1. Trial at a Glance
RADIANT was a phase 3 randomized, parallel-group, quadruple-masked oncology trial evaluating erlotinib versus placebo after surgery, with or without adjuvant chemotherapy, in participants with EGFR-positive non-small cell lung carcinoma. The registered primary endpoint was disease-free survival (DFS), a time-to-event outcome defined from randomization to documented NSCLC relapse or death in the absence of relapse.
| Characteristic | RADIANT |
|---|---|
| Trial name | RADIANT |
| NCT ID | NCT00373425 |
| Therapeutic area | Oncology |
| Condition | Non-small Cell Lung Cancer |
| Phase | 3 |
| Status | Completed |
| Start | 2006-09 |
| Primary completion | 2013-04 |
| Lead sponsor | OSI Pharmaceuticals |
| Sponsor type | Industry |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Quadruple |
| Primary purpose | Treatment |
| Enrollment | 1,252 |
| Arms | 2 |
| Interventions | Erlotinib; placebo |
| Primary endpoints registered | 2 |
| Primary endpoint type | Time-to-event |
| Formal statistical analyses posted | 2 |
| Primary analyses with estimate + CI | 2 |
| Primary method | Log-rank test |
| Effect measure | Hazard ratio |
| Hypothesis type | Superiority |
2. Clinical Question
The central statistical question was whether disease-free survival differed between participants randomized to erlotinib and those randomized to placebo. The registry defines DFS as the time from randomization until the first documented NSCLC relapse by radiological examination and/or biopsy, or death in the absence of relapse.
Population
Participants enrolled in the phase 3 study with non-small cell lung cancer and EGFR-positive tumors, following surgery with or without adjuvant chemotherapy.
Intervention
Erlotinib.
Comparator
Placebo.
Primary question
Does randomized assignment to erlotinib change the time to NSCLC relapse or death compared with placebo?
3. Trial Design
RADIANT used randomized allocation and a parallel-group design. The registry identifies the masking scheme as quadruple, meaning the study was designed with masking across multiple participant and study personnel roles rather than as an open-label comparison. The statistical comparison was therefore anchored to randomized treatment assignment.
The primary analyses used the randomized cohort full analysis set, consisting of all randomized participants. This is important for interpreting the treatment effect: participants were analyzed according to their randomized group rather than being reclassified according to later treatment exposure.
| Design feature | Registry description | Statistical relevance |
|---|---|---|
| Allocation | Randomized | Supports a treatment-group comparison anchored to random assignment. |
| Design model | Parallel | Each participant contributes follow-up under one randomized treatment group. |
| Masking | Quadruple | Reduces the potential for knowledge of treatment assignment to influence trial conduct or assessment. |
| Primary purpose | Treatment | The statistical objective was a superiority comparison of treatment groups. |
| Analysis population | Randomized cohort full analysis set | The primary comparison retained the randomized treatment groups. |
4. Endpoints
Primary endpoint: Disease-Free Survival
The registry contains two posted primary DFS analyses corresponding to different data cut-off dates. The endpoint definition is consistent across the two entries. DFS was assessed every 3 months during the active phase and every 6 months during long-term follow-up up to 5 years and yearly thereafter.
| Registered endpoint | Time frame | Definition |
|---|---|---|
| Disease Free Survival (DFS) | Every 3 months during active phase and every 6 months during long-term follow-up up to 5 years and yearly thereafter until the data cut-off date of 08 April 2013; maximum follow-up 64 months. | Time from randomization until the first day NSCLC relapse is documented by radiological examination and/or biopsy, or death in the absence of relapse. Participants without a DFS event were censored on the last adequate radiological assessment date. |
| Disease Free Survival (DFS) | Every 3 months during active phase and every 6 months during long-term follow-up up to 5 years and yearly thereafter until the data cut-off date of 11 June 2014; maximum follow-up 78 months. | Time from randomization until the first day NSCLC relapse is documented by radiological examination and/or biopsy, or death in the absence of relapse. Participants without a DFS event were censored on the last adequate radiological assessment date. |
The registry therefore provides two analyses of the same registered endpoint at different follow-up cut-offs. The later analysis extends the maximum reported follow-up from 64 months to 78 months.
5. Results
Two formal primary DFS analyses are posted in the registry. Both compare erlotinib with placebo using an unstratified two-sided log-rank test and report a hazard ratio. The hazard ratio is explicitly defined in the registry as erlotinib relative to placebo.
