This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record. Numerical trial results on this page are restricted to the ClinicalTrials.gov record.
1. Trial at a Glance
STEAM was a randomized, open-label, parallel-group phase 2 trial evaluating three first-line treatment strategies in metastatic colorectal cancer: concurrent FOLFOXIRI plus bevacizumab, sequential FOLFOXIRI plus bevacizumab, and FOLFOX plus bevacizumab.
| Feature | STEAM |
|---|---|
| Phase | Phase 2 |
| Condition | Colorectal Neoplasms |
| Brief title | Sequential and Concurrent FOLFOXIRI/Bevacizumab Regimens Versus FOLFOX/Bevacizumab in First-Line Metastatic Colorectal Cancer |
| Design | Randomized, parallel-group, unmasked |
| Primary purpose | Treatment |
| Enrollment | 280 |
| Arms | 3 |
| Primary endpoints | 2: one binary endpoint and one time-to-event endpoint |
| Results posted | Yes |
| Lead sponsor | Hoffmann-La Roche |
| Sponsor type | Industry |
| ClinicalTrials.gov | NCT01765582 |
2. Clinical Question
The statistical question was whether treatment strategies incorporating FOLFOXIRI plus bevacizumab differed from FOLFOX plus bevacizumab in first-line metastatic colorectal cancer, using overall response during first-line therapy and progression-free survival during first-line therapy as the two registered primary endpoints.
Population
Participants with colorectal neoplasms enrolled in a first-line metastatic colorectal cancer treatment trial.
Intervention strategies
Concurrent FOLFOXIRI plus bevacizumab and sequential FOLFOXIRI plus bevacizumab.
Comparator
FOLFOX plus bevacizumab.
Primary question
How do the randomized treatment strategies compare with respect to first-line overall response and progression-free survival?
3. Trial Design
Concurrent FOLFOXIRI + Bevacizumab
- FOLFOXIRI
- Bevacizumab
Sequential FOLFOXIRI + Bevacizumab
- Sequential FOLFOXIRI
- Bevacizumab
FOLFOX + Bevacizumab
- FOLFOX
- Bevacizumab
The intervention data list 5-fluorouracil, bevacizumab, capecitabine, irinotecan, folinic acid, and oxaliplatin as study interventions. The registry describes the overall study as a three-arm randomized parallel design.
4. Endpoints
| Registered endpoint | Type | Time frame | Registry definition |
|---|---|---|---|
| Percentage of Participants With Overall Response During First-Line Therapy (ORR1) | Binary | Randomization up to disease progression during first-line therapy or death, whichever occurs first | ORR1 was the percentage of participants with complete response (CR) or partial response (PR) during first-line therapy as assessed by the investigator according to RECIST v1.1. The registry definition further specifies CR as disappearance of all extranodal target lesions with all pathological lymph nodes decreased to <10 mm in short axis. |
| Progression-Free Survival During First-Line Therapy (PFS1) | Time-to-event | Randomization up to disease progression during first-line therapy or death, whichever occurs first | PFS1 was defined as time from randomization to the first occurrence of disease progression during first-line therapy, as assessed by the Investigator using RECIST v1.1, or death from any cause, whichever occurs first. |
5. Primary Results
ClinicalTrials.gov contains formal statistical analyses for both registered primary endpoints. Both analyses use the intention-to-treat population, defined by the registry as all randomized participants regardless of whether they received any dose of study treatment.
Overall Response During First-Line Therapy (ORR1)
Odds ratio for overall response
90% two-sided CI: 0.96–2.71 · P = 0.132
Arm A: Concurrent FOLFOXIRI + Bevacizumab vs Arm C: FOLFOX + Bevacizumab
The registry analyzed ORR1 using a Cochran-Mantel-Haenszel test. The comparison was between Arm A, concurrent FOLFOXIRI plus bevacizumab, and Arm C, FOLFOX plus bevacizumab. The analysis was stratified by extent of metastatic disease—liver-limited disease versus non-liver-limited disease—and tumor location—right versus left—after correction post-randomization.
An odds ratio of 1.62 means that the estimated odds of achieving an overall response were 1.62 times as high in Arm A as in Arm C under the analyzed comparison. Equivalently, the estimated odds were higher in the concurrent FOLFOXIRI plus bevacizumab group, although an odds ratio is not the same as a risk ratio or a difference in response percentages.
