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
CAPRISA 008 was a randomized, parallel-group, unmasked phase 2/3 prevention trial evaluating 1% tenofovir gel. The registry reports 372 enrolled participants, two arms, a binary primary endpoint, nine posted outcome measures, and eight posted statistical analyses.
| Feature | CAPRISA 008 |
|---|---|
| Trial name | CAPRISA 008 |
| Clinical question | Implementation effectiveness and safety of tenofovir gel provision through family planning services |
| Condition | HIV |
| Phase | Phase 2/3 |
| Status | Completed |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | None |
| Primary purpose | Prevention |
| Enrollment | 372 |
| Intervention | 1% tenofovir gel |
| Lead sponsor | Centre for the AIDS Programme of Research in South Africa |
| ClinicalTrials.gov | NCT01691768 |
2. Clinical Question
The trial evaluated whether provision of 1% tenofovir gel through family planning services could achieve gel-use patterns similar enough to the comparator setting to satisfy a prespecified non-inferiority objective, while also evaluating HIV incidence, pregnancy incidence, adherence, HPV incidence, tenofovir detection, and viral load among HIV seroconverters.
Population
Participants enrolled in the CAPRISA 008 randomized trial for HIV prevention, with analysis populations defined according to randomization and outcome-specific eligibility in the registry.
Intervention
1% tenofovir gel provided through the intervention service pathway.
Comparator
The control arm against which the intervention arm was compared for gel use and the secondary incidence and adherence outcomes.
Primary question
Whether gel use in women attending family planning intervention services was similar to, but no more than 20% lower than, gel use in the comparator group.
3. Trial Design
CAPRISA 008 used a randomized, parallel-group, unmasked design with two arms and a total enrollment of 372. The primary purpose was prevention. The trial began in October 2012 and reached primary completion in April 2015.
1% tenofovir gel
- 1% tenofovir gel was the registered intervention.
- The primary endpoint measured returned used applicators per month.
- Participants in the intervention arm followed two- and three-monthly schedules as described in the primary endpoint analysis.
Comparator pathway
- The control arm was the randomized comparator.
- The primary endpoint compared gel use between intervention and control.
- The registry notes a monthly appointment schedule for the intervention arm in contrast to two- and three-monthly intervals in the control arm.
4. Analysis Populations
The primary analysis explicitly distinguished an intention-to-treat population from a per-protocol population. This distinction is especially important because the primary hypothesis was non-inferiority.
| Population | Definition / role in registry-reported registry analysis |
|---|---|
| Intention-to-treat | All participants who were randomized, met pre-randomization eligibility criteria, and had post-enrollment follow-up data. Used for the primary analysis. |
| Per protocol | Excluded all subsequent data collected from participants who were not dispensed product for more than 120 days. |
| Adherence subset | Participants who returned used gel applicators and also reported sex during follow-up. |
| HPV subset | Participants in the ITT population who had HPV-negative results at randomization. |
| Viral-load population | All participants who became HIV infected. |
| Tenofovir-level population | All participants with drug levels measured at 12 months of follow-up. |
5. Primary Endpoint
| Endpoint | Registry definition / frame | Analysis |
|---|---|---|
| Mean Number of Returned Used Applicators Per Month (i.e in 30 Days) | Between 2012 to 2015, up to 28 months | Mixed-effects model |
The ClinicalTrials.gov record describes the primary endpoint as the mean number of returned used applicators per month. Because participants had different visit intervals, the analysis text describes estimating the monthly number from returned applicators relative to the number of days since the prior visit.
6. Statistical Methodology
Linear mixed-effects modeling
The primary analysis used a mixed-effects model, specifically described in the registry-reported analysis text as a univariate linear mixed model. This is appropriate to the longitudinal structure of repeated monthly gel-use measurements because observations from the same participant are not statistically independent.
The model allows repeated observations from the same participant to be related rather than treating every monthly observation as if it came from a different independent person.
For the ITT analysis, the registry text states that the least-square mean from the intervention arm was the minuend and the least-square mean from the control arm was the subtrahend. Thus the reported difference follows the direction intervention minus control.
