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
REDUCE was a randomized, parallel-group, quadruple-masked phase 3 prevention trial evaluating dutasteride against placebo in men at increased risk of prostate cancer. The registry reports three primary endpoint analyses based on biopsy-detectable prostate cancer over Years 1-4, together with a secondary analysis of mean change from baseline in testosterone at Month 48.
| Feature | REDUCE |
|---|---|
| Phase | Phase 3 |
| Condition | Neoplasms, Prostate |
| Primary purpose | Prevention |
| Design | Randomized, parallel-group, quadruple-masked |
| Enrollment | 8231 |
| Arms | 2 |
| Interventions | Dutasteride (drug); Placebo (drug) |
| Primary endpoints | 3 registered binary endpoints |
| Results | Posted |
| ClinicalTrials.gov | NCT00056407 |
| Lead sponsor | GlaxoSmithKline |
2. Clinical Question
The central statistical question was whether dutasteride reduced the occurrence of biopsy-detectable prostate cancer compared with placebo during the Years 1-4 assessment period in men enrolled in this prevention trial.
Population
Men enrolled in a phase 3 prevention study for prostate cancer in a population described by the registry as being at increased risk.
Intervention
Dutasteride 0.5 mg.
Comparator
Placebo.
Primary question
Does dutasteride reduce the occurrence of biopsy-detectable prostate cancer relative to placebo during the Years 1-4 assessment period?
3. Trial Design
Dutasteride 0.5 mg
- Dutasteride (drug)
- Included in the randomized comparison against placebo
- Primary efficacy analyses used the efficacy population
Placebo
- Placebo (drug)
- Included in the randomized comparison against dutasteride
- Primary efficacy analyses used the efficacy population
Randomization provides the basis for comparing the two intervention groups while reducing systematic allocation differences.
The registry identifies the trial as quadruple-masked, reducing the opportunity for knowledge of treatment assignment to influence trial conduct or assessment.
Participants were assigned to one of two treatment groups rather than crossing between treatments.
The trial was designed around prevention of prostate cancer rather than treatment of established cancer.
Biopsy ascertainment
The registered primary endpoints were based on study biopsies consisting of 10 biopsy samples (cores) obtained in a pre-defined pattern. Biopsies were read at the central pathology laboratory, which processed the majority, 94%, of biopsies. Biopsy cases positive for prostate cancer or precancerous lesions, including high-grade prostatic intraepithelial neoplasia (HGPIN) or typical small acinar proliferation (ASAP), and prostate surgeries were reviewed by the lead pathologist.
4. Primary Endpoints
The registry contains three related primary endpoint definitions. All use the same underlying clinical outcome—participants with biopsy-detectable prostate cancer—but define the risk set differently. That distinction is statistically important because changing which participants are considered at risk changes the denominator and therefore the estimand being summarized.
| Primary endpoint | Time frame | Risk-set definition |
|---|---|---|
| Number of Participants With Biopsy-detectable Prostate Cancer at Years 2 and 4 (Crude Rate Approach) | Years 1-2, Years 3-4, and Overall (Years 1-4) | All participants at risk at the beginning of each time period. |
| Number of Participants With Biopsy-detectable Prostate Cancer at Years 2 and 4 (Modified Crude Rate Approach) | Years 1-2, Years 3-4, and Overall (Years 1-4) | Participants who either were diagnosed with prostate cancer during the study or had an end-of-time-period biopsy. For each period, N is the number with at least 1 biopsy in the final 6 months or a positive biopsy during the period. |
| Number of Participants With Biopsy-detectable Prostate Cancer at Years 2 and 4 (Restricted Crude Rate Approach) | Years 1-2, Years 3-4, and Overall (Years 1-4) | Participants who had a biopsy during the specified time period. |
How the three approaches differ
The crude rate approach includes all participants at risk at the beginning of each period. The modified crude rate approach uses a more specific definition based on a positive biopsy or a biopsy near the end of the relevant period. The restricted crude rate approach limits the risk set to participants who had at least one biopsy during the relevant period.
