← Clinical Trials
Hepatocellular Carcinoma Phase 3 Time-to-Event Analysis NCT03755791

COSMIC-312: Complete Statistical Analysis of Cabozantinib Plus Atezolizumab in Advanced Hepatocellular Carcinoma

An independent statistical analysis of the randomized phase 3 COSMIC-312 trial evaluating cabozantinib plus atezolizumab versus sorafenib in participants with advanced hepatocellular carcinoma who had not received previous systemic anticancer therapy.

Trial period: 2018-06-10 to 2021-11-30  ·  Randomized phase 3  ·  837 participants enrolled
Scope of this record

This page separates reported trial results from statistical interpretation. Numerical results on this page are restricted to the ClinicalTrials.gov record. The registry provides the official trial record.

Registry note: 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

COSMIC-312 was a randomized, open-label, parallel-group phase 3 treatment trial in participants with advanced hepatocellular carcinoma who had not received previous systemic anticancer therapy. The registry reports three arms and two primary time-to-event endpoints: progression-free survival and overall survival.

837
Enrolled
Phase 3 trial
3
Arms
Parallel design
0.63
PFS HR
99% CI 0.44–0.91
0.99
OS HR
96% CI 0.78–1.24
FeatureCOSMIC-312
PhasePhase 3
ConditionHepatocellular Carcinoma
PopulationParticipants with advanced hepatocellular carcinoma who have not received previous systemic anticancer therapy
DesignRandomized, parallel-group, unmasked treatment trial
AllocationRandomized
Primary purposeTreatment
Enrollment837
Primary endpointsProgression Free Survival (PFS) and Overall Survival (OS)
Primary endpoint typeTime-to-event
Results postedYes
ClinicalTrials.govNCT03755791
Lead sponsorExelixis
Sponsor typeIndustry

2. Clinical Question

The central statistical question was whether the experimental combination of cabozantinib plus atezolizumab differed from sorafenib with respect to two clinically important time-to-event outcomes in participants with advanced hepatocellular carcinoma who had not received previous systemic anticancer therapy.

Population

Participants with advanced hepatocellular carcinoma who had not received previous systemic anticancer therapy.

Experimental comparison

Cabozantinib plus atezolizumab versus sorafenib for the primary PFS and OS comparisons.

Comparator

Sorafenib served as the control arm for the two posted primary endpoint analyses.

Primary questions

How does the experimental combination compare with sorafenib for progression-free survival and overall survival?

The three-arm structure is important when interpreting the registry record. The posted primary analyses in the ClinicalTrials.gov record are specifically the comparison of the experimental cabozantinib + atezolizumab arm with the sorafenib control arm. The statistical analyses posted on ClinicalTrials.gov do not report a primary endpoint comparison involving the single-agent cabozantinib arm.

3. Trial Design

01
Randomize837 enrolled
02
3 ArmsParallel-group design
03
TreatmentCabozantinib + atezolizumab, cabozantinib, or sorafenib
04
Follow-upPFS and OS
05
AnalysisLog-rank test and hazard ratio
EXPERIMENTAL ARM

Cabozantinib + atezolizumab

  • Cabozantinib
  • Atezolizumab
  • Primary PFS and OS comparisons against sorafenib were reported for this arm
SINGLE-AGENT ARM

Cabozantinib

  • Cabozantinib
  • The registry-reported primary statistical analyses do not report a formal PFS or OS comparison involving this arm
CONTROL ARM

Sorafenib

  • Sorafenib
  • Control group for the two registry-reported primary endpoint comparisons
Allocation
Randomized allocation.
Design model
Parallel-group design.
Masking
None; the trial is recorded as unmasked.
Primary purpose
Treatment.

The trial started on 2018-06-10 and had a primary completion date of 2021-11-30. The ClinicalTrials.gov record lists the status as ACTIVE_NOT_RECRUITING.

