This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
ESPRIT was a randomized, parallel-group phase 3 trial evaluating whether adding recombinant interleukin-2 to antiretroviral therapy affected HIV disease progression and death in HIV-positive patients. The registry reports 4,150 enrolled participants, two treatment groups, and a median follow-up of 7.6 years.
| Feature | ESPRIT |
|---|---|
| Trial name | ESPRIT |
| ClinicalTrials.gov identifier | NCT00004978 |
| Phase | Phase 3 |
| Condition | HIV Infections |
| Design | Randomized, parallel-group |
| Masking | None |
| Primary purpose | Treatment |
| Enrollment | 4,150 |
| Intervention | Recombinant interleukin-2 (rIL-2) |
| Comparator | No rIL-2 |
| Primary endpoint type | Binary registered endpoint analyzed as a time-to-event outcome |
| Results posted | Yes |
| Statistical analyses posted | 9 |
2. Clinical Question
The central question was whether treatment with recombinant interleukin-2, in patients taking antiretroviral therapy, changed the time to a new or recurrent HIV disease progression event or death compared with no rIL-2.
Population
HIV-positive patients taking antiretroviral therapy and enrolled in the ESPRIT randomized trial.
Intervention
Recombinant interleukin-2 (rIL-2).
Comparator
No rIL-2.
Primary question
Does rIL-2 change the hazard of a new or recurrent HIV disease progression event, including death?
3. Trial Design
rIL-2
- Recombinant interleukin-2
- Background antiretroviral therapy
- Follow-up through study end
No rIL-2
- No recombinant interleukin-2
- Background antiretroviral therapy
- Follow-up through study end
The trial was randomized and parallel-group, with no masking. The registry reports a start date of March 2000 and primary completion in November 2008. The primary endpoint was followed from randomization through study end, with a median follow-up of 7.6 years.
4. Endpoints
The registry contains one registered primary endpoint and nine posted statistical analyses covering the primary and secondary outcomes.
| Endpoint | Registry definition / measure | Time frame | Analysis |
|---|---|---|---|
| New or Recurrent HIV Disease Progression Event Including Death | Participants who die or experience at least one: any CDC Category C 1993 AIDS-defining events or invasive aspergillosis, bartonellosis, Chagas disease, Herpes zoster, visceral Leishmaniasis, Hodgkin's lymphoma, non-Hodgkin's lymphoma (all cell types), microsporidiosis, nocardiosis, disseminated Penicillium marneffii, extrapulmonary Pneumocystis carinii, or Rhodococcus equi disease. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| New or Recurrent Serious HIV Disease Progression Event Including Death | Serious HIV disease progression event or death. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| Number of Participants Who Died From Any Cause | Death from any cause. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| Participants With a New Disease Progression Event or Death | New disease progression event or death. | From randomization through 15 November 2008; median of 7.6 years follow-up | Cox proportional-hazards model |
| Absolute CD4 Cell Counts Averaged Throughout Followup | CD4+ cell counts averaged over all participants with at least one CD4+ measurement recorded during follow-up. | From randomization through study end; median of 7.6 years follow-up | Constrained longitudinal data analysis |
| Number of Participants With Changes in Anti-retroviral Treatment (ART) | First change in antiretroviral treatment. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| Grade 4 Signs and Symptoms | First grade 4 event. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| Hepatic, Metabolic, and Cardiac Conditions | Time to the relevant hepatic, metabolic, or cardiac condition. | From randomization through study end; median of 7.6 years follow-up | Cox proportional-hazards model |
| Pattern of Use of Prophylaxis for Opportunistic Infections | Medication use recorded at the last follow-up visit attended among participants attending at least one follow-up visit. | Last follow-up visit; median of 7.6 years follow-up | Chi-squared test |
5. Statistical Methodology
Primary time-to-event analysis
The primary endpoint was analyzed using an unadjusted Cox proportional-hazards regression model in the intention-to-treat population. The reported effect measure was a hazard ratio comparing rIL-2 with no rIL-2.
The hazard ratio is exp(β). For ESPRIT, the registry reports an HR of 0.94 for rIL-2 versus no rIL-2.
