← Clinical Trials
Ovarian Cancer Phase 3 Time-to-Event Analysis NCT00262847

GOG-0218: Complete Statistical Analysis of Bevacizumab in Ovarian Cancer

An independent statistical analysis of the randomized phase 3 GOG-0218 trial of carboplatin and paclitaxel with or without bevacizumab in patients with stage III or stage IV ovarian epithelial, primary peritoneal, or fallopian tube cancer.

Trial status: COMPLETED  ·  Enrollment: 1873  ·  Start: 2005-09  ·  Primary completion: 2010-11
Scope of this record

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

GOG-0218 was a randomized, double-masked, parallel-group phase 3 treatment trial evaluating carboplatin and paclitaxel with or without bevacizumab in patients with stage III or stage IV ovarian epithelial, primary peritoneal, or fallopian tube cancer. The registry reports one primary endpoint, progression-free survival, with formal Cox proportional-hazards analyses comparing the control arm with two bevacizumab-containing comparisons.

1873
Enrollment
Randomized trial
3
Arms
Parallel design
0.765
PFS HR
Arm I vs Arm III
<0.001
PFS P-value
Arm I vs Arm III
FeatureGOG-0218
Trial nameGOG-0218
ClinicalTrials.gov identifierNCT00262847
PhasePhase 3
StatusCOMPLETED
Therapeutic areaOncology
PopulationPatients with stage III or stage IV ovarian epithelial, primary peritoneal, or fallopian tube cancer
AllocationRandomized
Design modelParallel
MaskingDouble
Primary purposeTreatment
Enrollment1873
Lead sponsorNational Cancer Institute (NCI)
Sponsor typeNIH
Primary endpoint typeTime-to-event
Results postedYes
Statistical analyses posted4

2. Clinical Question

The registry describes a phase 3 randomized treatment comparison involving carboplatin and paclitaxel with or without bevacizumab. The primary statistical question was whether progression-free survival differed between the control regimen and each of the two specified bevacizumab-containing comparisons.

Population

Patients with stage III or stage IV ovarian epithelial, primary peritoneal, or fallopian tube cancer. The registry's condition list includes multiple histologic disease categories.

Control comparison

Arm I: placebo, paclitaxel, and carboplatin.

Bevacizumab comparisons

Arm II: placebo, paclitaxel, carboplatin, and bevacizumab. Arm III: paclitaxel, carboplatin, and the intervention recorded in the registry as “Bevaci”.

Primary question

For progression-free survival, how do the two prespecified Arm I comparisons differ in estimated hazard of progression or death?

3. Trial Design

Allocation
Randomized allocation
Design model
Parallel-group design
Masking
Double
Primary purpose
Treatment
Phase
Phase 3
Enrollment
1873 participants
01
Randomize1873 enrolled
02
Three armsParallel treatment groups
03
FollowTime-to-event endpoint
04
CompareCox proportional-hazards model
05
EstimateHazard ratio, CI, P-value

The registry identifies GOG-0218 as a randomized, parallel phase 3 trial with double masking. The available trial data do not provide a randomization ratio, arm-level randomized sample sizes, a factorial structure, crossover information, an interim-analysis description, or a missing-data/imputation strategy. Those design features are therefore not inferred here.

Study arms represented in the posted analyses

ARM I

Placebo + chemotherapy

  • Placebo
  • Paclitaxel
  • Carboplatin
ARM II

Placebo + chemotherapy + bevacizumab

  • Placebo
  • Paclitaxel
  • Carboplatin
  • Bevacizumab
ARM III

Chemotherapy + bevacizumab

  • Paclitaxel
  • Carboplatin
  • The registry's analysis record identifies the intervention as “Bevaci”.
OTHER INTERVENTIONS

Additional registered components

  • Laboratory Biomarker Analysis
  • Quality-of-Life Assessment

4. Endpoints

The registry lists one primary endpoint: progression-free survival. It defines the endpoint as the median progression-free survival from study entry until first disease progression, death, or date of last contact, with follow-up up to 6 years.