4. Endpoints

Primary endpointRegistry time frameRegistry definitionPosted analysis
Progression Free Survival (PFS) for the Experimental Arm Versus the Control Arm in the PFS Intent to Treat (PITT) Population From the date of first participant randomization up to 28 months PFS was defined as the time from randomization to the earlier of either the date of radiographic progression defined as a 20% increase in the sum of the longest diameters of target lesions, or the unequivocal appearance of new lesions, or progression of non-target disease per Blinded Independent Radiology Committee (BIRC) or the date of death due to any cause per Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. Log-rank test; hazard ratio
Overall Survival (OS) for the Experimental Arm Versus the Control Arm in the ITT Population From the date of first participant randomization up to 36 months OS was defined as the time from randomization to death due to any cause. Log-rank test; hazard ratio

The registry therefore defines both primary outcomes as time-to-event endpoints. PFS uses progression or death as the event, while OS uses death from any cause. The distinction is important: PFS can end because a participant progresses before death, whereas OS requires death to occur for the event to be observed.

Endpoint terminology: The registry identifies the PFS analysis population as the PFS Intent to Treat (PITT) Population and the OS analysis population as the ITT Population. The ClinicalTrials.gov record does not provide additional numerical definitions of the PITT population beyond the first 372 participants randomized to the experimental and control arms.

5. Analysis Populations

PopulationDefinition reported in the ClinicalTrials.gov recordPrimary endpoint
PFS Intent to Treat (PITT) Included the first 372 participants randomized to the experimental cabozantinib + atezolizumab arm and control sorafenib arm. PFS
Intent-to-treat (ITT) Included all participants randomized to the experimental cabozantinib + atezolizumab arm and control sorafenib arm. OS

This distinction is one of the most important statistical features of the posted results. The PFS analysis is not described using the same population wording as the OS analysis. The PFS PITT population contains the first 372 participants randomized to the two compared arms, whereas the OS ITT analysis includes all participants randomized to those two arms.

Do not silently combine populations: An effect estimate calculated in the PFS PITT population should not be presented as though it were estimated from the full ITT population. Likewise, the OS HR of 0.99 should not be interpreted as if it were calculated from the first 372 participants.

6. Statistical Methodology

Log-rank testing

The registry reports the log-rank test as the statistical method for both posted primary analyses. The log-rank test is designed for comparing time-to-event distributions between groups while accounting for the timing of events and censoring rather than reducing follow-up to a single binary outcome.

Conceptually, the test compares the number of events observed in each treatment group with the number that would be expected under a null hypothesis of no difference in the survival experience between groups. The comparison is accumulated across event times.

Core idea
Observed events  vs.  expected events under H0

For a time-to-event endpoint, participants contribute information while they are under observation. Participants who have not experienced the event by the end of their available follow-up can be censored rather than treated as if they experienced the event.

Hazard ratio

The effect measure reported for both primary analyses is the hazard ratio (HR). A hazard ratio compares the estimated instantaneous event rates between treatment groups over the analyzed follow-up.

Interpretation of the hazard ratio
HR < 1  →  lower estimated instantaneous event rate in the experimental group

An HR is a relative time-to-event measure. It is not a percentage of patients who benefit, not an absolute risk difference, and not a statement that every participant experiences the same proportional change in event risk.

Intention-to-treat analysis

The OS analysis was reported in the ITT population. The intention-to-treat principle keeps participants associated with their randomized group for the efficacy comparison. This preserves the central advantage of randomization: treatment assignment is established before subsequent outcomes are observed.

For PFS, the ClinicalTrials.gov record instead specify a PITT population consisting of the first 372 participants randomized to the experimental and control arms. That population definition should remain attached to the PFS estimate when the result is interpreted.

Time-to-event structure

Both primary endpoints begin at randomization. The statistical object is therefore not simply whether an event occurred, but when the event occurred, together with whether a participant remained event-free at the end of observable follow-up.

The registry does not report median PFS, median OS, Kaplan-Meier estimates at particular time points, numbers of events for the two primary analyses, or a formal description of a Cox proportional-hazards model. Those quantities are therefore not added here.

