Docetaxel Versus Standard Care in STAMPEDE

Why the Docetaxel Comparison Matters

The STAMPEDE docetaxel comparison asked whether adding early docetaxel chemotherapy to standard hormone therapy improved outcomes for men beginning long-term treatment for advanced prostate cancer. Its importance is not just the numerical survival result: the interpretation depends on which patients were enrolled, what treatment was compared, and which survival measure was used.

The comparison was part of STAMPEDE’s multi-arm platform, whose original structure was described in the earlier material on STAMPEDE’s original six-arm structure. Here the focus is the docetaxel contrast, especially among participants whose disease was metastatic at entry. The result is evidence about randomized treatment strategies in that trial population; it is not automatically an estimate for every person with prostate cancer or every later treatment setting.

Treatment and Enrolled Population

Participants assigned to the docetaxel strategy received docetaxel in addition to standard care based on long-term androgen-deprivation therapy (ADT). The regimen used six cycles of docetaxel, given every three weeks, with prednisolone. The control strategy was standard care based on long-term ADT without docetaxel as part of the randomized comparison. Thus, the contrast concerns adding early docetaxel to hormone therapy, not docetaxel versus no treatment and not docetaxel as a stand-alone treatment.

STAMPEDE recruited men with high-risk locally advanced or metastatic prostate cancer who were starting long-term hormone therapy. The metastatic result therefore applies most directly to participants with metastatic, hormone-sensitive disease at trial entry. It should not be conflated with results for nonmetastatic disease, for cancer that has already become resistant to castration, or for patients receiving a different treatment sequence. Those distinctions affect both prognosis and the meaning of a treatment comparison.

The principal clinical endpoint considered here is OS. A death from any cause counts as an event, so the endpoint does not require investigators to decide whether a death was caused by prostate cancer or treatment. This is an important strength of OS, although follow-up must be sufficiently long and deaths must be ascertained adequately. Failure-free survival can provide earlier information about disease progression or treatment failure, but it is a different endpoint and does not replace the direct interpretation of OS.

What the Hazard Ratio Summarizes

A common model for comparing time to death is the Cox proportional-hazards model. For treatment indicator \(Z\), where \(Z=1\) denotes assignment to docetaxel plus standard care and \(Z=0\) denotes assignment to standard care, the model is

Here \(h(t\mid Z)\) is the hazard at time \(t\), \(h_0(t)\) is an unspecified baseline hazard, and \(\beta\) is the treatment coefficient. The treatment HR is \(\exp(\beta)\). The model’s proportional-hazards assumption says that the ratio of the two hazards is constant over time:

This equation justifies interpreting \(1-\mathrm{HR}\) as a relative reduction in instantaneous hazard when the model is appropriate. It does not turn the HR into a risk ratio, a difference in the probability of death by a specified time, or a difference in expected survival time. Those quantities require other summaries and, for absolute estimates, the survival curves or equivalent data.

Checking the Uncertainty on the Log Scale

Hazard ratios are positive and their confidence intervals are commonly calculated on the logarithmic scale, where the coefficient \(\beta=\log(\mathrm{HR})\) can be approximated by a normal distribution. If a 95% interval is formed as \(\hat{\beta}\pm1.96SE\), exponentiating its endpoints returns an interval for the HR. This provides a simple check on the scale and approximate uncertainty of a reported result.

For a reported HR \(r\) with interval \((L,U)\), an approximate standard error can be recovered as

This calculation is a check based on rounded published values, not a substitute for the original analysis. The exact model, stratification, and unrounded estimates may produce small differences.

Absolute Survival and Relative Effects

Relative measures such as the HR are useful for summarizing a comparison, but clinicians and patients often also want an absolute survival difference at a meaningful time point. That requires estimated survival probabilities at that time. The median alone does not reveal the absolute difference at, for example, three years, and it does not describe the full shape of either survival curve.

A further alternative is restricted mean survival time (RMST) up to a prespecified time \(\tau\). It is the area under the survival curve from time 0 to \(\tau\), and a difference in RMST is interpreted as a difference in mean time alive during that restricted period. RMST can be useful when hazards are not proportional, but it answers a different question from a constant HR and depends on the chosen \(\tau\). Median survival, fixed-time survival, HR, and RMST are complementary summaries rather than interchangeable versions of one number.

Assumptions, Diagnostics, and Subgroup Interpretation

Random assignment supports a causal comparison of the assigned treatment strategies in the eligible trial population, provided the analysis preserves that assignment contrast and follow-up is handled appropriately. A survival analysis also relies on adequate event-time measurement and censoring that is independent of the event time, at least conditional on the variables and model used. If loss to follow-up is related to prognosis in a way the analysis does not address, the estimated survival contrast may be biased.

The proportional-hazards assumption should be assessed rather than treated as automatic. Useful checks include plotting estimated log cumulative hazards or examining scaled Schoenfeld residuals for a time pattern. Crossing survival curves or a visibly changing separation can also signal that a single HR may conceal time-varying effects. These checks do not prove the assumption; they help identify departures that warrant additional summaries, such as survival probabilities at specified times or RMST.

The metastatic subgroup estimate is directly relevant to men with metastatic disease at entry, but subgroup findings need cautious interpretation. An estimate within one subgroup does not by itself establish that treatment works differently across disease strata. That claim requires an interaction assessment, preferably specified and interpreted with the trial design in mind. When effects may differ across strata, report stratum-specific estimates and an interaction assessment. A pooled HR is an overall contrast across the randomized strata whose precise interpretation depends on the analysis model; it is not generally a simple or fixed weighted average of the stratum-specific HRs.

The main practical failure modes are therefore interpretive as well as statistical:

• Calling a hazard reduction a risk reduction. The HR concerns instantaneous event rates among those still at risk, not the proportion who die over the whole study.
• Reading the median difference as extra life for every participant. Medians describe the 50% point of each survival distribution, not individual treatment gains.
• Ignoring the treatment context. The result compares docetaxel added to ADT with ADT-based standard care in the enrolled setting; it does not isolate a chemotherapy effect independent of that context.
• Generalizing beyond the population. The metastatic subgroup finding does not alone establish the same effect in nonmetastatic or castration-resistant disease, or under a different treatment sequence.
• Assuming the HR is constant without checking. If the hazards are not proportional, one reported HR may be an incomplete summary, so time-specific or RMST estimates may be needed.

A Practical Reading Sequence

When reading a survival result from a multi-arm trial, separate the treatment contrast, target population, endpoint, and effect measure before interpreting the p-value. This prevents a statistically clear result from being overstated as a universal claim.

Why the Distinctions Matter

The STAMPEDE docetaxel result is clinically informative because it links a randomized treatment strategy to an OS advantage in the metastatic hormone-sensitive population studied. Its interpretation is strongest when the numerical result remains attached to that population and treatment context. The HR summarizes a relative comparison; medians and fixed-time probabilities describe survival on other scales. None should be substituted for another without stating the change in question being answered.

A common mistake is to say, “Docetaxel reduced the chance of death by 24%, so each patient lived 24% longer.” The first clause incorrectly treats a hazard ratio as a probability reduction, and the second turns a group-level contrast into an individual guarantee. A better statement is: “In the metastatic subgroup, assignment to docetaxel plus standard care was associated with an estimated HR for death of 0.76 versus standard care, corresponding to a 24% lower hazard under the proportional-hazards model; the reported median OS was approximately 60 versus 45 months.” This wording keeps the population, comparator, model-based relative measure, and median summary distinct.