SUNBEAM: Complete Statistical Analysis of Ozanimod in Relapsing Multiple Sclerosis

1. Trial at a Glance

SUNBEAM was a randomized, parallel, quadruple-masked phase 3 treatment trial in multiple sclerosis. The registry reports 1346.0 enrolled participants and three study arms involving ozanimod, interferon beta-1a, and placebo components used for masking.

2. Clinical Question

Population

The registered condition was Multiple Sclerosis. The primary analysis population was the ITT population, defined as all randomized participants who received at least 1 dose of study medication.

Intervention

The study evaluated Ozanimod, with analyses comparing the 1 mg and 0.5 mg doses against interferon beta-1a.

Comparator

The active comparator was Interferon Beta-1a. The intervention description also includes placebo to ozanimod and placebo to interferon beta-1a, consistent with the quadruple-masked parallel design.

Primary question

The primary statistical question was whether the adjusted annualized relapse rate during the 12-month treatment period differed between interferon beta-1a and each ozanimod dose. The registry classifies both primary comparisons as superiority analyses.

3. Trial Design

The registry describes a randomized, parallel, quadruple-masked design. The statistical analyses use several distinct model families because the outcomes have different distributions: relapse counts and MRI lesion counts use rate-based regression, disability progression uses time-to-event analysis, categorical MRI outcomes use a stratified CMH test, and continuous functional and quality-of-life outcomes use ANCOVA.

Serious adverse events

These counts describe serious adverse events by arm as reported in the registry. They do not, by themselves, establish whether differences between arms are statistically significant or whether individual adverse events differ in mechanism, severity, or duration.

4. Endpoints

Primary endpoint: Adjusted Annualized Relapse Rate

Time frame: 12 months.

The relapse rate was based on confirmed relapses. A relapse was defined as new or worsening neurological symptoms attributable to MS and preceded by a relatively stable or improving neurological state for at least 30 days. Symptoms had to persist for > 24 hours and not be attributable to confounding clinical factors. Relapses were confirmed when accompanied by objective neurological worsening based on examination by the blinded evaluator, consistent with an increase of ≥ 0.5 on the overall EDSS score relative to the most recent EDSS assessment, or 2 points on one functional system scale, or 1 point on ≥ two functional system scales.

The relapse rate was calculated as the total number of relapses divided by the total number of days in the study × 365.25. ARR was adjusted for region, baseline age, and baseline number of gadolinium-enhancing lesions; the natural log transformation of time on study was included as an offset term.

Secondary endpoints

• Adjusted mean number of new or enlarging hyperintense T2-weighted brain MRI lesions per scan over 12 months, with MRI scans assessed at Month 6 and Month 12.
• Adjusted mean number of gadolinium-enhancing brain MRI lesions at Month 12.
• Time to onset of disability progression confirmed after 3 months and after 6 months, from first dose through the end of the 12-month treatment period.
• Percentage of participants who were gadolinium-enhancing lesion-free at Month 12.
• Percentage of participants who were T2 lesion-free at Month 12.
• Change from baseline to Month 12 in MSFC score including the LCLA test.
• Mean change from baseline in MSQOL-54 physical-health and mental-health composite summary scores.
• Percent change from baseline in normalized brain volume at Month 12.

5. Statistical Methodology

Primary analysis populations and comparisons

Both primary analyses used the ITT population: all randomized participants who received at least 1 dose of study medication. The two comparisons were interferon beta-1a versus ozanimod 1 mg and interferon beta-1a versus ozanimod 0.5 mg.

Covariate adjustment

The primary ARR models adjusted for region, baseline age, and baseline number of gadolinium-enhancing lesions. The natural log transformation of time on study served as an offset. This is important because relapse is not simply a fixed-time binary outcome: participants contribute person-time, and the offset allows the model to express the event rate relative to time observed.

Missing-data handling

For the Month 12 gadolinium-enhancing lesion-free and T2 lesion-free outcomes, missing Month 12 MRI data were treated as non-response. For MSQOL-54 outcomes, missing post-baseline data were imputed using a mixed-effects regression model with random slope and intercept. The MSFC analyses were restricted to participants with available baseline and Month 12 MSFC z-scores.