7. Primary Result: Progression-Free Survival

The first posted primary analysis compared progression-free survival between the cabozantinib + atezolizumab experimental arm and the sorafenib control arm in the PFS Intent to Treat (PITT) population. The registry defines the analysis window as from the date of first participant randomization up to 28 months.

PFS hazard ratio

0.63

99% two-sided CI: 0.44–0.91   ·   P = 0.0012

Analysis population: PFS Intent to Treat (PITT) population

FeatureReported result
ComparisonCabozantinib + atezolizumab vs sorafenib
EndpointProgression Free Survival
Time frameFrom the date of first participant randomization up to 28 months
Analysis populationFirst 372 participants randomized to the experimental and control arms
Statistical methodLog-rank test
Effect measureHazard ratio
HR0.63
Confidence interval99% two-sided CI 0.44–0.91
P-value0.0012
Clinical Biostats interpretation

An HR of 0.63 means that the estimated instantaneous rate of progression or death in the cabozantinib + atezolizumab group was approximately 63% of the corresponding rate in the sorafenib group, under the time-to-event analysis represented by the reported hazard ratio. Equivalently, 0.63 corresponds to an estimated 37% lower instantaneous hazard relative to the control group.

The HR does not mean that 37% of participants avoided progression, that PFS was 37% longer for every participant, or that the absolute probability of progression or death was reduced by 37 percentage points. Those interpretations require absolute survival estimates, which are not reported in the ClinicalTrials.gov record.

The 99% two-sided confidence interval of 0.44–0.91 describes statistical uncertainty around the reported HR under the analysis framework. It does not describe the range of treatment effects experienced by individual patients. Its width also shows that the point estimate of 0.63 should not be treated as an exact or immutable treatment effect.

The P-value of 0.0012 addresses the evidence against the relevant null hypothesis under the reported testing framework; it does not measure the magnitude of the treatment effect. A P-value cannot be substituted for the HR or its confidence interval.

Because PFS is a time-to-event endpoint, censoring and the definition of progression matter. The registry specifically defines PFS using radiographic progression, new lesions, progression of non-target disease, or death. The ClinicalTrials.gov record does not provide enough information to evaluate the proportional-hazards assumption or to reconstruct a Kaplan-Meier curve.

8. Primary Result: Overall Survival

The second posted primary analysis compared overall survival between the cabozantinib + atezolizumab experimental arm and the sorafenib control arm in the ITT population. The registry defines OS as the time from randomization to death due to any cause and gives a time frame extending up to 36 months from the date of first participant randomization.

OS hazard ratio

0.99

96% two-sided CI: 0.78–1.24   ·   P = 0.9056

Analysis population: ITT population

FeatureReported result
ComparisonCabozantinib + atezolizumab vs sorafenib
EndpointOverall Survival
Time frameFrom the date of first participant randomization up to 36 months
Analysis populationAll participants randomized to the experimental and control arms
Statistical methodLog-rank test
Effect measureHazard ratio
HR0.99
Confidence interval96% two-sided CI 0.78–1.24
P-value0.9056
Clinical Biostats interpretation

An HR of 0.99 is very close to 1.00. In the reported analysis, the estimated instantaneous rate of death in the cabozantinib + atezolizumab group was approximately 99% of the corresponding rate in the sorafenib group.

This does not mean that the two groups had identical survival times for every participant. A hazard ratio of 0.99 is a summary relative measure; it does not establish equality of individual survival experiences, nor does it provide median survival or survival probabilities at particular time points.

The 96% two-sided confidence interval of 0.78–1.24 spans 1.00. That interval indicates uncertainty in the estimated relative hazard and includes values corresponding to a lower estimated hazard as well as values corresponding to a higher estimated hazard in the experimental group.

The P-value of 0.9056 is a measure associated with the statistical test, not a measure of effect size. A large P-value should not be translated into a probability that the null hypothesis is true, nor does it prove that the treatment groups have exactly the same OS.