The 90% two-sided confidence interval of 0.96–2.71 describes uncertainty around the estimated odds ratio. Because the interval includes 1, the data are compatible with no difference in the underlying odds as well as with a range of positive effects. The interval also shows that the estimate is not highly precise.
The P-value of 0.132 describes the compatibility of the observed result with the null hypothesis under the specified statistical test; it does not measure the size or clinical importance of the effect. A P-value is therefore not a substitute for the odds ratio and its confidence interval.
This was an ITT analysis and a stratified categorical comparison. The stratification factors are important because the reported odds ratio is adjusted through the Cochran-Mantel-Haenszel framework rather than being a simple unadjusted odds ratio.
Progression-Free Survival During First-Line Therapy (PFS1)
Hazard ratio for progression or death
90% two-sided CI: 0.53–0.88 · P = 0.005
Arm C: FOLFOX + Bevacizumab vs Arms A + B: pooled FOLFOXIRI + Bevacizumab
The PFS1 analysis compared Arm C with the pooled FOLFOXIRI plus bevacizumab arms, A and B. The registry reports a log-rank test with a hazard ratio as the effect measure. The analysis was stratified by extent of metastatic disease and tumor location after correction post-randomization.
The reported hazard ratio of 0.68 is for FOLFOX + bevacizumab versus pooled FOLFOXIRI + bevacizumab. Because the comparison is expressed in that direction, an HR below 1 indicates a lower estimated instantaneous rate of progression or death for the FOLFOX + bevacizumab group relative to the pooled FOLFOXIRI + bevacizumab groups. Expressed in the reverse direction for educational interpretation, the pooled FOLFOXIRI + bevacizumab strategy corresponds to an estimated hazard approximately 32% lower than FOLFOX + bevacizumab, because 1 − 0.68 = 0.32.
The 90% two-sided confidence interval of 0.53–0.88 gives the uncertainty range around the reported hazard-ratio estimate. It does not describe the range of individual patients' outcomes and should not be interpreted as a range of individual risk reductions.
The P-value of 0.005 measures evidence against the null hypothesis under the specified log-rank framework. It does not quantify the magnitude of the treatment effect. The hazard ratio and confidence interval provide that effect-size and precision information.
Because PFS1 is a time-to-event endpoint, the analysis is also affected by censoring and by the assumptions underlying interpretation of a hazard ratio. A hazard ratio summarizes relative event rates over follow-up; it is not itself a statement about median PFS, a fixed-time survival probability, or the probability that an individual patient will benefit.
6. Statistical Methodology
Intention-to-treat analysis
Both posted primary analyses used the ITT population. The registry defines this population as all randomized participants regardless of whether they received any dose of study treatment.
This matters because randomization establishes the treatment groups before treatment exposure and subsequent clinical events occur. An ITT analysis therefore preserves the randomized comparison rather than redefining participants according to treatment received. It also means that the efficacy analysis is not restricted only to participants who completed or tolerated treatment.
For STEAM, the registry explicitly defines the ITT population as all randomized participants, regardless of whether they received any dose of study treatment.
Cochran-Mantel-Haenszel test
The ORR1 comparison used the Cochran-Mantel-Haenszel framework. This method is designed for categorical outcomes when the comparison should account for one or more stratification variables.
In this trial, the analysis was stratified by extent of metastatic disease and tumor location. The resulting odds ratio therefore represents a stratified comparison rather than simply pooling every participant into one unadjusted 2×2 table.
The purpose of stratification is to compare treatment groups while accounting for the specified categorical strata. It does not mean that the treatment effect is identical within every possible clinical subgroup.
Log-rank test
PFS1 was analyzed using the log-rank test. This is a standard method for comparing time-to-event distributions between randomized groups while accounting for the timing of events and right-censored observations.
Unlike a simple comparison of proportions, a time-to-event analysis uses information about when progression or death occurs. Participants who have not experienced an event by their available follow-up can contribute information until their censoring time.
Hazard ratio
The hazard ratio is the reported effect measure for PFS1. It compares the estimated instantaneous event rates between the groups in the direction specified by the registry analysis.
