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
ASCLEPIOS I was a randomized, parallel-group, quadruple-masked phase 3 trial asking whether ofatumumab, an anti-CD20 monoclonal antibody given by subcutaneous injection, reduces the annualized relapse rate compared with the oral disease-modifying therapy teriflunomide in patients with relapsing multiple sclerosis.
| Feature | ASCLEPIOS I |
|---|---|
| Phase | Phase 3 |
| Condition | Relapsing multiple sclerosis |
| Design | Randomized, parallel-group, quadruple-masked, double-dummy, active-controlled |
| Arms | Ofatumumab 20 mg (OMB 20 mg) vs teriflunomide 14 mg (TER 14 mg) |
| Enrollment | 930 |
| Primary endpoint | Annualized relapse rate (ARR), baseline up to 2.5 years |
| Primary analysis | Negative binomial regression with log link and log time-in-study offset; rate ratio |
| Hypothesis | Superiority |
| Study dates | Start 2016-09-20; primary completion 2019-07-05 |
| Status | Completed; results posted |
| ClinicalTrials.gov | NCT02792218 |
| Sponsor | Novartis Pharmaceuticals (industry) |
2. Clinical Question
The central question was whether ofatumumab produces a lower rate of confirmed relapses than teriflunomide, an established active therapy, in people with relapsing multiple sclerosis. Because the comparator is an effective treatment rather than placebo, any difference estimated here is a head-to-head contrast between two active strategies, not the absolute effect of ofatumumab over no treatment.
Population
Patients with relapsing multiple sclerosis enrolled in a phase 3 treatment study.
Intervention
Ofatumumab 20 mg subcutaneous injection plus teriflunomide-matching placebo capsules.
Comparator
Teriflunomide 14 mg capsules plus matching placebo of the ofatumumab subcutaneous injections.
Primary question
Is the annualized rate of confirmed relapses lower with ofatumumab than with teriflunomide over up to 2.5 years of follow-up?
3. Trial Design
Ofatumumab
- Ofatumumab 20 mg subcutaneous injection
- Teriflunomide-matching placebo capsules
- Serious adverse events: 51/465 affected / at risk
Teriflunomide
- Teriflunomide 14 mg capsule
- Matching placebo of ofatumumab subcutaneous injections
- Serious adverse events: 39/462 affected / at risk
A trial designed to be read alongside a twin study
The ClinicalTrials.gov record states that this study was not powered for some secondary endpoints as a stand-alone study, and that it was pre-specified in the protocol to combine its data with study COMB157G2302. Disability outcomes, which accrue slowly and require many participants to measure precisely, were therefore planned as pooled analyses across the two studies. The primary endpoint (ARR) and the MRI and biomarker endpoints were analysed within ASCLEPIOS I itself.
4. Endpoints
Primary endpoint
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Annualized Relapse Rate (ARR) | The number of confirmed relapses in a year, calculated as the total number of relapses for all participants in the treatment group divided by the total participant-years of time in study. A confirmed relapse was accompanied by a clinically relevant change in the EDSS performed by the Independent EDSS rater: an increase of at least 0.5 points on the EDSS score, or an increase of 1 point on two functional scores or 2 points on one functional score (excluding bowel/bladder or cerebral functional system changes). Comparisons were made to the previous rating (the last EDSS rating that did not occur during a relapse). | Baseline up to 2.5 years |
Secondary endpoints with posted statistical analyses
| Endpoint | Time frame | Unit / effect measure |
|---|---|---|
| 3-month Confirmed Disability Worsening (3mCDW) based on EDSS — pooled and study COMB157G2301 | Baseline, every 3 months up to 2.5 years | Percentage of participants / hazard ratio |
| 6-month Confirmed Disability Worsening (6mCDW) based on EDSS — pooled and study COMB157G2301 | Baseline, every 3 months up to 2.5 years | Percentage of participants / hazard ratio |
| 6-month Confirmed Disability Improvement (6mCDI) based on EDSS — pooled and study COMB157G2301 | Baseline, every 3 months up to 2.5 years | Percentage of participants / hazard ratio |
| Number of Gd-enhancing T1 lesions per MRI scan | Baseline, yearly up to 2.5 years | Lesions per scan / rate ratio |
| Number of new or enlarging T2 lesions on MRI per year (annualized lesion rate) | Baseline, yearly up to 2.5 years | T2 lesions per year / rate ratio |
| Neurofilament light chain (NfL) concentration in serum | Month 3, 12 and 24 | pg/mL / geometric mean ratio |
| Annualized rate of brain volume loss based on percent brain volume change from baseline | Baseline, months 12 and 24 | Percentage of brain volume loss / mean difference |
| 3mCDW and 6mCDW > 8 weeks after onset of treatment — pooled | Baseline up to 2.5 years | Percentage of participants / hazard ratio |
Note that the registry labels each within-study disability analysis as "Study COMB157G2301"; this is the study code for ASCLEPIOS I, while "Pooled Data" combines it with COMB157G2302.
