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
onset 9 was a randomized, parallel-group, phase 3 trial comparing faster-acting insulin aspart with insulin aspart (NovoRapid/NovoLog) in people with type 2 diabetes. The primary endpoint was change in glycosylated haemoglobin (HbA1c) from week 0 to week 16, analyzed using an ANOVA model and evaluated under a non-inferiority framework.
| Feature | onset 9 |
|---|---|
| Trial name | onset 9 |
| ClinicalTrials.gov identifier | NCT03268005 |
| Phase | Phase 3 |
| Status | Completed |
| Therapeutic area | Endocrinology & Metabolism |
| Condition | Diabetes; Diabetes Mellitus, Type 2 |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Quadruple |
| Primary purpose | Treatment |
| Enrollment | 1264 |
| Arms | 2 |
| Start | 2017-09-19 |
| Primary completion | 2019-01-07 |
| Lead sponsor | Novo Nordisk A/S |
| Sponsor type | Industry |
2. Clinical Question
The primary clinical question was whether change in HbA1c from week 0 to week 16 with faster-acting insulin aspart was non-inferior to the corresponding change with NovoRapid in people with type 2 diabetes.
Population
People with diabetes mellitus, type 2, enrolled in the phase 3 onset 9 trial.
Intervention
Faster-acting insulin aspart.
Comparator
Insulin aspart, identified in the trial description as NovoRapid/NovoLog.
Primary question
Is faster-acting insulin aspart non-inferior to NovoRapid with respect to change in HbA1c from week 0 to week 16?
3. Trial Design
Faster-acting insulin aspart
- Faster-acting insulin aspart was the intervention being compared.
- The statistical analysis compared this arm with NovoRapid.
NovoRapid
- Insulin aspart was the comparator.
- The trial description identifies NovoRapid/NovoLog as the already available comparator medicine.
The registry identifies insulin degludec and metformin among the trial interventions. Metformin use at baseline was also included as a factor in the primary HbA1c analysis.
4. Endpoints
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Change in Glycosylated Haemoglobin (HbA1c) | Change from baseline (week 0) in glycosylated haemoglobin (HbA1c) was evaluated at week 16. The endpoint was evaluated based on data from the in-trial observation period. The in-trial observation period was from date of randomisation and until last trial-related participant-site contact. | Week 0, week 16 |
The registry lists one primary endpoint. The primary endpoint is a change-from-baseline measure, so the treatment comparison concerns the difference between the two randomized groups in their modeled change in HbA1c.
5. Primary Statistical Analysis
The posted primary analysis used an analysis of variance (ANOVA) model. Change from baseline in HbA1c was analyzed after multiple imputation assuming treatment according to randomization.
The model included treatment, region, and metformin use at baseline (Yes/No) as factors, with baseline HbA1c as a covariate.
The analysis population was defined as the overall number of participants analyzed being the number of participants with available data. The statistical analysis also identifies multiple imputation for missing data, with treatment handled according to randomization.
| Analysis component | Posted approach |
|---|---|
| Outcome | Change in HbA1c from week 0 to week 16 |
| Method | ANOVA |
| Effect measure | Treatment difference |
| Analysis population | Participants with available data |
| Missing-data approach | Multiple imputation |
| Treatment handling | According to randomization |
| Factors | Treatment, region, baseline metformin use |
| Covariate | Baseline HbA1c |
| Hypothesis type | Non-inferiority |
6. Primary Result: Change in HbA1c
The posted analysis compared faster-acting insulin aspart with NovoRapid for change in HbA1c from week 0 to week 16.
Treatment difference in HbA1c change
95% CI: −0.11 to 0.03 · P = 0.310
Outcome unit: percentage of HbA1c
| Measure | Faster Aspart vs NovoRapid |
|---|---|
| Estimated treatment difference | −0.04 |
| 95% confidence interval | −0.11 to 0.03 |
| Confidence interval type | Two-sided 95% |
| P-value | 0.310 |
| Hypothesis framework | Non-inferiority |
| Non-inferiority margin | 0.4% |
| Upper confidence-limit criterion | 0.03 is below 0.4% |
The estimated treatment difference was −0.04 percentage of HbA1c for faster-acting insulin aspart relative to NovoRapid. In the direction represented by the reported difference, the estimate is slightly below zero. The estimate describes the modeled difference in average change between the randomized treatment groups; it does not describe an individual patient's HbA1c response.