These are not merely alternative labels. They address a central issue in prevention trials that rely on scheduled or available diagnostic procedures: the observed cancer rate depends partly on who is actually observed through biopsy. Comparing the three approaches therefore provides information about the stability of the estimated treatment effect under different definitions of the analysis population.
5. Primary Results
The registry reports formal statistical analyses for all three primary endpoints. Each analysis compares placebo with dutasteride 0.5 mg and reports a relative risk reduction for the overall Years 1-4 assessment.
5.1 Crude Rate Approach
Relative risk reduction
95% CI: 15.6%–30.3% · P < 0.0001
Overall assessment across Years 1-4.
A reported relative risk reduction of 23.3% indicates that, under the crude-rate analysis, the estimated relative occurrence of biopsy-detectable prostate cancer was lower with dutasteride than with placebo. The estimate is a relative measure; it does not state the absolute number of cancers prevented or the absolute difference in event probability.
The 95% confidence interval of 15.6% to 30.3% describes the precision of the estimated relative reduction under this analysis. It does not mean that there is a 95% probability that the true effect lies inside this particular interval. Rather, it is the interval produced by the specified statistical procedure under repeated-sampling reasoning.
The P-value of <0.0001 addresses the statistical evidence against the null hypothesis under the specified superiority analysis. It does not measure the magnitude or clinical importance of the effect. A P-value should therefore be interpreted together with the effect estimate and confidence interval.
The registry identifies the reported method as Mantel-Cox and normalizes this to a Cox proportional-hazards model. Because the effect measure reported here is relative risk reduction rather than a hazard ratio, the 23.3% estimate should not simply be relabeled as an HR. Interpretation also depends on the censoring and risk-set structure underlying the analysis.
5.2 Modified Crude Rate Approach
Relative risk reduction
95% CI: 15.5%–30.0% · P < 0.0001
Overall assessment across Years 1-4.
The modified crude-rate analysis gives a relative risk reduction of 23.1%, very close to the 23.3% estimate from the crude-rate approach. The important statistical point is not that the two estimates are numerically identical—they are not—but that the estimated treatment effect remains similar when the registry changes the definition of the risk set.
The 95% confidence interval extends from 15.5% to 30.0%. This gives a quantitative indication of the uncertainty around the 23.1% point estimate. The interval is narrower or wider only in relation to the particular information and analysis used; it should not be interpreted as a measure of biological variability among individual participants.
The P-value of <0.0001 provides evidence against the null hypothesis specified for this superiority analysis. It is not a measure of the probability that dutasteride works, nor does it establish the size of the effect. The confidence interval is more informative for the range of relative effects compatible with the analysis.
Because the modified crude-rate approach includes participants based on a positive diagnosis or an end-of-period biopsy, the interpretation is tied to that ascertainment rule. Missing or absent biopsy information can affect who enters the analyzed risk set, so the population definition is part of the estimand rather than a technical detail that can be ignored.
5.3 Restricted Crude Rate Approach
Relative risk reduction
95% CI: 15.2%–29.8% · P < 0.0001
Overall assessment across Years 1-4.
The restricted crude-rate analysis estimates a 22.8% relative risk reduction. Its 95% confidence interval, 15.2% to 29.8%, is similar to the intervals from the other two primary analyses. The consistency of the point estimates—23.3%, 23.1%, and 22.8%—shows that the reported relative effect changes only modestly across the three prespecified risk-set approaches.
The restricted approach is specifically based on participants who had at least one biopsy during the relevant period. Consequently, the analysis is conditional on biopsy observation in a way that the broad crude approach is not. This is important when interpreting the estimate: the statistic describes the treatment comparison under the registry's restricted ascertainment definition rather than an abstract cancer risk independent of observation.