EndpointRegistry definitionTime frameType
Progression-free Survival Median progression-free survival (PFS). Onset of progression could be based on radiographic (RECIST) criteria or rising CA-125 (GCIG criteria). From study entry until first disease progression, death or date of last contact, up to 6 years Time-to-event

Secondary endpoint represented in the posted analyses

EndpointTime frameType
Overall Survival From study entry to death or last contact, up to 6 years Time-to-event

The registry reports four statistical analyses: two for the primary progression-free survival endpoint and two for overall survival as a secondary endpoint. All four use a Cox proportional-hazards model and report hazard ratios.

5. Statistical Methodology

Why a time-to-event framework is appropriate

Both progression-free survival and overall survival are time-to-event endpoints. Patients can have different lengths of follow-up, and some patients may reach the end of available follow-up without experiencing the event. The registered endpoint definition explicitly allows a patient to contribute follow-up until progression, death, or the date of last contact.

Kaplan-Meier estimation

Kaplan-Meier estimation is the standard descriptive framework for a time-to-event endpoint such as progression-free survival or overall survival. It estimates the probability of remaining event-free over time while accommodating right-censored observations.

Kaplan-Meier concept
S(t) = ∏ti ≤ t (1 − di/ni)

Here, di is the number of events at time ti and ni is the number of individuals at risk immediately before that time.

Cox proportional-hazards model

All four posted statistical analyses use a Cox proportional-hazards model. The model estimates a relative hazard between treatment groups while using the observed event times and censoring information.

Cox model interpretation
h(t|X) = h0(t) exp(βX)

For a binary treatment indicator, the hazard ratio is exp(β). A hazard ratio below 1 corresponds to a lower estimated instantaneous hazard in the numerator treatment group relative to the comparator, under the fitted model.

Stratified analysis

The registry analysis notes state that the proportional-hazards model was stratified by stage of disease and size of residual disease following initial staging surgery. Stratification allows the baseline hazard to differ across these strata while estimating the treatment hazard ratio across the comparison.

Superiority hypothesis

The posted analyses are classified as superiority analyses. The purpose is therefore to estimate and test whether the treatment groups differ, rather than to establish that one treatment is no worse than another within a prespecified non-inferiority margin.

6. Primary Results: Progression-Free Survival

The registry contains two formal primary-endpoint analyses. Both use a Cox proportional-hazards model stratified by stage of disease and size of residual disease following initial staging surgery.

Arm I vs Arm II

Progression-free survival hazard ratio

0.954

95% CI: 0.844–1.078   ·   P = 0.448

Comparison: Arm I (Placebo, Paclitaxel, Carboplatin) vs Arm II (Placebo, Paclitaxel, Carboplatin, Bevacizumab)

Primary endpointComparisonMethodHR95% CIP-value
Progression-free Survival Arm I vs Arm II Stratified Cox proportional-hazards model 0.954 0.844–1.078 0.448
Clinical Biostats interpretation

The estimated hazard ratio of 0.954 means that, under the fitted stratified Cox model, the estimated instantaneous hazard of progression or death for Arm I relative to Arm II was approximately 95.4% of the corresponding hazard in Arm II. Equivalently, the point estimate is 1 − 0.954 = 0.046, or about a 4.6% lower estimated hazard for Arm I relative to Arm II.

The hazard ratio is a relative time-to-event measure. It does not mean that 95.4% of patients avoided progression, that 4.6% of patients benefited, or that the absolute difference in progression-free survival was 4.6 percentage points.

The 95% confidence interval, 0.844 to 1.078, expresses uncertainty around the estimated hazard ratio. Because the interval includes 1, the data are compatible with a range of relative hazards on both sides of the no-difference value.

