← Clinical Trials
Heart Failure Phase 3 Completed NCT03057951

EMPEROR-Preserved: Complete Statistical Analysis of Empagliflozin in Heart Failure

An independent statistical review of the randomized, double-blind phase 3 EMPEROR-Preserved trial comparing 10 mg empagliflozin with placebo in patients with heart failure, focusing on the registered time-to-event and continuous outcomes and the statistical models used to analyze them.

Trial period: 2017-03-02 to 2021-04-26  ·  Enrollment: 5988  ·  Sponsor: Boehringer Ingelheim
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 on this page are restricted to the ClinicalTrials.gov record.

1. Trial at a Glance

EMPEROR-Preserved was a randomized, parallel, double-blind phase 3 trial in heart failure. The registry reports 5988 enrolled participants and two study arms: 10 mg empagliflozin and placebo.

5988
Enrolled
Phase 3
2
Study arms
Empagliflozin vs placebo
0.79
Primary HR
95.03% CI 0.69–0.90
0.0003
Primary P-value
Superiority analysis
FeatureEMPEROR-Preserved
Trial nameEMPEROR-Preserved
NCT identifierNCT03057951
Therapeutic areaCardiology
ConditionHeart Failure
PhasePhase 3
AllocationRandomized
Design modelParallel
MaskingDouble
Primary purposeTreatment
Enrollment5988
InterventionsEmpagliflozin and placebo
Primary endpoint typeTime-to-event
Results postedYes
Outcome measures posted10
Statistical analyses posted10
Lead sponsorBoehringer Ingelheim

2. Clinical Question

The registered primary question concerns the time to first event of adjudicated cardiovascular death or adjudicated hospitalization for heart failure, comparing 10 mg empagliflozin with placebo from randomization through completion of the planned treatment phase.

Population

Patients enrolled in the phase 3 EMPEROR-Preserved trial for the condition of heart failure. The ClinicalTrials.gov record reports 5988 enrolled participants.

Intervention

10 mg empagliflozin.

Comparator

Placebo.

Primary question

How does 10 mg empagliflozin compare with placebo for the time to first adjudicated cardiovascular death or adjudicated hospitalization for heart failure?

3. Trial Design

01
Randomize5988 enrolled
02
Parallel armsEmpagliflozin vs placebo
03
Double blindMasked treatment assignment
04
Follow-upTime-to-event and continuous outcomes
05
AnalysisCox, joint frailty, MMRM and random-coefficient models
ARM · 10 mg EMPAGLIFLOZIN

Empagliflozin

  • 10 mg empagliflozin
  • Randomized treatment arm
  • Compared with placebo
  • Included in the Randomised Set for the registered efficacy analyses
ARM · PLACEBO

Placebo

  • Placebo
  • Randomized comparator arm
  • Compared with 10 mg empagliflozin
  • Included in the Randomised Set for the registered efficacy analyses
Design interpretation. Randomization establishes the comparison structure, while double masking is intended to reduce the influence of treatment assignment on trial conduct and outcome assessment. The ClinicalTrials.gov record identifies the allocation as randomized, the design model as parallel, and the masking as double.

4. Endpoints

The registry reports 10 outcome measures and 10 statistical analyses. The primary endpoint is a time-to-event composite; the secondary analyses span recurrent hospitalization, renal function, cardiovascular and all-cause mortality, diabetes onset, patient-reported outcome, and all-cause hospitalization.