Again, the P-value of <0.0001 should not be confused with effect size. The estimate and its confidence interval describe magnitude and precision, while the P-value addresses evidence against the null hypothesis under the specified testing framework.
| Primary analysis | Relative risk reduction | 95% CI | P-value |
|---|---|---|---|
| Crude rate | 23.3 | 15.6–30.3 | <0.0001 |
| Modified crude rate | 23.1 | 15.5–30.0 | <0.0001 |
| Restricted crude rate | 22.8 | 15.2–29.8 | <0.0001 |
What the three primary estimates show statistically
The three estimates form a narrow range from 22.8 to 23.3. That pattern is useful because the three analyses do not use exactly the same risk-set definition. The treatment effect therefore appears stable across the three registry-defined approaches to biopsy ascertainment.
At the same time, these estimates should not be interpreted as three independent replications of the same hypothesis. They are closely related analyses of the same randomized trial and the same underlying clinical outcome. Treating the three P-values as though they came from unrelated experiments would overstate the amount of independent evidence.
6. Secondary Endpoint Result: Testosterone
Mean Change From Baseline in Testosterone at Month 48
The registry reports a secondary endpoint assessing mean change from baseline in testosterone at Month 48. The analysis population was the efficacy population using last observation carried forward (LOCF), and the registry notes that the numbers presented were less than the overall population because data were not available for all participants.
Difference in adjusted means
95% CI: 17.3%–20.4% · P < 0.001
Baseline to Month 48.
| Measure | Result |
|---|---|
| Analysis population | Efficacy Population (LOCF) |
| Method reported | General linear model, t-test |
| Effect measure | Difference in adjusted means |
| Estimate | 18.8 |
| 95% CI | 17.3–20.4 |
| P-value | <0.001 |
| Adjusted means | Placebo: -6.1; Dutasteride: 12.7 |
The adjusted mean difference of 18.8 percentage points is the difference between the adjusted means of -6.1 in the placebo arm and 12.7 in the dutasteride arm. The direction and magnitude of this estimate therefore describe a separation in adjusted mean percentage change from baseline between the two groups at Month 48.
The 95% confidence interval of 17.3 to 20.4 indicates relatively narrow statistical uncertainty around the estimated adjusted mean difference. The P-value of <0.001 provides evidence against the null hypothesis used for the superiority comparison, but it does not itself describe the size of the mean difference.
The missing-data issue is especially important here because the registry explicitly states that data were unavailable for all participants and that the presented numbers were therefore smaller than the overall population. LOCF replaces a missing later observation with the participant's last available observation. That assumption can affect both the estimated mean and its uncertainty, particularly when dropout is related to treatment, time, or the underlying outcome.
7. Safety Results
The registry reports serious adverse events by treatment arm. These figures describe the number of affected participants relative to the number at risk in each arm.
| Safety measure | Placebo | Dutasteride 0.5 mg |
|---|---|---|
| Participants with serious adverse events | 837/4126 | 748/4105 |
Placebo
837 participants with serious adverse events among 4126 at risk.
Dutasteride 0.5 mg
748 participants with serious adverse events among 4105 at risk.
The serious-adverse-event figures should be interpreted separately from the efficacy analyses. They are counts of affected participants relative to the stated at-risk populations and are not the same estimand as the primary time-to-event comparison of biopsy-detectable prostate cancer.
8. Statistical Methodology
Cox proportional-hazards model / Mantel-Cox method
The registry reports Mantel-Cox as the method for all three primary analyses and normalizes the method to the Cox proportional-hazards model. This places the primary analyses within the survival-analysis family even though the registered endpoint itself is expressed as the number of participants with biopsy-detectable prostate cancer.
A Cox model evaluates treatment effects through the relative event rate over follow-up. In a conventional Cox interpretation, the proportional-hazards assumption means that the relative hazard between treatment groups is treated as approximately stable over time. The registry's reported effect measure for these analyses, however, is relative risk reduction, not a hazard ratio, so the numerical estimates should be reported exactly as the registry presents them.
