IMvigor130: Complete Statistical Analysis of Atezolizumab in Urothelial Carcinoma

1. Trial at a Glance

IMvigor130 was a randomized, parallel, double-masked phase 3 trial with three treatment arms and 1213 enrolled participants. The trial evaluated atezolizumab both in combination with gemcitabine plus carboplatin or cisplatin and as monotherapy, with placebo plus the same platinum-based chemotherapy as the comparator.

2. Clinical Question

The statistical question was whether adding atezolizumab to platinum-based chemotherapy, or using atezolizumab alone, changed prespecified time-to-event outcomes relative to placebo plus platinum-based chemotherapy in participants with untreated locally advanced or metastatic urothelial carcinoma.

The primary comparisons were not all identical. One primary endpoint compared placebo plus chemotherapy with atezolizumab plus chemotherapy for investigator-assessed PFS. Two primary endpoints addressed OS: one compared the combination arm with placebo plus chemotherapy, and another compared atezolizumab monotherapy with placebo plus chemotherapy.

3. Trial Design

4. Randomization, Stratification, and Analysis Populations

The posted primary analyses used intention-to-treat populations. The registry specifies stratification in the primary analyses, with the exact factors differing according to the comparison.

This distinction is important because the monotherapy comparison does not use exactly the same analysis population definition as the combination comparison. An analysis population is part of the estimand's context: the hazard ratio describes the randomized comparison within the population actually specified for that analysis, not automatically every participant enrolled in the overall study.

5. Primary Endpoints

All three registered primary endpoints are time-to-event endpoints. This means that the analysis must account for both the timing of events and participants who have not experienced the event by the end of their available follow-up.

6. Statistical Methodology

Kaplan-Meier framework for time-to-event outcomes

The trial's PFS and OS endpoints are naturally expressed as time from randomization to an event. Kaplan-Meier estimation is the standard descriptive framework for such data because participants may have different lengths of follow-up and some observations are censored before the event occurs.

Log-rank test

The registry reports log-rank testing for the primary investigator-assessed PFS comparison and for the primary OS comparison involving atezolizumab monotherapy. A log-rank test evaluates whether the observed time-to-event experience differs between randomized groups over follow-up, while accounting for censoring.

Cox proportional-hazards model

The primary OS comparison of atezolizumab plus chemotherapy versus placebo plus chemotherapy used a Cox proportional-hazards model. The effect measure reported was a hazard ratio with a two-sided 95% confidence interval.

Stratified analysis

The posted analyses incorporated stratification factors. For the chemotherapy-containing primary comparisons, these included enrollment stage, PD-L1 status, Bajorin risk score or presence of liver metastases, and investigator choice of chemotherapy. For the monotherapy comparison, the registry identifies PD-L1 status, Bajorin risk score or presence of liver metastases, and investigator choice of chemotherapy.

Wald / z-test

Several secondary event-free-rate analyses used a Wald / z-test, reported in the registry as a Z-test. These analyses express the treatment comparison as a difference in event-free rates at a specified time point rather than as a hazard ratio.

Intention-to-treat analysis

The primary efficacy analyses were based on ITT populations as defined for each comparison. The central statistical advantage is preservation of the treatment assignment created by randomization: participants remain classified according to their randomized arm for the specified efficacy analysis rather than being reassigned according to treatment received.

7. Primary Results: Investigator-Assessed Progression-Free Survival

The first primary endpoint compared placebo plus gemcitabine and carboplatin/cisplatin with atezolizumab plus gemcitabine and carboplatin/cisplatin. The posted analysis used a log-rank test in the ITT population with stratification.

8. Primary Results: Overall Survival With Atezolizumab Plus Chemotherapy

The second primary endpoint compared atezolizumab plus gemcitabine and carboplatin/cisplatin with placebo plus gemcitabine and carboplatin/cisplatin. The posted analysis used a Cox proportional-hazards model with ITT analysis and stratification.

9. Primary Results: Overall Survival With Atezolizumab Monotherapy

The third primary endpoint compared atezolizumab monotherapy with placebo plus gemcitabine and carboplatin/cisplatin. The posted analysis used a log-rank test in the specified Stage 2 ITT population and included a one-sided testing note in the analysis text, while the reported confidence interval was two-sided.

10. Secondary Efficacy Results

The registry contains additional formal analyses of PFS, OS event-free rates, and patient-reported deterioration endpoints. These analyses are reported separately from the three primary endpoints because they use different endpoint definitions, time frames, analysis populations, or effect measures.

IRF-PFS

IRF-PFS was assessed from randomization to first documented disease progression or death from any cause, with a time frame of up to 35 months. The analysis used the ITT population, stratification, and a Cox proportional-hazards model.

Year-1 OS Event-Free Rate: Combination Comparison

The reported estimate is a difference in event-free rates rather than a hazard ratio. The 95% confidence interval extends from a negative difference to a positive difference, so the estimate should be interpreted as an uncertain time-specific contrast rather than as a fixed treatment effect.

