← Clinical Trials
HIV Prevention Phase 2/3 Randomized NCT00448669

TDF2: Complete Statistical Analysis of Tenofovir/Emtricitabine in HIV Prevention

An independent statistical review of the randomized TDF2 trial evaluating tenofovir disoproxil fumarate plus emtricitabine, with condoms and risk counseling, versus placebo with condoms and risk counseling for HIV prevention.

Trial status: Completed  ·  Enrollment: 1219  ·  Start: March 2007  ·  Primary completion: March 2011
Scope of this record

This page separates reported trial results from statistical interpretation. Numerical results on this page are restricted to the information reported in the ClinicalTrials.gov record. ClinicalTrials.gov provides the official trial registry record.

1. Trial at a Glance

TDF2 was a randomized, parallel-group, triple-masked phase 2/3 prevention trial in HIV infections. The trial compared oral tenofovir disoproxil fumarate plus emtricitabine with placebo, with condoms and risk counseling provided in both groups.

1219
Enrollment
2 randomized arms
2
Arms
Parallel design
62.2%
Reported efficacy
95% CI 21.5%–83.4%
0.03
HIV incidence P-value
Two-sided
FeatureTDF2
Trial nameTDF2
Brief titleBotswana TDF/FTC Oral HIV Prophylaxis Trial
PhasePhase 2/3
ConditionHIV Infections
Primary purposePrevention
DesignRandomized, parallel-group, triple-masked
Enrollment1219
Primary endpointsPercentage of participants with adverse drug reactions; HIV incidence
Results postedYes
Statistical analyses posted4
Lead sponsorCenters for Disease Control and Prevention
ClinicalTrials.govNCT00448669

2. Clinical Question

The central question was whether assignment to tenofovir disoproxil fumarate plus emtricitabine, together with condoms and risk counseling, was associated with a lower rate of HIV infection than assignment to placebo with condoms and risk counseling.

Population

Participants enrolled in the TDF2 Botswana TDF/FTC oral HIV prophylaxis trial for the prevention of HIV infection.

Intervention

Tenofovir disoproxil fumarate 300 mg plus emtricitabine 200 mg, with condoms and risk counseling.

Comparator

Placebo oral tablet, with condoms and risk counseling.

Primary question

Does TDF-FTC reduce the rate of HIV infection compared with placebo under the randomized trial design?

3. Trial Design

01
Enroll1219 participants
02
Randomize2 parallel arms
03
MaskTriple-masked trial
04
FollowMonthly visits
05
AnalyzeHIV incidence and safety
INTERVENTION

TDF-FTC, condoms, risk counseling

  • Tenofovir disoproxil fumarate 300 mg
  • Emtricitabine 200 mg
  • Condoms
  • Risk counseling
COMPARATOR

Placebo, condoms, risk counseling

  • Placebo oral tablet
  • Condoms
  • Risk counseling

The registry describes a randomized parallel design with triple masking. Randomization is important statistically because treatment assignment is determined independently of subsequent outcomes, supporting a comparison of outcome rates between the assigned groups.

Masking is part of the design, not an analysis method. Triple masking can reduce opportunities for knowledge of treatment assignment to influence participant behavior, clinical assessment, or other aspects of trial conduct. The ClinicalTrials.gov record identifies the masking level but do not specify which three groups were masked.

4. Primary Endpoints

Registered endpointTime frameDefinition / assessmentAnalysis
Percentage of Participants With Adverse Drug Reactions in the Tenofovir/Emtricitabine and Placebo Arms Monthly, for up to 3 years Study visits were scheduled every 30 days until completion, with interim evaluation for illness. Participants reported adverse effects at monthly and interim visits. Cox proportional-hazards model
HIV Incidence in the Tenofovir/Emtricitabine and Placebo Arms Monthly, for up to 3 years Monthly HIV testing was performed during study visits, with testing of all participants at study completion using an ELISA. Cox proportional-hazards model

Although both registered primary endpoints are described in the registry as binary endpoints, the posted statistical analyses treated the observed outcomes using a time-to-event framework. That distinction matters: HIV incidence over follow-up is not simply a single yes/no comparison at one fixed date when participants can become infected at different times.

5. Analysis Populations

The registry provides explicit analysis-population information for the HIV-incidence analysis and for safety. These populations are important because the interpretation of an effect estimate depends on who was included.