OS can also be influenced by events and treatments occurring after the initial progression event. The ClinicalTrials.gov record does not describe subsequent therapy, crossover, or other post-progression treatment patterns, so those factors cannot be quantified or incorporated into the interpretation here.

9. Reading the Two Primary Results Together

The two primary estimates illustrate why a clinical trial with multiple time-to-event endpoints should not be reduced to a single statistic. The reported PFS HR is 0.63, whereas the reported OS HR is 0.99. These estimates address different events and were calculated in different analysis populations.

DimensionPFSOS
Endpoint eventRadiographic progression, new lesions, progression of non-target disease, or deathDeath due to any cause
Analysis populationPITT population: first 372 participants randomized to experimental and control armsITT population: all participants randomized to experimental and control arms
Time frameUp to 28 months from first participant randomizationUp to 36 months from first participant randomization
MethodLog-rank testLog-rank test
Effect measureHazard ratioHazard ratio
Estimate0.630.99
Confidence interval99% two-sided: 0.44–0.9196% two-sided: 0.78–1.24
P-value0.00120.9056

Statistically, these are not contradictory measurements of the same endpoint. PFS asks whether the time to progression or death differs, while OS asks whether the time to death differs. A difference in one endpoint does not mathematically require the same difference in another endpoint.

The different confidence levels are also important. The PFS estimate is accompanied by a 99% two-sided confidence interval, while the OS estimate uses a 96% two-sided confidence interval. Confidence intervals with different confidence levels should not be visually or numerically compared as though they were constructed under identical uncertainty conventions.

Do not compare HRs as if they were ordinary percentages. The difference between 0.63 and 0.99 is informative, but it does not mean that the PFS effect was "36 percentage points larger" than the OS effect. Hazard ratios are multiplicative relative measures, not percentage-point outcomes.

10. What the Hazard Ratio Does — and Does Not — Mean

PFS example

The reported PFS HR of 0.63 corresponds to an estimated instantaneous progression-or-death rate about 37% lower in the experimental group relative to the control group. The phrase "instantaneous rate" is deliberate: a hazard is a rate at a particular point in time among participants who remain at risk, not the probability that an individual participant will experience an event.

OS example

The reported OS HR of 0.99 is close to the null value of 1.00. It represents an estimated instantaneous death rate approximately 1% lower in the experimental group than in the control group. That relative estimate should be read together with its 96% confidence interval of 0.78–1.24 rather than in isolation.

Why the confidence interval matters

A point estimate is only one estimate from the observed data. The confidence interval communicates how much uncertainty surrounds it under the statistical framework. For PFS, the 99% interval is 0.44–0.91; for OS, the 96% interval is 0.78–1.24. The intervals provide substantially more information about precision than the point estimates alone.

Why the P-value is not the effect size

The PFS P-value of 0.0012 and OS P-value of 0.9056 answer a testing question. They do not quantify clinical magnitude. Effect size is represented here by the HR, while statistical uncertainty is represented by the confidence interval. Keeping these three quantities conceptually separate prevents a common mistake in interpreting clinical-trial results.

11. Statistical Methods Explained

Why was a log-rank test used?

Both primary outcomes are time-to-event endpoints. A log-rank test is designed for comparing event-time distributions while incorporating the timing of events and censoring. It therefore fits the structure of PFS and OS better than a simple comparison of proportions at one arbitrary time point.

What does an HR of 0.63 mean for PFS?

It means the estimated instantaneous rate of progression or death was approximately 63% of the corresponding rate in the control group. Another way to express the same relative estimate is that the experimental group had an approximately 37% lower estimated hazard. It does not mean a 37% absolute reduction in the probability of progression.

What does an HR of 0.99 mean for OS?

An HR of 0.99 is close to the null value of 1.00. Under the reported model-based effect measure, the estimated instantaneous death rate was approximately 99% of the control-group rate. The associated 96% confidence interval of 0.78–1.24 shows that the point estimate should not be treated as an exact measurement.

Why does the PFS analysis use a PITT population?