Year-1 OS Event-Free Rate: Monotherapy Comparison

Year-1 PFS Event-Free Rate

The registry expresses this secondary result as a difference in PFS event-free rate, not as a hazard ratio. The estimate of -8.29 percentage points is accompanied by a 95% CI from -14.45 to -2.13. The sign of a difference depends on the order in which the groups are defined, so it should always be read together with the comparison labels.

Patient-Reported Global Health Status Deterioration

These endpoints measured time to deterioration in global health status using the EORTC QLQ-C30 score, with a time frame of up to approximately 73 months. The first comparison used a PRO-evaluable population with the registry's specified baseline and follow-up requirements; the second used an evaluable population with a baseline and at least one post-baseline assessment.

Patient-Reported Physical Function Deterioration

These analyses evaluated time to deterioration in physical function as measured by the EORTC QLQ-C30 score, with follow-up up to approximately 73 months. The combination comparison used the PRO-evaluable population, while the monotherapy comparison used the registry's evaluable population.

11. Reading the Secondary Results as a Statistical Set

The secondary analyses illustrate why a trial-results page should distinguish endpoint type from the numerical significance of an individual P-value. The registry reports several different estimands: hazard ratios for time-to-deterioration outcomes and IRF-PFS, and differences in event-free rates for specified one-year outcomes.

12. Safety Results

The ClinicalTrials.gov record reports serious adverse events by randomized arm as affected participants divided by the corresponding at-risk population.

The safety denominators are explicitly reported in the ClinicalTrials.gov record and differ across arms. That makes the numerator alone insufficient for comparing the frequency of serious adverse events. The affected and at-risk counts should be read together.

13. Statistical Methods Explained

Why was a log-rank test used?

Both PFS and OS are time-to-event outcomes, so participants may experience an event at different times or remain event-free at their last observation. The log-rank test compares the observed event patterns between randomized groups while incorporating the available follow-up and censoring information.

What does a hazard ratio of 0.82 mean?

For the primary investigator-assessed PFS comparison, an HR of 0.82 represents an estimated instantaneous progression-or-death rate about 18% lower in the atezolizumab-plus-chemotherapy group relative to the placebo-plus-chemotherapy group under the reported analysis. It does not mean that 18% more patients were cured or that each patient's individual risk fell by exactly 18%.

Why is the confidence interval important?

The estimate alone does not show how precisely the treatment effect has been estimated. For example, the primary PFS HR of 0.82 has a 95% CI of 0.70–0.96. The interval communicates uncertainty around the estimated relative hazard under the statistical model and sampling framework.

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

A P-value measures evidence against a null hypothesis under a specified statistical model and testing procedure. It is affected by both the observed effect and the amount of information in the analysis. Effect size is conveyed by measures such as the HR or difference in event-free rate, while the confidence interval provides information about precision.

Why does the monotherapy analysis need special attention to its analysis population?

The monotherapy OS endpoint was defined using an ITT population restricted to participants concurrently enrolled in Stage 2 and those randomized at the time of approval of Protocol WO30070 v6. Therefore, the HR of 0.98 is a result for that specified analysis population rather than an automatically applicable estimate for every participant enrolled in IMvigor130.

What is the difference between a hazard ratio and an event-free-rate difference?

A hazard ratio summarizes a relative time-to-event comparison, whereas an event-free-rate difference compares the proportion remaining event-free at a specified time point. The Year-1 PFS analysis reports a difference of -8.29, while the primary PFS analysis reports an HR of 0.82. These quantities are not interchangeable.

Why does the registry mention one-sided testing when the confidence interval is two-sided?

The monotherapy OS analysis notes one-sided testing, while its reported CI is explicitly two-sided 95%. A hypothesis test and a confidence interval are related but distinct statistical reporting elements. The safest interpretation is to preserve the registry's wording rather than silently converting one convention into another.

14. Confidence Intervals, P-Values, and Effect Measures

For the IMvigor130 primary endpoints, these distinctions are particularly important because the trial combines different formal methods. The primary PFS and monotherapy OS comparisons use log-rank testing, while the combination OS analysis uses a Cox proportional-hazards model. Secondary endpoints additionally use Wald / z-tests for specified event-free rates.

15. Stratification and the Role of Baseline Factors

Stratified analysis allows the treatment comparison to account for prespecified factors that define clinically or statistically relevant strata. In the registry-reported IMvigor130 analyses, these include PD-L1 status, Bajorin risk score or presence of liver metastases, investigator choice of chemotherapy, and, for the specified chemotherapy-containing analyses, enrollment stage.

Stratification should not be confused with subgroup analysis. A stratified primary analysis uses the strata as part of the treatment comparison. A subgroup analysis asks whether treatment effects appear different within particular subsets. The ClinicalTrials.gov record identifies stratification, but do not provide a complete set of subgroup interaction tests.