AnalysisPopulation described in the registry
HIV incidence Of the 1219 enrolled participants, 3 were excluded because they were HIV-infected at enrollment.
Safety Safety analyses were performed in the intention-to-treat cohort.
Condom-use analysis Of 1219, 19 were excluded: 3 HIV-infected at enrollment and 16 who never started drug. Of the remaining 1200, 24 reported no sex during the study. The remaining total of 1176 with at least 1 sex act were included in the analysis.

Why the HIV-at-enrollment exclusion matters

The primary HIV-incidence analysis excludes the 3 participants who were already HIV-infected at enrollment. This is directly relevant to the endpoint because incident HIV infection concerns infections occurring after entry into the at-risk study population. Including prevalent infection would mix two different quantities: HIV status already present at enrollment and HIV acquisition during follow-up.

6. Statistical Methodology

Cox proportional-hazards model

The registry reports a Cox regression for the HIV-incidence primary endpoint and for the adverse-drug-reaction primary endpoint. A Cox proportional-hazards model is appropriate when the analysis concerns the timing of events as well as whether events occur.

Conceptual Cox model
h(t | X) = h0(t) exp(βX)

The model describes the event hazard over time as a baseline hazard multiplied by a treatment-related relative factor. The corresponding treatment comparison is commonly expressed through a hazard ratio.

For the HIV endpoint, the registry reports the effect measure as efficacy rather than as a hazard ratio. Therefore, the reported value of 62.2 should be presented as the registry's efficacy estimate rather than converted into a hazard ratio or another effect measure.

Fisher exact test

Fisher's exact test was used for the comparison of adherence to study medication based on pill count over 36 months. This test is designed for categorical comparisons and is particularly useful when sample sizes or cell counts make large-sample approximations less attractive.

Logistic regression

Logistic regression was used for the secondary analysis of condom use at 12 months. The registry states that the model estimated the odds of reporting zero condomless sex acts.

Odds ratio
OR = odds in the modeled comparison group / odds in the reference group

An odds ratio of 1 would indicate equal odds under the model. Values above or below 1 represent relative changes in odds, not direct changes in probability.

Intention-to-treat analysis

The registry explicitly states that safety analyses were performed in the intention-to-treat cohort. The ITT principle analyzes participants according to their randomized assignment rather than changing assignment based on subsequent treatment behavior.

7. Primary Result: HIV Incidence

The primary efficacy endpoint was HIV incidence in the TDF-FTC and placebo groups over monthly follow-up for up to 3 years. The registry reports a Cox regression analysis comparing the TDF-FTC, condoms, and risk-counseling group with the placebo, condoms, and risk-counseling group.

Reported efficacy

62.2%

95% CI: 21.5%–83.4%   ·   P = 0.03

Two-sided 95% confidence interval; superiority hypothesis

Primary endpointAnalysisReported estimate95% CIP-value
HIV Incidence Cox proportional-hazards model Efficacy 62.2% 21.5%–83.4% 0.03

The registry's analysis notes state that the primary efficacy endpoint was the difference in rates of HIV infection between participants assigned to TDF-FTC and those assigned to placebo. The primary hypothesis was that TDF-FTC, as compared with placebo, would reduce the rate of HIV infection by at least 65%, with a predefined lower boundary for the 95% confidence interval of 10%.

Clinical Biostats interpretation

The reported 62.2% efficacy estimate summarizes the treatment effect using the effect-measure terminology reported by the registry. It should not be relabeled as a hazard ratio because the registry reports the Cox model but labels the effect measure as efficacy.

The 95% CI of 21.5%–83.4% communicates uncertainty around the estimated efficacy. It does not mean that individual participants experienced effects somewhere between 21.5% and 83.4%, nor does it describe the probability that the true effect lies inside that particular interval.

The P-value of 0.03 is evidence against the null hypothesis under the specified testing framework. It does not measure the magnitude of the effect. The effect magnitude is described by the 62.2% estimate, while its statistical precision is reflected by the confidence interval.

The registry also gives a prespecified efficacy target of at least 65% and a predefined lower confidence-bound criterion of 10%. Those design thresholds should be distinguished from the observed point estimate. A point estimate and a confidence-bound criterion answer different statistical questions.