Longitudinal CD4 analysis
Absolute CD4 cell counts averaged throughout follow-up were analyzed using constrained longitudinal data analysis (cLDA). The treatment difference was reported as a mean difference in final values, with rIL-2 minus no rIL-2 defining the direction of the estimate.
Categorical analysis
The pattern of prophylaxis use for opportunistic infections was analyzed using a chi-squared test. This is appropriate for comparing categorical distributions or proportions between randomized treatment groups when the outcome is defined categorically.
Intention-to-treat analysis
The primary disease-progression analysis was conducted using the intention-to-treat population. This means the treatment comparison is anchored to randomized assignment rather than subsequent treatment exposure.
Effect measures
| Endpoint family | Statistical method | Effect measure |
|---|---|---|
| Time-to-event outcomes | Cox proportional-hazards model | Hazard ratio |
| Longitudinal CD4 outcome | Constrained longitudinal data analysis | Mean difference |
| Binary prophylaxis-use outcome | Chi-squared test | P-value |
6. Primary Result: HIV Disease Progression or Death
The primary endpoint was analyzed from randomization through study end, with a median of 7.6 years of follow-up. The analysis used an intention-to-treat population and an unadjusted Cox proportional-hazards model.
Hazard ratio for new or recurrent HIV disease progression event including death
95% CI: 0.75–1.16 · P = .55
rIL-2 versus no rIL-2; two-sided 95% confidence interval.
The estimated hazard ratio of 0.94 means that the fitted Cox model estimated an instantaneous event rate approximately 6% lower with rIL-2 than with no rIL-2 over the analyzed follow-up. This is a relative hazard measure, not a 6% absolute reduction in the probability of disease progression or death.
The 95% confidence interval of 0.75–1.16 expresses uncertainty around the estimated hazard ratio. Because the interval includes 1, the reported estimate is compatible with both a lower and a higher hazard under the treatment comparison. The interval should not be interpreted as a range containing the effect for individual patients.
The P = .55 value addresses the statistical evidence against the model-based null hypothesis of no hazard difference. It does not measure the size, clinical importance, or probability of the treatment effect. The estimate, confidence interval, and study design should be considered together.
Because the analysis used a Cox proportional-hazards model, interpretation of a single hazard ratio also depends on the proportional-hazards framework. The registry identifies the model but does not provide a separate assessment of the proportional-hazards assumption in the information reported here.
7. Secondary Time-to-Event Results
Serious HIV disease progression or death
Hazard ratio
95% CI: 0.74–1.20 · P = .62
rIL-2 versus no rIL-2; ITT population.
The estimate is close to the primary endpoint estimate, while the confidence interval extends from below to above 1. This outcome therefore provides a similar statistical pattern but is a distinct endpoint and should not be treated as interchangeable with the primary composite.
Death from any cause
Hazard ratio
95% CI: 0.69–1.17 · P = .42
rIL-2 versus no rIL-2.
The point estimate corresponds to an approximately 10% lower estimated hazard of death under the fitted Cox model. The confidence interval includes 1 and spans both lower and higher hazards, illustrating why the point estimate should not be interpreted without its uncertainty interval.
New disease progression event or death through 15 November 2008
Hazard ratio
95% CI: 0.73–1.14 · P = .41
rIL-2 versus no rIL-2.
This endpoint uses the specified cutoff of 15 November 2008 rather than the generic study-end wording used for several other registry outcomes. The reported hazard ratio is again below 1, but the confidence interval crosses 1.
Change in antiretroviral treatment
Hazard ratio for first ART change
95% CI: 0.88–1.00 · P = .07
rIL-2 versus no rIL-2.
This analysis concerns time to the first change in antiretroviral treatment rather than time to HIV disease progression. Its hazard ratio should therefore be interpreted as a treatment-group comparison for a different clinical process.
Grade 4 signs and symptoms
Hazard ratio for first grade 4 event
95% CI: 1.07–1.41 · P = .003
rIL-2 versus no rIL-2; ITT analysis.
The hazard ratio of 1.23 indicates an estimated 23% higher instantaneous rate of a first grade 4 event in the rIL-2 group relative to the no-rIL-2 group under the Cox model.
