← Clinical Trials
HIV-1 Infection Phase 3 Non-Inferiority NCT00524368

ODIN: Complete Statistical Analysis of Darunavir/Ritonavir in HIV-1 Infection

An independent statistical analysis of the randomized phase 3 ODIN trial comparing darunavir/ritonavir 800 mg/100 mg once daily with darunavir/ritonavir 600 mg/100 mg twice daily in early treatment-experienced HIV-1 infected patients.

Trial status: COMPLETED  ·  Enrollment: 590  ·  Primary completion: 2009-08
Scope of this record

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

Independent analysis notice: 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

ODIN was a randomized, parallel, open-label phase 3 trial comparing two dosing schedules of darunavir/ritonavir in early treatment-experienced HIV-1 infected patients. The registered primary endpoint was virological response at Week 48, defined using the Time to Loss of Virologic Response (TLOVR) algorithm. The posted primary analysis was a non-inferiority comparison based on the difference in response proportions.

590
Enrolled
Phase 3 trial
2
Treatment arms
Parallel design
0.0019
Primary risk difference
800/100 mg q.d. minus 600/100 mg b.i.d.
−0.054
95% CI lower bound
Two-sided 95% CI
FeatureODIN
Trial nameODIN
ClinicalTrials.gov identifierNCT00524368
PhasePhase 3
StatusCOMPLETED
Therapeutic areaInfectious Disease
ConditionHuman Immunodeficiency Virus - Type 1
AllocationRANDOMIZED
Design modelPARALLEL
MaskingNONE
Primary purposeTREATMENT
Enrollment590
Lead sponsorTibotec Pharmaceuticals, Ireland
Start2007-10
Primary completion2009-08

2. Clinical Question

The clinical question was whether darunavir/ritonavir 800 mg/100 mg once daily could provide virological response at Week 48 that was non-inferior to darunavir/ritonavir 600 mg/100 mg twice daily in early treatment-experienced HIV-1 infected patients.

Population

Early treatment-experienced patients infected with Human Immunodeficiency Virus - Type 1.

Intervention

Darunavir/ritonavir 800 mg/100 mg once daily.

Comparator

Darunavir/ritonavir 600 mg/100 mg twice daily.

Primary question

Is the difference in Week 48 virological response consistent with non-inferiority of the once-daily regimen, using a maximum allowable difference of 12%?

3. Trial Design

01
Randomize590 participants
02
Parallel armsTwo dosing regimens
03
Follow48-week endpoint
04
AssessVirological response
05
AnalyzeITT statistical comparison
ARM · 800/100 mg ONCE DAILY

DRV/Rtv 800/100 mg Once Daily

  • Darunavir 800 mg/ritonavir 100 mg
  • Once-daily administration
  • Compared with the twice-daily regimen using the Week 48 virological response endpoint
ARM · 600/100 mg TWICE DAILY

DRV/Rtv 600/100 mg Twice Daily

  • Darunavir 600 mg/ritonavir 100 mg
  • Twice-daily administration
  • Comparator regimen for the non-inferiority analysis

The registry describes the allocation as randomized and the design model as parallel. Masking was listed as none, so the study was open-label. The primary purpose was treatment.

4. Endpoints

EndpointTime frameTypeRegistered definition / analysis
Virological Response at Week 48 (Number of Participants With Plasma Viral Load Less Than 50 Copies/mL) 48 Weeks Time-to-event As defined by the Time to Loss of Virologic Response (TLOVR) algorithm. The registry states that response and loss of response needed to be confirmed at 2 consecutive visits; participants who permanently discontinued were considered nonresponders after discontinuation; participants with intermittent missing viral load values were considered responders if the preceding and succeeding visits indicated response.

The registered primary endpoint is unusual in an important statistical sense: although the outcome is ultimately summarized as the number of participants with plasma viral load below 50 copies/mL at Week 48, the registry classifies the endpoint as time-to-event because the TLOVR algorithm incorporates the timing and confirmation of virologic response and loss of response.

