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 data reported in the ClinicalTrials.gov record.
1. Trial at a Glance
THEMIS was a randomized, quadruple-masked, parallel-group phase 3 trial comparing ticagrelor 60 mg with ticagrelor placebo in patients with type 2 diabetes mellitus. The primary endpoint was a composite of cardiovascular death, myocardial infarction, or stroke, analyzed as a time-to-event outcome using a Cox proportional-hazards model.
| Feature | THEMIS |
|---|---|
| Trial name | THEMIS |
| ClinicalTrials.gov identifier | NCT01991795 |
| Therapeutic area | Cardiology |
| Condition | Diabetes Mellitus, Type 2 |
| Phase | Phase 3 |
| Status | COMPLETED |
| Enrollment | 19271 |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | Quadruple |
| Primary purpose | Prevention |
| Interventions | Ticagrelor 60 mg; ticagrelor placebo |
| Primary endpoint | Composite of cardiovascular death, MI or stroke |
| Primary endpoint analysis | Cox proportional-hazards model; hazard ratio |
2. Clinical Question
The central statistical question was whether ticagrelor 60 mg differed from placebo with respect to the time to the first occurrence of the registered composite of cardiovascular death, myocardial infarction, or stroke in the THEMIS population.
Population
Participants in the THEMIS phase 3 study with the registered condition of type 2 diabetes mellitus.
Intervention
Ticagrelor 60 mg.
Comparator
Ticagrelor placebo.
Primary question
How does ticagrelor 60 mg compare with placebo for the time-to-event composite of cardiovascular death, MI or stroke?
3. Trial Design
Participants were allocated to study arms through randomization.
The trial used a parallel-group design with two study arms.
The registry identifies the study as quadruple masked.
The registered primary purpose was prevention.
Ticagrelor 60 mg
- Study intervention: ticagrelor 60 mg.
- Primary efficacy comparison: ticagrelor 60 mg versus ticagrelor placebo.
- Efficacy analyses used the full analysis set described in the registry.
Ticagrelor Placebo
- Study comparator: ticagrelor placebo.
- Primary efficacy comparison: ticagrelor placebo versus ticagrelor 60 mg.
- Efficacy analyses used the same full analysis set framework.
Trial timeline
Study start
The registry reports a study start date of February 10, 2014.
Primary completion
The registry reports January 25, 2019 as the primary completion date.
4. Analysis Populations
The primary efficacy analysis used the full analysis set. The registry defines this population as all randomized patients except the 51 patients who were randomized to study drug at a site prematurely closed by the sponsor.
| Population | Definition / role |
|---|---|
| Full analysis set | All randomized patients except the 51 patients randomized to study drug at a site prematurely closed by sponsor; used for the efficacy analyses reported here. |
| Safety analysis set | All patients who received at least 1 dose of randomized ticagrelor or placebo; used for safety endpoint analyses. |
5. Primary Endpoint
| Endpoint | Registry definition | Analysis |
|---|---|---|
| Composite of Cardiovascular (CV) Death, MI or Stroke | Participants with cardiovascular death, myocardial infarction or stroke. If no event, censoring occurs at the earliest of the primary analysis censoring date, last endpoint assessment date and non-CV death date. | Cox proportional-hazards model; hazard ratio |
The registered time frame was from randomisation to the primary analysis censoring date (PACD). The registry reports that the median time in study until PACD was 40 months.
This structure makes the primary outcome a time-to-event analysis rather than a simple comparison of the proportion of participants who experienced an event by an arbitrary fixed date.
6. Results: Primary Endpoint
The primary analysis compared ticagrelor 60 mg with ticagrelor placebo using a Cox proportional-hazards model in the full analysis set. The reported effect measure was the hazard ratio.
Composite of cardiovascular death, MI or stroke
95% CI: 0.81–0.99 · P = 0.0378
Two-sided confidence interval; superiority hypothesis.
A hazard ratio of 0.90 means that, under the fitted Cox model, the estimated instantaneous rate of experiencing the composite event was 0.90 times that of the comparator group over the analyzed follow-up. Expressed as a relative model-based quantity, this corresponds to an estimated 10% lower hazard.