The 95% confidence interval of 1.07–1.41 lies above 1, indicating that the reported estimate is separated from the null value under this confidence-interval framework. The P = .003 value provides evidence against the null hypothesis of no hazard difference, but it is not itself a measure of the magnitude or clinical importance of the difference.
This is a secondary safety-related endpoint, so its interpretation should remain distinct from the primary disease-progression endpoint. A statistically different safety outcome does not by itself determine the overall benefit-risk profile of the intervention.
Hepatic, metabolic, and cardiac conditions
Hazard ratio
95% CI: 0.76–1.22 · P = .74
rIL-2 versus no rIL-2.
The estimate is close to 1, while the confidence interval includes both values below and above 1. The P-value therefore should be read as evidence about the statistical null hypothesis rather than as evidence that the treatment has exactly no effect.
8. Secondary CD4 Cell Count Result
Absolute CD4 cell counts averaged throughout follow-up were analyzed among participants with at least one CD4+ measurement recorded during follow-up. The registry identifies constrained longitudinal data analysis as the statistical method.
Mean difference in final CD4 values
95% CI: 145–174 cells/mm3
rIL-2 minus no rIL-2; two-sided 95% confidence interval.
The reported mean difference of 159 cells/mm3 means that the longitudinal model estimated higher final CD4 values in the rIL-2 group by 159 cells/mm3 relative to the no-rIL-2 group, according to the treatment-difference definition reported in the registry.
The 95% confidence interval of 145–174 cells/mm3 describes uncertainty around the estimated mean difference. It does not describe the distribution of individual patient responses and should not be interpreted as saying that every patient experienced an increase within that range.
This endpoint illustrates an important distinction in clinical-trial statistics: a treatment can produce a measurable difference in a biological marker without that marker difference necessarily translating into the same magnitude or direction of effect on a clinical time-to-event endpoint.
9. Pattern of Prophylaxis for Opportunistic Infections
The registry reports a binary endpoint assessing medication use at the last follow-up visit attended among all participants attending at least one follow-up visit. The analysis used an intention-to-treat framework and a chi-squared test.
Chi-squared analysis
rIL-2 versus no rIL-2; last follow-up visit.
The reported P = .97 is a test result for the categorical comparison specified by the registry. A chi-squared P-value does not quantify the magnitude of the difference between treatment groups and should not be interpreted as a probability that the null hypothesis is true.
The registry does not provide a corresponding effect estimate or confidence interval for this endpoint in the posted analysis. Consequently, the P-value is the appropriate reported statistical result to summarize without converting it into an unreported effect size.
10. Serious Adverse Events by Treatment Group
The registry reports serious adverse events by treatment arm as the number affected among participants at risk.
| Treatment group | Affected | At risk |
|---|---|---|
| rIL-2 | 466 | 2,071 |
| No rIL-2 | 383 | 2,040 |
These counts describe the number affected and the number at risk reported by treatment group. They are not, by themselves, a time-to-event hazard ratio and should not be interpreted using the same framework as the Cox analyses.
11. Statistical Methods Explained
Why was a Cox proportional-hazards model used?
The primary endpoint and several secondary endpoints are time-to-event outcomes. A Cox model allows investigators to compare the instantaneous event rate between treatment groups while accounting for differing follow-up times and censoring. The ESPRIT registry analysis reports an unadjusted Cox proportional-hazards model for the primary endpoint.
What does an HR of 0.94 mean?
An HR of 0.94 means that the estimated instantaneous event rate in the rIL-2 group was 0.94 times that in the no-rIL-2 group under the fitted model. Equivalently, the point estimate corresponds to approximately a 6% lower estimated hazard. It does not mean that the probability of the event was exactly 6% lower.
Why is the confidence interval important?
The primary 95% confidence interval was 0.75–1.16. This interval shows the statistical uncertainty surrounding the point estimate. A point estimate alone can give an artificially precise impression; the interval shows that materially different hazard ratios remain compatible with the analysis.
Why is a longitudinal model appropriate for CD4 measurements?
CD4 cell counts are measured repeatedly over follow-up. A constrained longitudinal data analysis uses the longitudinal structure rather than treating repeated observations as independent single measurements. The registry reports a treatment difference based on a longitudinal model that considers CD4 measurements at follow-up visits.
