← Clinical Trials
Type 2 Diabetes Phase 4 Non-Inferiority NCT01897532

CARMELINA: Complete Statistical Analysis of Linagliptin in Type 2 Diabetes

An independent statistical analysis of the randomized, double-blind CARMELINA trial evaluating linagliptin versus placebo in patients with type 2 diabetes mellitus, with emphasis on the primary cardiovascular time-to-event endpoint, renal outcome, non-inferiority testing, Cox regression, confidence intervals, and the interpretation of hazard ratios.

CARMELINA  ·  NCT01897532  ·  Phase 4  ·  Completed
Scope of this record

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 and design details on this page are restricted to the ClinicalTrials.gov record.

1. Trial at a Glance

CARMELINA was a randomized, parallel, double-blind phase 4 clinical trial comparing linagliptin with placebo in patients with type 2 diabetes mellitus. The registered primary endpoint was a time-to-event composite of adjudication-confirmed cardiovascular death, non-fatal myocardial infarction, or non-fatal stroke.

6991
Enrolled
Total trial enrollment
2
Arms
Placebo vs linagliptin
1.02
Primary HR
95% CI 0.89–1.17
1.3
NI Margin
Prespecified margin
FeatureCARMELINA
Trial nameCARMELINA
NCT identifierNCT01897532
PhasePhase 4
StatusCompleted
Therapeutic areaEndocrinology
ConditionDiabetes Mellitus, Type 2
AllocationRandomized
Design modelParallel
MaskingDouble
Primary purposeTreatment
Enrollment6991
InterventionsPlacebo; Linagliptin
Lead sponsorBoehringer Ingelheim
Sponsor typeIndustry
Study datesStart: 2013-07-10; Primary completion: 2018-01-18

2. Clinical Question

The registered primary question can be framed statistically as a comparison of time to the first occurrence of a major cardiovascular event between patients randomized to linagliptin and patients randomized to placebo. The registry identifies non-inferiority as the hypothesis type for the first primary analysis.

Population

Patients enrolled in CARMELINA with the registered condition of diabetes mellitus, type 2.

Intervention

Linagliptin.

Comparator

Placebo.

Primary question

For the primary cardiovascular composite endpoint, is linagliptin non-inferior to placebo under the prespecified non-inferiority margin?

3. Trial Design

01
Enroll6991 patients
02
RandomizeTwo parallel arms
03
MaskDouble-blind design
04
ObserveTime-to-event outcomes
05
AnalyzeCox proportional hazards
ARM 1

Placebo

  • Placebo intervention
  • Comparator group
  • Serious adverse events: 1343/3485
ARM 2

Linagliptin

  • Linagliptin intervention
  • Experimental treatment group
  • Serious adverse events: 1293/3494
Design classification: the registry describes CARMELINA as randomized, parallel, double-blind, and treatment-oriented. The trial has two arms and 6991 enrolled participants. The ClinicalTrials.gov record does not provide additional allocation-ratio or baseline-characteristic details, so those quantities are not reconstructed here.

4. Primary Endpoint

The registered primary endpoint is a time-to-event composite endpoint. It measures time from randomization to the first occurrence of any adjudication-confirmed component of 3-point Major Adverse Cardiovascular Events (3-point MACE).

EndpointRegistered definitionTime frameType
3-point MACE Time to the first occurrence of any adjudication-confirmed component of the primary composite endpoint: cardiovascular death, non-fatal myocardial infarction, or non-fatal stroke. From randomization to individual end of observation; up to 4.3 years Time-to-event

The registry states that the percentage of observed patients with first occurrence of any of the adjudication-confirmed components was reported. The formal statistical analysis compares the placebo and linagliptin groups using a Cox proportional-hazards model.

5. Secondary Endpoint

CARMELINA also has a registered secondary time-to-event endpoint focused on renal outcomes.

EndpointDefinitionTime frameAnalysis
Renal composite endpoint Time to the first occurrence of any adjudication-confirmed component: renal death, sustained end stage renal disease (ESRD), or sustained decrease of 40% or more in estimated glomerular filtration rate (eGFR). From randomization to individual end of observation; up to 4.3 years Cox proportional-hazards model

6. Statistical Methodology

Cox proportional-hazards regression

The registry reports a Cox proportional-hazards regression model for the primary and secondary time-to-event analyses. The primary analysis model included randomized treatment and geographical region as factors. Breslow's method was used for dealing with ties.

