← Clinical Trials
Myocardial Infarction Phase 3 Placebo-Controlled NCT04509674

EMPACT-MI: Complete Statistical Analysis of Empagliflozin in Myocardial Infarction

A statistical review of the randomized, quadruple-masked phase 3 EMPACT-MI trial, which tested whether empagliflozin 10 mg, compared with placebo, lowers the risk of heart failure hospitalisation or death in people who have had a myocardial infarction.

Start: 2020-12-16  ·  Primary completion: 2023-11-05  ·  Status: Completed  ·  Sponsor: Boehringer Ingelheim
About this page

This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record. Reported results and statistical interpretation are kept visibly separate: numerical results appear in tables and result panels, and interpretation appears in clearly labelled boxes.

1. Trial at a Glance

EMPACT-MI was a randomized, parallel-group, placebo-controlled phase 3 trial evaluating whether the SGLT2 inhibitor empagliflozin could reduce the composite of first heart failure hospitalisation or all-cause death in patients after a myocardial infarction. The primary comparison produced a hazard ratio below 1 whose confidence interval included 1, so the trial did not demonstrate superiority on its primary endpoint.

6522
Enrolled
2 randomized arms
0.90
Primary HR
95% CI 0.76–1.06
0.2061
Primary P-value
Superiority hypothesis
1004
Days, maximum
Individual observation time
FeatureEMPACT-MI
PhasePhase 3
ConditionMyocardial infarction
DesignRandomized, parallel-group, placebo-controlled
MaskingQuadruple (participant, care provider, investigator and outcomes assessor)
ArmsEmpagliflozin 10 mg; placebo
Enrollment6522
Primary endpointComposite of time to first heart failure hospitalisation or all-cause mortality
Primary analysisCovariate-adjusted Cox proportional-hazards model in the randomised set
Key secondary analysesNegative binomial regression for total (first and recurrent) events; Cox model for time to cardiovascular death
DatesStart 2020-12-16; primary completion 2023-11-05
ClinicalTrials.govNCT04509674
SponsorBoehringer Ingelheim (industry)

2. Clinical Question

The registry title states the aim directly: to test whether empagliflozin can lower the risk of heart failure and death in people who had a heart attack. Statistically, this translates into a superiority comparison of a composite time-to-event endpoint between two randomized groups.

Population

People who had a myocardial infarction. The registry record summarised here does not list detailed eligibility criteria or baseline characteristics.

Intervention

Empagliflozin 10 mg (drug), taken as blinded study medication.

Comparator

Matching placebo (drug).

Primary question

Does empagliflozin reduce the hazard of the first occurrence of heart failure hospitalisation or death from any cause, compared with placebo?

3. Trial Design

01
Enrol6522 participants after MI
02
RandomizeEmpagliflozin 10 mg or placebo
03
Blinded treatmentQuadruple masking
04
Event follow-upUp to 1004 days
05
AnalysisCox and negative binomial models
EXPERIMENTAL ARM

Empagliflozin 10 mg

  • SGLT2 inhibitor at a single fixed dose
  • Analysed as randomized in the randomised set
  • Serious adverse events: 765 of 3234 at risk
CONTROL ARM

Placebo

  • Placebo (drug), identical blinded regimen
  • Analysed as randomized in the randomised set
  • Serious adverse events: 798 of 3229 at risk
Event-driven follow-up. Every participant was followed from randomisation (or from first study drug administration, if randomisation occurred after the first dose) until their individual day of trial completion, up to 1004 days. Because participants entered the trial at different times but finished around a common end, observation times differ substantially between individuals. This is exactly the situation in which time-to-event and rate-based methods, rather than simple proportions, are needed.

4. Analysis Population and Covariate Adjustment

All posted efficacy analyses used the randomised set (RS): all randomised patients, whether treated or not. This is an intention-to-treat style population, which keeps the comparison anchored to the randomized allocation.