What is the difference between ITT and an as-treated analysis?
ITT analysis preserves the randomized treatment assignment as the basis for comparison. This is especially important for the primary efficacy question because randomization establishes the treatment groups before later adherence, discontinuation, or other post-randomization events occur. The ESPRIT primary endpoint analysis is identified as ITT.
Why was a chi-squared test used for prophylaxis use?
The prophylaxis endpoint is binary and concerns categorical medication use at the last follow-up visit. A chi-squared test evaluates whether the observed categorical distribution differs between the treatment groups under the test's assumptions. Unlike the Cox analyses, it does not model time to an event.
12. Hazard Ratios Versus Mean Differences
ESPRIT illustrates two fundamentally different effect measures. The disease-progression and safety-related time-to-event analyses report hazard ratios, while the CD4 analysis reports a mean difference.
| Feature | Hazard ratio | Mean difference |
|---|---|---|
| Primary ESPRIT use | Time to disease progression, death, treatment change, or clinical event | Longitudinal CD4 cell count |
| Basic question | How do instantaneous event rates compare? | How do modeled mean values differ? |
| Null value | 1 | 0 |
| Interpretation | Relative event hazard | Absolute difference in the measured scale |
| ESPRIT example | Primary HR 0.94 | CD4 mean difference 159 cells/mm3 |
These effect measures cannot be directly compared numerically. A hazard ratio of 0.94 and a mean difference of 159 cells/mm3 answer different statistical questions.
13. Understanding the Primary Endpoint
The primary endpoint is a composite that includes death and a specified set of HIV disease progression events. Composite endpoints can increase the number of observed events and therefore contribute information to a time-to-event analysis, but interpretation depends on the individual components included in the composite.
Death component
Participants who die contribute to the composite endpoint regardless of whether another listed disease progression event occurred first.
Disease progression component
The registry specifies CDC Category C 1993 AIDS-defining events and a defined list of additional conditions.
Time origin
The primary endpoint is measured from randomization, establishing a common starting point for the randomized comparison.
Analysis population
The posted primary analysis uses the intention-to-treat population.
14. Censoring and Time-to-Event Interpretation
Time-to-event analyses are designed to incorporate participants whose event status is not observed through the entire study period. A participant who remains event-free until the end of available follow-up contributes information up to that point rather than being treated as though an event occurred.
For a time-to-event endpoint, the survival function represents the probability of remaining free of the event beyond time t. Cox regression instead focuses on the relative hazard between treatment groups.
The registry identifies the Cox method for the primary analysis but does not provide a detailed description of censoring rules in the posted statistical-analysis information. Interpretation should therefore remain at the level supported by the reported analysis.
15. Intention-to-Treat Analysis
The primary ESPRIT analysis is explicitly identified as an intention-to-treat analysis. The comparison is therefore based on randomized treatment assignment: rIL-2 versus no rIL-2.
ITT does not mean that every participant necessarily contributes the same amount of follow-up information. Time-to-event methods can incorporate varying observation times, with the statistical analysis accounting for the available follow-up through its event and censoring framework.
16. Multiple Endpoints and Statistical Interpretation
The registry reports one primary endpoint and multiple secondary outcomes. This creates an important distinction between the primary hypothesis and the collection of secondary analyses.
| Endpoint | Role | Reported result |
|---|---|---|
| New or recurrent HIV disease progression event including death | Primary | HR 0.94; 95% CI 0.75–1.16; P = .55 |
| New or recurrent serious HIV disease progression event including death | Secondary | HR 0.94; 95% CI 0.74–1.20; P = .62 |
| Death from any cause | Secondary | HR 0.90; 95% CI 0.69–1.17; P = .42 |
| New disease progression event or death | Secondary | HR 0.91; 95% CI 0.73–1.14; P = .41 |
| Absolute CD4 cell counts | Secondary | Mean difference 159; 95% CI 145–174 |
| Change in antiretroviral treatment | Secondary | HR 0.94; 95% CI 0.88–1.00; P = .07 |
| Grade 4 signs and symptoms | Secondary | HR 1.23; 95% CI 1.07–1.41; P = .003 |
| Hepatic, metabolic, and cardiac conditions | Secondary | HR 0.96; 95% CI 0.76–1.22; P = .74 |
| Pattern of prophylaxis use | Secondary | Chi-squared P = .97 |
The number of statistical tests means that individual secondary P-values should be interpreted in the context of the overall analysis plan. A nominally small P-value for one secondary endpoint does not automatically establish that the endpoint was a prespecified confirmatory claim unless the multiplicity framework supports that interpretation.