The P-value of 0.448 measures how compatible the observed result is with the statistical null hypothesis under the specified testing framework. It does not measure the size of the treatment effect, the probability that the null hypothesis is true, or the clinical importance of the estimate.

Because this is a Cox model, interpretation also depends on the proportional-hazards framework used for the analysis. The registry does not provide sufficient information here to independently assess that assumption from the underlying event-time data.

Arm I vs Arm III

Progression-free survival hazard ratio

0.765

95% CI: 0.676–0.866   ·   P < 0.001

Comparison: Arm I (Placebo, Paclitaxel, Carboplatin) vs Arm III (Paclitaxel, Carboplatin, Bevacizumab)

Primary endpointComparisonMethodHR95% CIP-value
Progression-free Survival Arm I vs Arm III Stratified Cox proportional-hazards model 0.765 0.676–0.866 <0.001
Clinical Biostats interpretation

The estimated hazard ratio of 0.765 means that, under the fitted stratified Cox model, the estimated instantaneous hazard of progression or death for Arm I relative to Arm III was 76.5% of the corresponding hazard in Arm III. Equivalently, 1 − 0.765 = 0.235, so the point estimate corresponds to approximately a 23.5% lower estimated hazard for Arm I relative to Arm III.

Because the analysis is explicitly labeled as a comparison of Arm I versus Arm III, the direction of the hazard ratio must be read from that ordering. A hazard ratio below 1 here means the estimated hazard is lower for Arm I than for Arm III; it should not automatically be described as a benefit of whichever intervention is named first without checking the comparison definition.

The 95% confidence interval is 0.676 to 0.866. It lies below 1, indicating that the confidence interval for the estimated relative hazard does not include the no-difference value.

The P-value is reported as <0.001. This is evidence against the specified null hypothesis under the posted superiority analysis, but the P-value is not an effect-size measure. The magnitude and precision of the estimated effect are conveyed more directly by the hazard ratio and its confidence interval.

As with the other Cox analysis, this result is subject to the assumptions and censoring structure of the proportional-hazards model. The registry does not provide the underlying patient-level event and censoring data needed to independently reconstruct the fitted model.

7. How to Read the Two Primary Comparisons Together

The two primary analyses answer two different pairwise questions. They should therefore not be collapsed into a single hazard ratio.

ComparisonHR95% CIP-valueStatistical description
Arm I vs Arm II0.9540.844–1.0780.448Point estimate near 1; CI includes 1
Arm I vs Arm III0.7650.676–0.866<0.001Point estimate below 1; CI excludes 1

The important statistical distinction is not simply that one P-value is smaller than the other. The comparisons have different point estimates, confidence intervals, and treatment-group definitions. The registry supports reporting each comparison on its own terms.

Direction matters: the reported hazard ratios are explicitly for Arm I versus Arm II and Arm I versus Arm III. If the comparison is reversed, the corresponding hazard ratio would be expressed on the reciprocal scale. The page therefore preserves the registry's stated group ordering rather than silently reversing the estimates.

8. Secondary Results: Overall Survival

Overall survival is listed as a secondary endpoint in the posted statistical analyses. The registry defines its time frame as from study entry to death or last contact, up to 6 years.

Arm I vs Arm II

Overall survival hazard ratio

1.069

95% CI: 0.912–1.255   ·   P = 0.410

Secondary endpointComparisonMethodHR95% CIP-value
Overall Survival Arm I vs Arm II Stratified Cox proportional-hazards model 1.069 0.912–1.255 0.410
Clinical Biostats interpretation

The hazard ratio of 1.069 means that the estimated instantaneous hazard of death for Arm I relative to Arm II was 1.069 times the corresponding hazard in Arm II under the fitted stratified Cox model. The point estimate is therefore approximately 6.9% higher on the hazard scale for Arm I relative to Arm II.

The 95% confidence interval, 0.912 to 1.255, crosses 1. The interval therefore includes both a lower and a higher relative hazard than the no-difference value.