Non-inferiority framework
The primary hypothesis was non-inferiority. The registry states that the study was designed to determine whether gel use in women attending family planning intervention services was similar to, but no more than 20% lower than, gel use among women in the comparator setting.
Non-inferiority is different from a conventional superiority test. The objective is not necessarily to show that the intervention produces a larger numerical outcome. Instead, the analysis asks whether the observed difference is sufficiently close to the comparator that a prespecified loss of efficacy or performance would not be exceeded.
A negative value therefore means the estimated intervention value was lower than the control value. The ClinicalTrials.gov record specifies the 20% non-inferiority criterion but do not provide a numerical absolute margin in the reported analysis record.
Secondary incidence-rate comparisons
HIV incidence rates, pregnancy incidence rates, and HPV incidence rates were compared using a z-test, reported in the registry method field. The corresponding effect measure was an incidence rate ratio, calculated as the intervention incidence rate divided by the control incidence rate.
Binary outcomes
Adherence above 80% and detectable tenofovir levels were reported using risk ratios. The registry method field describes log-binomial regression.
Continuous viral-load outcome
HIV viral load among HIV seroconverters was compared using a two-sided t-test. Unlike the incidence outcomes, the ClinicalTrials.gov record does not provide an effect estimate or confidence interval for this endpoint.
7. Primary Results
The registry contains two statistical analyses for the primary endpoint. One is the ITT analysis and the other is the per-protocol analysis. Both use a mixed-effects framework and report a two-sided 95% confidence interval.
Intention-to-treat primary analysis
Mean difference in final values
95% CI: −1.16 to 0.21
Mixed-effects model · Intervention minus control · Non-inferiority
The estimated intervention-minus-control difference was −0.47 used gel applicators per month. Because the estimate is negative, the estimated value in the intervention group was lower than the comparator under the reported model.
The 95% CI of −1.16 to 0.21 describes the statistical uncertainty around that estimated difference. It includes both negative and positive values, so the registry-reported interval is compatible with the intervention having a lower or higher mean outcome than the comparator within the interval's limits.
This confidence interval does not mean that 95% of individual participants had differences between −1.16 and 0.21. It describes uncertainty around the estimated group difference under the analysis framework.
The registry does not provide a p-value for this primary non-inferiority analysis. More importantly, a non-inferiority conclusion is not made simply by asking whether a conventional two-sided p-value is below 0.05. The relevant question is whether the confidence interval is compatible with a difference exceeding the prespecified non-inferiority margin. The ClinicalTrials.gov record states a criterion of no more than 20% lower use but do not provide a numerical absolute margin that would allow that comparison to be reproduced from the posted estimate alone.
The result also reflects the ITT population defined in the registry. In a non-inferiority setting, interpretation of both ITT and per-protocol analyses is important because deviations from assigned treatment or implementation can make groups appear more similar without establishing that the intervention truly retains the required level of performance.
Per-protocol primary analysis
Reported median difference in final values
95% CI: −0.98 to 0.48
Mixed-effects model · Per-protocol population · Non-inferiority
The per-protocol analysis reports a −0.25 difference, with a 95% CI from −0.98 to 0.48. The registry analysis notes that this population excludes subsequent data from participants who were not dispensed product for more than 120 days.
As with the ITT estimate, the negative point estimate indicates a lower intervention value under the stated intervention-minus-control direction. The confidence interval spans zero, meaning that the reported data are compatible with a range of intervention-versus-control differences.
The registry's effect-measure field labels this result Median Difference (Final Values), while the accompanying analysis text says that the investigators calculated the difference in means using a univariable linear mixed model. Those descriptions should not be silently treated as interchangeable; this page preserves the registry labeling and does not recalculate or reinterpret the reported estimate.
The absence of a reported p-value is not a deficiency by itself in a non-inferiority analysis. The central issue is the prespecified non-inferiority margin and whether the confidence interval excludes differences that would represent an unacceptable loss. Because the ClinicalTrials.gov record does not give the numerical absolute margin, the non-inferiority calculation cannot be independently reconstructed beyond the stated 20% criterion.