17. What the Primary Hazard Ratio Does — and Does Not — Mean
The primary HR of 0.94 means that the estimated instantaneous rate of the composite disease-progression event or death was 0.94 times the rate in the no-rIL-2 group under the fitted Cox model. Expressed as a relative comparison, the point estimate corresponds to approximately a 6% lower estimated hazard.
An HR of 0.94 does not mean that the absolute probability of disease progression or death was reduced by 6 percentage points. Absolute risk differences require treatment-group event probabilities at a specified time point or another absolute-risk measure.
The HR is not the probability that rIL-2 works, nor is the P-value the probability that the null hypothesis is true. Both are statistical quantities defined by the model and testing framework.
The 95% CI of 0.75–1.16 shows substantial uncertainty around the point estimate. The interval crosses 1, so the data are compatible with a lower hazard as well as a higher hazard under the modeled comparison.
18. Interpreting the CD4 Result Separately From Clinical Events
The CD4 result demonstrates why surrogate or biological measures and clinical outcomes should be considered separately. The registry reports a substantial modeled mean difference in absolute CD4 cell counts, while the primary clinical time-to-event analysis reports a hazard ratio of 0.94.
Biological measure
The CD4 endpoint describes a longitudinal laboratory measurement and is summarized as a mean difference.
Clinical event
The primary endpoint combines specified HIV disease progression events with death and is summarized using a hazard ratio.
Different estimands
The two estimates describe different treatment effects and cannot be substituted for one another.
Longitudinal structure
The CD4 analysis accounts for repeated follow-up measurements through constrained longitudinal data analysis.
19. Safety and Statistical Interpretation
The registry reports serious adverse events affecting 466 of 2,071 participants at risk in the rIL-2 group and 383 of 2,040 participants at risk in the no-rIL-2 group. It also reports a Cox analysis for grade 4 signs and symptoms.
| Safety-related measure | rIL-2 | No rIL-2 | Statistical result |
|---|---|---|---|
| Serious adverse events | 466/2,071 | 383/2,040 | Counts reported by arm |
| Grade 4 signs and symptoms | Time-to-first-event comparison | HR 1.23; 95% CI 1.07–1.41; P = .003 | |
| Hepatic, metabolic, and cardiac conditions | Time-to-event comparison | HR 0.96; 95% CI 0.76–1.22; P = .74 | |
The two safety summaries use different statistical structures. The serious-adverse-event figures are counts of affected participants among those at risk, whereas grade 4 signs and symptoms are analyzed as a time-to-first-event endpoint. Comparing the numbers directly as though they were the same estimand would be inappropriate.
20. Important Limitations and Interpretation Issues
- Composite primary endpoint: the primary outcome combines death with a specified set of disease progression events. The overall composite estimate does not identify whether every component behaved identically.
- Hazard-ratio interpretation: Cox regression produces a relative hazard estimate rather than an absolute risk difference. The registry does not report a separate assessment of the proportional-hazards assumption in the posted analysis information.
- Confidence-interval width: the primary 95% CI of 0.75–1.16 spans a meaningful range of possible relative effects, so the point estimate should not be interpreted in isolation.
- Multiple secondary endpoints: several secondary hypotheses were analyzed. Individual P-values should be interpreted according to the prespecified multiplicity framework rather than automatically treated as independent confirmatory tests.
- Different estimands: CD4 cell counts, time to clinical events, ART changes, grade 4 events, and prophylaxis use represent different clinical or biological processes and require different statistical interpretations.
- Longitudinal CD4 population: the CD4 analysis includes participants with at least one CD4+ measurement recorded during follow-up, which is a different analytic definition from the primary ITT time-to-event analysis.