Registry definition note: The registry definition continues beyond the text reproduced in the trial data. This page does not reconstruct or supplement the omitted portion from another source.

5. Statistical Methodology

Primary analysis: logistic regression

The posted primary analysis used logistic regression in the intention-to-treat population. The model included treatment as a factor and baseline viral load (log10) as a covariate. The reported effect measure was the difference in proportion of response, normalized here as a risk difference.

Primary comparison
Risk difference = P(response | DRV/Rtv 800/100 mg q.d.) − P(response | DRV/Rtv 600/100 mg b.i.d.)

The reported estimate was 0.0019. The registry reports a two-sided 95% confidence interval of −0.054 to 0.092.

Non-inferiority framework

The registry prespecified a maximum allowable difference of 12%. Non-inferiority of the once-daily regimen versus the twice-daily regimen was concluded if the lower limit of the two-sided 95% confidence interval for the difference exceeded −12%.

This is the key statistical decision rule. The question is not whether the observed difference is exactly zero. Rather, the question is whether the data are sufficiently compatible with the once-daily regimen being worse by more than the prespecified clinically acceptable margin.

Design assumptions for sample size

The registry analysis notes a planning assumption of a 70% response rate at 48 weeks in both treatment groups. It states that 306 participants per treatment arm were required to establish non-inferiority with a maximum allowable difference of 12%, a 1-sided significance level of 0.025, and 90% power.

Analysis population

The primary analysis used the intention-to-treat population. Permanent discontinuations were considered nonresponders after discontinuation.

Covariate adjustment

The primary logistic regression model included baseline viral load on the log10 scale as a covariate.

Secondary analyses

The registry reports three principal statistical model families across the posted secondary analyses: logistic regression for binary virologic response, ANCOVA for continuous outcomes, and Cox proportional-hazards models for time-to-event outcomes.

Endpoint familyMethodEffect measureAdjustment noted in registry
Binary virologic responseLogistic regressionRisk differenceTreatment factor and baseline plasma viral load
Continuous viral-load outcomesANCOVAMean difference / difference in least square meansBaseline log10 plasma viral load
Time to first virologic responseCox proportional-hazards modelHazard ratioTreatment factor and baseline plasma viral load
Time to loss of virologic responseCox proportional-hazards modelHazard ratioBaseline log10 viral load
CD4+ cell-count changeANCOVAMean differenceBaseline CD4 count and baseline viral load
FAHI score changeANCOVAMean differenceBaseline log10 viral load and baseline FAHI score

6. Results: Primary Endpoint

The primary endpoint was virological response at Week 48, defined as the number of participants with plasma viral load less than 50 copies/mL according to the TLOVR algorithm. The posted analysis compared DRV/Rtv 800/100 mg once daily with DRV/Rtv 600/100 mg twice daily using logistic regression in the ITT population.

Difference in proportion of response

0.0019

95% CI: −0.054 to 0.092   ·   P < 0.001

Non-inferiority margin: −12%

Primary endpointEffect measureEstimate95% CIP-valueHypothesis
Virological Response at Week 48, plasma viral load <50 copies/mL, TLOVR Risk difference 0.0019 −0.054 to 0.092 <0.001 Non-inferiority or equivalence
Clinical Biostats interpretation

What the estimate means: The estimated difference in the proportion of responders was 0.0019, defined as the once-daily regimen minus the twice-daily regimen. On the risk-difference scale, this estimate is close to zero, indicating that the estimated response proportions were very similar under the fitted analysis.

What it does not mean: A risk difference of 0.0019 does not mean that the two regimens produced identical outcomes in every patient. It is an estimate of the difference between the group-level response proportions under the specified analysis.

What the confidence interval says: The 95% two-sided confidence interval extends from −0.054 to 0.092. Thus, the data are compatible with the once-daily regimen having a response proportion as much as 0.054 lower or 0.092 higher than the twice-daily regimen under this statistical framework.

Why the p-value is not the effect size: The reported P < 0.001 is not a measure of how large the treatment difference is. In a non-inferiority analysis, the inferential question is tied to the prespecified margin and confidence interval rather than simply asking whether the observed difference differs from zero.