RoleRegistered endpointTime frameEndpoint type
Primary Time to First Event of Adjudicated Cardiovascular (CV) Death or Adjudicated Hospitalisation for Heart Failure (HHF) From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary Occurrence of Adjudicated Hospitalisation for Heart Failure (HHF) (First and Recurrent) From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary eGFR (CKD-EPI) cr Slope of Change From Baseline At baseline, week 4, 12, 32, 52, 76, 100, 124, 148, 172 and week 196, up to 1043 days. Continuous
Secondary Time to the First Event in the Composite Renal Endpoint: Chronic Dialysis, Renal Transplant, or Sustained Reduction in eGFR (CKD-EPI)cr From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary Time to First Adjudicated Hospitalisation for Heart Failure (HHF) From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary Time to Adjudicated Cardiovascular (CV) Death From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary Time to All-cause Mortality From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event
Secondary Time to Onset of Diabetes Mellitus (DM) in Patients With Pre-DM From randomization until completion of the planned treatment phase, to 1403 days. Time-to-event
Secondary Change From Baseline in Kansas City Cardiomyopathy Questionaire (KCCQ) Clinical Summary Score at Week 52 At baseline and at week 12, week 32 and week 52. Continuous
Secondary Occurrence of All-cause Hospitalisation (First and Recurrent) From randomization until completion of the planned treatment phase, up to 1403 days. Time-to-event

Primary endpoint definition

The registered primary endpoint is the Time to First Event of Adjudicated Cardiovascular (CV) Death or Adjudicated Hospitalisation for Heart Failure (HHF). The registry definition in the ClinicalTrials.gov record also describes an incidence-rate calculation per 100 patient-years and states that patients without a specific endpoint event were censored. The registry-reported definition text ends before the complete censoring rule is reproduced, so no additional censoring rule is inferred here.

5. Analysis Populations and Statistical Structure

The statistical analyses use more than one analysis population. The distinction matters because the population used for a time-to-event efficacy analysis is not necessarily the same as the population used for an on-treatment longitudinal measurement.

AnalysisAnalysis populationStatistical role
Primary composite endpoint Randomised Set (RS), including all randomised patients. Primary randomized efficacy comparison.
HHF, first and recurrent Randomised Set (RS), including all randomised patients. Recurrent-event analysis accounting for dependence with cardiovascular death.
eGFR slope Only patients included in the treated set (TS) and with available data for this endpoint. Longitudinal analysis of renal function trajectory.
Renal composite endpoint Randomised Set (RS), including all randomised patients. Time-to-event efficacy analysis.
First HHF Randomised Set (RS), including all randomised patients. Time-to-first-event analysis.
CV death Randomised Set (RS), including all randomised patients. Time-to-event analysis.
All-cause mortality Randomised Set (RS), including all randomised patients. Time-to-event analysis.
Diabetes onset in pre-DM Randomised Set (RS), including all randomised patients and with available data for this endpoint. Time-to-event analysis in the specified population.
KCCQ Clinical Summary Score Only patients included in the treated set (TS), including values obtained on treatment, with available data at the relevant timepoint. Repeated-measures analysis.
All-cause hospitalization, first and recurrent Randomised Set (RS), including all randomised patients. Recurrent-event analysis accounting for dependence with all-cause mortality.
Why this distinction matters: the primary endpoint is analyzed in the Randomised Set, preserving the randomized comparison. In contrast, the eGFR slope and KCCQ analyses use treated patients with available data. Those longitudinal estimates therefore answer a somewhat different statistical question from the primary randomized time-to-event analysis.

6. Primary Result: Cardiovascular Death or Hospitalisation for Heart Failure

The primary analysis compared placebo with 10 mg empagliflozin in the Randomised Set. The registry reports a Cox proportional-hazards model with treatment and prespecified baseline and regional covariates, and identifies the hypothesis as superiority.

Hazard ratio for first adjudicated CV death or HHF

0.79

95.03% CI: 0.69–0.90   ·   P = 0.0003

Comparison: 10 mg Empagliflozin vs Placebo

FeaturePrimary analysis
EndpointTime to First Event of Adjudicated Cardiovascular (CV) Death or Adjudicated Hospitalisation for Heart Failure (HHF)
Analysis populationRandomised Set (RS), including all randomised patients.
MethodCox proportional-hazards model
Effect measureHazard Ratio (HR)
Estimate0.79
Confidence interval95.03% two-sided CI: 0.69–0.90
P-value0.0003
HypothesisSuperiority
Alpha0.0497, resulting from interim analysis
Clinical Biostats interpretation

The estimated hazard ratio of 0.79 means that, under the fitted Cox model, the estimated instantaneous rate of experiencing the composite event was 0.79 times that of the comparison group. Expressed as a simple relative-hazard interpretation, this corresponds to an estimated 21% lower hazard for the empagliflozin group relative to placebo.

