IMblaze370: Complete Statistical Analysis of Cobimetinib Plus Atezolizumab in Metastatic Colorectal Cancer

1. Trial at a Glance

IMblaze370 was a randomized, open-label, parallel-group phase 3 trial with 363 participants and three treatment arms. The registered primary endpoint was overall survival, a time-to-event endpoint analyzed in the intention-to-treat population using stratified log-rank testing and hazard-ratio estimation.

2. Clinical Question

The registered study compared two atezolizumab-containing strategies with regorafenib in participants with metastatic colorectal adenocarcinoma. The primary statistical question was framed as a superiority comparison for overall survival.

3. Trial Design

4. Endpoints

The registry lists one primary endpoint: overall survival. Secondary analyses posted in the ClinicalTrials.gov record includes progression-free survival and investigator-assessed objective response.

5. Statistical Methodology

Intention-to-treat analysis

The primary and time-to-event secondary analyses were performed on the intention-to-treat (ITT) population. The ITT principle analyzes participants according to randomized treatment assignment rather than restricting the efficacy comparison to participants who remain on treatment.

Stratified log-rank test

The registry reports a stratified log-rank test for the primary overall-survival comparisons and for the stratified progression-free-survival comparisons. The registry-reported analysis notes identify two stratification factors: extended RAS mutation status and time since diagnosis of first metastasis.

Cox regression for hazard ratios

The registry notes state that hazard ratios for the stratified analyses were estimated using stratified Cox regression. For the unstratified analyses, hazard ratios were estimated using unstratified Cox regression.

Cochran-Mantel-Haenszel analysis

For objective response, the registry reports a stratified Cochran-Mantel-Haenszel analysis. The comparison was expressed as a difference in response rates, with a two-sided 95% confidence interval. The registry states that the stratification factors were extended RAS mutation status and time since diagnosis of first metastasis.

Hauck-Anderson confidence interval

For the response-rate differences, the registry notes state that the 95% confidence interval was constructed using the Hauck-Anderson method. This is distinct from the hypothesis-testing method: the Cochran-Mantel-Haenszel procedure provides the reported stratified comparison, while the Hauck-Anderson method is specified for construction of the confidence interval around the response-rate difference.

6. Statistical Methods Explained

Why use a stratified log-rank test?

The trial's primary endpoint is a time-to-event outcome, and the registry identifies two prespecified factors used for stratification: extended RAS mutation status and time since diagnosis of first metastasis. Stratification allows the survival comparison to account for these factors while preserving the randomized comparison within the defined strata.

What does an overall-survival hazard ratio of 1.00 mean?

For the reported regorafenib versus cobimetinib-plus-atezolizumab comparison, an HR of 1.00 is the estimated ratio of the instantaneous event rates between the two groups under the fitted stratified Cox model. An HR of exactly 1.00 does not mean that every participant had the same survival time. It summarizes a relative rate over the analyzed time-to-event data.

Why are confidence intervals important?

A point estimate is only one summary of the observed data. The 95% confidence interval communicates statistical uncertainty around the estimated effect under the specified analysis framework. For example, the reported stratified OS HR of 1.00 has a 95% CI of 0.73 to 1.38. The interval therefore gives substantially more information about precision than the point estimate alone.

Why does the p-value not measure effect size?

A p-value evaluates how compatible the observed data are with the null hypothesis under the specified testing procedure. It does not tell us how large the treatment effect is. The effect size is conveyed by the hazard ratio or response-rate difference, while the confidence interval describes uncertainty around that estimate.

Why use the Cochran-Mantel-Haenszel test for response?

Objective response is binary: a participant either meets the RECIST-defined complete-response or partial-response criterion or does not. The stratified Cochran-Mantel-Haenszel framework provides a way to compare response rates while accounting for the trial's specified stratification factors.

Why does censoring matter for overall survival?

Not every participant has an observed death by the analysis date. The registry therefore specifies censoring rules for participants who were not reported as dead and a separate rule for participants without post-baseline information. Survival methods use the information available up to the censoring time rather than treating censored participants as if their event time were known.