The P-value of 0.410 is not a measure of the size of the estimated difference. It indicates the degree of compatibility with the specified null hypothesis under the analysis framework, not the probability that one treatment is effective or ineffective.

Arm I vs Arm III

Overall survival hazard ratio

0.880

95% CI: 0.746–1.039   ·   P = 0.131

Secondary endpointComparisonMethodHR95% CIP-value
Overall Survival Arm I vs Arm III Stratified Cox proportional-hazards model 0.880 0.746–1.039 0.131
Clinical Biostats interpretation

The hazard ratio of 0.880 means that the estimated instantaneous hazard of death for Arm I relative to Arm III was 88.0% of the corresponding hazard in Arm III. The point estimate corresponds to approximately a 12.0% lower estimated hazard for Arm I relative to Arm III.

The 95% confidence interval, 0.746 to 1.039, includes 1. Thus, the interval includes the no-difference value and does not establish a statistically distinguishable relative hazard in this posted analysis at the conventional 1-versus-not-1 interpretation.

The P-value of 0.131 does not measure the magnitude of the observed hazard ratio. The hazard ratio and confidence interval provide the principal information about the estimated relative effect and its precision.

9. Overall Survival and Progression-Free Survival as Different Endpoints

GOG-0218 illustrates why progression-free survival and overall survival should be interpreted separately. Progression-free survival records the time to the first qualifying progression event or death, whereas overall survival records time to death or last contact.

Progression-free survival

The primary endpoint is defined from study entry until first disease progression, death, or date of last contact, up to 6 years. Progression may be based on radiographic RECIST criteria or rising CA-125 using GCIG criteria.

Overall survival

The secondary endpoint is defined from study entry to death or last contact, up to 6 years. Death is the event of interest rather than disease progression.

A treatment comparison can therefore produce different hazard-ratio estimates for PFS and OS because the endpoints capture different events and different portions of the clinical course. The registry results should be interpreted endpoint by endpoint rather than assuming that a PFS hazard ratio mechanically determines the OS hazard ratio.

10. Safety Results

The trial data provide serious adverse event counts by arm. These are reported as affected patients divided by patients at risk.

ArmSerious adverse eventsAffected / at risk
Arm I (Placebo, Paclitaxel, Carboplatin)Serious adverse events239/621
Arm II (Placebo, Paclitaxel, Carboplatin)Serious adverse events257/624
Arm III (Paclitaxel, Carboplatin, Bevaci)Serious adverse events287/619

The registry data do not provide a formal statistical comparison of serious adverse-event rates. Accordingly, the counts are reported descriptively rather than converted into an unreported P-value, risk ratio, odds ratio, or confidence interval.

Safety interpretation: the denominators in the serious-adverse-event record are the reported numbers at risk for this safety measure. They should not automatically be treated as the randomized enrollment in each arm, because the trial-level enrollment is reported separately as 1873.

11. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

Progression-free survival and overall survival are time-to-event endpoints with potentially unequal follow-up and censoring. A Cox model provides a way to compare the instantaneous event hazards between randomized groups without requiring the baseline hazard function itself to take a particular parametric form.

What does a hazard ratio of 0.765 mean?

For the registered Arm I-versus-Arm III analysis, an HR of 0.765 means that the estimated instantaneous hazard for Arm I was 76.5% of that for Arm III under the fitted model. On the hazard-ratio scale, the point estimate is 0.235 below 1, corresponding to approximately a 23.5% lower estimated hazard for Arm I relative to Arm III.

This is not the same as saying that 23.5% more patients survived, that survival probability was 23.5 percentage points higher, or that every patient experienced a 23.5% reduction in risk.

Why does the confidence interval matter?

A point estimate is only one estimate of the underlying relative treatment effect. The confidence interval describes the statistical uncertainty around that estimate under the model and study sampling framework. For Arm I versus Arm II, the 95% CI was 0.844–1.078; for Arm I versus Arm III, it was 0.676–0.866.

