PRODIGE 18: Complete Statistical Analysis of Cetuximab or Bevacizumab in Metastatic Colorectal Cancer

1. Trial at a Glance

PRODIGE 18 was a randomized, parallel-group phase 2 trial in patients with KRAS wild-type metastatic colorectal cancer whose disease had progressed after first-line treatment with bevacizumab. The trial evaluated chemotherapy associated with either cetuximab or bevacizumab, with progression-free survival at 4 months specified as the primary endpoint.

2. Clinical Question

The trial addresses a treatment-selection question in metastatic colorectal cancer after progression following first-line bevacizumab. The randomized comparison evaluated chemotherapy associated with either cetuximab or bevacizumab in patients described by the registry as having KRAS wild-type metastatic colorectal cancer.

The registry identifies oxaliplatin, folinic acid, 5-fluoro-uracil, irinotecan, bevacizumab, and cetuximab as interventions. It does not provide the detailed regimen schedule or treatment-cycle structure in the record summarized here, so those details are not reconstructed.

3. Trial Design

Why randomization matters

Randomization creates the framework for comparing outcomes between treatment groups while reducing systematic differences in treatment assignment. It does not guarantee that every baseline characteristic will be numerically identical between groups, particularly in a trial with 133 enrolled participants. Its principal statistical role is to make treatment assignment independent of prognostic factors in expectation under the randomization process.

What the open-label design means

Because masking was classified as none, knowledge of assigned treatment can potentially affect aspects of trial conduct, such as treatment decisions, reporting, or assessment. For a time-to-event endpoint such as PFS, objective progression criteria can reduce some sources of subjectivity, but the ClinicalTrials.gov record does not describe the detailed assessment or adjudication procedures.

4. Endpoints

Understanding the primary endpoint

The endpoint combines two possible events: disease progression according to RECIST v1.1 criteria or death. A participant who has experienced either event by the relevant analysis time contributes an event to the PFS analysis. A participant who remains alive without progression contributes follow-up information and is censored at the last follow-up according to the registry definition.

The phrase “PFS at 4 months” is important. It focuses attention on whether participants remain alive without progression at a specified time point rather than only on a median PFS or a single overall hazard ratio. Those are different summaries of the same underlying time-to-event process.

5. Statistical Methodology

The ClinicalTrials.gov record identifies progression-free survival at 4 months as the primary endpoint and provides its time-to-event definition. It does not report formal statistical analyses in the posted statistical-analyses section.

How a PFS endpoint is typically analyzed

Because PFS is a time-to-event endpoint, a conventional analysis would generally use methods that retain information about when progression or death occurs rather than reducing every participant to a simple binary outcome without regard to follow-up time.

Kaplan-Meier estimation

A standard descriptive approach for PFS is the Kaplan-Meier estimator. It estimates the probability of remaining event-free over time while accounting for right-censored observations.

For a primary endpoint specifically defined at 4 months, the Kaplan-Meier curve can be used to estimate the probability of remaining progression-free at 4 months. A direct comparison of the two 4-month estimates would answer a different question from comparing the entire PFS distributions over follow-up.

Log-rank testing

For randomized treatment groups with a time-to-event endpoint, the log-rank test is a conventional method for comparing the survival distributions. It uses the ordering of event times and compares observed and expected events between groups across follow-up.

Cox proportional-hazards modeling

A Cox proportional-hazards model is another conventional method for estimating a relative treatment effect for a time-to-event endpoint. Its familiar effect measure is the hazard ratio. However, the registry does not report a posted Cox analysis for PRODIGE 18, so a specific hazard ratio should not be attributed to this trial from the information available here.

Binary 4-month analysis versus full time-to-event analysis

The wording “PFS at 4 months” can be represented as a binary outcome: progression or death by 4 months versus no progression and survival at 4 months. Such an analysis can be summarized using proportions and a risk difference, risk ratio, or odds ratio. However, converting a time-to-event endpoint into a binary endpoint discards information about the exact timing of events before 4 months.