8. Secondary Endpoint Results
The registry posts statistical analyses for six secondary outcome measures. The results below reproduce the estimates, confidence intervals, and p-values reported in the ClinicalTrials.gov record without recalculation.
HIV Incidence Rates
Incidence rate ratio
95% CI: 0.40–2.35 · P = 0.928
Wald / z-test · Superiority framework · ITT population
An incidence rate ratio of 0.96 means that the intervention incidence rate was estimated at 0.96 times the control incidence rate because the registry defines the intervention incidence rate as the numerator and the control incidence rate as the denominator.
The point estimate is close to 1, but the 95% CI of 0.40 to 2.35 is wide. Thus the estimate is imprecise and is compatible with a substantially lower or substantially higher incidence rate in the intervention group.
The reported P = 0.928 is a test result under the specified superiority analysis. It does not measure the size or clinical importance of the effect, and it should not be interpreted as the probability that the treatment has no effect.
The registry specifically states that a log-rank test was not used because survival curves were not presented; instead, incidence rates were compared with a z-test. This is a distinction between an incidence-rate comparison and a time-to-event analysis.
Pregnancy Incidence Rates
Incidence rate ratio
95% CI: 0.45–2.04 · P = 0.895
Wald / z-test · Superiority framework · ITT population
The pregnancy incidence rate ratio of 0.96 uses the intervention rate as numerator and the control rate as denominator. A value of 1 would correspond to equal incidence rates under this definition.
The 95% CI of 0.45 to 2.04 is substantially wider than a narrow interval around 1 would be. It therefore indicates considerable uncertainty about the relative pregnancy incidence rate.
The reported P = 0.895 is evidence from the specified two-sided superiority test, not an effect-size measure. A nonsignificant p-value should not be converted into a statement that the two incidence rates are proven identical.
Percentage of Participants Achieving Adherence >80%
Risk ratio
95% CI: 0.94–1.24 · P = 0.304
Log-binomial regression · Superiority framework
A risk ratio of 1.08 means that the proportion achieving adherence above 80% was estimated to be 1.08 times the corresponding proportion in the control group within the specified analysis subset.
The 95% CI of 0.94 to 1.24 includes 1.00. The interval therefore includes the possibility of a modestly lower proportion in the intervention group as well as a higher proportion.
The P = 0.304 does not quantify how large the adherence difference is. The risk ratio and confidence interval are the more direct measures of the estimated relative difference and its precision.
This analysis uses a subset of the ITT population rather than every randomized participant: participants had to return used gel applicators and report sex during follow-up. That restriction is important when interpreting the result because the estimand applies to that defined subset rather than automatically to every randomized participant.
Human Papillomavirus Incidence Rates
Incidence rate ratio
95% CI: 0.06–1.32 · P = 0.097
Wald / z-test · Superiority framework · HPV-negative subset
The HPV incidence rate ratio of 0.33 means the estimated incidence rate in the intervention group was one-third of the control-group incidence rate under the registry's numerator/denominator definition.
However, the 95% CI of 0.06 to 1.32 is wide and crosses 1.00. The data therefore allow a range of relative incidence rates, including values below and above the null ratio.
The P = 0.097 is not a measure of the size of the apparent association. Nor should the point estimate of 0.33 be presented without its confidence interval: the interval demonstrates how uncertain that point estimate is.
This endpoint was restricted to participants who had HPV-negative results at randomization. It therefore evaluates incident HPV outcomes in that defined subset rather than representing an unrestricted estimate across all randomized participants.
Detectable Tenofovir Levels From Vaginal Samples at 12 Months
Risk ratio
95% CI: 0.70–1.18 · P = 0.462
Log-binomial regression · Participants with drug levels at 12 months
The risk ratio of 0.91 compares the probability of detectable tenofovir levels between the intervention and control groups among participants with drug levels measured at 12 months.
The 95% CI of 0.70 to 1.18 includes 1.00. Thus the interval includes both a lower and a higher intervention-group probability relative to control.