A relative risk reduction describes proportional reduction in risk relative to the comparator. It is distinct from an absolute risk difference and should not automatically be interpreted as a hazard ratio.
Risk-set construction
The most important methodological feature of the primary endpoint analyses is the treatment of biopsy observation. The crude, modified crude, and restricted crude approaches use different definitions of who contributes to the risk set. This is consequential because prostate cancer can only be identified through the specified biopsy-based ascertainment process.
In the crude approach, participants are included according to their status at the beginning of each time period. The modified crude approach requires either a prostate cancer diagnosis during the study or an appropriate end-of-period biopsy. The restricted approach limits the population to participants who had at least one biopsy in the relevant period.
Central pathology review
Biopsy results were read at the central pathology laboratory, which processed 94% of biopsies. Cases positive for prostate cancer or precancerous lesions and prostate surgeries were reviewed by the lead pathologist. Central review is statistically relevant because outcome misclassification can affect the comparison between randomized treatment groups; standardized outcome ascertainment can reduce one source of differential assessment.
t-test and general linear model
The secondary testosterone analysis used a general linear model and t-test, with the statistical method listed as a t-test. This is appropriate for comparing a continuous mean change between two treatment groups when the analysis is framed around adjusted means.
The registry reports adjusted means of 12.7 and -6.1, respectively, yielding the reported adjusted mean difference of 18.8.
Missing data and LOCF
The testosterone analysis used LOCF and explicitly notes that the number of observations was less than the overall population because data were not available for all participants. Missing-data handling is therefore part of the definition of the reported secondary estimate.
LOCF is simple to implement, but it can be sensitive to the assumption that the last observed value is a reasonable representation of the later unobserved value. It should not be treated as automatically equivalent to a modern model-based missing-data analysis or multiple imputation strategy.
9. Statistical Methods Explained
Why were three primary analysis approaches used?
The registry defines three related approaches to biopsy-detectable prostate cancer. They differ in how participants enter the risk set. Presenting all three helps examine whether the estimated relative treatment effect is sensitive to the way biopsy observation is handled.
Why does the risk set matter?
A risk set identifies the participants who can contribute information to a particular analysis at a particular point in time. If the analysis requires a biopsy during a period, participants without the relevant biopsy may not contribute in the same way as participants who were directly observed. Changing the risk-set definition therefore changes the statistical question being answered.
What does a 23.3 relative risk reduction mean?
It means that the estimated relative occurrence of the analyzed outcome was 23.3% lower under the crude-rate analysis for dutasteride compared with placebo. It does not mean that 23.3% of participants were prevented from developing prostate cancer, and it does not provide an absolute risk reduction.
Why isn't the relative risk reduction the same thing as a hazard ratio?
The registry reports Mantel-Cox methodology, normalized as a Cox proportional-hazards model, but it separately identifies the effect measure as relative risk reduction. A hazard ratio and a relative risk reduction are related concepts in time-to-event analysis but are not interchangeable labels for the same numerical quantity.
Why is the confidence interval important when the P-value is very small?
The P-value addresses evidence against the null hypothesis, whereas the confidence interval describes the statistical precision of the estimated effect. For the crude analysis, the estimate is 23.3 with a 95% confidence interval of 15.6 to 30.3. The interval therefore conveys substantially more information about the plausible magnitude of the estimated effect than the P-value alone.
Why does LOCF matter for the testosterone result?
The registry states that not all participants had available testosterone data and that the analysis used LOCF. A missing Month 48 observation can therefore be represented by the participant's last observed value. Because this treatment of missingness affects the outcome values entering the analysis, it can affect both the estimated mean difference and its uncertainty.
Why should the three primary P-values not be treated as three independent confirmations?