The HR does not mean that 21% of patients avoided the event, nor does it represent a 21-percentage-point absolute reduction in event probability. It is a relative, model-based measure of the event rate over time.

The 95.03% confidence interval from 0.69 to 0.90 describes statistical uncertainty around the estimated hazard ratio under the analysis framework. It does not describe the range of individual patient effects.

The P-value of 0.0003 addresses evidence against the null hypothesis under the prespecified testing framework. It does not measure the size or clinical importance of the treatment effect. Here, the registry specifically reports an interim-analysis-adjusted alpha of 0.0497, so the P-value should be understood in the context of that sequential design.

Because this is a Cox model, interpretation also depends on the usual proportional-hazards framework. The registry does not provide a time-varying hazard-ratio assessment in the ClinicalTrials.gov record, so the single HR should not be interpreted as proof that the relative hazard was identical at every point in follow-up.

7. Secondary Results: Recurrent Heart-Failure Hospitalisation

The registry separately analyzes the occurrence of adjudicated hospitalization for heart failure as a first and recurrent event. Rather than treating repeated hospitalizations as independent observations, the analysis uses a joint frailty model that accounts for dependence between recurrent HHF and cardiovascular death.

Hazard ratio for first and recurrent HHF

0.73

95.03% CI: 0.61–0.88   ·   P = 0.0009

Comparison: 10 mg Empagliflozin vs Placebo

Clinical Biostats interpretation

An HR of 0.73 corresponds to an estimated 27% lower hazard under the reported joint frailty model. This is not equivalent to saying that the number of hospitalizations was reduced by exactly 27%, because recurrent-event models incorporate repeated events and their dependence with mortality.

The 95.03% CI of 0.61–0.88 quantifies uncertainty around the model-based treatment effect. Its width reflects the precision of the estimate; it is not a prediction interval for how much an individual patient's hospitalization risk will change.

The P-value of 0.0009 indicates the strength of evidence against the reported null hypothesis under the analysis framework. It is not a measure of the magnitude of the HR.

The joint frailty approach is important because a patient can experience recurrent hospitalization, while cardiovascular death is a competing terminal outcome that prevents further observed hospitalizations. Treating every hospitalization as an isolated independent observation would not represent that structure adequately.

8. Secondary Result: eGFR Slope

The eGFR analysis is fundamentally different from the time-to-event analyses. Measurements were collected at baseline and multiple scheduled weeks, and the registry reports a random-coefficient model with a random intercept and random slope per patient.

Treatment-by-time interaction

1.363

99.9% CI: 0.861–1.865   ·   P < 0.0001

Outcome unit: mL/min/1.73 meters squared/year

FeatureeGFR slope analysis
EndpointeGFR (CKD-EPI) cr Slope of Change From Baseline
Time pointsBaseline, week 4, 12, 32, 52, 76, 100, 124, 148, 172 and week 196
Maximum stated follow-upUp to 1043 days
Analysis populationOnly patients included in the treated set (TS) and with available data for this endpoint.
MethodRandom intercept random coefficient model
Effect measureTreatment by time interaction
Estimate1.363
Confidence interval99.9% two-sided CI: 0.861–1.865
P-value<0.0001
Clinical Biostats interpretation

The reported 1.363 is a treatment-by-time interaction estimate rather than a hazard ratio. It describes a difference in the modeled longitudinal trajectory of eGFR between treatment groups, in the registry's stated unit of mL/min/1.73 meters squared/year.

It should therefore not be translated into a percentage reduction in an event hazard. The estimate is tied to the random-coefficient model, which allows each patient to have a patient-specific intercept and slope.

The 99.9% confidence interval of 0.861–1.865 reflects the confidence level reported by the registry and is intentionally not converted into a 95% interval. The P-value is <0.0001, while the analysis notes specify alpha = 0.001.