The 95% confidence interval of −0.11 to 0.03 describes statistical uncertainty around the estimated treatment difference. Because the interval includes zero, the result does not provide conventional evidence of a statistically significant difference between treatments at the two-sided 0.05 level. The confidence interval is also the key quantity for the registered non-inferiority question.
The registry specifies a 0.4% non-inferiority margin and states that non-inferiority was considered established if the upper limit of the 95% confidence interval for the treatment difference was less than or equal to 0.4%. Here, the upper limit is 0.03%, which is below that margin. Thus, under the registry's stated criterion, the non-inferiority condition was met.
The P-value of 0.310 should not be interpreted as the size of the treatment effect. A P-value addresses evidence against a specified null hypothesis; it does not tell us whether a treatment difference is clinically large or small. In a non-inferiority analysis, the confidence interval's relationship to the prespecified margin is particularly important.
7. Understanding the Non-Inferiority Analysis
Non-inferiority trials ask a different question from a conventional superiority trial. The objective is generally to determine whether the new treatment is not worse than the comparator by more than a prespecified clinically acceptable amount.
For onset 9, the posted upper confidence limit was 0.03%, which lies below the 0.4% non-inferiority margin.
The important statistical distinction is between non-inferiority and superiority. The reported two-sided P-value of 0.310 does not establish a statistically significant difference between treatments. That is compatible with the non-inferiority conclusion because the non-inferiority question is evaluated against the prespecified margin rather than requiring the confidence interval to exclude zero.
What the margin means
The 0.4% margin defines the largest treatment difference that the registry's non-inferiority framework permits while still considering faster-acting insulin aspart non-inferior to NovoRapid.
What the CI contributes
The confidence interval quantifies uncertainty around the treatment difference. Its upper limit determines how far the estimated difference could extend in the direction relevant to the non-inferiority criterion.
Why zero is different from the margin
Zero represents no estimated difference. The non-inferiority margin is a clinical/statistical boundary defining the maximum acceptable disadvantage under the trial framework.
Why P = 0.310 is not contradictory
A nonsignificant superiority comparison and a successful non-inferiority analysis can occur together because they answer different statistical questions.
8. Covariate Adjustment and the ANOVA Model
The primary analysis did not simply compare two unadjusted mean changes. The ANOVA model included treatment, region, and baseline metformin use as factors and baseline HbA1c as a covariate.
Why adjust for baseline HbA1c?
For a change-from-baseline outcome, baseline HbA1c is directly related to the endpoint definition. Including baseline HbA1c as a covariate can improve the statistical comparison by accounting for baseline differences and reducing unexplained variation in the outcome.
Why include treatment in the model?
Treatment is the primary comparison of interest. The treatment coefficient or contrast from the model provides the estimated difference between faster-acting insulin aspart and NovoRapid after accounting for the other model terms.
Why include region and metformin use?
The registry specifies region and baseline metformin use as factors. Including these variables makes the posted analysis more structured than a simple two-group comparison and accounts for these prespecified sources of variation in the treatment comparison.
The reported estimate of −0.04 is the treatment difference from the posted model, rather than a simple recalculation from unadjusted group means.
9. Multiple Imputation and Missing Data
The registry states that change from baseline in HbA1c was analyzed using ANOVA after multiple imputation, assuming treatment according to randomization. This is an important part of the analysis because longitudinal clinical trials can have participants without an observed endpoint at the planned assessment.
Multiple imputation creates multiple plausible completed datasets under an explicit statistical model for missing information. Each completed dataset is analyzed, and the resulting estimates are combined to reflect both within-imputation variability and uncertainty attributable to missingness.
Why imputation matters
A complete-case analysis would use only participants with observed endpoint data. Multiple imputation allows information from the analysis model to contribute while explicitly representing uncertainty about missing values.