Erlotinib vs Placebo
| Analysis population | Randomized cohort full analysis set (all randomized participants) |
|---|---|
| Groups compared | Erlotinib vs placebo |
| Endpoint | Disease-free survival |
| Method | Unstratified two-sided log-rank test |
| Effect measure | Hazard ratio, erlotinib to placebo |
| Hazard ratio | 0.90 |
| 95% CI | 0.741 to 1.104 |
| P-value | 0.3235 |
| Hypothesis | Superiority |
The hazard ratio of 0.90 means that the estimated instantaneous rate of a DFS event under the fitted time-to-event comparison was 0.90 times the corresponding rate for placebo. Expressed descriptively, this is an estimated hazard approximately 10% lower for erlotinib relative to placebo.
The hazard ratio is a relative time-to-event measure. It does not mean that 10% fewer participants experienced relapse, nor does it mean that every individual participant had a 10% reduction in risk.
The 95% confidence interval extends from 0.741 to 1.104. Because the interval includes 1, the registry's two-sided 0.05-level hypothesis-testing framework does not provide statistical evidence of a treatment difference at that analysis.
The P-value of 0.3235 measures the compatibility of the observed test statistic with the null hypothesis under the specified testing framework. It is not a measure of the size or clinical importance of the hazard ratio.
The analysis also used an intention-to-treat framework and a prespecified primary analysis triggered after at least 410 events had accrued. The registry notes that a statistically significant primary DFS result would have been required before testing key secondary efficacy variables under the hierarchical testing procedure.
Extended DFS Follow-up
| Analysis population | Randomized cohort full analysis set (all randomized participants) |
|---|---|
| Groups compared | Erlotinib vs placebo |
| Endpoint | Disease-free survival |
| Method | Unstratified two-sided log-rank test |
| Effect measure | Hazard ratio, erlotinib to placebo |
| Hazard ratio | 0.94 |
| 95% CI | 0.780 to 1.144 |
| P-value | 0.5620 |
| Hypothesis | Superiority |
In the extended analysis, the hazard ratio was 0.94. Under the same interpretation of the hazard ratio, this corresponds to an estimated instantaneous DFS-event rate approximately 6% lower with erlotinib than placebo.
The 95% confidence interval of 0.780 to 1.144 spans the null value of 1. The interval therefore remains compatible with both a reduction and an increase in the estimated hazard under the statistical model.
The P-value of 0.5620 is not evidence against the null hypothesis under the registry's two-sided testing framework. It should not be interpreted as the probability that the treatment is ineffective, nor as a probability that the null hypothesis is true.
Importantly, the later data cut-off does not simply constitute a larger sample. It represents additional follow-up and potentially different event information. Time-to-event analyses incorporate both events and censored follow-up, so extending observation changes the information available for estimation.
6. Comparing the Two Primary DFS Analyses
The two posted analyses provide a useful example of how a time-to-event estimate can change as follow-up matures. The first analysis uses a maximum follow-up of 64 months, while the later analysis extends maximum follow-up to 78 months. The point estimate moves from 0.90 to 0.94, while both confidence intervals include 1.
| Feature | 08 April 2013 | 11 June 2014 |
|---|---|---|
| Maximum follow-up | 64 months | 78 months |
| Hazard ratio | 0.90 | 0.94 |
| 95% CI | 0.741–1.104 | 0.780–1.144 |
| P-value | 0.3235 | 0.5620 |
| Statistical test | Unstratified two-sided log-rank | Unstratified two-sided log-rank |
| Analysis population | All randomized participants | All randomized participants |
The two estimates should not be treated as two independent randomized experiments. They are analyses of the same randomized trial at different follow-up cut-offs. The principal statistical distinction is therefore the amount and timing of follow-up information available at each analysis.
7. Statistical Methodology
Kaplan-Meier estimation
Disease-free survival is a time-to-event endpoint. A standard way to describe such an endpoint is with the Kaplan-Meier estimator, which estimates the probability of remaining event-free over time while retaining information from participants who are censored before experiencing an event.
Here, di is the number of events at time ti, while ni is the number at risk immediately before that event time.
The registry does not provide Kaplan-Meier median DFS values or specific survival probabilities at individual time points. Those quantities therefore are not reported here.
Log-rank test
Both formal primary analyses used an unstratified two-sided log-rank test. The log-rank test compares the observed and expected numbers of events between the randomized groups over the observed follow-up.