The analysis uses on-treatment data from treated patients. That population definition is important when comparing this result conceptually with the primary endpoint, which uses the Randomised Set.

9. Secondary Time-to-Event Results

The registry reports several additional time-to-event outcomes. Most use Cox proportional-hazards models with adjustment for age, baseline eGFR, region, baseline diabetes status, sex, baseline LVEF, and treatment.

EndpointMethodEffect95% CIP-value
Time to first event in composite renal endpoint Cox proportional-hazards model HR 0.95 0.73–1.24 0.7243
Time to first adjudicated HHF Cox proportional-hazards model HR 0.71 0.60–0.83 <0.0001
Time to adjudicated CV death Cox proportional-hazards model HR 0.91 0.76–1.09 0.2951
Time to all-cause mortality Cox proportional-hazards model HR 1.00 0.87–1.15 0.9893
Time to onset of diabetes mellitus in patients with pre-DM Cox proportional-hazards model HR 0.84 0.65–1.07 0.1539

Renal composite endpoint

The reported HR for the time to the first event in the composite renal endpoint was 0.95, with a 95% two-sided CI of 0.73–1.24 and P = 0.7243. The composite consisted of chronic dialysis, renal transplant, or sustained reduction in eGFR (CKD-EPI)cr.

First adjudicated HHF

The HR for time to first adjudicated HHF was 0.71, with a 95% two-sided CI of 0.60–0.83 and P < 0.0001. This is a time-to-first-event analysis and is therefore conceptually different from the recurrent-HHF analysis above.

Cardiovascular death

The HR for time to adjudicated cardiovascular death was 0.91, with a 95% two-sided CI of 0.76–1.09 and P = 0.2951.

All-cause mortality

The HR for time to all-cause mortality was 1.00, with a 95% two-sided CI of 0.87–1.15 and P = 0.9893. An HR of exactly 1.00 is the point estimate of no relative difference in the modeled hazard between the two groups.

Diabetes onset among patients with pre-DM

For time to onset of diabetes mellitus among patients with pre-DM, the HR was 0.84, with a 95% two-sided CI of 0.65–1.07 and P = 0.1539. This endpoint used the Randomised Set with available data for the endpoint.

Multiplicity matters. These secondary endpoints are multiple statistical questions rather than repeated measurements of one single hypothesis. The ClinicalTrials.gov record identifies superiority for the primary endpoint and provide hypothesis-type information for individual analyses, but do not provide a complete multiplicity hierarchy for all secondary endpoints. Their P-values should therefore be interpreted according to the trial's prespecified testing framework rather than as a collection of independent confirmatory tests.

10. Secondary Result: KCCQ Clinical Summary Score

The registry reports a longitudinal analysis of change from baseline in the Kansas City Cardiomyopathy Questionnaire Clinical Summary Score at week 52. Unlike the time-to-event endpoints, this analysis uses an MMRM framework.

Adjusted mean difference at Week 52

1.32

95% CI: 0.45–2.19   ·   P = 0.0028

Comparison: 10 mg Empagliflozin vs Placebo

Clinical Biostats interpretation

The adjusted mean difference of 1.32 represents the model-adjusted difference between treatment groups for change from baseline in the KCCQ Clinical Summary Score at week 52, using the registry's stated MMRM analysis.

The 95% CI of 0.45–2.19 describes uncertainty around that adjusted mean difference. It does not imply that every patient experienced an improvement within that range.

The P-value of 0.0028 quantifies evidence against the null hypothesis under the reported model. It does not tell us how important a 1.32-point difference is to an individual patient.

MMRM is appropriate for repeated measurements because it models the longitudinal structure rather than treating each patient-timepoint observation as an independent cross-sectional observation. The registry reports fixed effects for treatment-by-visit interaction and baseline KCCQ Clinical Summary Score-by-visit interaction, along with age, baseline eGFR, baseline LVEF, region, baseline diabetes status, sex, and week reachable.