Each model adjusted for the same prespecified set of baseline covariates in addition to treatment:

CovariateWhy a covariate like this is commonly included
Type-2 diabetes at baselineStrongly prognostic for cardiovascular and heart failure events; also relevant to an SGLT2 inhibitor
Geographical regionCaptures differences in care patterns and event ascertainment across sites
Age at baselineMajor determinant of mortality and hospitalisation risk
Estimated glomerular filtration rate at baselineKidney function is prognostic for heart failure and death
Left ventricular ejection fraction at baselinePost-infarction ventricular function is a key predictor of heart failure
Persistent or permanent atrial fibrillation at baselineAssociated with heart failure hospitalisation and mortality
Prior myocardial infarction at baselineMarks a population with more extensive coronary disease
Peripheral artery disease at baselineMarker of systemic atherosclerotic burden
Smoking at baselineEstablished cardiovascular risk factor

In a randomized trial, adjusting for prognostic baseline covariates is not needed to remove confounding; randomization already does that on average. Its purpose is to account for outcome variation explained by patient characteristics, which generally improves the precision of the treatment effect estimate. For non-collapsible measures such as the hazard ratio, it also means the estimate is conditional on the covariates, so an adjusted HR is not numerically interchangeable with an unadjusted one.

5. Endpoints

EndpointRoleRegistry definition / time framePosted analysis
Composite of time to first heart failure hospitalisation or all-cause mortalityPrimaryReported as the incidence rate of the first occurrence of hospitalisation for heart failure (HHF) or death, whichever is earliest: patients with an event during time at risk divided by total time at risk in that group, multiplied by 100 (per 100 patient-years). From randomisation or first study drug administration until individual day of trial completion; up to 1004 days.Cox regression; hazard ratio
Total number of HHF or all-cause mortalityKey secondaryEvents per 100 patient-years at risk; same time frameNegative binomial regression; adjusted event rate ratio
Total number of non-elective cardiovascular hospitalisations or all-cause mortalityKey secondaryEvents per 100 patient-years at risk; same time frameNegative binomial regression; adjusted event rate ratio
Total number of non-elective all-cause hospitalisations or all-cause mortalityKey secondaryEvents per 100 patient-years at risk; same time frameNegative binomial regression; adjusted event rate ratio
Total number of hospitalisations for MI or all-cause mortalityKey secondaryEvents per 100 patient-years at risk; same time frameNegative binomial regression; adjusted event rate ratio
Time to cardiovascular mortalitySecondaryPatients with events per 100 patient-years at risk; same time frameCox regression; hazard ratio

Two different questions are being asked. The primary endpoint counts only a patient's first event, so it answers "how quickly does the first heart failure hospitalisation or death occur?" The key secondary endpoints count all events, including repeat hospitalisations, so they answer "how much total burden of hospitalisation and death accumulates per unit of follow-up time?"

6. Primary Endpoint Result

Time to first heart failure hospitalisation or all-cause death

Hazard ratio, empagliflozin vs placebo

0.90

95% CI (two-sided): 0.76–1.06   ·   P = 0.2061

Covariate-adjusted Cox proportional-hazards model · Randomised set · Superiority hypothesis

ItemRegistry entry
Groups comparedPlacebo vs empagliflozin 10 mg (estimate expressed as empagliflozin vs placebo)
Statistical methodRegression, Cox
Effect measureHazard ratio (HR)
Estimate0.90
95% confidence interval0.76 to 1.06 (two-sided)
P-value0.2061
HypothesisSuperiority
Summary unitPatients with events per 100 patient-years at risk
Clinical Biostats interpretation

What the estimate means. An HR of 0.90 means that, under the fitted Cox model and conditional on the baseline covariates, the estimated instantaneous rate of a first heart failure hospitalisation or death was 10% lower in the empagliflozin group than in the placebo group over the analysed follow-up.

What it does not mean. It does not mean that 10% fewer patients had an event, that 10% of patients benefited, or that each individual's risk fell by 10%. The hazard ratio is a relative, model-based summary; it says nothing about absolute risk differences on its own.