For a two-group survival comparison, the null hypothesis can be expressed conceptually as equality of the underlying survival distributions. The RADIANT registry specifically states that the null hypothesis that the DFS distributions of the two arms were equivalent was tested using an unstratified two-sided log-rank test at the 0.05 level.
Hazard ratio
The principal effect measure was the hazard ratio, with erlotinib as the numerator and placebo as the denominator.
An HR below 1 indicates a lower estimated instantaneous event rate in the erlotinib group; an HR above 1 indicates a higher estimated instantaneous event rate, subject to the assumptions of the underlying time-to-event model.
A hazard ratio should not be confused with a risk ratio. A risk ratio compares probabilities over a specified period, whereas a hazard ratio describes a relative event rate within a time-to-event framework.
Intention-to-treat analysis
The primary analysis population was the randomized cohort full analysis set, defined as all randomized participants. The analysis notes explicitly identify an intention-to-treat analysis.
The key statistical benefit is preservation of the randomized comparison. Analyzing participants according to randomized treatment assignment protects the interpretability that comes from randomization and avoids redefining the treatment groups according to events that occur after randomization.
Censoring
Participants without a DFS event were censored at the date of their last adequate radiological assessment. Censoring allows participants with incomplete event follow-up to contribute information up to the point at which their event-free status is known.
For the usual interpretation of a Kaplan-Meier or Cox-type survival analysis, censoring assumptions matter. In particular, the statistical interpretation depends on censoring being appropriately handled and not introducing systematic information about the unobserved event process.
Two-sided testing
Both primary analyses used two-sided tests. The stated testing framework used an alpha level of 0.05 for the primary DFS analysis. A two-sided test considers departures from the null in either direction rather than prespecifying only one direction of departure.
Multiplicity and hierarchical testing
The first primary-analysis notes describe a hierarchical testing procedure. If the primary DFS analysis had been statistically significant in favor of erlotinib, the null hypothesis for key secondary efficacy variables would then have been tested under that hierarchy.
This is an important distinction from simply performing a collection of independent hypothesis tests. A hierarchical procedure establishes an ordered testing strategy so that later confirmatory hypotheses are contingent on earlier results.
8. Primary Analysis Population
The registry specifies that the primary analysis was based on the randomized cohort full analysis set, consisting of all randomized participants. The analysis notes also identify the analysis as intention-to-treat.
Why this matters
Randomization creates the basis for comparing outcomes across treatment groups. Retaining participants according to randomized assignment preserves that comparison.
What it does not guarantee
ITT analysis does not eliminate missing follow-up, censoring, treatment discontinuation, or other practical complications of longitudinal trials.
9. Prespecified Primary Analysis Timing
The first primary DFS analysis was planned when at least 410 events had accrued. This event-driven feature is characteristic of many time-to-event trials: statistical information is accumulated through observed events rather than solely through calendar time or the number of participants enrolled.
For a time-to-event endpoint, the amount of information depends heavily on the number and timing of observed events. Two trials with the same enrollment can therefore have very different statistical information if their event rates and follow-up differ.
The registry also states that if the primary DFS result had been statistically significant in favor of erlotinib, key secondary efficacy variables would have been tested under a hierarchical testing procedure. This establishes an important connection between the timing of the primary analysis and multiplicity control.
10. Statistical Methods Explained
Why is DFS analyzed as a time-to-event endpoint?
DFS records not only whether a participant experienced relapse or death, but also when the first event occurred. A participant who remains event-free for a longer period contributes different information from one who experiences an event shortly after randomization. Time-to-event methods preserve that timing information and accommodate censoring.
What does a hazard ratio of 0.90 mean?
In this trial's first primary analysis, the estimated hazard ratio of 0.90 compares the DFS event hazard for erlotinib with that for placebo. Descriptively, the estimated hazard was approximately 10% lower in the erlotinib group under the fitted time-to-event comparison. It does not mean that 10% fewer participants necessarily experienced an event.
Why does the confidence interval matter?
A point estimate alone does not communicate how precisely the treatment effect has been estimated. The 95% confidence interval of 0.741 to 1.104 describes the statistical uncertainty around the first hazard-ratio estimate under the specified model and sampling framework. Because it includes 1, the data are compatible with no hazard difference as well as with effects in either direction within the interval.
Why is the P-value not a measure of treatment effect?
The P-value evaluates the compatibility of the observed test statistic with the null hypothesis under the specified testing procedure. It does not tell us how large the treatment effect is. Effect size is described by the hazard ratio, while precision is described by the confidence interval.
Why use a log-rank test?