The HR does not mean that exactly 10% fewer participants experienced an event, nor does it provide an absolute risk difference. A hazard ratio describes relative event rates over time and must be interpreted in the context of censoring, follow-up, and the proportional-hazards model.
The 95% CI of 0.81–0.99 quantifies uncertainty around the estimated hazard ratio under the analysis framework. It does not describe the range of individual patient effects. The interval is relatively close to 1, so the point estimate should not be interpreted without considering its uncertainty.
The P-value of 0.0378 addresses the statistical evidence against the null hypothesis under the prespecified superiority framework. It is not a measure of the magnitude or clinical importance of the treatment effect. A smaller p-value would not, by itself, imply a larger treatment effect.
7. Secondary Endpoint Results
The registry also reports formal Cox analyses for cardiovascular death, myocardial infarction, ischaemic stroke, and all-cause death. Each uses the same broad time frame, analysis population definition, treatment comparison, and hazard-ratio framework as the primary efficacy analysis.
| Endpoint | HR | 95% CI | P-value | Hypothesis |
|---|---|---|---|---|
| CV Death | 1.02 | 0.88–1.18 | 0.7883 | Superiority |
| MI | 0.84 | 0.71–0.98 | 0.0294 | Superiority |
| Ischaemic Stroke | 0.80 | 0.64–0.99 | 0.0375 | Superiority |
| All-cause Death | 0.98 | 0.87–1.10 | 0.6846 | Superiority |
Cardiovascular death
Hazard ratio for CV death
95% CI: 0.88–1.18 · P = 0.7883
The point estimate of 1.02 is close to the null value of 1.00. The 95% CI of 0.88–1.18 spans 1, indicating uncertainty that includes both a lower and a higher hazard relative to the comparator under the model.
The p-value of 0.7883 is a test statistic under the stated superiority framework. It should not be read as the probability that the treatment has no effect, nor as evidence that the two treatments are exactly equivalent.
Myocardial infarction
Hazard ratio for MI
95% CI: 0.71–0.98 · P = 0.0294
An HR of 0.84 corresponds to an estimated 16% lower instantaneous hazard of MI under the Cox model. The 95% CI of 0.71–0.98 gives the uncertainty range reported by the registry and remains below 1.00.
The p-value of 0.0294 describes statistical evidence under the reported superiority analysis. It does not quantify the size of the treatment effect and does not establish an absolute reduction in MI events.
Ischaemic stroke
Hazard ratio for ischaemic stroke
95% CI: 0.64–0.99 · P = 0.0375
An HR of 0.80 corresponds to an estimated 20% lower instantaneous hazard of ischaemic stroke under the fitted model. The 95% CI of 0.64–0.99 describes uncertainty around that estimate and lies just below the null value at its upper boundary.
The p-value of 0.0375 should be interpreted as evidence against the null hypothesis within the reported superiority framework, not as a measure of effect magnitude.
All-cause death
Hazard ratio for all-cause death
95% CI: 0.87–1.10 · P = 0.6846
The estimated HR of 0.98 is close to 1.00. The 95% CI of 0.87–1.10 includes the null value and represents uncertainty compatible with both a modestly lower and a modestly higher hazard.
The p-value of 0.6846 does not measure the size of the observed effect. Nor does a nonsignificant p-value establish equivalence between the treatment groups.
8. Safety Results
The registry reports four prespecified safety-related time-to-event analyses: TIMI major bleeding, TIMI major or minor bleeding, PLATO major bleeding, and permanent discontinuation of study medication due to any bleeding event.
Serious adverse events
| Study arm | Affected / at risk | Percentage of at-risk participants |
|---|---|---|
| Ticagrelor 60 mg | 3049 / 9562 | 31.89%* |
| Ticagrelor Placebo | 3210 / 9531 | 33.68%* |
*Percentages are simple descriptive calculations from the affected and at-risk counts reported in the ClinicalTrials.gov record and are not reported registry estimates.
| Safety endpoint | HR | 95% CI | P-value |
|---|---|---|---|
| TIMI Major Bleeding Event (Primary Safety Objective) | 2.32 | 1.82–2.94 | <0.0001 |
| TIMI Major or Minor Bleeding Event | 2.49 | 2.02–3.07 | <0.0001 |
| PLATO Major Bleeding Event | 2.41 | 1.98–2.93 | <0.0001 |
| Permanent Discontinuation Due to Any Bleeding Event | 4.04 | 3.32–4.92 | <0.0001 |
TIMI major bleeding
Hazard ratio for TIMI major bleeding
95% CI: 1.82–2.94 · P < 0.0001
An HR of 2.32 indicates an estimated instantaneous hazard more than twice that of the comparator under the Cox model. The 95% CI of 1.82–2.94 represents the reported uncertainty around this relative estimate.