5. Statistical Methodology
Negative binomial regression for relapses and lesions
Relapses are counts, and participants spend different lengths of time in the study. The primary analysis modelled the number of relapses per participant with a negative binomial regression using a log link. The natural log of time-in-study entered as an offset, which converts a count model into a rate model and "annualizes" the relapse rate. The model was adjusted for treatment and region as factors, and for number of relapses in the previous year, baseline EDSS, baseline number of Gd-enhancing lesions and age at baseline as covariates.
Var(Yi) = μi + k·μi2 Rate ratio = exp(β1)
Here Yi is the relapse count, ti is time in study, and k is the dispersion parameter. When k = 0 the model reduces to Poisson regression; k > 0 allows some participants to relapse far more often than others.
The same family of model was used for Gd-enhancing T1 lesions per scan and for new or enlarging T2 lesions per year. MRI lesion counts are typically even more over-dispersed than relapse counts, with many participants having zero lesions and a few having many.
Cox proportional-hazards regression for disability endpoints
Confirmed disability worsening and improvement are time-to-event outcomes: the event is the first EDSS change that is confirmed at a later visit 3 or 6 months afterwards. Hazard ratios were estimated with Cox regression. Participants who leave the study or reach the end of follow-up without a confirmed event are censored, contributing information up to their last assessment. The registry reports the unit for these outcomes as a percentage of participants, but the posted comparisons are hazard ratios rather than differences in proportions.
Mixed model for repeated measures for serum NfL
NfL was measured at Months 3, 12 and 24 and analysed with a mixed model for repeated measures (MMRM). The treatment effect is reported as a geometric mean ratio, which arises when a biomarker with a right-skewed distribution is analysed on the log scale and the between-arm difference is back-transformed. MMRM uses all available visits from each participant and models the correlation between repeated measurements from the same person.
Random coefficient model for brain volume
The annualized rate of brain volume loss was analysed with a random coefficient model, a mixed model in which each participant has their own intercept and slope over time. The treatment effect is a difference in mean slopes, reported as a net mean difference.
Analysis populations
| Population | Where used |
|---|---|
| Full analysis set | Primary ARR analysis; most secondary efficacy analyses, including MRI, NfL, brain volume and within-study and pooled disability worsening. |
| Meta analysis set | Pooled 6-month confirmed disability improvement analysis combining ASCLEPIOS I and COMB157G2302. |
| Participants at risk (safety) | Serious adverse events: 465 in the OMB 20 mg arm and 462 in the TER 14 mg arm. |
A full analysis set generally follows the intention-to-treat principle, analysing randomized participants according to their assigned treatment, with only narrowly defined exclusions. That preserves the comparability created by randomization.
6. Primary Endpoint Result: Annualized Relapse Rate
Rate ratio for confirmed relapses (OMB 20 mg vs TER 14 mg)
95% CI: 0.375–0.655 · P < 0.001 · two-sided
Negative binomial regression · Full analysis set · Superiority hypothesis
| Outcome | Comparison | Effect measure | Estimate (95% CI) | P-value |
|---|---|---|---|---|
| Annualized Relapse Rate, baseline up to 2.5 years | OMB 20 mg vs TER 14 mg | Rate ratio | 0.495 (0.375–0.655) | <0.001 |
What the estimate means. A rate ratio of 0.495 means that, after adjustment for region, prior-year relapses, baseline EDSS, baseline Gd-enhancing lesions and age, the estimated rate of confirmed relapses with ofatumumab was about half the rate with teriflunomide: a 50.5% lower estimated relapse rate.
What it does not mean. It does not mean that half as many participants relapsed, that 50.5% of participants were relapse-free, or that each person's relapse frequency was halved. A rate ratio compares event rates per unit of time at the group level; relapses are pooled across participants and participant-years.
Precision. The 95% confidence interval of 0.375 to 0.655 lies entirely below 1. The data are compatible with an estimated relapse-rate reduction ranging from 34.5% (upper limit) to 62.5% (lower limit). The interval is asymmetric around 0.495 because it was computed on the log scale and back-transformed.