The log-rank test is designed for comparing time-to-event distributions between groups. It incorporates the timing of observed events and the available risk sets throughout follow-up rather than reducing the endpoint to a simple binary outcome at one arbitrary time point.
Why does the later analysis have a different hazard ratio?
The later analysis uses a different data cut-off and extends maximum follow-up from 64 months to 78 months. Additional follow-up can produce additional events, more censoring information, and different risk sets. The resulting hazard-ratio estimate can therefore differ from the earlier estimate even though the randomized groups and endpoint definition remain the same.
What does intention-to-treat contribute to the analysis?
The ITT framework maintains the original randomized comparison. It avoids redefining treatment groups based on post-randomization behavior and therefore helps preserve the causal interpretation associated with random allocation.
11. Understanding the Hazard Ratio More Precisely
The hazard ratio is often the most prominent number in a clinical trial's time-to-event results, but it requires careful interpretation.
| Hazard ratio | Descriptive interpretation |
|---|---|
| HR = 1 | No estimated difference in instantaneous event rates between groups. |
| HR < 1 | Lower estimated instantaneous event rate in the numerator group. |
| HR > 1 | Higher estimated instantaneous event rate in the numerator group. |
For RADIANT, the treatment ratio was explicitly defined as erlotinib relative to placebo. Consequently, an HR below 1 favors erlotinib with respect to the DFS event hazard, while an HR above 1 corresponds to a higher estimated hazard under erlotinib relative to placebo.
The first estimate, 0.90, and the later estimate, 0.94, are both below 1. However, the corresponding 95% confidence intervals extend across 1. The point estimates should therefore not be interpreted independently of their uncertainty intervals.
12. Confidence Intervals and Statistical Precision
The first primary analysis reported a 95% confidence interval from 0.741 to 1.104. The second reported a 95% confidence interval from 0.780 to 1.144.
Point estimate
The hazard ratio provides the single best estimate from the specified analysis.
Confidence interval
The interval shows the statistical uncertainty surrounding that estimate under the analysis framework.
Null value
For a hazard ratio, the null value is 1. An interval containing 1 is consistent with no treatment difference under the corresponding test.
Clinical interpretation
A confidence interval is not a range of individual patient outcomes. It describes uncertainty around the estimated population-level effect.
13. P-Values and Hypothesis Testing
The registry identifies both primary analyses as superiority tests. The first analysis reported P = 0.3235, and the later analysis reported P = 0.5620.
The registry specifies a two-sided alpha level of 0.05 for the primary DFS analysis. Under that framework, neither reported P-value reaches the stated threshold.
This distinction is especially important for time-to-event trials. A nonsignificant test can coexist with a point estimate below 1, because the estimate and its uncertainty must be considered together.
14. Multiplicity and Hierarchical Testing
The first primary-analysis notes describe a hierarchical testing procedure for key secondary efficacy variables. The sequence was conditional: only if the primary DFS analysis was statistically significant in favor of erlotinib would the null hypotheses for key secondary efficacy variables be tested under that hierarchical procedure.
This is an example of multiplicity management. When several hypotheses are tested, the probability of obtaining at least one apparently positive result can increase if no adjustment or structured testing strategy is used.
| Feature | RADIANT approach | Statistical implication |
|---|---|---|
| Primary hypothesis | DFS superiority | Primary efficacy conclusion was tied to the DFS analysis. |
| Test | Unstratified two-sided log-rank test | Comparison used a survival-distribution test. |
| Alpha | 0.05 | Specified two-sided significance framework. |
| Secondary testing | Hierarchical, conditional on significant primary DFS result favoring erlotinib | Later confirmatory testing was ordered rather than treated as an unrestricted collection of tests. |
15. Safety Results
The registry reports serious adverse events by several treatment or study categories. The affected and at-risk counts are reproduced below exactly as reported.
| Study group | Serious adverse events | At risk |
|---|---|---|
| RC: Erlotinib | 118 | 611 |
| RC: Placebo | 79 | 343 |
| BPC-NOLC: Erlotinib / Placebo | 21 | 134 |
| BPC-NOLC: Placebo Only | 1 | 11 |
| BPC-OLC: Erlotinib | 41 | 132 |
These safety counts should not be substituted for the randomized DFS analysis. Safety and efficacy answer different statistical questions, and the denominators and group labels reported by the registry must be retained when interpreting serious adverse-event counts.
16. Limitations and Interpretation Issues
- No median DFS values are reported here. The registry data posted on ClinicalTrials.gov for the formal analyses provide hazard ratios, confidence intervals and P-values, but do not provide median DFS estimates.