11. Secondary Result: Recurrent All-cause Hospitalisation

Hazard ratio for recurrent all-cause hospitalisation

0.93

95% CI: 0.85–1.01   ·   P = 0.1012

Comparison: 10 mg Empagliflozin vs Placebo

The registry used a joint frailty model accounting for dependence between recurrent all-cause hospitalization and all-cause mortality. The model included age, baseline eGFR, baseline LVEF, treatment, region, baseline diabetes status and sex.

Clinical Biostats interpretation

An HR of 0.93 corresponds to an estimated 7% lower hazard under the reported model. Because this is a recurrent-event analysis, that statement should not be interpreted as a simple 7% reduction in the total number of hospitalizations.

The 95% CI of 0.85–1.01 indicates uncertainty around the estimate. The interval crosses 1.00, while the P-value is 0.1012. The P-value is evidence information under the specified model; it is not an effect-size metric.

The use of a joint frailty model is statistically meaningful because all-cause mortality terminates a patient's ability to experience additional observed hospitalizations. The model therefore attempts to represent the dependence between the recurrent hospitalization process and the terminal mortality process.

12. Statistical Methodology

Cox proportional-hazards model

The registry identifies Cox proportional-hazards regression as the principal method for the primary endpoint and several secondary time-to-event endpoints. The primary model includes treatment, region, baseline diabetes status, age, sex, LVEF and baseline eGFR.

Conceptual Cox model
h(t | X) = h0(t) exp(β1X1 + β2X2 + ··· + βpXp)

The hazard ratio associated with treatment is obtained from the treatment coefficient. An HR below 1 indicates a lower estimated instantaneous event rate in the empagliflozin group relative to placebo under the model.

Covariate adjustment

The primary Cox analysis adjusts for region, baseline diabetes status, age, sex, LVEF and baseline eGFR in addition to treatment. The same general covariate set is described for several secondary Cox analyses.

Adjustment does not replace randomization. Rather, it incorporates prespecified prognostic information into the statistical model and can improve the precision of the estimated treatment contrast when the covariates explain some outcome variation.

Random-coefficient model for eGFR

The eGFR analysis uses a random coefficient model allowing a random intercept and random slope for each patient. Fixed effects include age, baseline eGFR, baseline LVEF, region, baseline diabetes status, sex, baseline-by-time interaction, treatment-by-time interaction and treatment.

This structure is useful for repeated measurements because patients can begin at different modeled baseline levels and can have different individual trajectories over time.

MMRM for KCCQ

The KCCQ analysis uses a Mixed Model Repeated Measures approach. The registry reports treatment-by-visit interaction and baseline KCCQ Clinical Summary Score-by-visit interaction among the fixed effects, together with baseline covariates and week reachable.

The treatment effect at a particular visit is therefore estimated within a longitudinal model rather than by simply subtracting two raw group means at week 52.

Joint frailty models

Two recurrent-event outcomes use joint frailty models. The HHF analysis accounts for dependence between recurrent HHF and cardiovascular death. The all-cause hospitalization analysis accounts for dependence between recurrent all-cause hospitalization and all-cause mortality.

The frailty component provides a way to represent unobserved patient-level propensity that can link repeated events within a patient and the terminal event process. This is materially different from applying an ordinary Cox model to each hospitalization as though repeated observations were independent.

13. Interim Analysis and Alpha Spending

The primary analysis notes state that the interim analysis resulted in an adjusted alpha of 0.0497. The registry also identifies interim analysis / alpha spending as a concept associated with the primary and some secondary analyses.

Why interim analysis changes alpha

Looking at accumulating trial data creates multiple opportunities to declare an effect. A sequential design therefore needs an error-control mechanism so that the overall probability of a false-positive conclusion remains consistent with the prespecified design.

What alpha = 0.0497 tells us

The registry explicitly reports alpha = 0.0497 for the primary Cox analysis as resulting from the interim analysis. The reported P-value should therefore be interpreted against the stated sequential-testing framework rather than assuming an unadjusted 0.05 threshold.