Because the analysis used a Cox model, interpretation also depends on the time-to-event framework and its assumptions, including the proportional-hazards assumption. The ClinicalTrials.gov record does not provide enough information to assess that assumption directly.

Why the efficacy estimate should not be turned into a risk ratio

A common statistical mistake is to treat every percentage reduction as interchangeable. The registry reports efficacy from a Cox regression. Without the underlying event-time data and the exact computational definition used for the registry's efficacy measure, it would be inappropriate to reconstruct a risk ratio, risk difference, or hazard ratio from the reported percentage.

8. Primary Safety Endpoint

The second registered primary endpoint was the percentage of participants with adverse drug reactions in the TDF-FTC and placebo arms, assessed monthly for up to 3 years. The registry reports a Cox regression with a two-sided analysis and a P-value of 0.003.

Adverse drug reactions

P = 0.003

Cox proportional-hazards model   ·   Two-sided

Safety analyses were performed in the intention-to-treat cohort.

Primary safety endpointMethodReported result
Percentage of Participants With Adverse Drug Reactions Cox proportional-hazards model P = 0.003
Clinical Biostats interpretation

The registry provides a P-value of 0.003 for the primary adverse-drug-reaction comparison, but it does not provide an effect estimate or confidence interval for this analysis in the ClinicalTrials.gov record. Therefore, the magnitude and precision of the between-group difference cannot be reconstructed from the posted analysis alone.

The P-value indicates evidence against the null hypothesis under the reported testing framework; it does not tell us how large the difference in adverse reactions was.

The use of a Cox model indicates that the analysis incorporated the timing of adverse events rather than treating every participant's safety experience as a single fixed binary observation at one arbitrary cutoff.

Because this was a safety endpoint, the finding should be interpreted independently of the HIV-incidence efficacy estimate. Statistical evidence of a difference in adverse reactions does not by itself establish whether that difference is clinically important.

9. Secondary Result: Condom Use

The trial also analyzed changes in condom use during the study. At 12 months, the registry defines the longitudinal dependent variable in terms of condomless vaginal sexual acts with casual and main partners among participants who reported at least one sexual partner in the previous 30 days. Logistic regression was used to estimate the odds of reporting zero condomless sex acts.

Odds ratio for reporting zero condomless sex acts

1.16

95% CI: 1.05–1.28   ·   P = 0.004

12-month analysis; two-sided 95% confidence interval

Secondary endpointPopulationMethodEstimate95% CIP-value
Changes in Condom Use During Study: Number of Participants With ≥1 Condomless Sex Acts 1176 participants with ≥1 sex act included in analysis Logistic regression OR 1.16 1.05–1.28 0.004
Clinical Biostats interpretation

An odds ratio of 1.16 means that the modeled odds of the specified outcome—reporting zero condomless sex acts—were estimated to be 1.16 times the odds in the comparison framework used by the regression.

This is an odds measure, not a probability ratio. An OR of 1.16 does not mean that the probability increased by exactly 16 percentage points or that it was exactly 16% higher.

The 95% CI of 1.05–1.28 describes uncertainty around the odds-ratio estimate. Because the interval is entirely above 1, the estimated direction of association is above the null value in this analysis.

The P-value of 0.004 measures evidence against the null hypothesis under the specified model. It does not quantify the practical importance of the behavioral difference.

The analysis is limited to the population defined by the registry: 1176 participants with at least 1 sex act after the stated exclusions. That makes the analysis population different from the full enrolled population of 1219.

10. Secondary Result: Adherence to Study Medication

Rates of adherence to study medication were assessed over 36 months using the percentage of pills taken. The registry reports Fisher's exact test for differences between the treatment groups based on pill count.

Fisher exact test

P = 0.79

36-month adherence analysis

Outcome unit: percentage of pills taken

Secondary endpointMethodP-value
Rates of Adherence to Study Medication Fisher exact test 0.79

The registry does not provide an effect estimate or confidence interval for this analysis in the ClinicalTrials.gov record. The P-value of 0.79 therefore should not be converted into a numerical difference in adherence.

Clinical Biostats interpretation

Fisher's exact test evaluates whether the categorical distribution differs between the two randomized groups under the specified analysis. A P-value of 0.79 does not demonstrate that adherence was identical; it indicates that this analysis did not provide evidence of a statistically detectable difference under its testing framework.