- No time-specific Kaplan-Meier probabilities are reported. The endpoint is time-to-event and is compatible with Kaplan-Meier estimation, but the ClinicalTrials.gov record does not provide specific survival probabilities at selected time points.
- Hazard-ratio interpretation depends on the time-to-event model. A hazard ratio is a relative event-rate measure rather than an absolute risk difference or a probability of remaining disease-free.
- Censoring matters. Participants without a DFS event were censored at their last adequate radiological assessment. The interpretation of survival estimates therefore depends on appropriate handling of censored follow-up.
- The two primary results are not independent experiments. The analyses use different data cut-off dates from the same randomized trial. Their estimates should be understood as sequential analyses with different amounts of follow-up information.
- Secondary endpoint conclusions are conditional. The first primary-analysis notes specify hierarchical testing of key secondary efficacy variables only if the primary DFS analysis was statistically significant in favor of erlotinib.
- Registry completeness is limited. The registry itself states that the posting carries limitations regarding currency or completeness and provides no warranty concerning the completeness of the posted information.
17. Registry-Specific Caveats
For statistical interpretation, this means that the registry should be read as the source of the posted analysis fields and endpoint definitions, while the reported numerical estimates should be interpreted within the limitations of the information actually displayed in the record.
18. Why This Trial Matters Statistically
RADIANT is a useful teaching example because the registry connects a randomized clinical trial with several core concepts in survival analysis. The primary endpoint is explicitly time-to-event, the formal comparison uses the log-rank test, the effect measure is a hazard ratio, and the primary population is based on all randomized participants.
| Statistical concept | How it appears in RADIANT |
|---|---|
| Randomization | Participants were randomized to erlotinib or placebo. |
| Blinding | The registry identifies quadruple masking. |
| Intention-to-treat analysis | The primary analysis used the randomized cohort full analysis set. |
| Time-to-event endpoint | DFS measures time from randomization to relapse or death. |
| Kaplan-Meier estimation | Provides the standard framework for estimating event-free survival over time. |
| Log-rank test | Used for the primary comparison of DFS distributions. |
| Hazard ratio | Used as the primary effect measure. |
| Confidence interval | 95% two-sided intervals accompany both primary hazard-ratio estimates. |
| P-value | Two-sided hypothesis tests produced P = 0.3235 and P = 0.5620. |
| Multiplicity | Key secondary efficacy variables were subject to a hierarchical testing procedure conditional on the primary result. |
19. A Statistical Reading of the RADIANT Results
The first primary DFS analysis estimated an HR of 0.90, with a 95% CI of 0.741 to 1.104 and P = 0.3235. The point estimate was below 1, but the confidence interval included the null value of 1.
The later analysis estimated an HR of 0.94, with a 95% CI of 0.780 to 1.144 and P = 0.5620. The additional follow-up did not produce a statistically significant two-sided treatment difference under the stated primary testing framework.
The point estimates changed from 0.90 to 0.94 as follow-up was extended. Both estimates were accompanied by confidence intervals crossing 1. The appropriate statistical interpretation is therefore based on the complete estimate-plus-uncertainty result rather than on the direction of the point estimate alone.
20. Related Tutorials
Learn more about the statistical methods used in this trial:
21. Related Statistical Calculators
22. Sources
- ClinicalTrials.gov: RADIANT, NCT00373425 . Official trial registry record containing the registered endpoints and posted statistical analyses.
- PubMed: PMID 26324372 .
Continue through the Clinical Biostats statistical library
Connect the trial's endpoint and analysis methods to deeper tutorials, statistical calculators, and additional clinical-trial analyses.
23. Record Summary
RADIANT provides a clear example of a randomized oncology trial centered on a time-to-event primary endpoint. The statistical analysis used the randomized cohort full analysis set, an intention-to-treat framework, an unstratified two-sided log-rank test, and a hazard ratio as the principal effect measure. The first posted DFS analysis produced an HR of 0.90 with a 95% CI of 0.741–1.104 and P = 0.3235. The later analysis, using extended follow-up, produced an HR of 0.94 with a 95% CI of 0.780–1.144 and P = 0.5620.
The most useful statistical reading of these results combines the hazard-ratio point estimate, its confidence interval, the P-value, the randomized analysis population, the censoring definition, and the prespecified testing hierarchy. The two posted analyses should be understood as analyses of the same randomized trial at different follow-up cut-offs rather than as independent treatment comparisons.