The P = 0.462 is the reported two-sided test result. It does not establish equivalence, and it does not imply that the intervention and control groups have identical probabilities.
HIV Viral Load Among HIV Seroconverters
Two-sided t-test
Outcome unit: log10 copies/ml
Population: all participants who became HIV infected
The viral-load analysis compared HIV viral load among participants who became HIV infected using a two-sided t-test. The ClinicalTrials.gov record reports P = 0.455, but do not provide an effect estimate or confidence interval.
Consequently, the p-value alone does not tell us the magnitude or precision of any difference in viral load. A complete quantitative interpretation would require the corresponding group summaries or an effect estimate with uncertainty.
The analysis population is also important: this is not an analysis of all randomized participants. It is restricted to participants who became HIV infected.
9. Results Summary
| Endpoint | Effect | 95% CI | P-value | Method |
|---|---|---|---|---|
| Primary endpoint — ITT | Mean difference −0.47 | −1.16 to 0.21 | Not reported | Mixed-effects model |
| Primary endpoint — per protocol | Median difference −0.25 | −0.98 to 0.48 | Not reported | Mixed-effects model |
| HIV incidence | Rate ratio 0.96 | 0.40–2.35 | 0.928 | Wald / z-test |
| Pregnancy incidence | Rate ratio 0.96 | 0.45–2.04 | 0.895 | Wald / z-test |
| Adherence >80% | Risk ratio 1.08 | 0.94–1.24 | 0.304 | Log-binomial regression |
| HPV incidence | Rate ratio 0.33 | 0.06–1.32 | 0.097 | Wald / z-test |
| Detectable tenofovir at 12 months | Risk ratio 0.91 | 0.70–1.18 | 0.462 | Log-binomial regression |
| HIV viral load among seroconverters | Not reported | Not reported | 0.455 | Two-sided t-test |
10. Statistical Methods Explained
Why was a mixed-effects model used for the primary endpoint?
The primary outcome was measured repeatedly over follow-up, so observations from the same participant can be correlated. A mixed-effects model allows the analysis to represent this within-participant structure rather than treating repeated measurements as independent observations. The registry-reported analysis text specifically describes a univariate linear mixed model.
What does the primary difference of −0.47 mean?
The registry-reported analysis defines the direction as intervention minus control. Therefore, −0.47 indicates an estimated intervention value 0.47 units lower than the comparator on the analyzed primary outcome scale. It does not mean that every participant returned 0.47 fewer applicators per month, nor does it by itself establish non-inferiority.
Why is the non-inferiority margin more important than a conventional p-value?
In a superiority test, the central question is often whether the data provide evidence against equality in favor of a treatment difference. In a non-inferiority trial, the question is whether the treatment is sufficiently close to the comparator. The CAPRISA 008 registry states a criterion of no more than 20% lower gel use. The confidence interval must therefore be considered relative to that prespecified tolerance. The ClinicalTrials.gov record does not provide the corresponding numerical absolute margin.
Why were both ITT and per-protocol analyses reported?
ITT preserves the randomized treatment assignment and therefore retains an important advantage of randomization. In non-inferiority trials, however, deviations from treatment or implementation can make groups appear artificially similar, so a per-protocol analysis provides a complementary check. The strongest interpretation comes from considering both populations according to the prespecified analysis plan.
What does a rate ratio of 0.96 mean?
A rate ratio of 0.96 means the intervention incidence rate was estimated to be 0.96 times the control incidence rate. Equivalently, the point estimate is 4% below the control rate, but that derived statement does not imply that the intervention definitely reduced incidence by 4%. The confidence interval must be considered, and for HIV incidence it extends from 0.40 to 2.35.
What is the difference between a risk ratio and a rate ratio?
A risk ratio compares probabilities or proportions over a specified population or period. A rate ratio compares incidence rates, which incorporate person-time. CAPRISA 008 uses risk ratios for outcomes such as adherence above 80% and detectable tenofovir levels, while HIV, pregnancy, and HPV outcomes are reported as incidence rate ratios.
Why does a confidence interval matter more than a p-value alone?