A time-to-event analysis preserves more of that information. The most appropriate statistical presentation depends on the prespecified analysis plan, which is not described in the posted statistical analyses for this record.

6. Statistical Methods Explained

Why is PFS a time-to-event endpoint?

PFS records not only whether progression or death occurred, but also the time from randomization until that event. This matters because two participants can both experience progression while having substantially different lengths of progression-free follow-up. Time-to-event methods preserve that timing information.

Why are patients censored?

A participant who is alive without progression at the last follow-up has demonstrated progression-free survival up to that point, but the exact future event time is unknown. Censoring allows that observed follow-up to contribute to the analysis without pretending that an unobserved event occurred at a known time.

What does “PFS at 4 months” mean?

It refers to the probability of being alive and progression-free at the 4-month time point under the endpoint definition. It is different from median PFS, which is the time at which the estimated progression-free survival probability reaches 0.50, when such a median can be estimated.

Why might Kaplan-Meier methods be useful here?

Kaplan-Meier estimation is designed for right-censored time-to-event data. It allows the estimated progression-free probability to change as progression or death events occur while retaining participants who have not yet experienced an event at their last observed follow-up.

What would a hazard ratio add?

A hazard ratio would summarize the relative instantaneous event rate between the two treatment groups under a Cox model. It would provide a different perspective from the 4-month PFS probability. The registry does not post a hazard-ratio estimate for PRODIGE 18, so no numerical HR is reported here.

Why does an open-label design matter statistically?

When participants and investigators know treatment assignment, knowledge of treatment can influence aspects of care or assessment. For an endpoint based on progression according to RECIST v1.1 or death, the extent of this influence depends on how progression was assessed and whether assessment was independently reviewed or otherwise standardized. Those procedures are not described in the registry information available here.

7. Planned Analysis

The registry says that the primary endpoint will measure progression-free survival at 4 months. PFS is defined from randomization to progression according to RECIST v1.1 criteria or death, with patients alive without progression censored at last follow-up.

For this type of endpoint, a conventional analysis would typically estimate progression-free survival over time using Kaplan-Meier methods and compare randomized groups using a time-to-event comparison such as the log-rank test. A Cox proportional-hazards model could provide a hazard ratio if that model were prespecified. Alternatively, because the endpoint is explicitly framed at 4 months, the analysis could report the estimated PFS probability at 4 months for each randomized group and an appropriate measure of between-group difference.

What an eventual numerical analysis would need to establish

A complete statistical interpretation would distinguish the estimated PFS probability at 4 months from the uncertainty around that estimate and from the between-group treatment effect. If a time-to-event model were used, its effect estimate and confidence interval would provide a relative comparison; if the endpoint were analyzed as a binary 4-month outcome, the corresponding absolute and relative measures would answer a somewhat different question.

The distinction is important because a 4-month PFS percentage describes the status of participants at one prespecified time point, whereas a hazard ratio summarizes relative event rates across the analyzed follow-up under a particular model. These quantities should not be treated as interchangeable.

8. Interpreting the Primary Endpoint

9. Design Features That Shape Interpretation

Randomization

Randomization is the central causal-design feature of PRODIGE 18. By randomly allocating participants to the parallel treatment strategies, the trial establishes a framework in which differences in outcomes can be attributed to treatment assignment more credibly than in an observational comparison.

Parallel treatment groups

The parallel design means participants remain associated with their randomized strategy rather than serving as their own randomized comparator. This makes between-group comparisons the principal basis for inference.

Open-label treatment

The absence of masking means treatment assignment was known. This is particularly relevant when considering endpoints that can involve clinical judgment. For PRODIGE 18, the registry specifies RECIST v1.1 for defining progression, but the available record does not describe whether progression assessments were independently reviewed or how disagreements were resolved.

Phase 2 context

The registry classifies PRODIGE 18 as phase 2. Phase 2 randomized trials can provide comparative evidence about treatment activity and clinical outcomes, while their statistical interpretation depends on the prespecified endpoint, sample size, analysis plan, and precision of the resulting estimates.