Why non-inferiority is the central issue: The prespecified lower boundary was −12%. The lower confidence limit of −0.054 is above −0.12, which satisfies the registry's stated criterion for concluding non-inferiority.

Analysis-population caution: The registry identifies the primary analysis population as ITT and specifies how permanent discontinuations and intermittent missing viral-load measurements were handled. Those rules are part of the estimand being reported and should not be silently replaced with a different missing-data strategy.

7. Results: Secondary Virologic Response at Week 48

Viral Load Less Than 400 Copies/mL

Difference in proportion of response

0.007

95% CI: −0.060 to 0.075   ·   P < 0.001

Non-inferiority margin: −12%

EndpointMethodEstimate95% CIP-valueHypothesis
Virologic Response at Week 48, viral load <400 copies/mL Logistic regression; risk difference 0.007 −0.060 to 0.075 <0.001 Non-inferiority or equivalence
Clinical Biostats interpretation

The estimated response difference was 0.007, again calculated as once daily minus twice daily. The 95% CI ranged from −0.060 to 0.075. Because the lower limit is above the prespecified −0.12 non-inferiority boundary, the registry's stated non-inferiority criterion is satisfied for this endpoint as well.

The confidence interval is more informative than the point estimate alone: it describes the statistical uncertainty around the treatment difference and shows that the interval remains entirely above the non-inferiority margin. The P < 0.001 should not be interpreted as a measure of the size of the difference.

8. Results: Continuous Viral-Load Outcomes

Change in log10 Viral Load From Baseline at Week 48

Difference between least square means

−0.003

95% CI: −0.188 to 0.182   ·   P = 0.977

The analysis used ANCOVA with treatment as a factor and baseline log10 plasma viral load as a covariate. The effect is the difference in least square means between the once-daily and twice-daily groups at Week 48.

Clinical Biostats interpretation

The estimated adjusted mean difference was −0.003 log10 copies/mL, which is very close to zero. The 95% CI of −0.188 to 0.182 describes the uncertainty around that adjusted mean difference.

The P = 0.977 does not establish that the two regimens are exactly equivalent. It indicates that this superiority analysis did not provide evidence of a difference from zero under the specified ANCOVA model. The endpoint's formal hypothesis type was recorded as superiority, which is distinct from the non-inferiority hypothesis used for the primary virological-response endpoint.

Time-averaged Difference of log10 Plasma Viral Load Over 48 Weeks

Difference between least square means

−0.03

95% CI: −0.169 to 0.115   ·   P = 0.711

This ANCOVA analysis used observed cases and included baseline log10 viral load as a covariate. The reported effect was the difference in least square means between the two treatment groups.

Clinical Biostats interpretation

The estimate of −0.03 indicates a small adjusted difference favoring the once-daily group on the reported direction of the outcome, but the confidence interval extends from −0.169 to 0.115, crossing zero. The P = 0.711 therefore does not provide evidence of superiority under the reported analysis.

Because this endpoint used observed cases, its interpretation also depends on which observations remained available for analysis. It should not be treated as interchangeable with the primary TLOVR-based endpoint.

9. Results: Time-to-Event Outcomes

Time to Reach First Virologic Response

Hazard ratio

0.990

95% CI: 0.824 to 1.191   ·   P = 0.917

The analysis used a Cox proportional-hazards model including treatment as a fixed factor and baseline plasma viral load as a covariate.

Clinical Biostats interpretation

A hazard ratio of 0.990 means that the estimated instantaneous rate of reaching first virologic response was approximately 0.990 times that in the twice-daily comparator under the fitted model. Equivalently, the point estimate is very close to 1.

The 95% CI of 0.824 to 1.191 spans 1, and P = 0.917 does not support a superiority claim. The confidence interval, rather than the point estimate alone, shows the range of relative effects that remain statistically compatible with the data under the model.

As with any Cox-model result, the hazard ratio is not a direct probability of response and should not automatically be translated into an absolute difference in response. Its interpretation also relies on the proportional-hazards framework used by the model.