A p-value summarizes evidence against a specified null hypothesis under a particular test. It does not tell the reader how large the treatment effect is. A confidence interval provides an estimated range of effects compatible with the statistical model and data. For CAPRISA 008, the wide intervals for the incidence-rate outcomes are particularly important because they show that the point estimates alone do not capture the uncertainty of the comparisons.
11. Non-Inferiority: The Key Statistical Issue
The primary endpoint makes CAPRISA 008 especially useful for teaching non-inferiority trial design. The registry states that the study was designed to provide 90% power to determine whether gel use through family planning intervention services was similar to, but no more than 20% lower than, gel use among women in the comparator setting. The registry also states that 700 women were estimated to be needed after accounting for anticipated loss to follow-up.
| Feature | Supplied trial information |
|---|---|
| Hypothesis type | Non-inferiority |
| Primary endpoint | Mean number of returned used applicators per month |
| Stated tolerance | No more than 20% lower gel use |
| Planned power | 90% |
| Estimated enrollment for power calculation | 700 women after accounting for anticipated loss to follow-up |
| Observed enrollment | 372 |
The planned enrollment and actual enrollment are different quantities. The registry's power statement describes an anticipated sample requirement, whereas the actual ClinicalTrials.gov record reports 372 enrolled participants. One should not substitute the observed enrollment into the original power calculation or infer achieved power from the final effect estimate.
12. Confidence Intervals and Effect Measures
CAPRISA 008 provides a useful demonstration of why the effect measure has to match the endpoint.
| Measure | What it compares | CAPRISA 008 example |
|---|---|---|
| Mean difference | Difference between estimated group means | Primary ITT estimate: −0.47 |
| Median difference | Difference in the reported median-based effect measure | Per-protocol field: −0.25 |
| Rate ratio | Intervention incidence rate divided by control incidence rate | HIV incidence: 0.96 |
| Risk ratio | Intervention probability/proportion divided by control probability/proportion | Adherence >80%: 1.08 |
These quantities answer different questions. A rate ratio cannot be read as though it were a risk ratio, and a difference in means cannot be converted into a percentage effect without additional information about the underlying scale.
13. Multiplicity and Secondary Outcomes
The ClinicalTrials.gov record identifies one primary endpoint and multiple secondary outcomes. The secondary analyses cover HIV incidence, pregnancy incidence, adherence, HPV incidence, detectable tenofovir, and viral load among seroconverters.
The registry states that power was not calculated for all secondary outcomes. This matters when interpreting a collection of secondary p-values. A secondary analysis can be informative without being individually powered to detect a prespecified effect, but a nominal p-value should not automatically be treated as evidence from a separately powered confirmatory hypothesis.
Primary analysis
The primary endpoint was explicitly tied to the non-inferiority objective and analyzed in both ITT and per-protocol populations.
Secondary analyses
Secondary outcomes addressed biological and implementation-related questions, including incidence rates, adherence, drug detection, and viral load.
The ClinicalTrials.gov record does not provide an alpha-allocation or formal multiplicity-adjustment procedure for the secondary endpoints. Accordingly, the reported secondary p-values should be presented as the registry reports them rather than treated as a newly reconstructed familywise-error-controlled testing hierarchy.
14. Why the Incidence Analyses Used a z-Test
The HIV, pregnancy, and HPV outcomes were analyzed as incidence rates. The registry explicitly states that the investigators used a z-test to compare incidence rates and did not use a log-rank test because they did not present survival curves.
A value of 1 represents equal incidence rates. Values below 1 favor a lower incidence rate in the intervention group; values above 1 indicate a higher intervention-group incidence rate.
This distinction illustrates a broader principle in clinical-trial statistics: the analysis method should follow the estimand and the way follow-up information is represented. An incidence-rate comparison and a time-to-event analysis are related but are not the same statistical question.
15. Safety Results
The ClinicalTrials.gov record reports serious adverse events by randomized arm. These are presented as affected participants divided by the number at risk.
| Safety measure | Intervention | Control |
|---|---|---|
| Serious adverse events | 13/189 | 17/183 |
The registry supplies the serious-adverse-event counts as affected participants over participants at risk. It does not provide a statistical comparison for this safety measure in the ClinicalTrials.gov record. Accordingly, the figures should be treated as descriptive safety information rather than as evidence of a formally tested difference.