What the confidence interval says. The 95% CI of 0.76 to 1.06 is compatible with effects ranging from a 24% lower hazard to a 6% higher hazard. Because the interval includes 1.00, the data are consistent with no difference between the groups, and the superiority hypothesis was not confirmed. The interval is also fairly wide at its lower end, so a moderate benefit is not excluded either. "Not statistically significant" is not the same as "shown to have no effect".

Why the P-value is not an effect size. P = 0.2061 describes how compatible the observed data are with a true HR of 1.00 under the model. It does not measure how large or clinically important the effect is. The same HR of 0.90 could produce a very small P-value in a larger trial with more events, or a larger P-value in a smaller one.

Cautions. The Cox model assumes that the ratio of hazards is roughly constant over follow-up; if the effect changed over time, a single HR summarises an average. The composite treats a heart failure hospitalisation and a death as equivalent first events, and death also prevents a later hospitalisation from being observed. Analysis was in the randomised set, so the estimate reflects assignment to treatment rather than the effect of taking every dose.

7. Key Secondary Endpoint Results: Total Events

The four key secondary endpoints count first and recurrent events and were analysed with a negative binomial model that included treatment and the same baseline covariates, with log(observation time) as an offset. Each estimate is an adjusted event rate ratio for empagliflozin versus placebo in the randomised set, with a two-sided 95% confidence interval.

Key secondary endpoint (total events)Adjusted rate ratio95% CIP-value
HHF or all-cause mortality0.870.68–1.100.2423
Non-elective CV hospitalisations or all-cause mortality0.920.78–1.070.2867
Non-elective all-cause hospitalisations or all-cause mortality0.870.7654–0.99780.0463
Hospitalisations for MI or all-cause mortality1.060.83–1.350.6311

The estimates point in both directions: three rate ratios are below 1, and the ratio for hospitalisation for MI or death is 1.06. Only the endpoint of non-elective all-cause hospitalisations or all-cause mortality has a confidence interval that excludes 1 (upper limit 0.9978), and it does so only narrowly.

Clinical Biostats interpretation

What a rate ratio means. An adjusted rate ratio of 0.87 means the model-estimated rate of events per unit of observation time was 13% lower with empagliflozin than with placebo. A ratio of 1.06 corresponds to a 6% higher estimated rate. Unlike the hazard ratio, these estimates use every qualifying event, so a patient with three hospitalisations contributes three events.

Precision. The interval for total HHF or death (0.68–1.10) is wider than the interval for the broader all-cause hospitalisation composite (0.7654–0.9978). Broader composites accumulate more events, which usually narrows the interval, but they also dilute a drug-specific mechanism with hospitalisations unrelated to heart failure.

The P = 0.0463 result. This is the only key secondary result below 0.05, and the upper confidence limit sits just below 1.00. The registry lists the hypothesis for these analyses as "other / not stated" and does not describe a testing hierarchy or multiplicity adjustment. In many cardiovascular trials, key secondary endpoints are tested only after the primary endpoint is significant; where that kind of strategy applies, a nonsignificant primary result means later P-values are nominal and descriptive, not confirmatory. With four key secondary endpoints examined, one nominal P-value near 0.05 should be read as hypothesis-generating.

What these results do not show. They do not show which component (hospitalisation or death) drives each composite, and they do not establish an effect on MI: a rate ratio of 1.06 with CI 0.83–1.35 is compatible with modest benefit, no effect or modest harm.