Randomized treatment assignment
The posted analysis assumes treatment according to randomization. This preserves the randomized treatment comparison within the stated analysis framework rather than redefining groups according to treatment actually received.
The registry does not provide additional details here about the imputation model or its assumptions. Consequently, the statistical interpretation should remain at the level of the reported multiple-imputation strategy rather than attributing a specific missing-data mechanism to the analysis.
10. Statistical Methods Explained
Why was ANOVA used?
The primary endpoint is a continuous change-from-baseline measurement, expressed as percentage of HbA1c. ANOVA provides a linear-model framework for comparing treatment groups while simultaneously incorporating categorical factors and a continuous baseline covariate.
What does a treatment difference of −0.04 mean?
The estimate indicates that the modeled change in HbA1c differed by −0.04 percentage of HbA1c for faster-acting insulin aspart relative to NovoRapid. It is a group-level adjusted treatment contrast, not a prediction of the response for an individual participant.
Why is the confidence interval more important than the P-value for non-inferiority?
The registry defines non-inferiority through the upper limit of the 95% confidence interval relative to the 0.4% margin. The observed upper limit was 0.03%, which is below the prespecified boundary. The P-value of 0.310 addresses a different question and does not determine whether the non-inferiority margin was crossed.
What does the 95% confidence interval of −0.11 to 0.03 tell us?
It describes uncertainty around the estimated treatment difference under the stated model and analysis framework. It includes zero, so the data do not establish a statistically significant superiority difference at the conventional two-sided level. At the same time, its upper boundary remains below the 0.4% non-inferiority margin.
Why include baseline HbA1c as a covariate?
Baseline HbA1c is closely connected to the change-from-baseline outcome. Covariate adjustment can account for baseline variation and make the treatment comparison more statistically efficient than an analysis that ignores the baseline measurement.
Why use multiple imputation?
The registry identifies multiple imputation as part of the primary analysis. Its purpose is to address missing endpoint information using a statistical model rather than restricting the analysis to participants with complete observed data alone.
Does non-inferiority mean the two treatments are identical?
No. Non-inferiority does not establish that the treatment effects are exactly equal. It establishes that, under the prespecified criterion, the uncertainty around the treatment difference is sufficiently bounded that a disadvantage larger than the 0.4% margin is not supported by the analysis.
11. Non-Inferiority Logic in Detail
The direction of the endpoint must be considered carefully in a non-inferiority analysis. The registry reports the treatment difference as faster aspart minus NovoRapid and defines the relevant criterion using the upper limit of the 95% confidence interval.
| Quantity | Reported value | Statistical role |
|---|---|---|
| Point estimate | −0.04 | Estimated treatment difference |
| 95% CI lower limit | −0.11 | Lower bound of uncertainty interval |
| 95% CI upper limit | 0.03 | Relevant boundary for the stated NI criterion |
| NI margin | 0.4% | Maximum permitted disadvantage under the stated framework |
| P-value | 0.310 | Evidence associated with the reported treatment comparison; not the NI criterion itself |
Because the upper confidence limit of 0.03% is below the 0.4% margin, the entire reported 95% confidence interval lies within the non-inferiority region on the side relevant to the registry's criterion. This is the central statistical result of the trial's posted primary analysis.
The non-inferiority conclusion should not be translated into a claim that the treatments have identical effects. Instead, it means that the analysis provides sufficient precision, under the prespecified margin, to exclude a treatment disadvantage as large as the non-inferiority boundary.
The distinction matters because the point estimate itself is close to zero, but the confidence interval is what determines how much uncertainty remains. A non-inferiority trial is therefore fundamentally a question about both the estimated difference and the precision of that estimate.
12. Safety Results
The registry reports serious adverse events by randomized treatment arm. These counts provide a focused safety summary and should be distinguished from the primary efficacy endpoint.
| Safety measure | Faster Aspart | NovoRapid |
|---|---|---|
| Serious adverse events, affected / at risk | 38 / 544 | 40 / 544 |
The reported serious adverse-event counts are affected participants over the corresponding at-risk populations. These data describe serious adverse events, not the full adverse-event profile, and they do not form part of the primary HbA1c non-inferiority analysis.