Why the p-value is not the effect size. P < 0.001 says that a rate ratio this far from 1 would be very unlikely if the two treatments had identical relapse rates. It does not say how large the difference is; that information is carried by the estimate and its interval.
Cautions. This is an active-comparator trial, so the ratio is relative to teriflunomide, not placebo. The ratio is adjusted for baseline covariates, which typically improves precision but means the estimate is conditional on those covariates. The time-in-study offset handles unequal follow-up, under the assumption that the relapse rate is roughly constant over time for a given participant. Finally, the negative binomial model assumes a specific mean-variance relationship; the dispersion parameter absorbs heterogeneity between participants but cannot correct for informative dropout.
7. Secondary Endpoint Results: Disability
The disability endpoints were prespecified for pooled analysis with COMB157G2302, and the registry states that this study was not powered to analyse them on its own. Both the pooled and the within-study estimates are posted.
| Outcome | Data source | HR (95% CI) | P-value |
|---|---|---|---|
| 3-month confirmed disability worsening | Pooled | 0.657 (0.500–0.863) | 0.003 |
| 3-month confirmed disability worsening | Study COMB157G2301 | 0.652 (0.445–0.956) | 0.029 |
| 6-month confirmed disability worsening | Pooled | 0.676 (0.498–0.917) | 0.012 |
| 6-month confirmed disability worsening | Study COMB157G2301 | 0.607 (0.396–0.930) | 0.022 |
| 6-month confirmed disability improvement | Pooled (meta analysis set) | 1.355 (0.952–1.952) | 0.092 |
| 6-month confirmed disability improvement | Study COMB157G2301 | 1.186 (0.709–1.983) | 0.516 |
| 3mCDW > 8 weeks after onset of treatment | Pooled | 0.641 (0.486–0.847) | 0.002 |
| 6mCDW > 8 weeks after onset of treatment | Pooled | 0.657 (0.481–0.898) | 0.008 |
For worsening, an HR below 1 favours ofatumumab. The pooled 3-month estimate of 0.657 corresponds to a 34.3% lower estimated hazard of confirmed worsening, and the pooled 6-month estimate of 0.676 to a 32.4% lower estimated hazard.
For improvement, the direction reverses: an HR above 1 means improvement events occurred at a higher estimated rate with ofatumumab. Both improvement intervals (pooled 0.952–1.952; within-study 0.709–1.983) include 1, so these data do not establish a difference in confirmed improvement.
Pooled versus within-study. The within-study point estimates for worsening (0.652 and 0.607) are close to the pooled estimates, but their intervals are wider, as expected from roughly half the data. This is exactly why pooling was prespecified: the within-study intervals only just exclude 1 (upper limits 0.956 and 0.930), whereas the pooled intervals are noticeably tighter.
Sensitivity analyses. The "> 8 weeks after onset of treatment" analyses exclude worsening events that began very early, before a treatment could plausibly have acted. Their estimates (0.641 and 0.657) are close to the main pooled estimates, which suggests that early events are not driving the result.
Cautions. A Cox hazard ratio summarises the effect as constant over follow-up (proportional hazards). The registry does not report tests of this assumption or the numbers of events per arm. Many secondary endpoints were tested, and the registry does not describe a multiplicity-control hierarchy, so secondary p-values are best read as supportive rather than independently confirmatory.
8. Secondary Endpoint Results: MRI Lesions
Gd-enhancing T1 lesions per scan: rate ratio
95% CI: 0.013–0.049 · P < .001
Negative binomial regression · Full analysis set
| Outcome | Time point | Rate ratio (95% CI) | P-value |
|---|---|---|---|
| Gd-enhancing T1 lesions per MRI scan | Baseline, yearly up to 2.5 years | 0.025 (0.013–0.049) | <.001 |
| New or enlarging T2 lesions per year | Month 12 | 0.26 (0.21–0.33) | <.001 |
| New or enlarging T2 lesions per year | Month 24 | 0.22 (0.15–0.34) | <.001 |
| New or enlarging T2 lesions per year | End of Study | 0.18 (0.15–0.22) | <.001 |
A rate ratio of 0.025 for Gd-enhancing lesions corresponds to a 97.5% lower estimated rate of enhancing lesions per scan. For new or enlarging T2 lesions, the rate ratios of 0.26, 0.22 and 0.18 correspond to 74%, 78% and 82% lower estimated annualized lesion rates at Month 12, Month 24 and End of Study. MRI lesion endpoints are more sensitive to inflammatory activity than clinical relapses, which is why their relative effects are often much larger than the effect on ARR. A very small rate ratio also reflects a low absolute lesion count in one arm; the registry's posted comparison does not by itself show how many lesions occurred per arm.