Core model concept
h(t | X) = h0(t) exp(βX)

In a Cox model, the treatment coefficient is transformed into a hazard ratio. For the treatment comparison here, the reported effect measure is the hazard ratio for linagliptin versus placebo.

Hazard ratio

The hazard ratio is a relative time-to-event measure. An HR of 1 would correspond to equal estimated hazards between the two groups under the model. An HR below 1 would indicate a lower estimated instantaneous event rate for linagliptin, while an HR above 1 would indicate a higher estimated instantaneous event rate.

Primary estimate
HR = 1.02    95% CI = 0.89–1.17

The reported point estimate is close to 1. The confidence interval extends below and above 1, so the interval represents uncertainty spanning both directions relative to the null value.

Handling tied event times

The registry states that Breslow's method was used for dealing with ties. Ties arise when multiple participants experience the event at the same recorded time. A tie-handling method allows the Cox model to incorporate such events rather than requiring every event time to be unique.

Geographical region

The reported Cox model included randomized treatment and geographical region as factors. This means the primary treatment comparison was not represented as a model containing treatment alone; geographical region was also incorporated into the fitted regression model.

7. Primary Results: 3-Point MACE

The registry contains two statistical analyses for the same primary endpoint. They share the same Cox proportional-hazards methodology, the same reported point estimate, and the same 95% confidence interval, but they correspond to different hypothesis-testing roles.

Non-Inferiority Analysis

Hazard ratio for first 3-point MACE

1.02

95% CI: 0.89–1.17   ·   P = 0.0002

Non-inferiority margin: 1.3

Clinical Biostats interpretation

The reported hazard ratio of 1.02 means that the estimated hazard of the first 3-point MACE event under the fitted Cox model was approximately 2% higher with linagliptin than with placebo. This is a model-based relative estimate; it does not mean that 2% more patients experienced an event, nor does it describe an individual's probability of experiencing MACE.

The 95% CI of 0.89–1.17 describes statistical uncertainty around the estimated hazard ratio. Importantly, the interval is entirely below the prespecified non-inferiority margin of 1.3. The non-inferiority question is therefore evaluated against the margin rather than against the null value of 1 alone.

The reported P = 0.0002 is associated with the non-inferiority hypothesis test and should not be interpreted as a measure of the size of the treatment effect. A p-value does not tell us whether the observed effect is clinically large or small; the HR and its confidence interval provide the effect estimate and its precision.

The registry specifies a one-sided alpha level of 2.5% for the first non-inferiority hypothesis. The reported confidence interval is two-sided at 95%, so the inferential quantities should not be treated as interchangeable representations of the one-sided testing procedure.

Additional Analysis of the Same Primary Endpoint

Hazard ratio for first 3-point MACE

1.02

95% CI: 0.89–1.17   ·   P = 0.6301

Hypothesis type: Other / not stated

Clinical Biostats interpretation

This second posted analysis has the same estimated hazard ratio, 1.02, and the same 95% CI of 0.89–1.17, but a different reported p-value, 0.6301, because the registry identifies the hypothesis type for this analysis as "Other / not stated" rather than non-inferiority.

The effect estimate still describes the fitted treatment comparison, while the confidence interval still describes uncertainty around that estimate. The p-value of 0.6301 should not be used to reinterpret the effect estimate as zero or to replace the prespecified non-inferiority framework. The appropriate interpretation depends on the hypothesis attached to the analysis.

For non-inferiority, the critical reference point is the prespecified margin of 1.3. A conventional superiority-style comparison against 1 asks a different statistical question from whether the treatment effect remains within the acceptable non-inferiority margin.

Primary analysisEstimate95% CIP-valueHypothesis
First primary analysisHR 1.020.89–1.170.0002Non-inferiority
Second primary analysisHR 1.020.89–1.170.6301Other / not stated

8. Why the Non-Inferiority Margin Matters

The registry states that the non-inferiority margin was chosen as 1.3. This is the key reference value for the primary non-inferiority hypothesis.

Non-inferiority logic
Upper 95% CI limit = 1.17  <  1.3 = non-inferiority margin

For a hazard ratio where values above 1 represent greater estimated hazard, the upper confidence-limit comparison with the prespecified margin is central to the non-inferiority interpretation.

The distinction between 1 and 1.3 is fundamental. A null hypothesis for equality or superiority is concerned with whether the treatment effect differs from 1. A non-inferiority hypothesis instead asks whether the treatment could be worse by more than the prespecified acceptable amount.