Safety and efficacy answer different questions. The primary statistical analysis concerns change in HbA1c, whereas serious adverse-event counts describe an important safety outcome. A numerical comparison of serious adverse-event counts alone does not establish a treatment-related difference without a prespecified statistical safety analysis and appropriate consideration of exposure, event timing, and event definitions.
13. What the Primary Estimate Does — and Does Not — Mean
The treatment difference of −0.04 is an adjusted estimate comparing the modeled change in HbA1c between faster-acting insulin aspart and NovoRapid. It does not mean that every participant experienced a −0.04 change, nor does it describe individual-level treatment response.
The 95% CI of −0.11 to 0.03 describes uncertainty around the treatment difference. It does not represent the range of HbA1c responses among individual participants. The interval contains zero but remains below the 0.4% non-inferiority margin at its upper boundary.
The reported P = 0.310 does not measure the magnitude or clinical importance of the treatment difference. A P-value is a measure related to statistical evidence under a specified hypothesis; it should be interpreted alongside the treatment estimate and confidence interval.
The central non-inferiority conclusion comes from comparing the upper confidence limit with the prespecified margin. The upper limit of 0.03% is below the 0.4% margin specified in the registry analysis description.
14. Analysis Population and Interpretation
The registry defines the analysis population as the number of participants with available data. The posted analysis also states that multiple imputation was used and that treatment was assumed according to randomization.
This combination is important when interpreting the result. The randomized assignment provides the basis for the treatment comparison, while the missing-data strategy addresses incomplete endpoint information. The resulting estimate therefore reflects the statistical model and imputation framework rather than a simple comparison of raw observed changes.
Randomization
Randomized allocation provides the framework for comparing the two treatment strategies without defining the groups according to later treatment behavior.
Covariate adjustment
Baseline HbA1c is incorporated as a covariate, while region and baseline metformin use enter as factors.
Missing data
Multiple imputation is part of the posted primary analysis, so the result is not simply a complete-case calculation.
Non-inferiority
The treatment comparison is judged against the prespecified 0.4% margin rather than only against zero difference.
15. What the Registry Reports — and What It Does Not Report
The registry reports one primary endpoint and one formal statistical analysis for that endpoint. The posted analysis includes the treatment difference, two-sided 95% confidence interval, P-value, ANOVA method, covariate adjustment, and multiple-imputation approach.
The registry record does not report median HbA1c values, group-specific mean changes, subgroup treatment effects, longitudinal HbA1c trajectories, or additional formal endpoint analyses in the information summarized for this page. Those quantities therefore do not form part of the statistical interpretation here.
The registry reports 52 outcome measures, but the posted statistical-analysis record identifies one primary endpoint analysis. The presence of additional registered outcome measures does not by itself establish that each has a posted formal statistical comparison.
16. Limitations and Interpretation Issues
- Non-inferiority depends on the margin: the conclusion is tied to the prespecified 0.4% margin. The margin is therefore an essential part of interpreting the result rather than a secondary technical detail.
- Confidence-interval interpretation: the interval of −0.11 to 0.03 supports the stated non-inferiority criterion, but it does not prove exact equality between treatments.
- Missing-data assumptions: multiple imputation requires assumptions about the information used to represent missing endpoint values. The registry summary does not provide the detailed imputation model.
- Model dependence: the treatment difference is an adjusted model-based estimate incorporating treatment, region, baseline metformin use, and baseline HbA1c.
- Analysis population: the registry describes the analyzed population as participants with available data while also describing multiple imputation. The precise implementation details beyond the posted description are not reported here.
- Safety interpretation: serious adverse-event counts are reported, but the available summary does not provide a formal comparative safety analysis.
- Outcome scope: the posted primary statistical analysis concerns HbA1c change from week 0 to week 16. Other registered outcomes should not be treated as statistically established simply because they are listed as outcome measures.