7. Primary Results: Overall Survival

Four primary endpoint analyses are posted in the ClinicalTrials.gov record. They comprise two treatment comparisons evaluated with stratified analysis and corresponding unstratified analyses.

Regorafenib vs Cobimetinib + Atezolizumab — Stratified Analysis

Regorafenib vs Atezolizumab — Stratified Analysis

Unstratified Analyses

The registry also reports unstratified log-rank analyses, with hazard ratios estimated using unstratified Cox regression. These analyses provide a useful sensitivity perspective because they remove the stratification from the formal model.

8. Secondary Results: Progression-Free Survival

Progression-free survival was analyzed in the ITT population using the same general time-to-event framework. The registry defines PFS as time from randomization up to disease progression or death due to any cause, with investigator assessment according to RECIST v1.1.

Regorafenib vs Cobimetinib + Atezolizumab — Stratified Analysis

Regorafenib vs Atezolizumab — Stratified Analysis

Unstratified PFS Analyses

9. Secondary Results: Objective Response

The registry reports investigator-assessed objective response as the percentage of evaluable participants in the ITT population with measurable disease at baseline who achieved a complete response or partial response according to RECIST Version 1.1.

Regorafenib vs Cobimetinib + Atezolizumab

Regorafenib vs Atezolizumab

10. Results Overview

The posted analyses show a consistent statistical structure across the major efficacy outcomes: overall survival and progression-free survival were analyzed as time-to-event endpoints, while objective response was analyzed as a binary endpoint. Both stratified and unstratified analyses are available for OS and PFS.

11. Serious Adverse Events

The ClinicalTrials.gov record reports serious adverse events by treatment arm as affected participants divided by the corresponding number at risk.

The reported serious-adverse-event figures should be read together with their denominators. The combination arm has 71 affected participants among 183 at risk, whereas the regorafenib and atezolizumab arms each have 90 participants in the registry-reported safety summary. The affected counts alone therefore should not be used to compare event frequency across arms.

12. Stratification and Why It Matters

The registry identifies extended RAS mutation status and time since diagnosis of first metastasis as stratification factors for the stratified efficacy analyses. These variables are incorporated into the statistical comparison rather than simply listed as descriptive baseline characteristics.

Stratification is especially useful in a randomized multi-arm study when important prognostic factors are known in advance. The statistical objective is not to claim that the factors cause the outcome, but to perform the treatment comparison within a framework that respects the prespecified strata.

13. Multiplicity and Multiple Comparisons

IMblaze370 has one registered primary endpoint but the registry-reported primary analyses include two treatment comparisons against regorafenib, each evaluated with stratified and unstratified approaches. The secondary endpoint analyses add PFS and objective response comparisons.

14. Statistical Interpretation of the Hazard Ratios

15. Stratified vs Unstratified Results

The availability of both stratified and unstratified analyses provides an instructive comparison of model specifications.

The point estimates are similar across the two approaches in all four comparisons. This does not make stratified and unstratified analyses statistically identical. Rather, it shows that the numerical direction and approximate magnitude of the estimated hazard ratios are similar in the ClinicalTrials.gov record.

The stratified analysis remains particularly important for interpreting the prespecified primary comparison because it incorporates the documented stratification factors. The unstratified results are best understood as a separate analysis rather than as a replacement for the stratified analysis.

16. Time-to-Event Analysis in This Trial

Both OS and PFS are time-to-event endpoints. Their defining feature is that the analysis uses not only whether an event occurred, but also the amount of observed follow-up before the event or censoring.

17. Analysis Populations

The registry distinguishes the analysis populations by endpoint. The primary and time-to-event analyses were performed on the ITT population. Objective response was analyzed among evaluable participants in the ITT population with measurable disease at baseline, as determined by the investigator.

This distinction is statistically important. Restricting a response analysis to participants with measurable disease changes the analysis population relative to the full randomized population. That does not make the response endpoint invalid; it simply means its estimate answers a narrower question than an efficacy analysis performed across all randomized participants.