Non-inferiority is not "proving the treatments are identical." A non-inferiority result means that the data are sufficiently compatible with the treatment effect remaining within the prespecified non-inferiority boundary. It does not establish that the two interventions have exactly the same effect.

9. Secondary Results: Renal Composite Endpoint

The registry reports a formal Cox proportional-hazards analysis for the secondary renal composite endpoint.

Hazard ratio for first renal composite event

1.04

95% CI: 0.89–1.22   ·   P = 0.6918

Renal death, sustained ESRD, or sustained decrease of 40% or more in eGFR

Clinical Biostats interpretation

The reported hazard ratio of 1.04 indicates that the estimated hazard of the first renal composite event was approximately 4% higher with linagliptin than with placebo under the fitted Cox model. This is a relative model-based estimate and does not represent a 4-percentage-point increase in the probability of a renal event.

The 95% CI of 0.89–1.22 includes 1, indicating that the interval is compatible with both a lower and a higher hazard relative to placebo. The interval is also narrower than the non-inferiority margin of 1.3, but the ClinicalTrials.gov record identifies this secondary analysis as "Other / not stated" rather than as the primary non-inferiority hypothesis.

The P = 0.6918 is not an effect-size measure. It should be interpreted in the context of the hypothesis attached to this secondary analysis and the fact that secondary endpoints can have a different inferential role from the primary endpoint.

Secondary endpointHR95% CIP-valueMethod
Renal death, sustained ESRD, or sustained decrease of 40% or more in eGFR 1.04 0.89–1.22 0.6918 Cox proportional-hazards model

10. Analysis Population

The primary statistical analyses were conducted in the Treated Set (TS). The registry defines this as all patients treated with at least one dose of trial medication, with additional registry wording concerning medication dispensing and site visits.

Treated Set

The primary analysis population identified by the registry for the posted cardiovascular analyses.

Why this matters

The analysis population determines which participants contribute to the reported estimate and therefore forms part of the definition of the statistical result.

The ClinicalTrials.gov record does not provide a separate intention-to-treat analysis population definition or a per-protocol population for the posted results. Those populations are therefore not substituted for the reported Treated Set.

11. One-Sided Testing and Sequential Hypotheses

The registry explicitly identifies one-sided testing in the analysis text. The first hypothesis, non-inferiority of the primary endpoint, was to be tested at a one-sided alpha level of 2.5%.

Why one-sided?

Non-inferiority is directional: the relevant concern is whether the treatment is worse than placebo by more than the prespecified acceptable margin.

Why 2.5%?

The registry explicitly states a one-sided alpha level of 2.5% for the first non-inferiority hypothesis.

The registry analysis note further states that in case of significance the next set of hypotheses was to be considered. The registry-reported text does not provide the complete subsequent hypothesis sequence, so no additional hierarchy is reconstructed here.

Statistical caution: the presence of two posted analyses for the primary endpoint does not mean that their p-values answer the same question. The first is explicitly identified as non-inferiority; the second is classified as "Other / not stated." The hypothesis attached to an analysis is essential to interpreting its p-value.

12. Kaplan-Meier Estimation and Time-to-Event Thinking

The primary and secondary endpoints are time-to-event outcomes. In this setting, a participant contributes information not only through whether an event occurred but also through the amount of observation available before the event or the end of observation.

Kaplan-Meier concept
S(t) = P(T > t)

The survival function represents the probability of remaining event-free beyond time t. For CARMELINA, the relevant event can be the first qualifying cardiovascular or renal composite event.

The registry's formal method is the Cox proportional-hazards model. A Kaplan-Meier framework is useful for describing time-to-event data, but the posted registry result here is expressed as a hazard ratio rather than as a reported median event time or a specific Kaplan-Meier survival probability.

The registry-reported CARMELINA data do not report median event times or Kaplan-Meier percentages at specific time points. Those quantities are therefore not added or reconstructed.

13. Safety Results

The ClinicalTrials.gov record provides serious adverse events by treatment arm. The reported values are affected participants divided by participants at risk.

Safety measurePlaceboLinagliptin
Serious adverse events1343/34851293/3494
Serious adverse events · affected participants
Placebo
1343
Linagliptin
1293

The numerator/denominator presentation is preferable here to calculating an additional percentage because the task specifies that reported numbers should not be recomputed or converted. The two arms also have different reported numbers at risk, so the raw affected-participant counts should not be interpreted as directly comparable risks without accounting for their denominators and the underlying safety-analysis framework.