16. Statistical Interpretation of the Primary Evidence
The ITT primary estimate is −0.47, with the intervention group defined as the minuend and the control group as the subtrahend. This establishes the direction and magnitude of the reported model-based difference, but the point estimate alone does not establish whether the difference is acceptable under the non-inferiority criterion.
The 95% CI of −1.16 to 0.21 indicates meaningful uncertainty around the point estimate. Because it crosses zero, the interval includes both a negative and positive intervention-control difference. For a non-inferiority question, however, zero is not the sole reference point: the relevant boundary is the prespecified non-inferiority margin.
No p-value is reported for the primary non-inferiority analyses. This is important because attempting to infer a conventional superiority conclusion from the confidence interval would answer a different question from the one the trial was designed to address.
The ITT and per-protocol analyses answer related but distinct questions. The ITT analysis maintains the randomized assignment framework, whereas the per-protocol analysis removes specified post-randomization observations. In a non-inferiority trial, both perspectives can be important because deviations from the intended treatment pathway can bias a comparison toward similarity.
17. What the Secondary Results Do — and Do Not — Show
The secondary endpoints illustrate several different forms of statistical evidence. The HIV incidence rate ratio is 0.96, the pregnancy incidence rate ratio is 0.96, the adherence risk ratio is 1.08, the HPV incidence rate ratio is 0.33, and the detectable-tenofovir risk ratio is 0.91. These estimates should not be placed on a single "better versus worse" scale because they describe different outcomes and use different effect measures.
The HPV incidence estimate of 0.33, for example, is numerically farther from 1 than the HIV incidence estimate of 0.96. But its confidence interval, 0.06–1.32, is also wide and includes 1. The correct statistical reading therefore requires both the point estimate and uncertainty interval.
Similarly, the adherence risk ratio of 1.08 has a 95% CI of 0.94–1.24. The point estimate alone might suggest a higher proportion achieving adherence above 80%, but the interval includes values below and above 1. The reported P = 0.304 provides the corresponding hypothesis-test result but does not change the size of the estimated effect.
18. Missing Data and Censoring Considerations
The ClinicalTrials.gov record does not specify an imputation method for missing primary-endpoint measurements. It also does not provide a detailed missing-data sensitivity-analysis strategy. The primary ITT definition requires post-enrollment follow-up data, while the per-protocol definition excludes subsequent data after a specified product-dispensing criterion.
For a longitudinal mixed-effects analysis, assumptions about the mechanism generating missing observations can affect interpretation. A mixed model can use available repeated measurements under its statistical assumptions, but the model itself does not prove that missingness is ignorable. Because no further imputation or sensitivity-analysis details are reported here, this page does not attribute a specific missing-data assumption to the trial beyond what is explicitly stated in the analysis record.
19. Study Timeline
Trial start
CAPRISA 008 began as a randomized phase 2/3 prevention trial evaluating 1% tenofovir gel provision through family planning services.
Study follow-up
The primary endpoint and several secondary outcomes were assessed between 2012 and 2015, with the primary endpoint allowing follow-up up to 28 months.
Primary completion
The registry lists April 2015 as the primary completion date.
Results posted
ClinicalTrials.gov reports results, including nine posted outcome measures and eight statistical analyses.
20. Important Limitations and Interpretation Issues
- Recall bias: the registry states that the two- to three-month clinic intervals in the control arm are likely to introduce recall bias, unlike the monthly intervals in the intervention arm.
- Non-inferiority margin: the registry states a criterion of no more than 20% lower gel use but does not supply a numerical absolute margin in the provided analysis record.
- Planned versus actual enrollment: the registry describes an estimated requirement of 700 women for 90% power, whereas the ClinicalTrials.gov record reports 372 enrolled participants. These should not be conflated.
- ITT versus per protocol: the primary endpoint was analyzed in both populations, and the populations have different post-randomization definitions.