Primary testing framework
Reported alpha = 0.0497   ·   Reported P = 0.0003

The important statistical point is that an interim analysis changes the reference testing framework. The registry explicitly connects the primary analysis alpha to the interim analysis.

14. Why the Hazard Ratio Is the Natural Primary Effect Measure

The primary endpoint records time until the first occurrence of either adjudicated cardiovascular death or adjudicated hospitalization for heart failure. Time-to-event data contain more information than a simple event/no-event indicator because patients can have different lengths of follow-up.

A Cox model uses that timing information while accommodating right censoring. A patient who has not experienced the primary event by the end of observed follow-up does not simply disappear from the analysis; that patient's available event-free follow-up contributes information up to the censoring time.

Hazard-ratio interpretation
HR = 0.79  →  estimated hazard ratio for empagliflozin vs placebo

A hazard ratio below 1 indicates a lower estimated instantaneous event rate in the empagliflozin group under the fitted model. It is not an absolute risk difference, probability, or number-needed-to-treat measure.

The primary estimate of 0.79 is therefore best understood together with its 95.03% CI of 0.69–0.90 and P = 0.0003. Each quantity answers a different question: effect magnitude, uncertainty, and evidence against the null hypothesis.

15. Statistical Methods Explained

Why was a Cox proportional-hazards model used for the primary endpoint?

The primary endpoint is explicitly a time-to-event endpoint: time from randomization until the first adjudicated cardiovascular death or adjudicated hospitalization for heart failure. Cox regression is designed for this setting because it estimates a relative hazard while allowing follow-up to differ between patients and incorporating censoring.

What does an HR of 0.79 mean?

It means that the fitted model estimates the hazard in the empagliflozin group at 0.79 times the hazard in the placebo group. A simple transformation, 1 − 0.79, gives a 21% lower estimated hazard. It does not mean a 21% absolute reduction in event probability.

Why is the primary confidence interval 95.03% rather than 95%?

The registry reports a two-sided 95.03% confidence interval for the primary analysis and states that alpha = 0.0497 resulted from the interim analysis. The confidence level therefore reflects the sequential testing framework reported in the registry rather than an assumed generic 95% level.

Why is the recurrent HHF analysis different from the first-HHF analysis?

First HHF counts only the first hospitalization event. The recurrent-HHF endpoint explicitly includes first and recurrent events. Repeated events from the same patient are correlated, and cardiovascular death terminates future observed HHF events. The joint frailty model accounts for this dependence rather than treating every hospitalization as an independent observation.

Why was an MMRM used for the KCCQ endpoint?

KCCQ Clinical Summary Score was measured repeatedly at baseline, week 12, week 32 and week 52. MMRM models those repeated measurements jointly and allows the treatment effect to be evaluated within the longitudinal structure of the data.

Why does the eGFR analysis use a random coefficient model?

The eGFR endpoint is a trajectory rather than a single event. The registry's random coefficient model allows each patient to have a random intercept and random slope, while treatment-by-time and baseline-by-time interactions capture differences in modeled trajectories over follow-up.

Why should the secondary P-values not automatically be treated as separate confirmatory conclusions?

The trial evaluates multiple secondary outcomes in addition to the primary endpoint. Multiplicity can increase the chance of false-positive findings when many hypotheses are tested. The ClinicalTrials.gov record does not provide a complete endpoint hierarchy or multiplicity procedure for every secondary analysis, so the individual P-values should be interpreted in the context of the prespecified trial testing framework.

16. Safety Results

The ClinicalTrials.gov record reports serious adverse events by treatment arm as affected patients over patients at risk.

Safety measurePlacebo10 mg Empagliflozin
Serious adverse events 1543 / 2989 1436 / 2996

These figures are presented exactly as reported in the registry: affected patients divided by patients at risk. They should not be confused with the efficacy analysis populations or with the denominators used for individual longitudinal endpoints.