14. Trial Timeline

2013-07-10

Trial start

CARMELINA began enrollment on 2013-07-10.

Up to 4.3 years

Registered observation window

The primary cardiovascular and secondary renal endpoints were defined from randomization to individual end of observation, up to 4.3 years.

2018-01-18

Primary completion

The registry records 2018-01-18 as the primary completion date.

Completed

Results posted

The ClinicalTrials.gov record identifies the study as completed and indicates that results were posted.

15. What the Hazard Ratio Does — and Does Not — Mean

Effect size

For the primary endpoint, an HR of 1.02 means the fitted model estimated the instantaneous hazard of first 3-point MACE to be approximately 2% higher with linagliptin than placebo. This does not mean that 2% more participants experienced MACE, and it does not mean that every participant had a 2% higher individual risk.

Confidence interval

The 95% CI of 0.89–1.17 describes uncertainty around the estimated hazard ratio. It is not a range containing 95% of individual patient effects, nor does it directly provide the probability that the true HR lies within the interval.

P-value

The two primary analyses report different p-values, 0.0002 and 0.6301, despite the same HR and confidence interval. This illustrates why a p-value cannot be interpreted independently of the hypothesis being tested. The first analysis is explicitly non-inferiority; the second is classified as "Other / not stated."

Cox-model caution

The hazard ratio is conditional on the Cox proportional-hazards modeling framework. A single HR summarizes the relative hazard under that model; it is not a direct description of cumulative incidence or a time-specific absolute risk difference.

16. Understanding the Primary Result as a Non-Inferiority Trial

The central statistical feature of CARMELINA's primary analysis is not simply whether the HR differs from 1. Instead, the prespecified non-inferiority margin of 1.3 defines how much higher the hazard could be before the result would fail the stated non-inferiority criterion.

QuantityValueStatistical role
Null/equality reference1No difference in hazard under the HR scale
Observed HR1.02Point estimate of relative hazard
95% CI lower limit0.89Lower uncertainty boundary
95% CI upper limit1.17Upper uncertainty boundary
Non-inferiority margin1.3Maximum prespecified relative hazard compatible with non-inferiority framework
One-sided alpha2.5%Significance level for first non-inferiority hypothesis

This distinction is one of the most important lessons from the trial. A treatment can fail to demonstrate superiority against 1 while still meeting a non-inferiority objective against a larger prespecified margin. Conversely, a confidence interval that crosses 1 does not automatically invalidate a non-inferiority claim if its upper boundary remains below the non-inferiority margin.

17. Statistical Methods Explained

Why was a Cox proportional-hazards model used?

The primary and secondary endpoints are time-to-event outcomes, with observation extending from randomization to each participant's individual end of observation. Cox regression is designed for this type of censored survival data and provides a hazard ratio as a relative treatment-effect measure.

What does an HR of 1.02 mean?

An HR of 1.02 means that the estimated instantaneous hazard under the fitted model was approximately 2% higher for linagliptin than placebo. It does not mean a 2-percentage-point increase in event probability, nor does it imply that every patient experienced the same relative effect.

Why is the non-inferiority margin 1.3 more important than simply asking whether the HR is above or below 1?

The equality value of 1 and the non-inferiority margin answer different questions. The value 1 represents equal hazards. The margin of 1.3 represents the prespecified boundary for how much worse the treatment could be while still satisfying the non-inferiority objective.

Why can the same HR have different p-values?

The ClinicalTrials.gov record shows HR 1.02 with a 95% CI of 0.89–1.17 in both primary analyses, but the p-values are 0.0002 and 0.6301. The analyses have different stated hypothesis roles: the first is non-inferiority, while the second is "Other / not stated." A p-value is therefore inseparable from the null hypothesis and testing procedure that generated it.

What does the 95% confidence interval tell us?

The interval 0.89–1.17 represents uncertainty around the estimated hazard ratio under the specified statistical framework. It includes the equality value of 1 but remains below the non-inferiority margin of 1.3. This distinction is central to interpreting the primary result.

Why does Breslow's method appear in the analysis?

The registry states that Breslow's method was used for dealing with ties in the Cox model. When multiple events occur at the same recorded time, a tie-handling method is needed because the standard Cox partial-likelihood formulation assumes distinct event times.