This endpoint was analyzed in the safety analysis set, defined as patients who received at least 1 dose of randomized ticagrelor or placebo. The time frame was from randomisation to 7 days following the last dose of study medication, with a maximum exposure duration of 59 mo.
TIMI major or minor bleeding
Hazard ratio for TIMI major or minor bleeding
95% CI: 2.02–3.07 · P < 0.0001
The HR of 2.49 represents a substantially higher estimated instantaneous hazard of the combined major-or-minor bleeding endpoint in the ticagrelor group relative to placebo under the fitted model. The 95% CI of 2.02–3.07 indicates the precision reported for that estimate.
As with the efficacy endpoints, this is a relative time-to-event measure. It is not an absolute percentage increase in patients experiencing bleeding.
PLATO major bleeding
Hazard ratio for PLATO major bleeding
95% CI: 1.98–2.93 · P < 0.0001
The HR of 2.41 indicates a higher estimated instantaneous hazard of PLATO major bleeding under the Cox model. The confidence interval of 1.98–2.93 provides the reported uncertainty around the estimate and remains above 1.00.
The p-value of <0.0001 describes strong statistical evidence against the superiority null hypothesis as formulated for this analysis. It does not indicate that the effect is 0.0001 in magnitude or that the probability of a particular clinical conclusion is less than 0.01%.
Permanent discontinuation because of bleeding
Hazard ratio for permanent discontinuation due to any bleeding event
95% CI: 3.32–4.92 · P < 0.0001
An HR of 4.04 means that the estimated instantaneous hazard of permanent discontinuation due to any bleeding event was approximately four times that of the placebo group under the fitted Cox model. The 95% CI of 3.32–4.92 quantifies the reported uncertainty.
This endpoint is different from a bleeding-event endpoint itself: it measures discontinuation attributed to bleeding. Consequently, it should not be interpreted as though the HR were the hazard ratio for all bleeding events.
9. Statistical Methodology
Cox proportional-hazards model
The registry identifies Regression, Cox as the reported statistical method for the primary and secondary time-to-event analyses. The statistical method is a Cox proportional-hazards model, a survival-analysis model that estimates a relative hazard between treatment groups while accounting for differing follow-up times and censoring.
For a two-group treatment comparison, the hazard ratio is related to the exponentiated treatment coefficient: HR = exp(β). The model compares instantaneous event rates rather than simply comparing cumulative event proportions.
Hazard ratio
The hazard ratio is the central effect measure reported for every formal statistical analysis reported in the ClinicalTrials.gov record. An HR below 1 indicates a lower estimated instantaneous event rate in the ticagrelor group relative to placebo; an HR above 1 indicates a higher estimated instantaneous event rate.
| HR value | Model-based interpretation |
|---|---|
| HR = 1.00 | No estimated difference in instantaneous hazard between groups. |
| HR < 1.00 | Lower estimated instantaneous hazard in the ticagrelor group. |
| HR > 1.00 | Higher estimated instantaneous hazard in the ticagrelor group. |
Kaplan-Meier estimation
A time-to-event endpoint such as the THEMIS primary composite can be described using Kaplan-Meier estimation. The method estimates the probability of remaining event-free over time while allowing participants to contribute information until their event or censoring time.
Here, di is the number of events at time ti, and ni is the number at risk immediately before that time.
The ClinicalTrials.gov record specifically identifies the formal statistical method as Cox regression and does not provide a separate Kaplan-Meier estimate in the posted analysis fields. Kaplan-Meier estimation is therefore best understood here as the complementary descriptive survival-analysis framework rather than as an additional reported numerical result.