The three analyses are highly related analyses of the same randomized population and the same biopsy-detectable prostate cancer outcome. Their similar estimates provide useful sensitivity information about risk-set definitions, but the analyses are not independent experiments. Statistical interpretation should therefore distinguish consistency across analysis definitions from independent replication.
10. Understanding the Primary Endpoint Across Time
| Assessment | Time frame | Purpose |
|---|---|---|
| Years 1-2 | Years 1-2 | Assess biopsy-detectable prostate cancer during the first defined period. |
| Years 3-4 | Years 3-4 | Assess biopsy-detectable prostate cancer during the subsequent defined period. |
| Overall | Years 1-4 | Provide the overall assessment reported by the registry. |
The registry provides the overall estimates for the three primary analyses. The corresponding endpoint definitions retain the Years 1-2 and Years 3-4 assessment periods, but the posted effect estimates shown here are explicitly identified as being for the overall assessment.
This distinction is useful when reading the results. An overall estimate summarizes the specified Years 1-4 assessment, whereas the separate time periods describe how participants were organized for endpoint ascertainment. One should not infer period-specific effect estimates from the overall estimates.
11. Interpreting Effect Size, Precision, and Statistical Evidence
Effect size
The primary relative risk reduction estimates range from 22.8 to 23.3 across the three approaches. These are relative measures, not absolute risk differences.
Precision
The 95% confidence intervals are 15.2–29.8, 15.5–30.0, and 15.6–30.3 for the three approaches, respectively.
Statistical evidence
Each primary analysis reports P < 0.0001. The secondary testosterone analysis reports P < 0.001.
Estimand
The exact interpretation depends on the endpoint definition and the risk-set approach used by each analysis.
A useful statistical reading separates three questions. First, what is the estimated effect? For the primary endpoint, the estimates are relative risk reductions of 23.3, 23.1, and 22.8. Second, how precisely is that effect estimated? The confidence intervals provide that information. Third, how much evidence is there against the null hypothesis? The P-values address that question.
These quantities answer different questions. A small P-value does not imply a large effect, and a confidence interval should not be replaced by the P-value when communicating uncertainty. For a clinical prevention trial, the absolute event rates would also be important for translating a relative effect into an absolute benefit, but the registry information used here does not provide the corresponding event counts for the primary analyses.
12. Analysis Populations and Generalizability
The crude-rate primary analysis uses an efficacy population defined as all randomized participants with a negative entry biopsy, as determined by the central pathology laboratory, who received at least 1 dose of study drug. The modified crude and restricted crude approaches further define their populations according to biopsy and prostate-cancer ascertainment during the specified periods.
| Analysis | Population definition |
|---|---|
| Crude rate | All randomized participants with a negative entry biopsy, as determined by the central pathology laboratory, who received at least 1 dose of study drug; all participants at risk at the beginning of each time period were included. |
| Modified crude rate | Participants who either were diagnosed with prostate cancer during the study or had an end-of-time-period biopsy, with the period-specific risk sets defined by the registry. |
| Restricted crude rate | Participants who had at least one biopsy during the relevant time period. |
| Testosterone | Efficacy Population using LOCF; the registry notes that data were unavailable for all participants. |
These distinctions are important when generalizing results. The primary analyses are not simply summaries of every enrolled participant regardless of eligibility for the relevant analysis. The entry-biopsy requirement and the subsequent biopsy-based definitions determine which participants contribute to each endpoint.
13. Limitations
Biopsy-dependent ascertainment
The primary outcome is biopsy-detectable prostate cancer. Consequently, ascertainment depends on participants having the relevant biopsy observations. The registry's use of three crude-rate definitions reflects this issue directly.
Different risk-set definitions answer slightly different questions
The three primary approaches are closely related but not identical. The crude, modified crude, and restricted approaches use different criteria for inclusion in the risk set. Their similarity is reassuring as a sensitivity feature, but the estimates should not be treated as mathematically interchangeable.