For STEAM's PFS1 analysis, the numerator group is Arm C and the denominator group is the pooled Arms A + B. The direction of the comparison must therefore be checked before interpreting an HR below 1.
Confidence intervals
Both primary analyses report 90% two-sided confidence intervals. The confidence interval is an uncertainty statement about the estimated effect under the statistical model and sampling framework. It should not be read as the probability that the true effect lies inside the particular interval after the study has been completed.
Superiority testing
The registry classifies both primary analyses as superiority hypotheses. This is different from a non-inferiority design: the objective is to evaluate evidence of a difference favoring one randomized strategy rather than to demonstrate that a treatment is not worse than a prespecified margin.
7. Stratification and the Direction of the Comparisons
One of the most important statistical details in STEAM is that the two primary analyses do not use the same treatment comparison.
| Endpoint | Comparison | Method | Effect measure |
|---|---|---|---|
| ORR1 | Arm A: Concurrent FOLFOXIRI + Bevacizumab vs Arm C: FOLFOX + Bevacizumab | Cochran-Mantel-Haenszel | Odds ratio = 1.62 |
| PFS1 | Arm C: FOLFOX + Bevacizumab vs Arms A + B: pooled FOLFOXIRI + Bevacizumab | Log-rank test | Hazard ratio = 0.68 |
This distinction prevents a common statistical reading error. The numerical direction of an effect estimate depends on which group appears first in the comparison. The ORR1 estimate is reported as Arm A relative to Arm C, whereas the PFS1 hazard ratio is reported as Arm C relative to pooled Arms A and B.
8. Safety Results
The ClinicalTrials.gov record reports serious adverse events by randomized arm as affected participants divided by participants at risk.
| Arm | Serious adverse events | Affected / at risk |
|---|---|---|
| Arm A: Concurrent FOLFOXIRI + Bevacizumab | 39 participants | 39/91 |
| Arm B: Sequential FOLFOXIRI + Bevacizumab | 42 participants | 42/90 |
| Arm C: FOLFOX + Bevacizumab | 43 participants | 43/90 |
These figures should be read as descriptive safety counts. The ClinicalTrials.gov record does not provide a formal statistical comparison of serious adverse-event rates, confidence intervals, or P-values, so no inferential safety comparison is presented here.
Arm A
39 of 91 participants were affected by serious adverse events.
Arm B
42 of 90 participants were affected by serious adverse events.
Arm C
43 of 90 participants were affected by serious adverse events.
What is not shown
The ClinicalTrials.gov record does not report a formal between-arm safety hypothesis test.
9. Statistical Methods Explained
Why use a Cochran-Mantel-Haenszel test for ORR1?
ORR1 is a binary endpoint: a participant either achieved the registry-defined complete or partial response or did not. The Cochran-Mantel-Haenszel method allows the categorical comparison to account for prespecified or specified strata. In STEAM, those strata were extent of metastatic disease and tumor location.
What does an odds ratio of 1.62 mean?
The ORR1 odds ratio is 1.62 for concurrent FOLFOXIRI plus bevacizumab versus FOLFOX plus bevacizumab. In odds terms, the estimated odds of response were 1.62 times as large in Arm A as in Arm C. This is not equivalent to saying that the response probability was 1.62 times as large.
Why is the confidence interval important?
The 90% two-sided CI of 0.96–2.71 shows the uncertainty surrounding the ORR1 estimate. Because the interval extends across 1, the data do not exclude a null odds ratio under this confidence-interval framework. The width of the interval also matters: an estimate should not be interpreted without considering its precision.
Why use a log-rank test for PFS1?
PFS1 is a time-to-event endpoint, so the timing of progression or death matters. The log-rank test compares the event-time distributions while incorporating the available follow-up from participants who are censored rather than treating every participant as if the endpoint were simply a binary outcome at a fixed date.
What does an HR of 0.68 mean when the comparison is reversed?
The registry reports the HR for FOLFOX plus bevacizumab versus pooled FOLFOXIRI plus bevacizumab. Therefore, the value below 1 describes a lower estimated hazard in the numerator group, FOLFOX plus bevacizumab. If the same relative comparison is expressed in the opposite direction, the interpretation changes. The key is to identify the numerator and denominator before interpreting the number.
Why is the P-value not the effect size?