8. Cardiovascular Mortality

Hazard ratio for cardiovascular death, empagliflozin vs placebo

1.03

95% CI: 0.81–1.31   ·   P = 0.8124

Covariate-adjusted Cox proportional-hazards model · Randomised set

Clinical Biostats interpretation

An HR of 1.03 is close to 1.00: the estimated hazard of cardiovascular death was 3% higher in the empagliflozin group under the model. The 95% CI of 0.81 to 1.31 is compatible with anything from a 19% lower to a 31% higher hazard, so the trial neither shows nor rules out a meaningful effect on cardiovascular mortality. The large P-value (0.8124) reflects that the point estimate is near the null, not that equivalence has been demonstrated; equivalence would require a prespecified margin and an interval lying entirely within it. Deaths from non-cardiovascular causes act as competing events for this endpoint, which the cause-specific Cox model treats as censoring.

9. Safety: Serious Adverse Events

ArmParticipants with serious adverse eventsParticipants at risk
Empagliflozin 10 mg7653234
Placebo7983229

The number of participants with at least one serious adverse event was similar in the two arms, with slightly fewer in the empagliflozin group against almost identical denominators. These counts are descriptive. A crude proportion does not account for differing exposure time, and serious adverse event counts in a cardiovascular outcome trial overlap with efficacy outcomes, since hospitalisations for heart failure or MI can themselves be reported as serious adverse events. The registry does not post a formal statistical comparison of serious adverse events.

10. Statistical Methodology

Cox proportional-hazards model (primary and CV death)

The primary endpoint and time to cardiovascular death were analysed with Cox regression including treatment and nine baseline covariates. The Cox model estimates the hazard ratio without specifying the shape of the baseline hazard, using only the ordering of event times across patients who remain at risk.

Model form
h(t | x) = h0(t) · exp(βtrt·Treatment + β1·Diabetes + β2·Region + … + β9·Smoking)

The hazard ratio for treatment is exp(βtrt). Patients who complete the trial without an event are censored at their individual completion day and contribute follow-up time up to that point.

Negative binomial regression (total events)

Recurrent-event endpoints were analysed with negative binomial regression. Like Poisson regression, it models event counts with an offset for time at risk, but it adds a dispersion parameter so that the variance can exceed the mean. That matters because hospitalisations cluster: some patients are admitted repeatedly while most have none, and ignoring this overdispersion would produce confidence intervals that are too narrow.

Model form
log E[Yi] = log(Ti) + β0 + βtrt·Treatment + covariates,   Var(Yi) = μi + k·μi2

Yi is the number of events for patient i, Ti is observation time (the offset), and k is the dispersion parameter. The adjusted rate ratio is exp(βtrt).

Incidence rates per 100 patient-years

The registry defines the descriptive summary for the primary endpoint as patients with a first event divided by total time at risk, multiplied by 100. This exposure-adjusted rate handles unequal follow-up better than a simple percentage, but it assumes a roughly constant rate over time and is a descriptive summary, not the adjusted treatment comparison.

Randomised set

All posted efficacy analyses were performed in all randomised patients, whether treated or not. This preserves the balance created by randomization and answers the policy-level question of what happens when patients are assigned to empagliflozin rather than placebo.

11. Multiplicity and Endpoint Hierarchy

EMPACT-MI had a single primary endpoint tested for superiority, four key secondary total-event endpoints, and cardiovascular mortality as a further secondary endpoint. The registry does not describe how type I error was controlled across these analyses.

AnalysisRoleInterpretation
First HHF or all-cause deathPrimary (superiority)Confirmatory test; CI includes 1, so superiority not shown
Four total-event compositesKey secondarySupportive; one nominal P-value below 0.05 should be read in light of the primary result
Time to CV deathSecondaryDescriptive; estimate close to 1 with a wide interval
Serious adverse eventsSafetyCounts by arm; no formal test posted
Why this matters: when the primary endpoint does not reach significance, a secondary result with P < 0.05 cannot rescue the trial's main conclusion. Testing several endpoints increases the chance that at least one crosses 0.05 by chance, which is why prespecified hierarchies or alpha-splitting rules exist.

12. Statistical Methods Explained

Why does a hazard ratio of 0.90 with a confidence interval of 0.76–1.06 count as a neutral trial?