The registry explicitly defines the PFS PITT population as the first 372 participants randomized to the experimental and control arms. That population definition is part of the result. Using the phrase "PFS in all randomized participants" would therefore misstate the registry-reported analysis population.

Why is OS described as an ITT analysis?

The registry states that the OS ITT population included all participants randomized to the experimental and control arms. An ITT framework preserves treatment assignment from randomization, which helps maintain the comparability created by randomization when analyzing efficacy.

Why is a confidence interval more informative than a P-value alone?

The P-value describes evidence relative to a null hypothesis under the testing procedure. The confidence interval additionally describes the range of effect estimates compatible with the statistical uncertainty represented by that interval. For example, the OS interval of 0.78–1.24 shows that the HR estimate is uncertain enough to include 1.00.

Can an HR be interpreted as a risk ratio?

No. A hazard ratio is a time-to-event measure involving the instantaneous event rate among participants who remain at risk. A risk ratio compares cumulative probabilities over a specified period. The two measures can differ, especially when event timing and censoring are important.

12. Understanding PFS Versus OS

COSMIC-312 provides a useful illustration of why progression-free survival and overall survival should be treated as distinct statistical endpoints.

PFS captures earlier events

PFS counts progression or death as the endpoint event. It can therefore register a treatment-group difference before many participants have died.

OS has a different event

OS requires death from any cause. A participant who progresses but remains alive has experienced a PFS event but not an OS event.

Different populations

The registry-reported PFS analysis uses the first 372 participants randomized to the experimental and control arms, while the OS analysis uses all randomized participants in those arms.

Different time frames

The registry specifies up to 28 months for PFS and up to 36 months for OS from the date of first participant randomization.

This distinction means that the PFS HR of 0.63 should not be presented as though it were a surrogate measurement of the OS HR of 0.99. The two estimates are based on different event definitions, different analysis populations, and different registry time frames.

13. Safety Results

The ClinicalTrials.gov record reports serious adverse events by arm as affected participants divided by participants at risk. These figures should be read as the registry's reported counts and denominators rather than converted into a different statistical measure.

ArmSerious adverse eventsAt risk
Experimental: Cabozantinib + Atezolizumab226429
Single-Agent Cabozantinib88188
Control: Sorafenib85207

The ClinicalTrials.gov record does not provide a formal statistical comparison of serious adverse-event rates, confidence intervals, exposure-adjusted incidence rates, or adverse-event definitions beyond the category "serious adverse events." Those analyses are therefore not added.

Safety and efficacy use different statistical questions. The primary efficacy analyses are time-to-event comparisons based on randomized treatment assignment. Serious adverse events are reported here as arm-level affected/at-risk counts. The two types of evidence should not be collapsed into a single efficacy-versus-safety statistic.

14. Trial Design Features That Affect Statistical Interpretation

Randomization

Randomization is the foundational design feature for the primary treatment comparison. It creates the basis for comparing outcomes between treatment groups without requiring the statistical analysis to assume that observed baseline differences were the cause of outcome differences.

The registry identifies the allocation as RANDOMIZED. The ClinicalTrials.gov record does not report the randomization ratio or stratification factors, so neither is inferred here.

Three-arm structure

The trial has three arms, but the two posted primary statistical analyses in the ClinicalTrials.gov record compare only the experimental cabozantinib + atezolizumab arm with the sorafenib control arm. The single-agent cabozantinib arm is part of the overall trial design but is not assigned a primary PFS or OS comparison in the ClinicalTrials.gov record.

Open-label design

The masking field is recorded as NONE. This means treatment assignment was not masked in the trial design recorded in the ClinicalTrials.gov record. For objective time-to-event endpoints, the endpoint definition and assessment process remain important components of interpretation, particularly for the progression component of PFS.

Multiplicity

The trial has two registered primary endpoints, PFS and OS, and three treatment arms. The ClinicalTrials.gov record does not state an alpha-allocation procedure, hierarchical testing sequence, multiplicity adjustment, or hypothesis-testing hierarchy. Accordingly, no specific multiplicity strategy is attributed to the trial here.