6. Results: Primary Endpoint

Adjusted Annualized Relapse Rate — Ozanimod 1 mg

Adjusted Annualized Relapse Rate — Ozanimod 0.5 mg

7. Results: Secondary Endpoints

MRI lesion counts

The T2-lesion analyses used MRI scans assessed at Month 6 and Month 12 and included the natural log transformation of available MRI scans over 12 months as an offset. The registry reports the rate ratio for these analyses as Ozanimod / IFN beta-1a. This direction differs from the primary endpoint's displayed group-comparison wording, so the reported effect-measure definition should be used when interpreting these results.

Disability progression

All four analyses used the ITT population and Cox proportional-hazards models adjusted for region, baseline age, and baseline EDSS score. The confidence intervals are comparatively wide, particularly for the 6-month-confirmed endpoints, so the point estimates should not be interpreted without their uncertainty intervals.

Gadolinium-enhancing lesion-free status

Missing Month 12 MRI data were treated as non-responders. The CMH analyses were stratified by region and EDSS category per IVRS. The reported effect is a difference in percentages, normalized here as a risk difference; it is not a rate ratio or hazard ratio.

T2 lesion-free status

These analyses also treated missing Month 12 MRI results as non-responders and used CMH stratification by region and EDSS category per IVRS. Because each confidence interval includes zero, the reported interval is compatible with a range that includes no between-group percentage difference.

MSFC score including the LCLA test

The analysis population was the ITT population with available baseline and Month 12 MSFC z-scores. ANCOVA adjusted for region, EDSS category per IVRS, and baseline MSFC score.

MSQOL-54 physical and mental health composite summaries

The registry states that these analyses used ANCOVA with adjustment for region, EDSS category per IVRS, and the corresponding baseline summary score. Missing post-baseline data were imputed using a mixed-effects regression model with random slope and intercept.

Normalized brain volume

The endpoint was percent change from baseline in normalized brain volume at Month 12. The registry reports no effect estimate or confidence interval for these two analyses. It states that the data were non-normal and therefore analyzed with rank ANCOVA, adjusted for region and EDSS category per IVRS, using the specified ranked baseline and percent-change quantities.

8. Statistical Methods Explained

Why was Poisson regression used for annualized relapse rate?

Relapses are event counts observed over participant-specific amounts of study time. Poisson regression provides a natural framework for modeling event rates, and the registry explicitly included the natural log of time on study as an offset. This distinguishes a rate analysis from a simple comparison of average counts, because participants who contribute different amounts of observation time can be placed on a common rate scale.

What does a rate ratio of 0.518 mean?

A rate ratio of 0.518 indicates that the modeled relapse rate associated with the ozanimod 1 mg comparison was 0.518 times the comparator rate under the specified model. It is a relative measure. It does not tell us the absolute number of relapses prevented, and it should not be read as a percentage of individual participants who avoided relapse.

Why was negative binomial regression used for MRI lesion counts?

MRI lesion outcomes are counts, but count data need not have the variance structure assumed by a basic Poisson model. The registry identifies negative binomial regression for the MRI lesion outcomes, providing a count-rate model that can accommodate greater dispersion. The use of the number of available scans as an offset also recognizes that the amount of MRI observation can vary.

What does a hazard ratio measure for disability progression?

A hazard ratio compares the instantaneous event rates between groups over the observed follow-up period, conditional on the model. It is not the ratio of median times and is not a direct probability ratio. These analyses used Cox proportional-hazards models adjusted for region, baseline age, and baseline EDSS score.

Why use the Cochran-Mantel-Haenszel test for lesion-free outcomes?

The lesion-free outcomes are binary at Month 12, while the registry also specifies stratification by region and EDSS category per IVRS. The CMH framework provides a stratified comparison rather than treating every participant as if the stratification factors were irrelevant. The resulting difference in percentages is therefore interpreted in the context of those strata.

Why was rank ANCOVA used for normalized brain volume?

The registry states that the brain-volume-loss data were non-normal. Rank ANCOVA addresses that distributional issue by analyzing ranked quantities while retaining adjustment for region and EDSS category per IVRS. The important distinction is that the reported p-values arise from this rank-based analysis; the registry does not provide corresponding effect estimates or confidence intervals for these two comparisons.