Why doesn't the P-value measure effect size?

A P-value addresses compatibility with a null hypothesis under a specified statistical model and testing framework. It is affected by both the estimated effect and the amount of information in the analysis. Two studies can therefore have similar effect estimates but different P-values, or statistically different P-values despite similar point estimates.

Why was the Cox analysis stratified?

The registry states that the proportional-hazards model was stratified by stage of disease and size of residual disease following initial staging surgery. Stratification permits different baseline hazards across those strata while preserving a common treatment-effect parameter for the comparison.

What does the registry's six-year time frame mean?

The registered PFS endpoint follows patients from study entry until first progression, death, or last contact, up to 6 years. The OS endpoint follows patients from study entry to death or last contact, up to 6 years. The time frame is therefore part of the endpoint definition rather than a statement that every patient was followed for exactly 6 years.

12. Stratification and the Interpretation of the Cox Model

The posted analyses specify stratification by two baseline or surgical-disease characteristics: stage of disease and size of residual disease following initial staging surgery.

Stratification factorRole in analysis
Stage of diseaseDefines strata for the proportional-hazards analysis.
Size of residual disease following initial staging surgeryDefines additional strata for the proportional-hazards analysis.

A stratified Cox model is different from simply putting a covariate into the model as an ordinary regression coefficient. With stratification, the baseline hazard is allowed to vary by stratum while the treatment comparison is estimated across the strata.

Conceptual interpretation
hs(t|X) = h0s(t) exp(βX)

The subscript s indicates that each stratum can have its own baseline hazard, while β represents the treatment effect summarized by the hazard ratio.

13. What the Primary Endpoint Does — and Does Not — Tell Us

Progression-free survival is not a simple proportion

The registry's primary endpoint is a time-to-event measure. It incorporates both when progression or death occurs and the censoring information available for patients who have not experienced the event by their last contact.

The HR is not an absolute benefit

An HR of 0.765 is a relative hazard estimate. It does not directly tell us the median PFS, the proportion progression-free at a particular calendar time, or the number of additional progression-free months for an individual patient.

A P-value is not clinical magnitude

The P-value of <0.001 for Arm I versus Arm III indicates strong statistical evidence against the specified null hypothesis in the posted analysis. It does not quantify how clinically important the hazard ratio is.

The comparison direction must be preserved

Because the registry explicitly defines the analyses as Arm I versus Arm II and Arm I versus Arm III, the sign and interpretation of the reported hazard ratio depend on that ordering. Reversing the groups would change the numerical HR to its reciprocal.

14. Limitations

15. Registry Data Cutoff Caveat

Data-freeze limitation: the registry states that the results for both Overall Survival (OS) and Progression Free Survival (PFS) were based on the analysis performed on a data set frozen on August 26, 2011.

This distinction is important when interpreting the posted estimates. A statistical analysis is tied to the dataset available at its specified data cutoff. Later evidence, if any, would represent a different information set and should not be silently combined with the estimates reported here.

16. No Unsupported Design Features Added

Several concepts are important in clinical-trial statistics but are not documented in the registry-reported GOG-0218 data. The page therefore does not assign them a role in the trial.

Design topicWhat the ClinicalTrials.gov record supports
Non-inferiority marginNot reported in the ClinicalTrials.gov record; the posted analyses are classified as superiority.
CrossoverNot reported in the ClinicalTrials.gov record.
Factorial designNot reported; the design model is parallel.
Interim analysisNot reported in the ClinicalTrials.gov record.
Missing-data / imputation methodNot reported in the ClinicalTrials.gov record.
Bayesian methodsNot reported; the posted method is a Cox proportional-hazards model.
Multiplicity adjustmentNo procedure is provided in the ClinicalTrials.gov record.

This distinction is intentional. A complete statistical analysis should explain what is documented without filling gaps with assumptions about the protocol or publication history.