Interim analysis

The ClinicalTrials.gov record does not identify an interim-analysis procedure, alpha-spending approach, stopping boundary, or information fraction. No such procedure is inferred from the existence of posted primary results.

Missing data and imputation

The ClinicalTrials.gov record does not describe missing-data handling, censoring rules beyond the registered PFS event definition, or imputation procedures. For time-to-event endpoints, censoring is inherently part of the analysis structure, but the exact censoring conventions used in this trial are not reported here.

Bayesian methods

No Bayesian method is identified in the ClinicalTrials.gov record. The reported primary analyses use the log-rank test and hazard ratio.

15. What the Confidence Intervals Tell Us

The confidence intervals are particularly useful because the point estimates alone can create a false sense of precision.

EndpointHRConfidence levelTwo-sided confidence intervalNull value 1.00 inside interval?
PFS0.6399%0.44–0.91No
OS0.9996%0.78–1.24Yes

For a hazard ratio, 1.00 is the conventional null value corresponding to equal estimated hazards between the two groups. The PFS interval from 0.44 to 0.91 lies below 1.00, whereas the OS interval from 0.78 to 1.24 crosses 1.00.

That observation is descriptive of the reported intervals. It should not be extended into an assertion about the probability that the true effect is above or below 1.00, because a confidence interval is not a posterior probability distribution.

A useful mental model
Point estimate = estimated effect   |   CI = uncertainty around estimate   |   P-value = evidence against null

These quantities are related, but they answer different questions. A strong statistical interpretation uses all of them rather than treating any single number as a complete description of the result.

16. Why the PFS and OS Populations Matter

The analysis population is not a footnote. It defines which randomized participants contribute to the estimate.

PFS: first 372 participants

The registry specifies that the PITT population included the first 372 participants randomized to the experimental and control arms.

OS: all randomized participants

The OS ITT population included all participants randomized to the experimental and control arms.

Because the populations differ, the PFS and OS HRs should not be interpreted as though they came from the same analytical dataset. This is particularly important when comparing estimates numerically. A change in the observed HR can reflect the endpoint, follow-up, population, and event structure as well as the underlying treatment effect.

17. What Is Not Reported in the Supplied Registry Data

A complete statistical analysis should also distinguish between information that is available and information that is not reported in the ClinicalTrials.gov record. ClinicalTrials.gov does not report the following quantities in the ClinicalTrials.gov record:

Leaving these quantities out is preferable to reconstructing them from memory or importing numbers from external publications when the page's evidence rule is restricted to the ClinicalTrials.gov record.

18. Important Limitations and Interpretation Issues

19. Why This Trial Matters Statistically

COSMIC-312 is a useful teaching case because it demonstrates how several fundamental principles of clinical-trial statistics interact within a single randomized study.

ConceptHow it appears in COSMIC-312
RandomizationThe trial uses randomized allocation.
Parallel-group designThe study has three parallel treatment arms.
Time-to-event endpointsPFS and OS are both registered as time-to-event primary endpoints.
Log-rank testThe registry reports a log-rank method for both primary analyses.
Hazard ratioHR is the reported effect measure for both primary endpoints.
Confidence intervalsPFS uses a 99% two-sided CI and OS uses a 96% two-sided CI.
P-valuesThe registry reports P = 0.0012 for PFS and P = 0.9056 for OS.
ITT principleOS is analyzed in the ITT population.
Analysis-population definitionPFS uses a specifically defined PITT population consisting of the first 372 participants randomized to the experimental and control arms.
Endpoint interpretationPFS and OS measure different clinical events and should not be treated as interchangeable.
Three-arm designThe overall trial contains cabozantinib + atezolizumab, single-agent cabozantinib, and sorafenib arms.
Safety reportingSerious adverse events are reported as affected participants over participants at risk for each arm.

The trial is particularly valuable for understanding why a statistical result is more than a P-value. The estimand-like question begins with the population and endpoint definition, the design determines how groups are made comparable, the time-to-event method accounts for event timing and censoring, the hazard ratio summarizes relative effect, and the confidence interval communicates uncertainty.

