years. |
20. Planned Analysis: A Practical Statistical Workflow
The registry's planned methodology can be viewed as a sequence of linked statistical operations rather than as a single test.
This workflow illustrates why a clinical-trial statistical analysis should not be reduced to a p-value. The Kaplan-Meier estimator describes the time-to-event experience, the log-rank test addresses the prespecified group comparison, the Cox model quantifies the relative treatment effect, and the confidence interval describes its statistical precision.
The sequential population testing adds another layer: the registry specifies how the evidence from the overall population relates to the subsequent ≥50% population analysis. That hierarchy is part of the design and should be preserved when interpreting future results.
21. Clinical Interpretation vs Statistical Interpretation
Statistical interpretation
The registered analysis is designed to compare time-to-event outcomes between randomized treatment groups using Kaplan-Meier estimation, stratified log-rank testing, and Cox proportional-hazards models, with prespecified one-sided type I error rates.
Clinical interpretation
The clinical meaning of ANVIL will depend on the eventual magnitude, precision, timing, and consistency of the observed DFS and OS treatment effects, together with the definition of the endpoints and the characteristics of the enrolled population.
22. Statistical Interpretation of the Endpoint Definitions
One of the most important distinctions in ANVIL is the difference between DFS and OS.
| Question | DFS | OS |
|---|---|---|
| What starts the clock? | Randomization | Randomization |
| What ends the event-free period? | Recurrence, any new lung cancer including opposite-lung cancer, or death from any cause | Death from any cause |
| Assessment horizon | Up to 10 years | Up to 10 years |
| Planned estimator | Kaplan-Meier | Kaplan-Meier |
| Planned model | Cox proportional hazards | Cox proportional hazards |
| Primary comparison | Stratified log-rank test | Stratified log-rank test |
The distinction matters because a treatment can affect the timing of recurrence without producing an identical effect on mortality. Conversely, OS captures the ultimate mortality endpoint but can reflect events and treatments occurring after the initial randomized intervention. The two endpoints therefore provide complementary information rather than interchangeable measurements.
23. Record Summary
ANVIL is a phase 3 randomized, parallel, open-label treatment trial with 903 participants and two arms. The registry identifies nivolumab after surgery and chemotherapy as the intervention strategy and observation as the comparator in patients with stage IB-IIIA non-small cell lung cancer.
The two primary endpoints are disease-free survival and overall survival, both assessed up to 10 years. The registered statistical plan uses Kaplan-Meier estimation, stratified log-rank testing, and Cox proportional-hazards models, with one-sided type I error rates of 1.5% for the overall population and 1% for the ≥50% population. The latter analysis is performed only if the primary test in all-comers is not statistically significant.
The registry also specifies adjusted confidence intervals and subset analyses covering randomization stratification factors and known prognostic factors such as performance status, age, and gender. Because no formal statistical analyses or numerical primary-endpoint results are posted in the registry, the current statistical record is centered on the registered design and analysis framework rather than observed treatment-effect estimates.
24. Sources
- ClinicalTrials.gov: ANVIL (NCT02595944).
Continue through the Clinical Biostats clinical trial library
Clinical trial analyses connect study design, endpoints, statistical models, and interpretation into a consistent framework for understanding randomized evidence.