- Follow-up definition: most endpoints use study end with a median of 7.6 years of follow-up, while the disease progression or death endpoint also has a specific cutoff of 15 November 2008.
- Safety denominators: serious adverse events are reported using affected and at-risk counts by arm; these figures should not be treated as equivalent to the Cox-model estimands.
21. Why This Trial Matters Statistically
ESPRIT is a useful teaching case because it combines randomized treatment comparison with several distinct statistical estimands. The trial illustrates how the same intervention can be evaluated through clinical time-to-event outcomes, repeated laboratory measurements, categorical outcomes, and safety events.
| Concept | How it appears in ESPRIT |
|---|---|
| Randomization | Participants were randomly allocated to rIL-2 or no rIL-2. |
| Intention-to-treat analysis | The primary Cox analysis was conducted in the ITT population. |
| Time-to-event endpoints | Disease progression, death, ART changes, grade 4 events, and other clinical conditions were analyzed over follow-up. |
| Cox proportional-hazards model | Used for the primary and several secondary time-to-event outcomes. |
| Hazard ratio | Primary HR 0.94; additional HR estimates ranged from 0.90 to 1.23 among the posted Cox analyses. |
| Confidence intervals | Reported for the primary and several secondary effect estimates. |
| Constrained longitudinal data analysis | Used for repeated CD4 cell-count measurements. |
| Mean difference | CD4 treatment difference was 159 cells/mm3, with a 95% CI of 145–174. |
| Chi-squared testing | Used for the categorical prophylaxis-use endpoint. |
| Safety analysis | Serious adverse events were reported as affected participants among those at risk by treatment group. |
22. Statistical Methods Explained in Context
The most important lesson from ESPRIT is that the statistical method follows the structure of the endpoint. Time-to-event outcomes require a framework that handles follow-up and censoring; repeated CD4 measurements require a longitudinal model; and categorical medication use can be compared with a categorical test.
This mapping is more informative than treating "statistical analysis" as a single method applied uniformly to every endpoint.
23. Overall Statistical Interpretation
The primary ESPRIT analysis estimated a hazard ratio of 0.94 for new or recurrent HIV disease progression including death, with a 95% confidence interval of 0.75–1.16 and P = .55. The point estimate was below 1, but the confidence interval crossed the null value of 1.
Several secondary time-to-event analyses showed similar point estimates below 1, including all-cause death with HR 0.90, disease progression or death with HR 0.91, and serious HIV disease progression or death with HR 0.94. The grade 4 signs and symptoms analysis differed in direction, with an HR of 1.23 and a 95% CI of 1.07–1.41.
The CD4 analysis provides a different statistical perspective: the reported longitudinal mean difference was 159 cells/mm3, with a 95% CI of 145–174. This demonstrates that a treatment-associated difference in a laboratory measure can coexist with a different statistical pattern for a clinical time-to-event endpoint.
Taken together, the ESPRIT results are best understood by keeping these estimands separate. The primary clinical endpoint, secondary clinical events, laboratory measurements, categorical medication-use outcome, and safety outcomes answer different questions and should not be collapsed into a single numerical interpretation.
24. Related Tutorials
Learn more about the methods used in this trial:
25. Related Calculators
26. Sources
- ClinicalTrials.gov: NCT00004978.
- Publication: PubMed record: PMID 15206062.
- Publication: PubMed record: PMID 19828532.
- Publication: PubMed record: PMID 26908541.
- Publication: PubMed record: PMID 24870935.
- Publication: PubMed record: PMID 22911697.
Continue through the Clinical Biostats statistical library
Connect the trial's endpoints and statistical methods to deeper tutorials and practical statistical calculators.
27. Record Summary
ESPRIT provides a useful statistical case study in which a randomized phase 3 comparison was evaluated across several different endpoint structures. The primary outcome was a time-to-event composite analyzed with an ITT Cox model, while CD4 measurements were analyzed longitudinally and prophylaxis use was analyzed categorically. The primary hazard ratio was 0.94 with a 95% CI of 0.75–1.16 and P = .55. The posted secondary analyses demonstrate why clinical-trial interpretation requires attention to the estimand, endpoint definition, analysis population, effect measure, confidence interval, and statistical model rather than relying on a single P-value.