Equally important, the PFS and OS results show why an analyst should not assume that all efficacy endpoints move together. A treatment can have one observed time-to-event estimate and a different estimate for another endpoint because the endpoints capture different biological and clinical events.

20. A Practical Framework for Reading COSMIC-312

Step 1: Identify the question

Start with the population, experimental treatment, comparator, and endpoint rather than the P-value.

Step 2: Identify the analysis population

For PFS, note the PITT definition. For OS, note the ITT definition.

Step 3: Read the effect estimate

Interpret HR 0.63 and HR 0.99 as relative time-to-event measures, not percentages of patients.

Step 4: Read the interval

Ask how much uncertainty surrounds each HR and whether the interval includes the null value of 1.00.

Step 5: Read the P-value

Use the P-value as evidence relative to a null hypothesis, not as a measure of treatment magnitude.

Step 6: Check what is missing

Look for medians, absolute survival estimates, event counts, subgroup analyses, and model assumptions before making broader conclusions.

21. Statistical Interpretation vs Clinical Interpretation

Statistical interpretation

The registry-reported primary analyses report a PFS HR of 0.63 with a 99% two-sided CI of 0.44–0.91 and P = 0.0012, and an OS HR of 0.99 with a 96% two-sided CI of 0.78–1.24 and P = 0.9056.

Clinical interpretation

The clinical meaning requires consideration of the endpoint definitions, absolute event probabilities, duration of benefit, adverse events, and other treatment-pathway information. Those additional quantities are not reported in the ClinicalTrials.gov record used for this page.

This distinction is deliberate. Statistical evidence describes the observed comparison and its uncertainty. Clinical interpretation requires a broader assessment of what the endpoint means to patients and how the treatment's benefits and harms compare in context.

22. Longitudinal Trial History

2018-06-10 · Trial start

Study initiated

The COSMIC-312 trial began on June 10, 2018, according to the ClinicalTrials.gov record.

2021-11-30 · Primary completion

Primary completion date

The ClinicalTrials.gov record lists November 30, 2021 as the primary completion date.

Results posted · Registry analysis

Primary statistical results

The registry contains two posted primary statistical analyses: one for PFS and one for OS, both using the log-rank test and hazard ratio.

23. Related Tutorials

Learn more about the statistical methods used in this trial:

24. Related Calculators

25. Sources

Continue with Clinical Biostats statistical methods

Use the trial's endpoints as a practical starting point for learning survival analysis, hazard ratios, confidence intervals, log-rank testing, and intention-to-treat analysis.

26. Record Summary

COSMIC-312 is a randomized phase 3, three-arm, parallel-group trial in advanced hepatocellular carcinoma without previous systemic anticancer therapy. The ClinicalTrials.gov record reports two primary time-to-event analyses comparing cabozantinib plus atezolizumab with sorafenib. PFS was analyzed in the specified PITT population of the first 372 participants randomized to the experimental and control arms, producing an HR of 0.63 with a 99% two-sided CI of 0.44–0.91 and P = 0.0012. OS was analyzed in the ITT population of all participants randomized to the experimental and control arms, producing an HR of 0.99 with a 96% two-sided CI of 0.78–1.24 and P = 0.9056.

The most important statistical lesson is that these estimates must be interpreted in their proper context. PFS and OS use different event definitions, the registry specifies different analysis populations, and the confidence intervals use different confidence levels. The HR communicates a relative time-to-event effect; the confidence interval communicates uncertainty around that estimate; and the P-value addresses evidence relative to a null hypothesis. None of these quantities alone describes absolute clinical benefit.

Clinical Biostats methodology: A trial-results page should reconstruct the statistical story without filling gaps with unsupported numbers. For COSMIC-312, that means preserving the registry's endpoint definitions, analysis populations, log-rank methods, hazard ratios, confidence intervals, P-values, and reported safety counts while clearly identifying analyses and design features that are not contained in the ClinicalTrials.gov record.