Safety interpretation: the serious-adverse-event figures provide an arm-level safety count and denominator, but the ClinicalTrials.gov record does not provide a detailed breakdown by event type, severity category, treatment exposure duration, or formal between-group safety comparison. Those additional quantities should not be inferred from the reported counts.

17. Primary Endpoint Statistical Walkthrough

Step 1 · Randomization

Establish the treatment comparison

The trial uses randomized allocation with two parallel treatment arms: placebo and 10 mg empagliflozin.

Step 2 · Follow-up

Observe time to the first composite event

The primary endpoint follows patients from randomization until completion of the planned treatment phase, up to 1403 days.

Step 3 · Model

Fit a Cox proportional-hazards model

The registry specifies treatment, region, baseline diabetes status, age, sex, LVEF and baseline eGFR as model terms.

Step 4 · Estimate

Express the treatment contrast as an HR

The reported hazard ratio is 0.79, with a 95.03% two-sided confidence interval of 0.69–0.90.

Step 5 · Test

Interpret evidence under the interim framework

The reported P-value is 0.0003, with alpha = 0.0497 resulting from the interim analysis.

18. Comparing First-Event and Recurrent-Event Analysis

One of the most instructive features of EMPEROR-Preserved's statistical analysis is that hospitalization appears in both first-event and recurrent-event formulations.

QuestionEndpointMethodEstimate
What is the hazard of the first adjudicated HHF? Time to First Adjudicated Hospitalisation for Heart Failure Cox proportional-hazards model HR 0.71 (95% CI 0.60–0.83), P < 0.0001
What is the treatment effect when first and recurrent HHF events are considered together with CV death? Occurrence of Adjudicated Hospitalisation for Heart Failure (HHF) (First and Recurrent) Joint frailty model HR 0.73 (95.03% CI 0.61–0.88), P = 0.0009

These are not redundant analyses. The first-HHF endpoint asks about the first hospitalization event, whereas the recurrent endpoint incorporates additional hospitalization events and explicitly models dependence with cardiovascular death. Consequently, their estimates should not be expected to be identical.

19. Comparing Time-to-Event and Longitudinal Outcomes

The trial also illustrates why statistical method should follow endpoint structure rather than being selected merely because one method is common in clinical trials.

Time-to-event

The primary composite, first HHF, cardiovascular death, all-cause mortality, diabetes onset and renal composite endpoint are analyzed with time-to-event methods. Cox regression is the principal reported approach.

Repeated continuous outcome

KCCQ Clinical Summary Score is measured repeatedly and analyzed using MMRM, with treatment-by-visit interaction and baseline score-by-visit interaction.

Trajectory outcome

eGFR slope is analyzed with a random-coefficient model that allows patient-specific intercepts and slopes.

Recurrent events

First and recurrent HHF and first and recurrent all-cause hospitalization use joint frailty models to account for dependence with mortality.

20. Confidence Intervals and P-values

The trial's results provide an opportunity to distinguish three related but different statistical quantities.

QuantityExample from EMPEROR-PreservedWhat it communicates
Point estimate Primary HR = 0.79 The estimated treatment effect under the specified model.
Confidence interval 95.03% CI 0.69–0.90 Statistical uncertainty around the primary HR estimate under the stated confidence framework.
P-value P = 0.0003 Evidence against the reported null hypothesis under the specified testing framework.

A P-value should not be read as the probability that the null hypothesis is true. Likewise, a confidence interval should not be interpreted as a range containing a particular patient's treatment effect. The point estimate, interval and P-value answer different statistical questions.

21. What the Primary HR Does — and Does Not — Mean

Effect size

The primary HR of 0.79 indicates a lower estimated hazard for the empagliflozin group relative to placebo under the fitted Cox model. In simple relative terms, it corresponds to a 21% lower estimated hazard.

Not an absolute risk reduction

The HR does not tell us the absolute probability that a patient will experience the composite endpoint, nor does it give the percentage of patients who personally benefited. Absolute risk measures require event probabilities at a specified time point or another absolute-effect framework.

