PROMISE: Complete Statistical Analysis of Antiretroviral Strategies in HIV Prevention

1. Trial at a Glance

PROMISE was a randomized, double-blind, parallel phase 3 prevention trial evaluating antiretroviral strategies intended to reduce mother-to-child transmission of HIV and assess maternal and infant outcomes.

2. Clinical Question

The PROMISE trial addressed several related prevention and maternal-health questions across antepartum, postpartum, and maternal-health components. The statistical structure is therefore more complex than a single two-arm comparison.

3. Trial Design

4. Primary Endpoints

The registry identifies seven primary endpoints spanning the antepartum, postpartum, and maternal-health components. Their statistical structures differ substantially: some are binary outcomes measured at a defined visit or birth, while others are time-to-event outcomes requiring follow-up over time.

Endpoint definitions

Confirmed infant HIV infection in the antepartum component was defined as HIV nucleic acid test positivity from a specimen drawn at birth (Day 0–5) or Week 1 (Day 6–14), confirmed by HIV NAT positivity from a second specimen collected at a different time point.

Confirmed infant HIV infection in the postpartum component was defined as infant HIV NAT positivity from any post-randomization visit after the Week 1 visit, confirmed by HIV NAT positivity from a second specimen collected at a different time point. The analysis was conducted at the mother-infant pair level, with the worst outcome for multiple births counted as a single event.

Adverse pregnancy outcomes were defined as a composite including stillbirth, preterm delivery (<37 weeks), low birth weight (<2,500 grams), and congenital anomalies.

Maternal progression was defined as progression to an AIDS-defining illness or death. AIDS-defining illness referred to WHO Clinical Stage 4 illnesses identified in the protocol, with events reviewed and confirmed by an endpoint review group.

5. Analysis Populations and Unit of Analysis

A major statistical feature of PROMISE is that the unit of analysis differs by endpoint. The registry explicitly describes mother-level, mother-infant pair-level, and time-to-event analyses, while maintaining the intent-to-treat principle for the reported primary efficacy comparisons.

These distinctions matter because a clinical trial can preserve randomized treatment assignment while using different observation units for different outcomes. A mother-infant pair is not equivalent to an individual infant, and a binary outcome at birth is not equivalent to an event observed over variable follow-up time.

6. Statistical Methodology

Fisher exact test for binary endpoints

The registry reports Fisher exact testing for several pairwise comparisons of antepartum safety, obstetrical complications, and adverse pregnancy outcomes. Fisher's exact test is appropriate for a two-group comparison of a binary outcome when the exact distribution of the observed contingency table is used rather than relying on a large-sample approximation.

Risk difference for antepartum infant HIV infection

The principal posted estimate for the antepartum infant HIV endpoint was a risk difference comparing the combined Arm B and Arm C group with Arm A.

Hazard ratios for time-to-event endpoints

The postpartum infant HIV and maternal-health primary endpoints were reported using hazard ratios. A hazard ratio compares the instantaneous event rates between randomized groups over the observed follow-up under the time-to-event model.

Log-rank testing

The maternal-health progression endpoint used a log-rank test, and the postpartum maternal adverse-event endpoint also used a log-rank analysis. The log-rank framework compares event-time distributions while incorporating censoring during follow-up.

Two-sided Wald / Z tests

The registry reports two-sided Z tests for the secondary endpoint evaluating overall and HIV-free infant survival through 104 weeks of age. The analyses compare the reported proportional probability between specified antepartum treatment groups.

Repeated confidence intervals and interim monitoring

The antepartum infant HIV analysis used a repeated confidence interval based on a Lan-DeMets approach with an O'Brien-Fleming type I error spending function to preserve an experiment-wise type I error rate of 5%. The postpartum infant HIV confidence interval was also described as a repeated confidence interval.

7. Primary Results: Antepartum Infant HIV Infection

The registry reports a prespecified comparison in which Antepartum Arms B and C were combined and compared with Arm A. The analysis included Periods 1 and 2, used the intent-to-treat principle, and summarized the worst outcome for multiple births at the maternal-infant level.