10. What Is Known From the Registry — and What Is Not

This distinction is statistically important. A completed trial and a trial with a defined primary endpoint are not the same thing as a registry record containing a posted numerical analysis of that endpoint. The available record establishes the design and endpoint framework, but it does not supply an estimate, confidence interval, or p-value for the primary endpoint.

11. Limitations

• No posted statistical analysis: the ClinicalTrials.gov record does not contain a formal posted statistical analysis for the primary endpoint.
• No numerical PFS estimate: the available registry information does not report a numerical estimate for PFS at 4 months.
• No confidence interval or p-value: corresponding measures of statistical uncertainty and hypothesis testing are not posted in the statistical-analyses section.
• Limited methodological detail: the registry information identifies the primary endpoint definition but does not specify the detailed statistical analysis plan, including the exact comparative test or model used.
• No subgroup results: the available record does not provide subgroup-specific efficacy estimates for interpretation.
• No safety results: the available record does not provide serious adverse-event counts by randomized arm.
• No detailed treatment schedule: the registry identifies the interventions but does not provide the complete regimen structure in the information summarized here.
• No reported baseline comparison: detailed baseline characteristics by randomized group are not included in the available record.

These limitations affect what can be concluded quantitatively. They do not prevent a useful examination of the trial's design and endpoint, but they do prevent a numerical assessment of the treatment contrast from the registry's posted statistical-analysis section.

12. Why This Trial Matters Statistically

PRODIGE 18 is a useful teaching example because it illustrates how a randomized phase 2 study can be structured around a clinically meaningful time-to-event endpoint while also showing why trial design information and numerical results are separate components of statistical evidence.

13. Statistical Methods in This Trial: A Deeper View

Why the 4-month time point is informative

A fixed time point can make a clinical endpoint straightforward to communicate: the analysis asks whether participants remain alive and progression-free at a defined point after randomization. This can be particularly useful when the trial question is centered on early disease control.

However, the fixed time point should not be confused with the complete PFS distribution. Participants can progress before or after 4 months, and those event times contain information that is not completely represented by a single 4-month percentage.

Why RECIST v1.1 is part of the statistical definition

Progression is not merely an administrative date. The registry defines progression using RECIST v1.1 criteria. The criteria therefore determine when one of the two components of the composite PFS event is considered to have occurred. Changes in the operational definition of progression can change the observed PFS endpoint even if the underlying disease experience were unchanged.

Why death is included with progression

Including death as an event prevents a participant who dies without a previously documented progression from being treated as if they remained progression-free indefinitely. This is a fundamental feature of PFS definitions: death is considered an event even when radiographic progression has not been recorded first.

Why randomization does not eliminate uncertainty

Randomization addresses systematic treatment-assignment bias, but a finite sample still produces statistical uncertainty. With 133 enrolled participants, the observed treatment-group difference would still be subject to sampling variability. Confidence intervals and other inferential quantities are therefore essential when numerical results are available.

Why the absence of a posted p-value matters

A p-value would address the compatibility of the observed data with a specified null hypothesis under the statistical model and testing procedure. It would not measure the size or clinical importance of a treatment effect. Because no formal statistical analysis is posted in the registry record, there is no trial-specific p-value to interpret here.

14. Statistical Interpretation Without a Numerical Result

The absence of these numerical quantities means that the statistical story of PRODIGE 18, as represented in the registry, is primarily a story about design and endpoint construction rather than an interpretable numerical estimate of treatment effect.

15. Trial Timeline

16. Population and Intervention Framework

The registry lists oxaliplatin, folinic acid, 5-fluoro-uracil, irinotecan, bevacizumab, and cetuximab among the interventions. The information available here does not provide enough detail to reconstruct which chemotherapy components were paired with each targeted agent or the precise dosing schedule.

17. Endpoint Interpretation: PFS at 4 Months

The composite nature of the endpoint is particularly important. A participant can reach the PFS event because of radiographic progression or because of death. Consequently, PFS is not equivalent to a tumor-response endpoint: a patient can have stable or responding disease at one assessment and later experience death, which would still constitute a PFS event.