A P-value describes the compatibility of the observed data with a null hypothesis under a specified statistical procedure. It does not tell us how large the treatment difference is. In STEAM, the effect sizes are represented by the OR of 1.62 for ORR1 and the HR of 0.68 for PFS1, while their confidence intervals communicate uncertainty around those estimates.
Why does ITT matter in a randomized trial?
The ITT population preserves the randomized treatment assignment for the efficacy comparison. In STEAM, the registry explicitly defines ITT as all randomized participants regardless of whether they received any dose of study treatment. This avoids redefining the treatment groups after randomization based on subsequent treatment exposure.
10. Trial Termination and Statistical Interpretation
The registry caveat states that STEAM was terminated on 12 November 2015 because the primary objective was not met. This is an important part of the statistical record because the reason and timing of trial termination can affect how the observed results are interpreted.
Trial start
The registry lists the study start date as 23 January 2013.
Trial terminated
The registry limitation states that the trial was terminated because the primary objective was not met.
Primary completion
the ClinicalTrials.gov record lists 14 March 2016 as the primary completion date.
Termination does not itself tell us the magnitude or direction of an endpoint estimate. Those questions must be answered from the actual statistical analyses. In STEAM, the posted primary analyses provide an ORR1 odds ratio of 1.62 and a PFS1 hazard ratio of 0.68, each with a 90% two-sided confidence interval and P-value.
11. Primary Endpoint Interpretation as a Statistical Pair
ORR1: binary evidence
The ORR1 analysis asks whether the odds of achieving a RECIST-defined response differed between the concurrent FOLFOXIRI plus bevacizumab arm and the FOLFOX plus bevacizumab arm.
PFS1: time-to-event evidence
The PFS1 analysis asks whether the timing of progression or death differed between FOLFOX plus bevacizumab and the pooled FOLFOXIRI plus bevacizumab arms.
The two endpoints capture different dimensions of treatment effect. ORR1 reduces the clinical response experience to a binary response classification over the registered first-line treatment period. PFS1 incorporates the timing of disease progression or death and therefore uses more of the longitudinal event information.
| Statistical feature | ORR1 | PFS1 |
|---|---|---|
| Endpoint type | Binary | Time-to-event |
| Analysis population | ITT | ITT |
| Comparison | Arm A vs Arm C | Arm C vs pooled Arms A + B |
| Method | Cochran-Mantel-Haenszel test | Log-rank test |
| Effect measure | Odds ratio | Hazard ratio |
| Estimate | 1.62 | 0.68 |
| Confidence interval | 90% two-sided: 0.96–2.71 | 90% two-sided: 0.53–0.88 |
| P-value | 0.132 | 0.005 |
| Stratification | Metastatic disease extent and tumor location | Metastatic disease extent and tumor location |
12. What the Effect Estimates Do — and Do Not — Mean
The ORR1 estimate of 1.62 means the estimated odds of response were 1.62 times as high in Arm A as in Arm C. It does not mean that 62% more participants responded, nor does it give a 62-percentage-point difference in response probability.
The PFS1 HR of 0.68 is reported for Arm C relative to pooled Arms A and B. It therefore represents a lower estimated instantaneous rate of progression or death for the numerator group. It does not mean that 32% of participants avoided progression, that each participant had exactly a 32% reduction in risk, or that the median PFS differed by 32%.
The 90% confidence intervals describe uncertainty around the corresponding effect estimates. They do not describe the range of outcomes among individual participants. The ORR1 interval crosses 1, while the PFS1 interval does not, but these observations should be interpreted together with the prespecified trial design, endpoint hierarchy, analysis population, and the fact that the trial was terminated.
The P-values of 0.132 and 0.005 answer a hypothesis-testing question under their respective statistical procedures. They should not be interpreted as probabilities that a treatment works, probabilities that the null hypothesis is true, or measures of effect magnitude.
13. Limitations
- Trial termination: the registry states that the trial was terminated because the primary objective was not met. This is part of the context in which the posted analyses should be interpreted.
- Different primary comparisons: the ORR1 and PFS1 analyses use different treatment comparisons. Directly comparing their numerical effect estimates without accounting for comparison direction would be misleading.
- Confidence-interval level: the posted primary analyses use 90% two-sided confidence intervals rather than the more commonly encountered 95% intervals.