Time-to-event censoring
The primary endpoint definition specifies that participants without an event are censored at the earliest of the primary analysis censoring date, the last endpoint assessment date, and the non-CV death date. This is a critical part of the analysis because participants can contribute partial follow-up rather than requiring every participant to be observed for the same duration.
Intention-to-treat principle
The registry's analysis text identifies intention-to-treat analysis as an analysis concept for the efficacy endpoints. The reported efficacy population is the full analysis set described above, which excludes the 51 randomized patients from the prematurely closed site.
This distinction is important when reading the result: the reported HR of 0.90 is not based on a simple per-protocol comparison. It is based on the registry-defined efficacy analysis population and the randomized treatment groups.
10. Confidence Intervals and Statistical Precision
A confidence interval should be read together with its point estimate. For example, the primary HR is 0.90, while its 95% CI is 0.81–0.99. The point estimate is the single best estimate produced by the reported model, while the interval conveys sampling uncertainty around it.
| Endpoint | Point estimate | 95% CI width | Null value 1.00 inside CI? |
|---|---|---|---|
| Primary composite | 0.90 | 0.81–0.99 | No |
| CV Death | 1.02 | 0.88–1.18 | Yes |
| MI | 0.84 | 0.71–0.98 | No |
| Ischaemic Stroke | 0.80 | 0.64–0.99 | No |
| All-cause Death | 0.98 | 0.87–1.10 | Yes |
| TIMI Major Bleeding | 2.32 | 1.82–2.94 | No |
| TIMI Major or Minor Bleeding | 2.49 | 2.02–3.07 | No |
| PLATO Major Bleeding | 2.41 | 1.98–2.93 | No |
| Permanent Discontinuation Due to Bleeding | 4.04 | 3.32–4.92 | No |
The confidence intervals for CV death and all-cause death include 1.00, whereas those for the primary composite, MI, ischaemic stroke, and the reported bleeding endpoints do not. This describes the statistical intervals reported in the registry; it should not be converted into a claim that an endpoint with a confidence interval crossing 1.00 has "no effect."
11. Statistical Methods Explained
Why was a Cox proportional-hazards model used?
The endpoints in the posted analyses are time-to-event outcomes. Participants can experience the event at different times and can be censored before an event is observed. Cox regression is designed for this setting and produces a hazard ratio that compares the modeled instantaneous event rates between treatment groups.
What does an HR of 0.90 mean?
An HR of 0.90 means that the fitted model estimates the instantaneous event rate in the ticagrelor group to be 0.90 times that in the placebo group over the analyzed follow-up. It can be expressed as an estimated 10% lower hazard, but it is not a 10-percentage-point reduction in event probability.
Why is the confidence interval important?
The point estimate alone gives only one value. The 95% CI of 0.81–0.99 for the primary endpoint shows how much statistical uncertainty surrounds the HR estimate. A reader should therefore avoid treating 0.90 as though it were an exact treatment effect.
Why does the p-value not measure effect size?
A p-value summarizes evidence against a null hypothesis under a specified statistical model. It depends on the estimated effect, its uncertainty, and the amount of information in the analysis. It therefore cannot be used as a substitute for the HR or its confidence interval.
What is the difference between an efficacy population and a safety population?
The efficacy analysis uses the registry-defined full analysis set, while the safety analyses use patients who received at least 1 dose of randomized ticagrelor or placebo. This means the populations answer somewhat different statistical questions: efficacy is tied to the randomized comparison as defined by the registry, while safety is tied to exposure.
Why is censoring important?
Not every participant necessarily has an observed endpoint event during the analysis period. The primary endpoint explicitly defines censoring rules. Cox regression incorporates those different follow-up times rather than requiring every participant to have the same observation period.
12. Interpreting the Primary Result in Context
The primary HR of 0.90 is a relative measure. Its most direct interpretation is that the model estimated a lower instantaneous hazard for the composite endpoint in the ticagrelor group than in the placebo group. The 95% CI, 0.81–0.99, communicates the uncertainty around that estimate.
Several details are important when translating this result into a statistical conclusion:
- The endpoint is a composite of cardiovascular death, MI, and stroke.