17. Why This Trial Matters Statistically

GOG-0218 is a useful teaching example because the ClinicalTrials.gov record contains a clear progression from trial design to endpoint definition to formal time-to-event modeling.

ConceptHow it appears in GOG-0218
RandomizationThe trial is explicitly randomized.
Parallel designThe registry identifies a parallel design model with three arms.
BlindingThe registry identifies the masking as double.
Time-to-event endpointProgression-free survival is the registered primary endpoint.
Kaplan-Meier estimationProvides the natural descriptive framework for the registered time-to-event outcomes, although the ClinicalTrials.gov record does not report Kaplan-Meier estimates.
Hazard ratioAll four posted analyses use hazard ratio as the effect measure.
Cox regressionThe registry reports a Cox proportional-hazards model for every posted analysis.
StratificationThe Cox models are stratified by stage of disease and size of residual disease following initial staging surgery.
Confidence intervalsBoth primary analyses report 95% confidence intervals around the hazard ratio.
P-valuesBoth primary analyses report P-values for the superiority comparisons.
Multiple pairwise comparisonsThe single primary endpoint has two formal posted treatment comparisons.
Safety denominatorsSerious adverse events are reported as affected patients divided by patients at risk for each arm.

The statistical lesson is that a hazard ratio is only meaningful when its endpoint, event definition, comparison direction, model, stratification variables, confidence interval, and P-value are all read together.

18. A Practical Reading of the Reported Evidence

For the primary endpoint, the two posted comparisons are not statistically interchangeable. The Arm I-versus-Arm II analysis has an HR of 0.954 with a 95% CI of 0.844–1.078 and P = 0.448. The Arm I-versus-Arm III analysis has an HR of 0.765 with a 95% CI of 0.676–0.866 and P < 0.001.

The secondary overall-survival analyses show HR 1.069 for Arm I versus Arm II and HR 0.880 for Arm I versus Arm III. Their respective 95% confidence intervals are 0.912–1.255 and 0.746–1.039, with P-values of 0.410 and 0.131.

These estimates illustrate an important principle in clinical-trial interpretation: a single trial can produce different effect estimates for different endpoints and different pairwise comparisons. The appropriate statistical description is therefore to preserve the endpoint definition and comparison specified by the registry rather than combining the results into a single summary measure.

19. Related Tutorials

Learn more about the methods used in this trial:

20. Related Calculators

21. Sources

Continue through the Clinical Biostats statistical pathway

Use the trial's endpoint and modeling concepts to move from the clinical question to the underlying survival-analysis methods and statistical calculations.

22. Record Summary

GOG-0218 provides a clear example of randomized clinical-trial time-to-event analysis. The ClinicalTrials.gov record describes a phase 3, randomized, double-masked, parallel trial with 1873 participants and three arms. Progression-free survival is the registered primary endpoint, defined from study entry until first disease progression, death, or last contact, up to 6 years.

The four posted statistical analyses use stratified Cox proportional-hazards models and hazard ratios. For progression-free survival, the Arm I-versus-Arm II comparison produced HR 0.954 (95% CI 0.844–1.078; P = 0.448), while the Arm I-versus-Arm III comparison produced HR 0.765 (95% CI 0.676–0.866; P < 0.001). For overall survival, the corresponding HRs were 1.069 (95% CI 0.912–1.255; P = 0.410) and 0.880 (95% CI 0.746–1.039; P = 0.131).

The key statistical lesson is to interpret these results within their exact comparison, endpoint definition, model, stratification, confidence interval, and P-value. The hazard ratio is a relative time-to-event measure, not an absolute survival difference, and the registry's data-freeze caveat indicates that the reported OS and PFS results were based on a dataset frozen on August 26, 2011.

Clinical Biostats methodology: This page deliberately separates documented registry results from statistical explanation. Where the ClinicalTrials.gov record does not report an arm-level estimate, event count, subgroup result, design parameter, or analysis procedure, that information has not been reconstructed or inferred.