18. What the P-values Do — and Do Not — Tell Us

A p-value does not tell us the probability that the null hypothesis is true, the probability that the treatment has no clinical effect, or the magnitude of an observed treatment effect. Those questions require the effect estimate, its confidence interval, and the clinical context of the endpoint.

19. Confidence Intervals as Measures of Precision

For hazard ratios, 1 is the null value because an HR of 1 corresponds to equal estimated instantaneous event rates. For a difference in response rates, 0 is the null value because a difference of zero means no difference in the estimated response rates.

The confidence intervals in the ClinicalTrials.gov record therefore provide an immediate visual indication of whether the corresponding null value is contained within the reported interval. They also show how much uncertainty surrounds each point estimate.

20. Trial Timeline

21. What Is and Is Not Reported in the Supplied Registry Data

The ClinicalTrials.gov record is unusually useful for statistical teaching because they contain formal analyses, effect measures, confidence intervals, P-values, analysis populations, and methodological notes. At the same time, several common trial-results components are not included in the ClinicalTrials.gov recordset.

22. Limitations

• Registry-level summary: the ClinicalTrials.gov record provides statistical-analysis summaries rather than the underlying participant-level dataset.
• No survival medians: median OS and median PFS values are not included in the ClinicalTrials.gov record and therefore are not reported here.
• No Kaplan-Meier reconstruction: event and censoring times are not provided, so survival curves cannot be validly reconstructed from the available summaries.
• Hazard-ratio assumptions: Cox regression is model-based. The ClinicalTrials.gov record does not document an assessment of the proportional-hazards assumption.
• Multiple comparisons: the ClinicalTrials.gov record does not identify an alpha-allocation or multiplicity-adjustment procedure. Individual P-values should therefore not be assigned an undocumented familywise interpretation.
• Different analysis populations: response was analyzed in evaluable ITT participants with measurable disease at baseline, whereas OS and PFS were analyzed in the ITT population.
• Unstratified analyses: the unstratified estimates are useful as additional analyses, but they do not replace the prespecified stratified framework described in the registry notes.
• Safety scope: only serious adverse-event affected/at-risk counts are reported. A complete safety profile cannot be inferred from these figures.
• Missing-data methods: no imputation strategy is reported in the ClinicalTrials.gov record. The OS endpoint does provide explicit censoring rules for participants without an observed death.
• Generalizability: the ClinicalTrials.gov record does not provide a detailed baseline table, so the representativeness of the randomized population cannot be evaluated here beyond the registered disease population.

23. Why This Trial Matters Statistically

IMblaze370 provides a compact teaching example of how a multi-arm randomized trial can use different statistical methods for different endpoint types while maintaining a coherent treatment-comparison framework.

24. Related Tutorials

Learn more about the methods used in this trial:

25. Related Calculators

26. Sources

• ClinicalTrials.gov: NCT02788279 — IMblaze370.
• Linked publication: PubMed record — PMID 37741832.
• Linked publication: PubMed record — PMID 36310331.
• Linked publication: PubMed record — PMID 31003911.

27. Record Summary

IMblaze370 demonstrates the statistical structure of a randomized phase 3 oncology trial with three parallel treatment arms, a registered time-to-event primary endpoint, ITT efficacy analysis, prespecified stratification, stratified and unstratified Cox-based hazard-ratio estimates, and secondary binary-response analyses using the Cochran-Mantel-Haenszel framework. The ClinicalTrials.gov record also illustrate why treatment comparisons should be interpreted through the combined lens of effect estimates, confidence intervals, P-values, analysis populations, stratification, and endpoint definitions.

The most important statistical distinction is between the types of estimands being reported. OS and PFS are time-to-event outcomes summarized through hazard ratios, whereas objective response is a binary outcome summarized through a difference in response rates. The reported serious-adverse-event counts provide a separate safety perspective and should not be conflated with the efficacy estimates.