- The analysis is time-to-event, not simply a comparison of event percentages.
- The effect measure is a hazard ratio, not a risk ratio or absolute risk difference.
- The efficacy population is the registry-defined full analysis set.
- The primary hypothesis type is superiority.
- The reported confidence interval is two-sided at 95%.
- The median time in study until the primary analysis censoring date was 40 months.
The statistical story is therefore more specific than "ticagrelor reduced cardiovascular events by 10%." The directly supported statement is that the Cox model estimated an HR of 0.90 for the composite endpoint, with a two-sided 95% CI of 0.81–0.99 and a p-value of 0.0378, using the registry-defined full analysis set.
The distinction matters because a hazard ratio is not an absolute effect. To state an absolute number needed to treat, absolute risk reduction, or fixed-time event-rate difference would require corresponding absolute event estimates that are not contained in the ClinicalTrials.gov record.
13. Understanding the Secondary Results
The individual endpoint estimates illustrate why a composite outcome should not be interpreted as though it were identical to every component.
MI
The reported HR was 0.84 with a 95% CI of 0.71–0.98 and P = 0.0294.
Ischaemic stroke
The reported HR was 0.80 with a 95% CI of 0.64–0.99 and P = 0.0375.
CV death
The reported HR was 1.02 with a 95% CI of 0.88–1.18 and P = 0.7883.
All-cause death
The reported HR was 0.98 with a 95% CI of 0.87–1.10 and P = 0.6846.
The component results are not interchangeable with the primary composite. A composite can show an estimated effect even when individual components have different estimates because the composite counts the first occurrence of any qualifying event.
14. Safety Versus Efficacy: A Statistical Contrast
THEMIS provides a useful statistical example of why efficacy and safety should be examined as separate dimensions. The primary efficacy HR was 0.90, while the reported bleeding HRs were all above 2.00.
| Analysis domain | Endpoint | HR | 95% CI |
|---|---|---|---|
| Efficacy | CV death, MI or stroke | 0.90 | 0.81–0.99 |
| Efficacy | MI | 0.84 | 0.71–0.98 |
| Efficacy | Ischaemic stroke | 0.80 | 0.64–0.99 |
| Safety | TIMI major bleeding | 2.32 | 1.82–2.94 |
| Safety | TIMI major or minor bleeding | 2.49 | 2.02–3.07 |
| Safety | PLATO major bleeding | 2.41 | 1.98–2.93 |
| Safety | Permanent discontinuation due to bleeding | 4.04 | 3.32–4.92 |
This is statistically important because a treatment effect cannot be summarized adequately by a single efficacy endpoint. The trial data contain both benefit-oriented and harm-oriented time-to-event outcomes, and their effect measures point in different directions relative to the null value of 1.00.
The appropriate statistical reading is not to select one HR and ignore the others. Instead, each endpoint should be interpreted according to its definition, analysis population, follow-up, effect measure, and uncertainty. The ClinicalTrials.gov record shows a primary efficacy HR below 1 and substantially higher bleeding hazards in the prespecified safety analyses.
15. The Role of the Analysis Population
The registry states that the full analysis set included all randomized patients except 51 patients randomized to study drug at a site prematurely closed by the sponsor. This is a concrete deviation from simply saying "all randomized patients."
That distinction matters because the numerical HR of 0.90 is tied to the actual analysis population used in the registry. A reanalysis that included or excluded different participants could, in principle, produce a different estimate even if the treatment assignments and endpoint definitions remained unchanged.
The registry identifies intention-to-treat analysis as an analysis concept, while specifically defining the population used for the posted primary efficacy analysis.
16. What the Hazard Ratio Does and Does Not Tell You
What it tells you
The relative difference in modeled instantaneous event rates between the treatment groups under the Cox proportional-hazards framework.
What it does not tell you
The absolute probability that an individual patient will experience the event, the absolute risk difference, or the number needed to treat.
What the CI adds
The uncertainty around the estimated HR under the statistical framework used for the analysis.
What the p-value adds
Evidence against the specified null hypothesis; it is not a measure of treatment-effect magnitude.
This distinction is especially important when comparing endpoints with different HRs. An HR of 0.80 and an HR of 2.40 are both relative measures, but they describe different endpoints and different directions of association relative to the comparator.