Precision

The 95.03% CI of 0.69–0.90 provides the uncertainty interval reported by the registry for the primary analysis. A confidence interval is more informative than the point estimate alone because it shows the statistical precision of the estimated treatment contrast.

Testing

The P-value of 0.0003 is evidence information under the prespecified hypothesis-testing framework. It is not a measure of the clinical magnitude of the treatment effect and should not be compared across endpoints as though smaller P-values represented larger effects.

22. Limitations

23. Why This Trial Matters Statistically

EMPEROR-Preserved is a useful teaching case because its registry analyses span several important branches of modern clinical-trial statistics. The primary endpoint uses randomized time-to-event analysis, while secondary endpoints demonstrate why recurrent events and repeated continuous measurements require different modeling strategies.

Statistical conceptHow it appears in EMPEROR-Preserved
RandomizationRandomized, parallel phase 3 design with two treatment arms.
BlindingDouble masking.
Time-to-event analysisPrimary composite and multiple secondary endpoints.
Cox regressionPrimary endpoint and several secondary time-to-event outcomes.
Hazard ratioPrimary and multiple secondary time-to-event effect measures.
Covariate adjustmentAge, baseline eGFR, baseline LVEF, region, baseline diabetes status, sex and treatment are used in several models.
Interim analysisPrimary analysis alpha reported as 0.0497 resulting from interim analysis.
Recurrent eventsJoint frailty models for recurrent HHF and recurrent all-cause hospitalization.
MMRMKCCQ Clinical Summary Score measured longitudinally through week 52.
Random coefficientseGFR trajectory modeled with random intercept and random slope per patient.
Different analysis populationsRandomised Set for major efficacy time-to-event analyses versus treated-set analyses for eGFR and KCCQ.

24. Overall Statistical Interpretation

The primary result is a Cox-model hazard ratio of 0.79 for time to first adjudicated cardiovascular death or adjudicated hospitalization for heart failure, with a 95.03% two-sided confidence interval of 0.69–0.90 and P = 0.0003. The registry identifies the hypothesis as superiority and reports an alpha of 0.0497 resulting from interim analysis.

The secondary analyses show why a clinical trial cannot be reduced to one number. First HHF is analyzed with Cox regression, recurrent HHF with a joint frailty model, eGFR trajectory with a random-coefficient model, and KCCQ with MMRM. The renal composite, cardiovascular death, all-cause mortality, diabetes onset and recurrent all-cause hospitalization each have their own statistical estimand and uncertainty.

From a statistical perspective, the central lesson is that endpoint definition, follow-up structure, analysis population and model must be interpreted together. A hazard ratio for a first event is not interchangeable with a recurrent-event hazard ratio; a treatment-by-time interaction is not a hazard ratio; and a longitudinal adjusted mean difference should not be interpreted as an event-risk measure.

25. Related Tutorials

Learn more about the methods used in this trial:

26. Related Calculators

27. Sources

Continue with Clinical Biostats

Explore statistical tutorials, calculators, and additional clinical-trial analyses covering the methods used across modern randomized studies.

28. Record Summary

EMPEROR-Preserved provides a compact example of how statistical analysis must adapt to the structure of clinical-trial endpoints. The primary randomized comparison uses a Cox proportional-hazards model for a time-to-first composite endpoint, with an interim-analysis-adjusted alpha. Secondary analyses extend the statistical framework to recurrent hospitalization through joint frailty models, renal-function trajectories through a random-coefficient model, and repeated patient-reported outcomes through MMRM.

The reported primary HR of 0.79, its 95.03% CI of 0.69–0.90, and P = 0.0003 summarize one prespecified treatment comparison, but understanding the trial statistically requires more than the headline result. The analysis populations, endpoint definitions, censoring structure, covariate adjustment, recurrent-event dependence, repeated measurements, and interim-testing framework all affect how the reported estimates should be interpreted.

Clinical Biostats methodology: A trial-results page should not merely repeat the abstract. The goal is to reconstruct the statistical story of the trial in a standardized format while clearly separating reported evidence from educational interpretation.