Why the comparison combined Arms B and C

The registry identifies the combination of Arm B and Arm C minus Arm A as the a priori planned comparison. Statistically, combining two intervention groups creates a single contrast against Arm A rather than treating the three arms as three independent pairwise confirmatory tests.

8. Primary Results: Postpartum Infant HIV Infection

The postpartum infant HIV endpoint was analyzed as a time-to-event outcome through the site-recommended duration of breastfeeding, complete cessation of breastfeeding, or 18 months of age, whichever came first.

9. Primary Results: Maternal Progression to AIDS-Defining Illness or Death

The maternal-health component evaluated progression to an AIDS-defining illness or death from study entry through July 7, 2015, with an average of 94 weeks of follow-up. The registry reports a log-rank analysis and a hazard ratio comparing continuation with discontinuation of triple antiretroviral therapy.

10. Other Primary Safety and Pregnancy Results

The registry contains additional primary endpoint analyses for maternal toxicity, obstetrical complications, adverse pregnancy outcomes, and postpartum adverse events. These analyses use categorical or time-to-event methods rather than a single common effect measure.

Antepartum maternal toxicities

The endpoint included Grade 3 or higher toxicities and selected Grade 2 hematologic, renal, and hepatic adverse events through the Week 1 postpartum study visit. Events were graded using the DAIDS AE Grading Table identified in the registry.

Antepartum obstetrical complications

Obstetrical complications included deaths, diagnoses, signs or symptoms, chemistry laboratory tests, or hematological laboratory tests with grades of 3 or worse, with the specified MedDRA pregnancy, puerperium, and perinatal classifications.

Adverse pregnancy outcomes

The composite was assessed at birth and included stillbirth, preterm delivery (<37 weeks), low birth weight (<2,500 grams), and congenital anomalies. Multiple gestation pregnancies were summarized at the unique mother level.

Postpartum maternal adverse events

The endpoint was the incidence of Grade 3 or higher adverse events and selected Grade 2 hematologic, renal, and hepatic adverse events through the recommended breastfeeding duration, complete cessation of breastfeeding, or 18 months of age, whichever came first. All mothers were included, with 15 mothers without post-randomization data censored at randomization.

11. Serious Adverse Events by Arm

The registry reports serious adverse events as affected participants divided by the number at risk for each maternal and infant treatment grouping.

These counts describe affected participants relative to the reported number at risk. They are not themselves effect estimates, confidence intervals, or hypothesis tests. A safety comparison requires attention to the denominator, follow-up, endpoint definition, and whether participants can contribute more than one event.

12. Secondary Endpoint Results: Infant Survival Through 104 Weeks

The registry reports secondary analyses of the probability of overall and HIV-free infant survival through 104 weeks of age, measured from birth. The analyses used two-sided Z tests and distinguish between the Periods 1&2 group and the period 2 group.

Overall survival

HIV-free survival

The Periods 1&2 group consists of infants born to mothers randomized to Arms A and B throughout the PROMISE study under both protocol versions. The period 2 group contains infants born to mothers randomized during the last protocol version. Because Arm C was available to all participants only under the last protocol version, the period 2 analysis allows comparisons involving Arm C within that protocol version.

Only one infant was included per pregnancy, specifically the infant with the earliest failure time. This is an important unit-of-analysis rule because multiple births can otherwise create statistical dependence between observations from the same pregnancy.

13. Interim Monitoring and Repeated Confidence Intervals

The antepartum infant HIV analysis explicitly incorporates interim-analysis methodology. The registry describes a repeated confidence interval using the Lan-DeMets approach with an O'Brien-Fleming type I error spending function and an experiment-wise type I error rate of 5%.

This is an important distinction between the point estimate and its uncertainty interval. The interval level reflects the trial's sequential monitoring structure, whereas the estimate itself describes the observed treatment contrast.