17. Primary Analysis Censoring
The primary efficacy time frame extends from randomisation to the primary analysis censoring date (PACD). The registry reports a median time in study until PACD of 40 months.
The endpoint definition also specifies that participants without an event are censored at the earliest of:
- the primary analysis censoring date;
- the last endpoint assessment date; and
- the non-CV death date.
18. Safety Follow-Up and Exposure Window
The four reported bleeding analyses use a different time frame from the primary efficacy endpoint: from randomisation to 7 days following the date of last dose of study medication. The registry reports a maximum duration of exposure of 59 mo.
| Feature | Primary efficacy | Safety bleeding analyses |
|---|---|---|
| Starting point | Randomisation | Randomisation |
| End of time frame | Primary analysis censoring date | 7 days following last dose of study medication |
| Reported median / maximum duration | Median time in study until PACD: 40 months | Maximum duration of exposure: 59 mo |
| Analysis population | Full analysis set | Safety analysis set |
| Statistical method | Cox proportional-hazards model | Cox proportional-hazards model |
Because the time frames and populations differ, the efficacy and safety HRs should not be interpreted as though they arose from an identical estimand. The endpoint definitions determine what event process each HR represents.
19. Multiple Endpoints and Interpretation
The trial data contain one registered primary endpoint and multiple secondary and other prespecified analyses. The primary endpoint is the composite of cardiovascular death, MI or stroke. The other reported analyses include individual efficacy components and several bleeding outcomes.
This creates an important statistical distinction between estimating effects and making confirmatory claims. Every additional hypothesis test introduces another opportunity to observe a small p-value by chance. The ClinicalTrials.gov record does not specify an alpha-spending, multiplicity-adjustment, hierarchical-testing, or gatekeeping procedure for these nine posted analyses.
20. Proportional-Hazards Assumption
The Cox model is described as a proportional-hazards model. Its standard interpretation assumes that the treatment hazard ratio is meaningfully represented by a common relative hazard over the analyzed time scale.
In practical statistical analysis, this assumption can be evaluated using graphical or formal diagnostic methods, such as examination of log-minus-log survival plots or tests based on time-varying effects. Those diagnostics are not contained in the registry analysis data, so no conclusion about the adequacy of the proportional-hazards assumption is made here.
When proportional hazards is a reasonable approximation, a single HR provides a compact summary of the relative event rate. When hazards vary materially over time, the single HR may obscure important temporal features of the treatment effect.
21. Limitations
- Hazard-ratio interpretation: The reported treatment effects are model-based hazard ratios, not absolute risk differences or fixed-time risk ratios.
- Proportional-hazards assumption: The ClinicalTrials.gov record identifies a Cox proportional-hazards model but do not provide diagnostic results establishing that the proportional-hazards assumption holds throughout follow-up.
- Composite endpoint: The primary endpoint combines cardiovascular death, MI, and stroke. The composite result should not be interpreted as though it were a result for each individual component.
- Component interpretation: The registry reports separate secondary analyses, but the ClinicalTrials.gov record does not include an interaction analysis testing whether treatment effects differ statistically among the components.
- Analysis population: The full analysis set excludes 51 randomized patients from a prematurely closed site, so the posted HR is tied to that registry-defined population.
- Safety population: Safety analyses use participants who received at least 1 dose, which differs from the efficacy population definition.
- Multiplicity: The trial data contain multiple formal endpoint analyses but do not provide enough information to reconstruct a complete multiplicity-adjustment strategy.
- Absolute effects: The ClinicalTrials.gov record does not provide absolute event probabilities or absolute risk differences for the endpoints analyzed here.
- Safety interpretation: Bleeding HRs describe specific prespecified bleeding outcomes and should not be generalized to every possible adverse event.