This distinction is important because absence of statistical evidence for a difference is not proof of equivalence. A confidence interval or an explicit equivalence/non-inferiority analysis would be needed to quantify how large a difference could reasonably remain compatible with the data.

11. Safety Results

The registry supplies serious adverse-event counts by randomized arm. These are reported as affected participants over participants at risk.

Safety measureTDF-FTC, condoms, risk counselingPlacebo, condoms, risk counseling
Serious adverse events 42 / 611 44 / 608
Serious adverse events — affected participants
TDF-FTC group
42
Placebo group
44

The ClinicalTrials.gov record reports the affected and at-risk counts but do not provide a formal effect estimate, confidence interval, or P-value for the serious-adverse-event comparison. Accordingly, this page does not calculate a relative risk, odds ratio, or risk difference from those counts.

Why avoid recomputing the safety effect? The task-specific data rules restrict reported numbers to the ClinicalTrials.gov record and prohibit recomputing or converting results. More importantly, a derived safety estimate would depend on the intended analysis population and definition of the safety endpoint. The posted serious-adverse-event counts are therefore presented descriptively.

12. Statistical Methods Explained

Why was a Cox proportional-hazards model used for HIV incidence?

HIV infection can occur at different times during follow-up. A time-to-event model can use information about when infection occurs rather than reducing the entire study to a single end-of-study binary indicator. The Cox model is semiparametric because it estimates the treatment-related relative effect without requiring a fully specified baseline hazard function.

What does the 62.2% efficacy estimate mean?

It is the efficacy estimate reported by the ClinicalTrials.gov statistical analysis. In context, it describes the reported treatment effect for HIV incidence comparing TDF-FTC with placebo. It should not be treated as a hazard ratio, risk ratio, or absolute percentage-point reduction because the registry labels the effect measure as efficacy.

Why is the confidence interval important?

The 95% CI of 21.5%–83.4% provides information about the statistical precision of the efficacy estimate. The interval is much more informative than the point estimate alone because it shows the range of effect values compatible with the statistical estimation framework at the stated confidence level.

Why doesn't a P-value measure effect size?

A P-value is influenced by both the magnitude of the observed departure from the null and the amount of information available for estimating that departure. It therefore cannot be used as a substitute for an effect estimate. In TDF2, the HIV-incidence P-value is 0.03, while the effect estimate is reported separately as 62.2% efficacy.

Why was logistic regression used for condom use?

The specified outcome was whether participants reported zero condomless sex acts, making a binary outcome suitable for logistic regression. The resulting odds ratio of 1.16 summarizes the modeled relative odds, not a direct difference in probabilities.

Why use Fisher's exact test for adherence?

The registry specifically reports Fisher's exact test for the comparison of adherence based on pill count. Fisher's exact test evaluates categorical contingency-table differences without relying on the same large-sample approximation used by a conventional chi-square test.

Why does intention-to-treat matter for safety?

The registry explicitly states that safety analyses were performed in the intention-to-treat cohort. An ITT framework preserves the original randomized assignment rather than redefining treatment groups according to later behavior. This makes the analysis population traceable to the randomized design.

13. Primary Hypothesis and Confidence-Boundary Logic

The registry gives an unusually specific description of the primary efficacy hypothesis. The stated hypothesis was that TDF-FTC, compared with placebo, would reduce the rate of HIV infection by at least 65%, with a predefined lower boundary for the 95% confidence interval of 10%.

Point estimate

The reported efficacy estimate was 62.2%. This is the central estimate produced by the posted analysis.

Confidence interval

The 95% CI was 21.5%–83.4%, showing the statistical uncertainty around the efficacy estimate.

Prespecified target

The registry describes a primary hypothesis of at least 65% reduction in the rate of HIV infection.

Lower boundary

The predefined lower boundary for the 95% confidence interval was 10%.

These quantities should not be conflated. The observed estimate answers "what effect was estimated?" The confidence interval answers "how precisely was that effect estimated?" The prespecified hypothesis and lower confidence-bound criterion describe how the trial intended to judge the evidence against a clinically specified effect.

Important statistical distinction: the 65% hypothesis, the 10% confidence-boundary criterion, the 62.2% estimate, and the P-value of 0.03 are different components of the statistical framework. A careful interpretation reports each separately rather than treating any one of them as a substitute for the others.