Time to Loss of Virologic Response

Hazard ratio

0.945

95% CI: 0.699 to 1.279   ·   P = 0.716

The Cox analysis included baseline log10 viral load as a covariate. The effect measure was the hazard ratio for time to loss of virologic response.

Clinical Biostats interpretation

The hazard ratio of 0.945 is close to 1. The 95% CI extends from 0.699 to 1.279, indicating substantial uncertainty around the estimated relative hazard. The P = 0.716 does not provide evidence of superiority in this secondary analysis.

A hazard ratio near 1 should not be read as proof that the treatment effects are identical. The appropriate interpretation is that the observed data did not demonstrate a superiority difference under the specified Cox model.

10. Results: CD4+ Cell Count and Functional Assessment

Change in CD4+ Cell Count From Baseline

Difference between least square means

−5.95

95% CI: −26.09 to 14.20   ·   P = 0.562

The analysis used ANCOVA in the ITT population with last observation carried forward (LOCF). The model included treatment as a factor and baseline CD4 count and baseline log10 viral load as covariates.

Clinical Biostats interpretation

The estimated adjusted difference in CD4+ cell-count change was −5.95 in the reported unit of 106/l, comparing once daily with twice daily. The 95% CI ranged from −26.09 to 14.20, crossing zero, with P = 0.562.

Because LOCF was used to impute intermittent missing values and values missing because of premature discontinuation, the result is conditional on that imputation strategy. LOCF is a missing-data method; it does not make the missing observations genuinely observed.

Change From Baseline in Total Functional Assessment of HIV Infection (FAHI) Score

Difference between least square means

0.55

95% CI: −2.97 to 4.06   ·   P = 0.761

This ANCOVA used the ITT population with LOCF. The model included treatment and baseline log10 viral load and baseline FAHI score as covariates.

Clinical Biostats interpretation

The adjusted mean difference was 0.55 points on the reported FAHI score scale, with a 95% CI from −2.97 to 4.06. The interval includes zero and P = 0.761 does not provide evidence of superiority under the reported model.

Again, the confidence interval is more informative than the P-value alone because it communicates the precision of the estimated difference. The result should also be interpreted in light of the LOCF rule used for missing observations.

11. Complete Posted Statistical Results

EndpointMethodEffectEstimate95% CIP-value
Virological Response at Week 48, viral load <50 copies/mL, TLOVR Logistic regression Risk difference 0.0019 −0.054 to 0.092 <0.001
Virologic Response at Week 48, viral load <400 copies/mL Logistic regression Risk difference 0.007 −0.060 to 0.075 <0.001
Change in log10 Viral Load From Baseline at Week 48 ANCOVA Mean difference −0.003 −0.188 to 0.182 0.977
Time to Reach First Virologic Response Cox proportional-hazards model Hazard ratio 0.990 0.824 to 1.191 0.917
Time to Loss of Virologic Response Cox proportional-hazards model Hazard ratio 0.945 0.699 to 1.279 0.716
Time-averaged Difference of log10 Plasma Viral Load Over 48 Weeks ANCOVA Mean difference −0.03 −0.169 to 0.115 0.711
Change in CD4+ Cell Count From Baseline ANCOVA Mean difference −5.95 −26.09 to 14.20 0.562
Change From Baseline in Total FAHI Score ANCOVA Mean difference 0.55 −2.97 to 4.06 0.761

The ClinicalTrials.gov record reports eight statistical analyses: one primary analysis and seven secondary analyses. The table above preserves the reported estimates, confidence limits, and P-values without recomputing or rounding them.

12. Serious Adverse Events

The ClinicalTrials.gov record reports serious adverse events by treatment arm as affected participants divided by participants at risk.

Treatment armSerious adverse eventsAffected / at risk
DRV/Rtv 800/100 mg Once Daily 16 participants 16/294
DRV/Rtv 600/100 mg Twice Daily 27 participants 27/296
Serious adverse events: affected participants
800/100 mg once daily
16
600/100 mg twice daily
27

The serious-adverse-event counts should be interpreted descriptively. The ClinicalTrials.gov record does not provide a formal statistical comparison of serious adverse events, confidence interval, or P-value, so this page does not construct one.