22. Why This Trial Matters Statistically
THEMIS is a useful teaching case because the registry data bring together the major building blocks of a modern randomized time-to-event analysis: randomization, masking, a composite endpoint, censoring rules, a defined efficacy population, Cox regression, hazard ratios, confidence intervals, superiority testing, individual component analyses, and safety time-to-event outcomes.
| Concept | How it appears in THEMIS |
|---|---|
| Randomization | The study uses randomized allocation in a parallel-group phase 3 design. |
| Blinding | The registry identifies the study as quadruple masked. |
| Time-to-event analysis | The primary and posted secondary analyses are analyzed from randomisation through defined censoring or exposure windows. |
| Cox regression | The registry reports Regression, Cox for the formal analyses. |
| Hazard ratio | Every posted formal analysis uses HR as the reported effect measure. |
| Confidence interval | All reported formal analyses provide two-sided 95% confidence intervals. |
| Superiority testing | The primary and posted secondary/safety analyses are identified as superiority hypotheses. |
| Composite endpoint | The primary endpoint combines CV death, MI and stroke. |
| Censoring | The primary endpoint has explicit censoring rules tied to PACD, endpoint assessment, and non-CV death. |
| Analysis populations | Efficacy uses the full analysis set; safety uses the safety analysis set. |
| Safety analysis | Multiple bleeding endpoints are evaluated using the same Cox modeling framework. |
| Component outcomes | CV death, MI, and ischaemic stroke are also analyzed separately. |
23. A Statistical Reading of the Entire Result Set
Viewed as a collection of estimates rather than as isolated p-values, the results show a coherent contrast between the efficacy and bleeding endpoints. The primary composite has an HR of 0.90; MI has an HR of 0.84; ischaemic stroke has an HR of 0.80; CV death has an HR of 1.02; and all-cause death has an HR of 0.98.
The bleeding analyses point in the opposite direction relative to the null: TIMI major bleeding has an HR of 2.32, TIMI major or minor bleeding an HR of 2.49, PLATO major bleeding an HR of 2.41, and permanent discontinuation due to any bleeding an HR of 4.04.
The important statistical lesson is that a randomized trial can generate several distinct treatment-effect estimates, each corresponding to a different estimand. There is no single HR that represents every outcome. The correct interpretation keeps the endpoint definition, direction of effect, confidence interval, p-value, population, and follow-up window attached to each estimate.
For the primary endpoint specifically, the registry provides enough information to state the model-based relative result precisely: HR 0.90, 95% CI 0.81–0.99, P = 0.0378. For the safety endpoints, the corresponding estimates are materially above 1.00. These estimates should be considered jointly when learning how randomized clinical-trial evidence is represented statistically, without collapsing distinct endpoints into a single number.
24. Statistical Concepts in This Trial
Learn more about the methods used in this trial:
25. Related Statistical Calculators
Explore calculators that reinforce the statistical methods illustrated by THEMIS:
26. Sources
- ClinicalTrials.gov: THEMIS (NCT01991795).
- PubMed: PMID 38658101.
- PubMed: PMID 36270493.
- PubMed: PMID 33558152.
- PubMed: PMID 27160944.
Continue through the Clinical Biostats statistical library
Use the related tutorials and calculators to explore the survival-analysis methods, effect measures, and clinical-trial concepts illustrated by THEMIS.
27. Record Summary
THEMIS provides a compact example of how a large randomized phase 3 trial can be represented through time-to-event statistical methods. The primary endpoint was a composite of cardiovascular death, MI or stroke, analyzed from randomisation to the primary analysis censoring date with a median time in study until PACD of 40 months. The registry-defined efficacy analysis used a full analysis set and Cox proportional-hazards regression, producing a primary HR of 0.90 with a two-sided 95% CI of 0.81–0.99 and P = 0.0378.
The secondary efficacy estimates were HR 1.02 for CV death, 0.84 for MI, 0.80 for ischaemic stroke, and 0.98 for all-cause death. The safety analyses used a safety analysis set and reported higher hazards for TIMI major bleeding, TIMI major or minor bleeding, PLATO major bleeding, and permanent discontinuation due to bleeding, with HRs of 2.32, 2.49, 2.41, and 4.04, respectively.
The most useful statistical lesson is that these estimates cannot be reduced to a single "trial effect." Each hazard ratio describes a specific endpoint under a specific population and follow-up definition. Correct interpretation therefore requires reading the endpoint definition, analysis population, censoring rules, hazard ratio, confidence interval, p-value, and hypothesis type together.