9. How to Read the Results as a Statistical Story

The SUNBEAM results illustrate why a clinical-trial conclusion should not be based on a single p-value. The primary endpoint is a rate outcome, so the natural effect measure is a rate ratio. The secondary endpoints deliberately span several statistical structures: counts per scan, time-to-event outcomes, binary MRI status, continuous functional scores, quality-of-life summaries, and a non-normal brain-volume outcome.

The primary analyses show two estimated rate ratios below 1.0, with two-sided 95% confidence intervals that do not include 1. The secondary MRI lesion-count analyses use the same general rate-ratio logic, whereas the disability analyses use hazard ratios. For the latter, the confidence intervals span 1.0 in all four reported comparisons. That is an important statistical distinction: an effect estimate can be below or above 1 without the confidence interval establishing a precise directional difference.

The categorical MRI outcomes are similarly informative because the effect measure is a risk difference rather than a relative rate or hazard. Some confidence intervals include zero and some do not. For continuous outcomes, the null value is likewise zero for a mean difference. This makes the scale of the null hypothesis part of correct interpretation: 1 for a ratio measure and 0 for a difference measure.

The handling of missing information also matters. Treating missing Month 12 MRI assessments as non-responders changes the definition of the binary endpoint, while the MSQOL-54 analyses use a mixed-effects regression model for missing post-baseline data. These are different assumptions and should not be treated as interchangeable missing-data strategies.

Finally, the multiplicity structure deserves attention. The registry reports one registered primary endpoint but two primary comparisons, plus numerous secondary endpoints and paired treatment comparisons. The reported p-values are useful evidence for their individual analyses, but a reader should distinguish an endpoint-level p-value from a global statement about the entire collection of hypotheses. The registry information presented here does not specify a multiplicity-adjustment procedure, alpha-spending scheme, or hierarchical testing sequence.

10. Limitations

• Relative rather than absolute primary effect: the primary results are reported as rate ratios. The registry does not provide the underlying absolute annualized relapse rates in the statistical-analysis records summarized here.
• Model dependence: rate ratios and hazard ratios are model-based summaries. Their interpretation depends on the specified regression model and covariate adjustment.
• Time-to-event assumptions: Cox regression relies on the proportional-hazards framework. The registry identifies the model but does not provide a proportional-hazards diagnostic in the reported analysis records.
• Multiple comparisons: the registry reports two primary comparisons and many secondary analyses. A multiplicity-adjustment procedure is not specified in the analysis information presented here.
• Missing MRI data: missing Month 12 MRI observations were treated as non-responders for the lesion-free endpoints, which is a specific assumption about missingness.
• Available-case components: several continuous and MRI analyses require participants with available measurements at specified time points, so the effective analysis population can differ from the full randomized population.
• Incomplete effect reporting for brain volume: the registry reports p-values for normalized brain-volume change but no corresponding estimate or confidence interval in the posted analyses.
• Serious adverse events: the registry provides affected/at-risk counts by arm, but the summarized data do not include a comparative hypothesis test or confidence interval for serious adverse events.

11. Why This Trial Matters Statistically

SUNBEAM is a useful teaching example because a single randomized trial incorporates several core families of clinical-trial methods. The primary outcome is a recurrent-event rate, leading to Poisson regression with a time offset and covariate adjustment. MRI lesion outcomes use negative binomial regression, binary MRI status uses stratified CMH testing, disability progression uses Cox regression, and functional and quality-of-life outcomes use ANCOVA.

The trial also demonstrates why endpoint definition is part of statistical analysis rather than merely clinical terminology. The primary relapse endpoint depends on a prespecified clinical confirmation process, while disability progression depends on confirmation after either 3 months or 6 months. Changing the event definition changes the estimand being analyzed.

Another important lesson is that the same treatment comparison can look different across estimands. A rate ratio, hazard ratio, risk difference, and mean difference all answer different questions. The numerical null value also changes with the scale. Reading the estimate without first identifying its estimand is therefore a common source of misinterpretation.

Finally, SUNBEAM provides a practical example of covariate adjustment. Region, baseline age, baseline EDSS, baseline MSFC, baseline summary scores, and baseline MRI-related measures appear in different models. Adjustment is not a universal statistical operation applied identically to every endpoint; it is part of the model specification for the particular outcome.

12. Related Tutorials

Learn more about the methods used in this trial:

13. Related Calculators

14. Sources

• ClinicalTrials.gov: NCT02294058.
• PubMed: PMID 40462564.
• PubMed: PMID 33454584.