13. Non-Inferiority: Why the Margin Matters

Non-inferiority trials answer a different question from conventional superiority trials. The primary ODIN analysis did not require the estimated risk difference to be positive or even exactly zero. Instead, it asked whether the once-daily regimen could be ruled out as being worse than the twice-daily regimen by more than the prespecified 12% margin.

ODIN decision rule
Non-inferiority if lower 95% CI > −0.12

Primary estimate = 0.0019; lower 95% CI = −0.054. Because −0.054 is greater than −0.12, the registry's stated non-inferiority criterion is met.

The margin is therefore central to interpretation. A confidence interval can cross zero and still support non-inferiority, because the relevant question is whether the lower bound crosses the non-inferiority boundary. Conversely, a result that is statistically different from zero would not automatically establish non-inferiority unless the confidence interval also satisfied the prespecified margin criterion.

What the margin represents

The −12% value is the maximum allowable loss in the response proportion specified by the trial's non-inferiority framework.

What the margin does not represent

It is not the observed treatment difference. The observed estimate was 0.0019, while −12% is the prespecified boundary used for the decision.

14. Statistical Methods Explained

Why was logistic regression used for the primary endpoint?

The primary outcome was summarized as virological response status at Week 48 under the TLOVR algorithm. Logistic regression is appropriate for modeling a binary response while allowing treatment to enter as a factor and baseline viral load to be incorporated as a covariate. ODIN's registry analysis specifically reports logistic regression and baseline viral-load adjustment.

Why was the effect reported as a risk difference?

The trial's reported effect measure was the difference in proportion of response, normalized as a risk difference. This makes the non-inferiority margin directly interpretable on the same absolute scale: the estimated difference and the −12% margin are both expressed as differences in response proportions.

Why was baseline viral load included as a covariate?

Covariate adjustment can account statistically for baseline variation that is prognostically relevant to the outcome. In ODIN, baseline viral load on the log10 scale was included in the primary logistic regression model and in several secondary models. The adjusted estimate should therefore not be interpreted as a simple unadjusted difference between raw response proportions.

Why is the confidence interval central to the non-inferiority conclusion?

Because the trial's decision rule is based on the lower confidence limit relative to the non-inferiority margin. The primary 95% CI was −0.054 to 0.092. Its lower bound is above −0.12, which is the criterion stated in the registry for concluding non-inferiority.

What does a hazard ratio of 0.990 mean?

For time to first virologic response, a hazard ratio of 0.990 means that the fitted Cox model estimated an instantaneous event rate approximately 0.990 times that of the comparator group. It does not mean that 99.0% of participants responded, nor does it directly provide an absolute probability of response by a particular time.

Why was ANCOVA used for viral load, CD4+ count, and FAHI outcomes?

These endpoints were continuous measures, so ANCOVA provides a framework for comparing adjusted mean outcomes between treatment groups while incorporating baseline covariates. ODIN used baseline viral load for viral-load outcomes, baseline CD4 count and viral load for CD4+ change, and baseline viral load and baseline FAHI score for FAHI change.

What does LOCF mean for the CD4+ and FAHI analyses?

Last observation carried forward means that the last available observation was used when an observation was missing, including values missing because of premature discontinuation in the reported ITT analyses. It is an imputation rule, not a recovery of the unobserved value. Consequently, the statistical result is conditional on that missing-data strategy.

15. Intention-to-Treat Analysis and Missing Data

The primary and secondary analyses were generally described in the ClinicalTrials.gov record as using the intention-to-treat population. For the primary TLOVR endpoint, participants who permanently discontinued were considered nonresponders after discontinuation. Participants with intermittent missing viral-load values were considered responders if the preceding and succeeding visits indicated response.