16. Time-to-Event Endpoints and Censoring

PFS and OS are not ordinary binary outcomes. A participant who has not progressed or died by the end of follow-up has not necessarily been observed indefinitely; that participant contributes information up to the point at which follow-up ends or censoring occurs.

For the primary PFS endpoint, the event was the first documented disease progression determined by the investigator using RECIST v1.1 or death from any cause, whichever occurred first. For OS, the event was death due to any cause.

17. Comparing the Three Primary Endpoints

These three results should be interpreted as three distinct prespecified primary comparisons rather than collapsed into a single overall treatment effect. The PFS endpoint concerns progression or death, whereas the two OS endpoints concern death from any cause. The two OS comparisons also involve different treatment contrasts and, for the monotherapy comparison, a different specified ITT population.

18. Limitations and Interpretation Issues

• Multiple primary endpoints: the registry identifies three primary endpoints. The ClinicalTrials.gov record does not specify the complete multiplicity-control strategy, so individual P-values should be interpreted in the context of the overall trial design rather than in isolation.
• Different treatment comparisons: the primary PFS and one primary OS analysis concern atezolizumab plus chemotherapy, while the other primary OS analysis concerns atezolizumab monotherapy.
• Different analysis populations: the monotherapy OS endpoint has a specifically restricted Stage 2 ITT definition that differs from the chemotherapy-containing primary analyses.
• Hazard-ratio assumptions: Cox-model interpretation relies on the model's proportional-hazards framework. The ClinicalTrials.gov record does not provide diagnostics demonstrating whether that assumption held over the entire follow-up period.
• Censoring: time-to-event analyses depend on how censored observations contribute information. Participant-level censoring data are not provided here.
• Secondary endpoints: several secondary outcomes use different populations and effect measures, including PRO-evaluable populations and one-year event-free-rate differences.
• One-sided versus two-sided reporting: the monotherapy OS analysis text notes one-sided testing while reporting a two-sided 95% CI. These components should be retained as reported rather than harmonized by assumption.
• Safety versus efficacy populations: the reported serious-adverse-event denominators differ from the total trial enrollment and should not be treated as efficacy analysis populations.

19. Why This Trial Matters Statistically

IMvigor130 is a useful teaching case because it places several core clinical-trial methods into a single randomized phase 3 framework. The trial contains multiple treatment contrasts, three primary time-to-event endpoints, stratified analyses, ITT populations, Cox and log-rank methods, event-free-rate comparisons, and patient-reported time-to-deterioration outcomes.

The statistical lesson is that a complex randomized trial should be read endpoint by endpoint. The design establishes the randomized comparisons, the analysis population defines who contributes to a particular estimate, the statistical method defines how event timing and censoring are handled, and the effect measure determines what the numerical result actually means.

20. A Practical Framework for Interpreting IMvigor130

21. Overall Statistical Interpretation

Taken together, these results illustrate why the phrase "the trial's hazard ratio" is statistically inadequate. IMvigor130 has several hazard ratios, each attached to a particular endpoint, comparison, population, time frame, and analysis method. The appropriate interpretation begins with those definitions rather than with the numerical estimate alone.

22. Statistical Concepts in This Trial

Learn more about the methods used in this trial:

23. Related Statistical Calculators

Use these statistical topics to explore the calculations underlying the reported analyses:

24. Sources

• ClinicalTrials.gov: IMvigor130, NCT02807636.
• Linked publication: PubMed record — PMID 38280376.
• Linked publication: PubMed record — PMID 38101433.
• Linked publication: PubMed record — PMID 38101431.
• Linked publication: PubMed record — PMID 37954037.
• Linked publication: PubMed record — PMID 32416780.

25. Record Summary

IMvigor130 provides a detailed example of a randomized phase 3 trial in which several treatment comparisons and endpoint types must be interpreted separately. The three registered primary endpoints are all time-to-event outcomes, but they use different treatment contrasts and, for the monotherapy OS comparison, a specifically restricted Stage 2 ITT population. The posted analyses use log-rank testing, Cox proportional-hazards modeling, stratification, and Wald / z-tests for selected secondary event-free-rate comparisons.

The primary investigator-assessed PFS analysis reported an HR of 0.82 (95% CI 0.70–0.96; P = 0.0073). The primary OS analysis for atezolizumab plus chemotherapy reported an HR of 0.85 (95% CI 0.73–1.0; P = 0.0230). The primary OS analysis for atezolizumab monotherapy reported an HR of 0.98 (95% CI 0.82–1.16; P = 0.3968). Secondary analyses included IRF-PFS, one-year OS and PFS event-free rates, and time to deterioration in global health status and physical function.

The most important statistical lesson is methodological rather than numerical: a treatment effect cannot be interpreted correctly without identifying the exact endpoint, analysis population, treatment contrast, statistical method, effect measure, confidence interval, and testing framework. IMvigor130 demonstrates this principle particularly clearly because its three primary endpoints do not represent a single homogeneous comparison.