14. Time-to-Event Analysis and Censoring

The HIV-incidence endpoint was assessed monthly for up to 3 years. Time-to-event analysis is useful in this setting because participants may have different lengths of observed follow-up, and infection can occur at different points during that follow-up.

Conceptual survival-analysis quantities
S(t) = P(T > t)

Here, T represents the time to the event of interest. Participants who have not experienced the event by the end of their observable follow-up contribute information up to their censoring time.

The ClinicalTrials.gov record does not provide the individual event times, censoring times, Kaplan-Meier estimates, or a Kaplan-Meier curve. Consequently, this page does not reconstruct a survival curve or report median time to HIV infection.

ClinicalTrials.gov does not report these quantities in the ClinicalTrials.gov record. The absence of a median time-to-event estimate here is therefore deliberate rather than an assumption about the trial's results.

15. Blinding and Randomization as Statistical Design Features

TDF2 was randomized and triple-masked. These features operate before and during outcome collection rather than being statistical tests themselves.

Randomization
Random assignment creates the basis for comparing outcomes between treatment groups without intentionally selecting treatment according to participant characteristics.
Parallel design
Each participant belongs to one randomized treatment group, allowing outcomes to be compared across the two arms over follow-up.
Triple masking
The registry identifies the trial as triple-masked. The ClinicalTrials.gov record does not specify the identities of the three masked parties.
Prevention purpose
The primary purpose was prevention, so the principal efficacy question concerns incident HIV infection during follow-up.

16. Multiplicity, Interim Analysis, and Bayesian Methods

The ClinicalTrials.gov record identifies two primary endpoints and four posted statistical analyses, but they do not report an interim-analysis procedure, an alpha-spending strategy, a multiplicity-adjustment procedure, or a Bayesian analysis.

Design topicWhat the ClinicalTrials.gov record supports
Primary endpoints2 registered primary endpoints
MultiplicityNo multiplicity-adjustment procedure is reported in the ClinicalTrials.gov record
Interim analysisNo interim-analysis method is reported in the ClinicalTrials.gov record
Bayesian methodsNo Bayesian method is reported
SuperiorityPrimary analyses are identified as superiority hypotheses

Because the ClinicalTrials.gov record does not describe an alpha-allocation strategy across the two primary endpoints, this page does not infer one. Likewise, the existence of multiple posted analyses should not automatically be interpreted as evidence of a formal multiplicity problem without knowing the prespecified testing hierarchy and error-control procedure.

17. Missing Data and Study Completion

The registry specifically notes that rates of study completion were lower than predicted because more participants withdrew from the study, mostly because of relocation or conflicting obligations.

Interpretation issue: withdrawal and incomplete follow-up can affect both the amount of information available and the assumptions required for time-to-event analyses. The ClinicalTrials.gov record does not describe an imputation method for the primary HIV-incidence analysis, so no specific missing-data or imputation procedure is attributed to the trial here.

For a time-to-event analysis, withdrawal does not necessarily mean that all information from a participant is discarded. A participant may contribute observed follow-up up to the point at which further outcome ascertainment is unavailable. The statistical consequences depend on the censoring mechanism and the assumptions underlying the analysis.

18. Secondary Analysis of Condom Use

The condom-use analysis illustrates an important distinction between a primary prevention outcome and a behavioral secondary endpoint. The HIV-incidence endpoint concerns acquisition of HIV, whereas the condom-use analysis models a behavioral outcome at 12 months.

FeatureHIV incidenceCondom-use analysis
RolePrimarySecondary
Time frameMonthly, for up to 3 years12 months
AnalysisCox proportional-hazards modelLogistic regression
Effect measureEfficacyOdds ratio
Estimate62.2%1.16
95% CI21.5%–83.4%1.05–1.28
P-value0.030.004

The different models reflect different questions. Cox regression uses event timing, while logistic regression models the odds of a binary outcome. The estimates therefore cannot be compared numerically as if they were the same statistical quantity.

19. Statistical Interpretation of the Main Results

Effect size

The principal HIV-incidence analysis reports 62.2% efficacy. This is the registry's reported effect measure for the comparison between TDF-FTC and placebo. It is not a hazard ratio or a direct percentage-point change in HIV incidence.