9. Secondary Endpoint Results: Serum NfL and Brain Volume
| Outcome | Time point / method | Effect measure | Estimate (95% CI) | P-value |
|---|---|---|---|---|
| Serum NfL concentration | Month 3, MMRM | Geometric mean ratio | 0.93 (0.89–0.98) | 0.011 |
| Serum NfL concentration | Month 12, MMRM | Geometric mean ratio | 0.73 (0.69–0.77) | <.001 |
| Serum NfL concentration | Month 24, MMRM | Geometric mean ratio | 0.77 (0.72–0.82) | <.001 |
| Annualized rate of brain volume loss | Baseline, months 12 and 24; random coefficient model | Mean difference (net) | 0.07 (−0.02 to 0.15) | 0.118 |
NfL is a blood marker of neuroaxonal injury. A geometric mean ratio of 0.93 at Month 3 means that the geometric mean NfL concentration was an estimated 7% lower with ofatumumab; by Month 12 the ratio was 0.73 (27% lower) and at Month 24 0.77 (23% lower). The pattern of a modest early difference that grows by Month 12 is consistent with a treatment effect that takes time to appear in a biomarker reflecting accumulated tissue damage, though the MMRM treats each visit as a separate contrast rather than modelling that trajectory.
For brain volume, the estimated net mean difference of 0.07 percentage points has a confidence interval from −0.02 to 0.15 that includes zero, with P = 0.118. These data therefore do not show a difference between arms in the annualized rate of brain volume loss over this follow-up. A non-significant result is not evidence of equivalence: the interval is compatible with a small difference in either direction.
10. Safety: Serious Adverse Events
| Measure | OMB 20 mg | TER 14 mg |
|---|---|---|
| Participants with serious adverse events (affected / at risk) | 51/465 | 39/462 |
Serious adverse event counts are descriptive. No formal statistical comparison of safety is posted, and safety counts are summarised among participants at risk, meaning those exposed to study treatment, rather than in the efficacy analysis population. A difference in counts of this size should not be read as a confirmed difference in risk without an appropriate estimate and interval, and serious adverse events need to be considered alongside the efficacy results rather than netted against them in a single number.
11. Statistical Methods Explained
Why was a negative binomial model used for relapses rather than Poisson regression?
A Poisson model assumes that the variance of the relapse count equals its mean. In multiple sclerosis, many participants have no relapses while a minority have several, so the variance is larger than the mean (over-dispersion). A Poisson model would then understate the standard error and give confidence intervals that are too narrow. The negative binomial model adds a dispersion parameter to capture this extra variability, which is why it is standard for ARR and MRI lesion counts.
What does the log time-in-study offset do?
Participants were followed for different lengths of time, up to 2.5 years. Including log(time in study) as an offset with its coefficient fixed at 1 turns the model for expected counts into a model for counts per year. Someone followed for two years is expected to have roughly twice as many relapses as someone followed for one year at the same underlying rate, and the offset builds that directly into the model.
Why is the rate ratio 0.495 not the same as "50% fewer patients relapsed"?
The rate ratio compares total relapses per participant-year between arms. A reduction in the proportion of participants who relapse, a reduction in repeated relapses among those who do, or both, can produce the same rate ratio. The proportion of relapse-free participants is a different estimand and would need its own analysis.
Why were disability outcomes pooled across two trials?
Confirmed disability worsening is relatively uncommon over 2.5 years, so a single trial sized for relapses would produce imprecise hazard ratios. Prespecifying a pooled analysis with the twin study COMB157G2302 before seeing the results increases the number of events and narrows the confidence interval, while avoiding the bias of deciding to pool after the fact. The within-study estimates are shown for transparency, but the pooled estimates are the intended analysis.
How should a hazard ratio above 1 for disability improvement be read?
For improvement, the "event" is a good outcome, so an HR above 1 favours the treatment with more improvement events. The pooled HR of 1.355 points in favour of ofatumumab, but its interval (0.952 to 1.952) includes 1 and P = 0.092, so the data are compatible with no difference. The direction of benefit always depends on whether the event is desirable or undesirable.
What does a geometric mean ratio of 0.73 for NfL mean?