- Secondary endpoints: the registry states that power was not calculated for all secondary outcomes.
- Outcome-specific populations: several secondary endpoints use subsets rather than all randomized participants.
- Wide confidence intervals: several secondary incidence-rate estimates have broad confidence intervals, limiting precision even when the point estimate is far from the null value.
- Missing-data details: the ClinicalTrials.gov record does not specify an imputation procedure or detailed sensitivity-analysis strategy for missing primary-endpoint measurements.
- Safety inference: serious adverse-event counts are reported by arm, but no formal statistical comparison is provided in the ClinicalTrials.gov record.
- Method-label consistency: the registry-reported primary per-protocol analysis field labels its effect measure as a median difference, while its accompanying analysis text describes a difference in means. The page preserves that distinction rather than silently resolving it.
21. Why This Trial Matters Statistically
CAPRISA 008 is a useful teaching case because the statistical problem is not simply "did the treatment work?" The primary objective is a non-inferiority question about implementation and gel use, while the secondary outcomes span incidence rates, binary adherence, biological drug detection, and continuous viral load.
| Statistical concept | How it appears in CAPRISA 008 |
|---|---|
| Randomization | Randomized, parallel-group design with 372 enrolled participants |
| Intention-to-treat | Primary analysis population includes randomized participants meeting the stated eligibility and follow-up criteria |
| Per-protocol analysis | Primary endpoint also analyzed after excluding specified subsequent data |
| Non-inferiority | Primary objective uses a criterion of no more than 20% lower gel use |
| Mixed-effects model | Primary longitudinal gel-use analysis |
| Confidence intervals | Reported for the primary difference and most secondary effect estimates |
| Rate ratio | HIV, pregnancy, and HPV incidence-rate comparisons |
| Risk ratio | Adherence above 80% and detectable tenofovir levels |
| Wald / z-test | Incidence-rate comparisons |
| t-test | HIV viral load among seroconverters |
| Outcome-specific populations | Adherence, HPV, drug-level, and viral-load analyses use defined subsets |
| Recall bias | Different clinic-appointment intervals between arms are identified by the registry as a limitation |
22. Statistical Methods in Context
Implementation outcome
The primary endpoint evaluates returned used applicators per month, making the statistical target a longitudinal measure of observed product use rather than HIV infection itself.
Biological outcomes
HIV and HPV incidence and detectable tenofovir levels provide additional outcome perspectives, each with its own analysis population and effect measure.
Non-inferiority logic
The question is whether intervention performance remains within an acceptable loss relative to control, not simply whether a superiority p-value crosses a conventional threshold.
Precision
The confidence intervals are essential because several point estimates are accompanied by substantial uncertainty.
23. Related Tutorials
Learn more about the methods used in this trial:
24. Related Calculators
25. Sources
- ClinicalTrials.gov: NCT01691768 — CAPRISA 008.
- Linked publication: PubMed — PMID 28348164.
- Linked publication: PubMed — PMID 25527071.
Continue with the underlying statistical methods
Explore the statistical concepts behind randomized trials, non-inferiority analyses, longitudinal models, incidence-rate comparisons, and binary-outcome measures.
26. Record Summary
CAPRISA 008 provides a particularly useful statistical example because its primary endpoint is an implementation measure analyzed under a non-inferiority framework, while its secondary outcomes use several distinct statistical estimands. The primary ITT analysis reports a difference of −0.47 with a 95% CI of −1.16 to 0.21, and the per-protocol analysis reports −0.25 with a 95% CI of −0.98 to 0.48. Secondary analyses report incidence rate ratios for HIV, pregnancy, and HPV; risk ratios for adherence and detectable tenofovir; and a two-sided t-test for viral load among HIV seroconverters.
The most important statistical lesson is that these results cannot be reduced to a single p-value or a single direction of effect. The primary non-inferiority question requires the confidence interval to be evaluated against the prespecified margin; secondary outcomes require attention to their distinct estimands and analysis populations; and the recall-bias concern identified by the registry is relevant to interpretation of the primary implementation outcome itself.