Precision

The 95% CI of 21.5%–83.4% shows that the estimated effect is subject to uncertainty. The width of the interval is important because the point estimate alone cannot communicate how precisely the treatment effect was estimated.

Statistical evidence

The HIV-incidence analysis reports P = 0.03. This is evidence against the null hypothesis under the stated testing framework, but it is not an effect-size measure.

Behavioral endpoint

The condom-use analysis reports OR 1.16 with a 95% CI of 1.05–1.28 and P = 0.004. The modeled outcome was reporting zero condomless sex acts, so interpretation must preserve that outcome definition.

Safety endpoint

The primary adverse-drug-reaction analysis reports P = 0.003 but no estimate or confidence interval in the ClinicalTrials.gov record. Therefore, the statistical evidence can be described, but the magnitude of the safety difference cannot be quantified from the posted analysis reported here.

20. Limitations

21. Why This Trial Matters Statistically

TDF2 is a useful teaching case because the trial connects randomized prevention research with several different statistical frameworks. The same study contains a time-to-event primary efficacy analysis, a time-to-event safety analysis, logistic regression for a behavioral endpoint, and Fisher's exact test for adherence.

ConceptHow it appears in TDF2
RandomizationRandomized parallel-group allocation to TDF-FTC or placebo.
BlindingTriple-masked trial.
Time-to-event analysisHIV incidence was analyzed with a Cox proportional-hazards model.
Cox modelUsed for the primary HIV-incidence and adverse-drug-reaction analyses.
Confidence intervalThe HIV-incidence efficacy estimate has a two-sided 95% CI of 21.5%–83.4%.
Prespecified boundaryThe registry describes a 10% lower boundary for the 95% CI.
Superiority testingThe primary analyses are identified as superiority hypotheses.
Intention-to-treatSafety analyses were performed in the ITT cohort.
Logistic regressionUsed to model the odds of reporting zero condomless sex acts at 12 months.
Odds ratioCondom-use analysis reported OR 1.16 with 95% CI 1.05–1.28.
Fisher exact testUsed for the 36-month adherence comparison.
GeneralizabilityThe registry cautions that findings may not be generalizable to other populations.

22. Related Tutorials

Learn more about the methods used in this trial:

23. Related Calculators

24. Important Statistical Takeaways

Different endpoints need different models

HIV incidence was analyzed with Cox regression, condom use with logistic regression, and adherence with Fisher's exact test.

Effect and evidence are different

The 62.2% efficacy estimate describes the reported effect, while P = 0.03 describes statistical evidence under the specified testing framework.

Confidence intervals add precision

The 21.5%–83.4% interval shows uncertainty around the HIV-incidence efficacy estimate and should be interpreted alongside the point estimate.

Analysis populations matter

The HIV-incidence and condom-use analyses use different populations, so their estimates should not be interpreted as though they came from identical sets of participants.

25. Sources

Continue through the Clinical Biostats statistical tutorials

Use the methods in TDF2 as a starting point for deeper study of survival analysis, regression, categorical-data methods, confidence intervals, and randomized clinical-trial design.

26. Record Summary

TDF2 provides a compact example of how a randomized prevention trial can require several complementary statistical methods. The primary HIV-incidence endpoint was evaluated with a Cox proportional-hazards model and reported an efficacy estimate of 62.2% with a 95% CI of 21.5%–83.4% and P = 0.03. The primary adverse-drug-reaction endpoint was also analyzed with Cox regression and reported P = 0.003, although the ClinicalTrials.gov record does not provide an effect estimate or confidence interval. Secondary analyses used logistic regression for condom use, yielding OR 1.16 with a 95% CI of 1.05–1.28 and P = 0.004, and Fisher's exact test for medication adherence, with P = 0.79.

The most important statistical lesson is that these numbers cannot be interpreted interchangeably. Efficacy, odds ratios, P-values, confidence intervals, and categorical-test results answer different questions. A careful analysis therefore keeps the endpoint definition, analysis population, model, effect measure, uncertainty, and prespecified hypothesis framework connected to one another.

Clinical Biostats methodology: A trial-results page should not merely reproduce reported numbers. The goal is to explain how the design and statistical methods determine what each result means, while keeping reported evidence separate from educational interpretation and avoiding unsupported reconstruction of results.