18. Censoring and Follow-Up

The registry explicitly states that patients alive without progression will be censored at the last follow-up. This establishes the basic censoring rule for the primary endpoint.

Censoring is not the same as an absence of an event forever. It means that the trial has observed the participant to a particular point without recording the event, after which the participant's subsequent event status is not incorporated into the observed follow-up in the same way. Correct interpretation therefore depends on the relationship between censoring and the underlying event process.

19. What a Complete Numerical Analysis Would Contain

For a randomized trial with PFS at 4 months as the primary endpoint, a statistically informative results presentation would normally distinguish several layers of evidence.

These measures are complementary rather than interchangeable. The 4-month PFS probability is directly aligned with the named primary endpoint, while a hazard ratio describes a broader relative time-to-event relationship. Confidence intervals add information about precision, and p-values address a hypothesis-testing question rather than measuring effect size.

20. Important Statistical Cautions

• A p-value is not an effect size. A small p-value, if one were reported, would not by itself indicate how large or clinically important the treatment difference was.
• A confidence interval is not a range containing individual patient effects. It describes uncertainty around an estimated population-level parameter under the relevant statistical framework.
• PFS at 4 months is not median PFS. The former describes status at a fixed time point; the latter is a summary of the time-to-event distribution.
• A hazard ratio is not a risk ratio. These quantities have different definitions and interpretations.
• Censoring requires careful interpretation. Participants who are censored have observed follow-up through their censoring time, not an assumed event-free lifetime.
• Open-label status can matter. Knowledge of treatment assignment can influence aspects of assessment and treatment, depending on how the endpoint is operationalized.
• Randomization is not the same as perfect balance. Random allocation provides a causal comparison framework but does not force identical baseline characteristics in a finite sample.

21. Why This Trial Is a Useful PFS Case Study

PRODIGE 18 illustrates an important distinction in clinical-trial statistics: a well-defined endpoint can be statistically sophisticated even when the registry record contains little numerical analysis.

The primary endpoint begins at a clearly defined time point—randomization—and combines two clinically meaningful events, progression and death. The registry also explicitly defines the censoring rule and the criteria used to define progression. Those details are essential because changing any of them can change the statistical estimand.

The trial's randomized parallel design then provides the framework for comparing the two treatment strategies. The statistical question is not simply whether progression occurred. It is whether the distribution of time to progression or death differs between the randomized groups, with particular attention to the 4-month time point specified by the primary endpoint.

That structure makes PRODIGE 18 a useful example for understanding how clinical questions become estimands, how estimands become statistical endpoints, and how endpoint definitions determine the appropriate analysis methods.

22. Limitations of the Available Statistical Evidence

This distinction is particularly important when reading registry records. “Completed” describes the trial's status; it does not itself establish a numerical efficacy conclusion. Similarly, a registered endpoint identifies what the investigators intended to measure, but not the magnitude or statistical significance of the eventual treatment comparison.

23. Sources

• ClinicalTrials.gov: PRODIGE 18 — NCT01442649.
• PubMed: PubMed record associated with PRODIGE 18.

24. Record Summary

PRODIGE 18 is a completed randomized phase 2, parallel-group trial enrolling 133 participants with KRAS wild-type metastatic colorectal cancer and progressive disease after first-line treatment with bevacizumab. The trial compared chemotherapy associated with either cetuximab or bevacizumab and specified progression-free survival at 4 months as its primary endpoint. PFS was defined as the time from randomization to progression according to RECIST v1.1 criteria or death, with patients alive without progression censored at their last follow-up.

From a statistical perspective, the trial provides a clear example of a randomized time-to-event endpoint. Kaplan-Meier estimation, log-rank comparison, and Cox modeling are conventional approaches for this type of outcome, although the registry record does not post a formal statistical analysis specifying which approach was used. The absence of posted numerical statistical analyses means that the registry supports interpretation of the design, estimand, event definition, and censoring framework, but does not provide a numerical treatment-effect estimate, confidence interval, or p-value for the primary endpoint.