18. Primary vs Secondary Statistical Questions

EndpointRoleHR95% CIP-value
First 3-point MACE Primary non-inferiority analysis 1.02 0.89–1.17 0.0002
First 3-point MACE Primary analysis; other / not stated 1.02 0.89–1.17 0.6301
First renal composite event Secondary 1.04 0.89–1.22 0.6918

The table illustrates why trial results should be organized by endpoint and hypothesis rather than by p-value alone. The primary cardiovascular endpoint has a formal non-inferiority analysis, while the renal endpoint is secondary and carries a different inferential role.

19. Limitations

20. Why This Trial Matters Statistically

CARMELINA is a useful teaching case because the central statistical question is more nuanced than a conventional superiority comparison. The trial combines randomization and blinding with a time-to-event endpoint, Cox regression, a hazard-ratio effect measure, one-sided non-inferiority testing, a prespecified non-inferiority margin, and a secondary renal composite endpoint.

ConceptHow it appears in CARMELINA
RandomizationRandomized allocation to placebo or linagliptin
BlindingDouble-blind design
Time-to-event analysisPrimary cardiovascular and secondary renal endpoints
Cox regressionFormal model for treatment comparisons
Hazard ratioReported effect measure for the time-to-event analyses
Confidence interval95% two-sided intervals around the reported HR estimates
Non-inferiorityPrimary hypothesis with margin of 1.3
One-sided testingFirst non-inferiority hypothesis tested at one-sided alpha of 2.5%
Tie handlingBreslow's method in the Cox model
Model adjustmentRandomized treatment and geographical region included as factors
Composite endpoint3-point MACE combines cardiovascular death, non-fatal MI, and non-fatal stroke
Renal compositeRenal death, sustained ESRD, or sustained decrease of 40% or more in eGFR

21. A Practical Reading of the CARMELINA Results

A statistically literate reading of the primary result should proceed in several steps.

  1. Identify the endpoint: the primary outcome is time to first adjudication-confirmed 3-point MACE.
  2. Identify the analysis population: the posted primary analyses use the Treated Set.
  3. Identify the model: Cox proportional-hazards regression was used, with randomized treatment and geographical region as factors and Breslow's method for ties.
  4. Read the effect estimate: HR 1.02 is close to the equality value of 1.
  5. Read the confidence interval: 0.89–1.17 gives the uncertainty around the HR estimate.
  6. Identify the hypothesis: the first analysis is explicitly non-inferiority.
  7. Compare with the margin: the upper confidence limit of 1.17 is below the prespecified non-inferiority margin of 1.3.
  8. Interpret the p-value in context: P = 0.0002 belongs to the stated non-inferiority hypothesis, whereas the separate posted analysis reports P = 0.6301 for a hypothesis classified as "Other / not stated."

This sequence prevents a common statistical mistake: treating every clinical-trial p-value as though it were testing the same null hypothesis.

22. Related Statistical Concepts

Learn more about the methods used in this trial:

23. Related Statistical Calculators

The trial's methods also connect naturally to practical time-to-event and non-inferiority calculations:

24. Sources

Continue through the Clinical Biostats statistical pathway

Explore the statistical methods behind randomized clinical trials, survival analysis, non-inferiority testing, confidence intervals, and practical sample-size and analysis calculations.

25. Record Summary

CARMELINA is a randomized, double-blind, parallel phase 4 trial with 6991 enrolled participants comparing linagliptin with placebo. Its registered primary endpoint is time to first adjudication-confirmed 3-point MACE, analyzed using a Cox proportional-hazards model with randomized treatment and geographical region as factors and Breslow's method for ties. The primary non-inferiority analysis reports HR 1.02 with a 95% CI of 0.89–1.17 and P = 0.0002, against a prespecified non-inferiority margin of 1.3 and a one-sided alpha level of 2.5%. A separate posted analysis of the same endpoint reports the same HR and confidence interval with P = 0.6301. The secondary renal composite analysis reports HR 1.04, 95% CI 0.89–1.22, and P = 0.6918.

Clinical Biostats methodology: The central statistical lesson from CARMELINA is that a clinical-trial result cannot be interpreted from its hazard ratio or p-value alone. The endpoint definition, analysis population, Cox model, confidence interval, hypothesis type, one-sided testing framework, and non-inferiority margin all determine what the reported result means.