Endpoint / analysisMissing-data rule reported
Primary TLOVR virological response Permanent discontinuation treated as nonresponse after discontinuation; intermittent missing viral-load values could be considered responsive when preceding and succeeding visits indicated response.
Virologic response <400 copies/mL Same ITT framework described in the ClinicalTrials.gov record.
CD4+ cell-count change LOCF for intermittent missing values and values missing because of premature discontinuation.
FAHI score change LOCF for intermittent missing values and values missing because of premature discontinuation.
Time-averaged viral-load difference Observed cases.

These are materially different strategies. A TLOVR endpoint incorporates response and loss-of-response rules, LOCF substitutes the last observed value, and an observed-case analysis uses available observations without the LOCF substitution described for the CD4+ and FAHI endpoints. Results across these analyses therefore should not be treated as if they were based on an identical missing-data estimand.

16. Cox Proportional-Hazards Models

ODIN used Cox proportional-hazards models for time to first virologic response and time to loss of virologic response. Both analyses incorporated baseline viral-load information as a covariate.

Hazard-ratio framework
HR = estimated instantaneous event rate in once-daily group ÷ estimated instantaneous event rate in twice-daily group

An HR below 1 indicates a lower estimated instantaneous event rate for the once-daily group under the fitted model; an HR above 1 indicates a higher estimated rate.

The proportional-hazards assumption is a model assumption underlying the usual interpretation of a single Cox hazard ratio. The ClinicalTrials.gov record does not report a formal proportional-hazards diagnostic, so this page does not claim that the assumption was verified.

Time-to-event endpointHR95% CIP-value
Time to Reach First Virologic Response0.9900.824 to 1.1910.917
Time to Loss of Virologic Response0.9450.699 to 1.2790.716

17. ANCOVA and Covariate Adjustment

ANCOVA appeared in the analyses of change in log10 viral load, time-averaged viral load, CD4+ cell count, and FAHI score. In each case, the treatment comparison was expressed as a difference in least square means.

OutcomeBaseline covariatesReported estimate95% CI
Change in log10 viral load Baseline log10 plasma viral load −0.003 −0.188 to 0.182
Time-averaged viral-load difference Baseline log10 plasma viral load −0.03 −0.169 to 0.115
Change in CD4+ cell count Baseline CD4 count and baseline log10 viral load −5.95 −26.09 to 14.20
Change in total FAHI score Baseline log10 viral load and baseline FAHI score 0.55 −2.97 to 4.06

The use of least square means means that the reported difference is model-adjusted rather than simply the difference between two raw arithmetic means. This is particularly relevant when baseline covariates are included in the model.

18. Non-Inferiority Versus Superiority in the Posted Analyses

The ClinicalTrials.gov record identifies different hypothesis types for different endpoints. The primary virological-response analysis and the secondary response analysis use a non-inferiority or equivalence framework. Several other endpoints are identified as superiority analyses.

EndpointHypothesis typePrimary interpretive question
Virological Response <50 copies/mLNon-inferiority or equivalenceDoes the lower confidence bound remain above the −12% margin?
Virologic Response <400 copies/mLNon-inferiority or equivalenceDoes the lower confidence bound remain above the −12% margin?
Change in log10 viral loadSuperiorityIs the adjusted mean difference distinguishable from zero?
Time to first virologic responseSuperiorityIs the hazard of reaching response different between groups?
Time to loss of virologic responseSuperiorityIs the hazard of loss of response different between groups?
Time-averaged viral loadSuperiorityIs the adjusted mean difference different from zero?
CD4+ cell-count changeSuperiorityIs the adjusted mean difference different from zero?
FAHI score changeSuperiorityIs the adjusted mean difference different from zero?

This distinction is important. A non-inferiority result and a nonsignificant superiority result are not interchangeable conclusions. They answer different statistical questions and use different decision criteria.

19. P-values and Confidence Intervals

The ODIN results illustrate why P-values and confidence intervals should be read together but not treated as synonyms.

P-value

A P-value addresses the compatibility of the observed data with a specified null hypothesis under the statistical model. It is not the magnitude of an effect.

Confidence interval

A confidence interval communicates uncertainty around the estimated effect and is particularly important here because the primary non-inferiority decision depends on its lower boundary.