14. Intention-to-Treat Analysis

The registry repeatedly identifies the intent-to-treat principle for the primary analyses. Under ITT, participants remain associated with the randomized treatment group for the primary comparison rather than being reassigned based on treatment received or subsequent adherence.

15. Statistical Methods Explained

Why was Fisher exact testing used for several primary endpoints?

Several primary endpoints were binary participant-level outcomes, including maternal toxicities, obstetrical complications, and adverse pregnancy outcomes. Fisher exact testing provides an exact test for the allocation of binary outcomes between two groups. The registry specifically identifies Fisher exact testing for the reported pairwise comparisons.

What does a risk difference of −1.3 mean?

A risk difference is calculated as the risk in one comparison group minus the risk in the reference group. Here, the reported −1.3 compares the combined Antepartum Arms B and C with Arm A. The negative sign indicates a lower estimated risk in the combined group on the reported scale.

Why is the confidence interval 96.5% rather than 95% for the antepartum HIV result?

The registry describes a repeated confidence interval based on the Lan-DeMets approach with an O'Brien-Fleming type I error spending function. Sequential monitoring requires adjustment of the inference so that the overall type I error remains controlled. The 96.5% interval reflects that repeated-analysis framework.

What does the postpartum HIV hazard ratio of 1.0 mean?

The point estimate of 1.0 indicates equal estimated instantaneous event rates between the two postpartum strategies under the reported hazard-ratio analysis. It does not mean that the underlying cumulative infection probabilities are necessarily identical at every time point, nor does it imply that uncertainty is absent. The corresponding 96% CI of 0.3–3.1 is wide.

Why is the maternal-health result reported with both a hazard ratio and a log-rank P-value?

The hazard ratio describes the relative event-rate contrast, while the log-rank test evaluates the evidence for a difference between the time-to-event distributions. They answer related but different statistical questions. A P-value is not a substitute for the effect estimate or its confidence interval.

Why does censoring matter for the postpartum analyses?

Time-to-event methods use each participant's available follow-up until an event or censoring. In the postpartum infant HIV analysis, 35 mother-infant pairs without post-randomization HIV test results were censored at randomization. The validity of time-to-event inference depends on appropriate handling and assumptions concerning censoring.

Why were only one infant per pregnancy used in the 104-week survival analyses?

The registry specifies that only one infant was included per pregnancy, selecting the infant with the earliest failure time. This avoids treating multiple infants from the same pregnancy as statistically independent observations and creates a defined pregnancy-level analysis rule.

16. Understanding the P-values

P-values appear throughout the posted primary and secondary analyses, but they should be interpreted together with the effect measure, confidence interval, endpoint definition, and comparison structure.

17. Secondary Analyses and Protocol Periods

The secondary infant-survival analyses are structured around protocol periods. This is not simply a matter of labeling: the available treatment arms changed between protocol versions, which affects which randomized comparisons can be made without mixing incompatible treatment availability.

This structure is a useful example of why protocol history can become part of statistical analysis. When treatment availability changes, combining all observations without accounting for the protocol period can create an inappropriate comparison.

18. Primary Endpoint Comparison Map

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

20. What the Risk Difference Does — and Does Not — Mean

21. Time-to-Event Analysis in PROMISE

Three primary outcomes use time-to-event concepts: postpartum confirmed infant HIV infection, postpartum maternal adverse events, and maternal progression to AIDS-defining illness or death. Secondary overall and HIV-free infant survival through 104 weeks also use a time-to-event framework.

The postpartum infant HIV analysis illustrates the importance of follow-up definition. Its observation window extends through the recommended breastfeeding duration, complete cessation of breastfeeding, or 18 months of age, whichever comes first. The maternal-health endpoint instead runs from study entry until July 7, 2015, with an average of 94 weeks of follow-up.

22. Safety Interpretation

Safety is represented by several distinct endpoint definitions rather than one global safety measure. The antepartum primary safety endpoint focuses on Grade 3 or higher toxicities and selected Grade 2 hematologic, renal, and hepatic adverse events through the Week 1 postpartum visit. The postpartum safety endpoint extends through the breastfeeding-related follow-up period or 18 months of age.