Relative effects do not provide absolute benefit
A relative risk reduction of 23.3 does not indicate the absolute difference in prostate cancer incidence between the two groups. Absolute measures would require corresponding event probabilities or counts under the relevant analysis definition.
Missing testosterone observations
The Month 48 testosterone analysis used LOCF and included fewer observations than the overall population because data were unavailable for all participants. The assumptions underlying LOCF can influence the estimated adjusted mean difference.
Multiplicity
Three primary endpoint analyses and a secondary endpoint are reported. Because the three primary analyses concern the same underlying outcome and population under different risk-set definitions, their P-values should not be interpreted as though they were three unrelated confirmatory tests. The ClinicalTrials.gov record does not specify an alpha-allocation strategy across these analyses.
Proportional-hazards considerations
The registry identifies the primary method as Mantel-Cox and normalizes it to a Cox proportional-hazards model. A Cox model relies on a proportional-hazards framework for its conventional hazard interpretation. The reported effect measure here is relative risk reduction, so the page does not substitute a hazard-ratio interpretation for the registry's stated measure.
Scope of reported results
The registry reports 26 outcome measures and 4 statistical analyses, including 3 primary-endpoint analyses. The analyses presented here focus on the posted primary analyses, the posted secondary testosterone analysis, and serious adverse events by arm.
14. Why This Trial Matters Statistically
REDUCE illustrates an important problem in clinical-trial biostatistics: the definition of an endpoint is inseparable from the definition of who can contribute information to that endpoint. When a disease is detected through scheduled diagnostic procedures, the statistical analysis must account for the observation process rather than treating every participant as equally observed at every time point.
The three primary approaches make this issue explicit. The crude-rate approach uses participants at risk at the beginning of each period. The modified approach incorporates positive biopsies or an end-of-period biopsy. The restricted approach requires a biopsy during the relevant period. Despite those differences, the reported overall relative risk reductions are closely grouped at 23.3, 23.1, and 22.8.
That pattern provides an instructive example of sensitivity analysis. A sensitivity analysis is most informative when it changes an important analytical assumption while retaining the same underlying scientific question. Here, the alternative risk-set definitions allow readers to see whether the estimated treatment effect depends strongly on how biopsy ascertainment is incorporated.
The trial also illustrates why effect measures should be named precisely. The statistical method is reported as Mantel-Cox and normalized as a Cox proportional-hazards model, while the effect measure is reported as relative risk reduction. Method and estimand are related, but they are not interchangeable. Careful statistical reporting preserves both pieces of information.
Finally, the secondary testosterone analysis demonstrates a different statistical problem. A continuous outcome can be summarized through adjusted means and a mean difference, but missing observations require an explicit strategy. Here, the registry reports LOCF and notes that data were unavailable for all participants. Thus, missing-data assumptions become part of the interpretation rather than an invisible preprocessing step.
15. Key Statistical Takeaways
| Question | What the registry reports | Statistical lesson |
|---|---|---|
| Was the primary effect consistent across approaches? | RRR 23.3, 23.1, and 22.8 | Similar estimates across alternative risk-set definitions indicate stability of the reported relative effect. |
| How precise were the primary estimates? | 95% CIs 15.6–30.3, 15.5–30.0, and 15.2–29.8 | Confidence intervals communicate uncertainty around the point estimates. |
| How strong was the statistical evidence? | P < 0.0001 for each primary analysis | P-values address evidence against a null hypothesis, not effect magnitude. |
| What was the secondary continuous outcome? | Adjusted mean difference 18.8 at Month 48 | Continuous outcomes require interpretation of mean differences and their assumptions separately from time-to-event outcomes. |
| How was missing testosterone data handled? | LOCF | Missing-data handling can influence an estimated treatment effect. |
16. Related Tutorials
Learn more about the methods used in this trial:
17. Related Calculators
Explore the statistical methods behind clinical trials
Clinical Biostats tutorials and statistical calculators provide practical explanations of study design, survival analysis, effect measures, missing data, and statistical inference.