For the primary endpoint, P < 0.001 accompanies an estimated risk difference of 0.0019 and a 95% CI of −0.054 to 0.092. The P-value should not be read as evidence that the treatment effect is large. The non-inferiority conclusion comes from comparing the lower confidence bound with the prespecified −0.12 margin.

20. Trial Timeline

2007-10 · Trial start

ODIN begins

The registered study start was October 2007.

2009-08 · Primary completion

Primary study completion

The registered primary completion date was August 2009.

Completed · Results posted

Statistical results available

The registry reports posted outcome measures and statistical analyses, including one primary-endpoint analysis and seven secondary analyses.

21. Limitations

22. Why This Trial Matters Statistically

ODIN is a useful teaching example because the primary question is fundamentally different from a conventional superiority trial. The central analysis is a non-inferiority comparison of response proportions, while the secondary analyses span binary, continuous, and time-to-event outcomes.

ConceptHow it appears in ODIN
RandomizationThe trial used randomized allocation.
Parallel designThe registry identifies the design model as PARALLEL.
Non-inferiorityThe primary endpoint used a −12% maximum allowable difference.
Confidence intervalsThe primary decision rule used the lower limit of a two-sided 95% confidence interval.
Logistic regressionUsed for the primary virological-response analysis and the <400 copies/mL secondary response analysis.
Risk differenceThe primary treatment effect was reported as a difference in response proportions.
ANCOVAUsed for continuous viral-load, CD4+, and FAHI outcomes.
Covariate adjustmentBaseline viral load and other baseline measures were incorporated into specified models.
Cox proportional-hazards modelUsed for time to first response and time to loss of response.
Intention-to-treatThe primary and secondary analyses were based on ITT populations as specified in the registry data.
Missing-data handlingTLOVR rules, LOCF, and observed-case analysis were used for different endpoints.

The statistical lesson is that the same randomized comparison can generate very different estimands. A risk difference in response, an adjusted mean difference, and a hazard ratio are not alternative ways of reporting one identical quantity. Each describes a different aspect of the treatment comparison.

23. A Statistical Reading of the Primary Result

Putting the primary analysis together

The primary risk-difference estimate was 0.0019, with a two-sided 95% CI of −0.054 to 0.092. The prespecified non-inferiority margin was −0.12. Because the lower confidence limit of −0.054 is above −0.12, the registry's stated criterion for non-inferiority is satisfied.

The important point is that this conclusion does not require the confidence interval to exclude zero. The interval may include both modest harm and modest benefit relative to the comparator while still excluding the possibility of harm greater than the prespecified non-inferiority margin.

The result therefore illustrates a central distinction between superiority and non-inferiority testing: the inferential boundary is determined by the clinical margin, not simply by zero.

24. Related Tutorials

Learn more about the methods used in this trial:

25. Related Statistical Calculators

26. Sources

Continue through Clinical Biostats

Explore the statistical methods and calculators connected with randomized clinical-trial analysis.

27. Record Summary

ODIN provides a compact example of how clinical-trial statistics change with the scientific question. Its primary Week 48 virological-response analysis used logistic regression with baseline viral-load adjustment and evaluated a risk difference against a prespecified −12% non-inferiority margin. The reported estimate was 0.0019, with a two-sided 95% CI of −0.054 to 0.092 and P < 0.001; the lower confidence bound was above the stated non-inferiority boundary.

The secondary analyses then used ANCOVA for continuous outcomes and Cox proportional-hazards models for time-to-event outcomes. These analyses illustrate why effect measures must be interpreted according to their statistical scale: a risk difference, adjusted mean difference, and hazard ratio answer different questions. The registry's missing-data rules also differ across endpoints, making it important to interpret each result in the context of its own analysis population and data-handling strategy.

Clinical Biostats methodology: A trial-results page should distinguish the reported numerical evidence from statistical explanation. For ODIN, the central teaching point is the non-inferiority framework: the conclusion depends on whether the confidence interval remains above the prespecified margin, not simply on whether a P-value is below a conventional superiority threshold.