- Time-to-event assumptions: interpretation of a hazard ratio depends on the statistical framework and censoring process. A single HR does not describe every feature of two survival distributions.
- Safety inference: serious adverse-event counts are reported descriptively by arm, but the ClinicalTrials.gov record does not provide a formal inferential comparison of these safety outcomes.
- Stratification: the CMH and log-rank analyses account for the specified stratification variables, but stratification does not establish that treatment effects are identical across all possible clinical subgroups.
- Multiplicity information: the ClinicalTrials.gov record identifies two primary endpoints and two posted primary analyses but does not provide a detailed multiplicity-adjustment scheme or alpha-allocation procedure.
- Interim-analysis information: the ClinicalTrials.gov record does not describe an interim-analysis schedule or formal alpha-spending procedure.
- Missing-data methodology: the ClinicalTrials.gov record does not report a specific imputation strategy. No imputation method has therefore been inferred.
- Bayesian methodology: no Bayesian method is identified in the ClinicalTrials.gov record.
- Generalizability: the ClinicalTrials.gov record identifies the condition and treatment strategies but does not provide a complete baseline-characteristic table from which broader external-validity conclusions could be drawn.
14. Why This Trial Matters Statistically
STEAM is a useful statistical teaching case because it combines a three-arm randomized design with two different primary endpoint types and two different analytical frameworks. The same trial therefore illustrates why the statistical method must follow the structure of the endpoint rather than treating every outcome as a simple comparison of percentages.
| Concept | How it appears in STEAM |
|---|---|
| Randomization | Participants were randomized in a three-arm parallel-group phase 2 trial. |
| Intention-to-treat | Both posted primary analyses used all randomized participants regardless of whether they received study treatment. |
| Binary endpoint | ORR1 classified participants according to complete or partial response during first-line therapy. |
| Cochran-Mantel-Haenszel test | ORR1 was analyzed with stratification by metastatic disease extent and tumor location. |
| Odds ratio | ORR1 was summarized with an odds ratio of 1.62 for Arm A versus Arm C. |
| Time-to-event endpoint | PFS1 measured time from randomization to progression or death. |
| Log-rank test | PFS1 was compared using the log-rank method. |
| Hazard ratio | PFS1 was summarized with an HR of 0.68 for Arm C versus pooled Arms A + B. |
| Stratified analysis | Both primary analyses incorporated metastatic disease extent and tumor location. |
| Confidence intervals | Both primary effect estimates were reported with 90% two-sided confidence intervals. |
| Superiority | The registry classifies the primary hypotheses as superiority hypotheses. |
| Endpoint interpretation | ORR1 and PFS1 answer different clinical and statistical questions and should not be treated as interchangeable measures. |
15. Related Tutorials
Learn more about the methods used in this trial:
16. Related Statistical Calculators
17. Sources
- ClinicalTrials.gov: STEAM, NCT01765582. Trial registration and posted outcome/statistical-analysis data.
- PubMed record: PMID 32816630.
- PubMed record: PMID 30552157.
Continue with the Clinical Biostats statistical pathway
Explore the statistical concepts behind randomized trials, binary endpoints, stratified categorical analyses, and time-to-event methods.
18. Record Summary
STEAM provides a compact example of how statistical interpretation depends on endpoint structure and comparison direction. The randomized phase 2 trial enrolled 280 participants across three parallel treatment arms and registered two primary endpoints: overall response during first-line therapy and progression-free survival during first-line therapy. The ORR1 analysis used a stratified Cochran-Mantel-Haenszel method and reported an odds ratio of 1.62 with a 90% two-sided CI of 0.96–2.71 and P = 0.132 for concurrent FOLFOXIRI plus bevacizumab versus FOLFOX plus bevacizumab. The PFS1 analysis used a log-rank method and reported an HR of 0.68 with a 90% two-sided CI of 0.53–0.88 and P = 0.005 for FOLFOX plus bevacizumab versus pooled FOLFOXIRI plus bevacizumab.
The statistical lesson is not simply to identify which numerical estimate is smaller. The analyst must first identify the endpoint, analysis population, comparison direction, statistical method, effect measure, confidence-interval level, and hypothesis type. For STEAM, those details materially change how the reported estimates should be understood.