NfL values are right-skewed, so they are analysed on the log scale. The difference between arms on the log scale, back-transformed, becomes a ratio of geometric means. A ratio of 0.73 means that at Month 12 the typical (geometric mean) NfL concentration with ofatumumab was 73% of that with teriflunomide, that is, 27% lower. It is a multiplicative comparison, not a difference in pg/mL.
Why is the Month 3 NfL result (P = 0.011) less informative than its p-value might suggest?
The Month 3 ratio of 0.93 is statistically significant but small, and its interval (0.89 to 0.98) sits close to 1. With many secondary endpoints and time points tested, individual p-values near conventional thresholds carry less weight than a consistent pattern across endpoints. The later ratios, 0.73 and 0.77, are both larger in magnitude and more precisely estimated.
12. Limitations
- Active comparator: all effects are relative to teriflunomide. They do not describe the effect of ofatumumab versus placebo or versus other therapies.
- Stand-alone power for secondary endpoints: the registry states that the study was not powered for some secondary endpoints on its own; within-study disability estimates are correspondingly imprecise.
- Multiplicity: many secondary outcomes and time points were analysed. The registry does not describe how type I error was controlled across them, so secondary p-values should not each be treated as confirmatory.
- Model assumptions: the Cox models assume proportional hazards, and the negative binomial models assume a constant rate over follow-up for each participant. The registry does not report diagnostics for these assumptions.
- Missing data and dropout: the offset and censoring handle unequal follow-up, but results assume that dropout is not related to unobserved relapse or disability risk. The registry does not describe sensitivity analyses for missing data.
- Absolute effects not shown: the posted comparisons are relative measures (rate ratios, hazard ratios, geometric mean ratios). Relative effects are most informative when read alongside absolute rates, which should be taken from the full record rather than inferred from the ratios.
- Surrogate endpoints: MRI lesions and NfL are biomarkers of disease activity. Very large relative effects on lesions do not translate one-to-one into effects on long-term disability.
- Follow-up duration: up to 2.5 years is short relative to the course of multiple sclerosis, and the brain volume result shows that some slower outcomes may not separate within this window.
13. Why This Trial Matters Statistically
ASCLEPIOS I is a useful teaching case because it combines several model families, each matched to a different outcome type, within a single randomized comparison, and because it shows prespecified pooling across twin trials in practice.
| Concept | How it appears in ASCLEPIOS I |
|---|---|
| Count-data regression | Negative binomial models for relapses and MRI lesions, with log time-in-study offset |
| Rate ratio | Primary effect measure for ARR (0.495) and for lesion endpoints |
| Covariate adjustment | Region, prior-year relapses, baseline EDSS, baseline Gd-enhancing lesions and age in the primary model |
| Time-to-event analysis | Confirmed disability worsening and improvement analysed with Cox regression |
| Hazard ratio direction | HR < 1 favours ofatumumab for worsening; HR > 1 favours it for improvement |
| Prespecified pooling | Disability endpoints combined with COMB157G2302 to reach adequate precision |
| Repeated measures | MMRM for serum NfL at Months 3, 12 and 24 |
| Log-scale analysis | Geometric mean ratios for a skewed biomarker |
| Random coefficient model | Participant-specific slopes for brain volume change |
| Double-dummy blinding | Injection and capsule placebos to mask different routes of administration |
| Non-significant is not equivalent | Brain volume difference with CI −0.02 to 0.15 |
14. Related Tutorials
Learn more about the methods used in this trial:
15. Related Calculators
16. Sources
- ClinicalTrials.gov: NCT02792218 — Efficacy and Safety of Ofatumumab Compared to Teriflunomide in Patients With Relapsing Multiple Sclerosis.
- PubMed: PMID 32757523.
- PubMed: PMID 35266417.
- PubMed: PMID 37606897.
- PubMed: PMID 39018512.
Explore the methods behind the results
Each trial analysis connects its endpoints and models to tutorials and calculators, so the statistical reasoning can be followed step by step.
17. Record Summary
ASCLEPIOS I compared ofatumumab with teriflunomide in 930 participants with relapsing multiple sclerosis. The primary negative binomial analysis estimated a relapse rate ratio of 0.495 (95% CI 0.375–0.655), and prespecified pooled Cox analyses with COMB157G2302 estimated lower hazards of 3- and 6-month confirmed disability worsening. MRI lesion rate ratios and serum NfL geometric mean ratios pointed in the same direction, while disability improvement and brain volume loss did not show clear differences. The most useful reading combines relative effect estimates, their confidence intervals, the model behind each outcome type and the design choices, such as double-dummy masking and prespecified pooling, that support causal interpretation.