17. Why This Trial Matters Statistically
onset 9 is a useful teaching example because its primary analysis illustrates several important principles of modern clinical-trial statistics within a single endpoint: randomized treatment comparison, covariate-adjusted ANOVA, multiple imputation, confidence-interval interpretation, and non-inferiority testing.
| Concept | How it appears in onset 9 |
|---|---|
| Randomization | Participants were randomized to faster-acting insulin aspart or NovoRapid. |
| Parallel design | The trial used a parallel-group design with 2 arms. |
| Blinding | The registry describes quadruple masking. |
| Continuous endpoint | The primary endpoint was change in HbA1c. |
| ANOVA | The primary endpoint was analyzed using ANOVA. |
| Covariate adjustment | Baseline HbA1c was included as a covariate. |
| Factor adjustment | Treatment, region, and baseline metformin use were included as factors. |
| Multiple imputation | Missing endpoint information was addressed through multiple imputation. |
| Confidence interval | The 95% CI was −0.11 to 0.03. |
| Non-inferiority | The upper confidence limit was compared with a 0.4% margin. |
| P-value | The posted treatment comparison had P = 0.310. |
The trial is particularly useful for distinguishing two ideas that are often confused: failure to demonstrate superiority and demonstration of non-inferiority. A P-value of 0.310 does not establish a statistically significant superiority difference, but the non-inferiority conclusion follows from the confidence interval's position relative to the prespecified 0.4% margin.
18. Statistical Interpretation of the Full Result
The primary estimate of −0.04 is close to zero relative to the stated non-inferiority margin. The 95% confidence interval extends from −0.11 to 0.03, indicating that the plausible range represented by the interval remains well within the 0.4% boundary on the upper side relevant to the non-inferiority assessment.
This illustrates why a point estimate alone is insufficient in a non-inferiority trial. If the estimate were close to zero but the confidence interval were very wide, the analysis might not be able to rule out a clinically important disadvantage. Conversely, a relatively precise confidence interval can support a non-inferiority conclusion even when it includes zero.
The statistical story in one comparison
Non-inferiority margin: 0.4% · P = 0.310
The upper confidence limit of 0.03% is below the prespecified 0.4% margin.
The most informative way to read this result is not simply "the P-value was nonsignificant." The primary analysis was designed around a non-inferiority question. The treatment estimate was −0.04, the uncertainty interval was −0.11 to 0.03, and the upper confidence limit was below the 0.4% margin. Those three quantities together describe the core statistical evidence for the registered non-inferiority assessment.
The result also demonstrates why confidence intervals are especially valuable for non-inferiority trials: they show both the estimated direction of the difference and how large a disadvantage remains compatible with the data under the stated model.
19. Related Tutorials
Learn more about the methods used in this trial:
20. Related Calculators
21. Sources
- ClinicalTrials.gov: NCT03268005 — onset 9.
- PubMed: PubMed record for PMID 32209647.
- PubMed: PubMed record for PMID 35290624.
Continue through Clinical Biostats
Connect this trial's statistical methods with deeper tutorials and practical statistical tools.
22. Record Summary
onset 9 provides a compact example of how a non-inferiority clinical-trial analysis should be interpreted. The phase 3 randomized trial compared faster-acting insulin aspart with NovoRapid in people with type 2 diabetes, with change in HbA1c from week 0 to week 16 as the registered primary endpoint. The posted analysis used ANOVA with treatment, region, and baseline metformin use as factors and baseline HbA1c as a covariate, together with multiple imputation.
The reported treatment difference was −0.04 percentage of HbA1c, with a two-sided 95% confidence interval of −0.11 to 0.03 and a P-value of 0.310. The registry specified a 0.4% non-inferiority margin and stated that non-inferiority was established when the upper confidence limit was less than or equal to that margin. Because the upper limit was 0.03%, the posted analysis satisfied that criterion.
The statistical lesson is broader than the numerical result itself. Non-inferiority analysis requires the treatment estimate, confidence interval, and prespecified margin to be interpreted together. The P-value does not measure effect size, the confidence interval does not describe individual patient responses, and non-inferiority does not mean that two treatments are exactly identical. Together, these distinctions provide the appropriate statistical framework for understanding the primary analysis of onset 9.