These endpoints differ in severity, timing, and statistical structure. Consequently, a single P-value should not be treated as a summary of overall safety. A complete interpretation considers the endpoint definition, denominator, follow-up period, event timing, and comparison structure.

23. Limitations

• Multiple components: PROMISE contains antepartum, postpartum, and maternal-health components with different treatment contrasts and endpoint definitions. Results should therefore be interpreted within their specific component.
• Multiple pairwise comparisons: several primary binary endpoints were analyzed through multiple pairwise comparisons among Antepartum Arms A, B, and C. The registry specifically identifies an a priori planned combined Arm B + C versus Arm A comparison for the antepartum infant HIV endpoint.
• Protocol periods: some treatment arms were available only under particular protocol versions. Secondary infant-survival analyses therefore distinguish Periods 1&2 from period 2.
• Different units of analysis: some outcomes use mothers, some use mother-infant pairs, and the infant survival analyses select one infant per pregnancy. These choices affect the interpretation of the corresponding estimates.
• Censoring: time-to-event analyses require assumptions about censoring. The registry explicitly identifies participants censored at randomization for incomplete post-randomization testing or follow-up.
• Wide confidence intervals: the postpartum infant HIV HR of 1.0 has a 96% CI of 0.3–3.1, while the maternal-health HR of 0.55 has a 95% CI of 0.14–2.08. Both intervals indicate substantial uncertainty.
• Incomplete effect-size reporting: several posted primary analyses provide P-values without a corresponding effect estimate and confidence interval in the registry statistical-analysis record. Those P-values should not be interpreted as effect sizes.
• Composite outcomes: adverse pregnancy outcomes combine several clinically distinct events, including stillbirth, preterm delivery, low birth weight, and congenital anomalies. A composite can reflect differences in one component more strongly than another.
• Sequential inference: the antepartum infant HIV result uses a repeated-confidence-interval framework, so ordinary fixed-sample interpretations of the confidence level would be inappropriate.

24. Why This Trial Matters Statistically

PROMISE is a useful statistical teaching case because it combines several common clinical-trial problems within one randomized program: multiple treatment components, different units of analysis, binary and time-to-event endpoints, protocol-period changes, interim monitoring, repeated confidence intervals, censoring, and intent-to-treat analysis.

25. Related Tutorials

Learn more about the methods used in this trial:

26. Related Calculators

27. Sources

• ClinicalTrials.gov: NCT01061151.
• PubMed: PMID 37535953.
• PubMed: PMID 37199427.
• PubMed: PMID 36170749.
• PubMed: PMID 35858874.
• PubMed: PMID 34750297.

28. Record Summary

PROMISE provides a detailed example of how a large randomized prevention trial can require several statistical frameworks rather than a single universal analysis. The primary analyses include a risk difference for antepartum infant HIV infection, hazard ratios for postpartum infant HIV infection and maternal progression, Fisher exact testing for several binary maternal and pregnancy outcomes, and log-rank testing for time-to-event endpoints. Secondary analyses add two-sided Z tests for infant overall and HIV-free survival through 104 weeks.

The statistical interpretation depends on preserving the distinction between effect size, uncertainty, and hypothesis testing. The antepartum infant HIV comparison reports a risk difference of −1.3 with a 96.5% confidence interval of −2.1 to −0.4 under a repeated-confidence-interval framework. The postpartum infant HIV analysis reports an HR of 1 with a 96% CI of 0.3–3.1. The maternal-health analysis reports an HR of 0.55 with a 95% CI of 0.14–2.08 and P = 0.37.

The broader statistical lesson is that randomized treatment comparisons must be interpreted in the context of the endpoint definition, analysis population, unit of analysis, censoring, protocol period, multiplicity, and inferential framework. PROMISE illustrates each of these issues within a single clinical-trial program.