The trial tested superiority, which requires evidence that the true HR is below 1. The two-sided 95% CI extends above 1.00, so the data remain compatible with no effect at the conventional 5% level. The point estimate favours empagliflozin, but the trial was not able to distinguish that estimate from chance variation with the prespecified level of confidence.

Why use a Cox model for the primary endpoint but a negative binomial model for the key secondary endpoints?

The primary endpoint stops counting at a patient's first event, which makes it a time-to-first-event outcome suited to the Cox model. The key secondary endpoints count every hospitalisation plus death, so the outcome is a count per unit of time. Counts of recurrent events are naturally modelled with rate models, and the negative binomial form accommodates the fact that repeated hospitalisations are concentrated in a minority of patients.

What does the log(observation time) offset do?

Participants were followed for different lengths of time, up to 1004 days. A patient followed for two and a half years has more opportunity for events than one followed for a year. Including log(observation time) as an offset fixes its coefficient at 1, so the model describes events per unit of time rather than raw counts, and the rate ratio compares like with like.

Why adjust for nine baseline covariates in a randomized trial?

Randomization balances covariates on average, so adjustment is not about bias correction. Adjusting for strongly prognostic variables such as age, ejection fraction, kidney function and diabetes explains part of the between-patient variation in outcomes, which usually improves statistical efficiency. Because the covariates were prespecified, adjustment does not give analysts room to choose the model after seeing the data.

Is a hazard ratio of 0.90 the same as a rate ratio of 0.87?

No. The hazard ratio compares instantaneous rates of a first event among patients still event-free, while the rate ratio compares average total event rates over follow-up, including repeat events. They are often similar in direction and size, but they estimate different quantities and are not interchangeable.

Why is P = 0.8124 for cardiovascular death not evidence that empagliflozin has no effect on it?

A large P-value only says the observed estimate is close to what would be expected if there were no effect. The confidence interval of 0.81–1.31 still includes clinically meaningful benefit and harm. Absence of evidence of a difference is not evidence of equivalence, which would require a predefined margin.

13. Limitations and Interpretation Issues

14. Why This Trial Matters Statistically

EMPACT-MI is a useful teaching case because a large, well-conducted trial produced a point estimate favouring treatment on its primary endpoint without confirming superiority, and then reported first-event and total-event analyses that answer related but distinct questions.

ConceptHow it appears in EMPACT-MI
Superiority testingPrimary HR 0.90 with CI 0.76–1.06: favourable estimate, hypothesis not confirmed
Cox proportional-hazards modelCovariate-adjusted HR for time to first HHF or death and for CV death
Negative binomial regressionTotal-event analyses with an observation-time offset and overdispersion
Hazard ratio vs rate ratioFirst-event and recurrent-event summaries of related composites
Covariate adjustmentNine prespecified prognostic baseline covariates in every model
Confidence intervalsDistinguish "no effect shown" from "no effect"
MultiplicityOne nominal P-value of 0.0463 among several secondary tests after a neutral primary
Exposure-adjusted ratesIncidence per 100 patient-years for unequal follow-up up to 1004 days
Composite endpoints and competing risksDeath and hospitalisation combined; death precludes later hospitalisation

15. Related Tutorials

Learn more about the methods used in this trial:

16. Related Calculators

17. Sources

Continue learning with Clinical Biostats

Connect the endpoints and methods of this trial to in-depth statistical tutorials, then try them out with a calculator.

18. Record Summary

EMPACT-MI is a clear example of how to read a neutral cardiovascular outcome trial. Its primary covariate-adjusted Cox analysis estimated a 10% lower hazard of first heart failure hospitalisation or death with empagliflozin, but the 95% CI of 0.76–1.06 included no effect and P = 0.2061. Total-event analyses using negative binomial regression gave rate ratios between 0.87 and 1.06, with a single nominal P-value below 0.05, and the hazard ratio for cardiovascular death was 1.03. Sound interpretation combines the direction and size of each estimate, the width of its interval, the distinction between first and total events, and the multiplicity context.

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