From the Clinical Question to an Arm-Level Finding
After nephrectomy for renal cancer, a patient may have no detectable disease but still face a risk of recurrence. RAMPART’s disease-free survival (DFS) question concerns whether assigned post-operative treatment strategies change the time until recurrence or death. The endpoint therefore gives clinical meaning to an eventual comparison between trial arms: it is not simply a comparison of tumour response in patients with measurable disease.
The three strategies are active surveillance, adjuvant durvalumab, and adjuvant durvalumab plus tremelimumab. Durvalumab is an immune checkpoint inhibitor targeting PD-L1; tremelimumab targets CTLA-4. A comparison of the two treatment arms evaluates the assigned combination strategy against durvalumab alone. It does not, by itself, isolate the effect of tremelimumab in every possible treatment setting.
A DFS result is specific to this composite endpoint, its event definitions, the population randomized after nephrectomy, and the strategies assigned. A difference in DFS is not automatically a difference in overall survival, a reduction in cancer-specific mortality, or evidence that treatment cures more patients. Recurrence and death are clinically important, but the composite does not say which component produced a difference unless component results are also examined.
What Each RAMPART Comparison Means
Randomization makes the assigned arms comparable on average at baseline. Under appropriate follow-up and analysis, an arm-level contrast can therefore estimate the effect of assignment to one strategy rather than another for the trial-eligible population. The primary interpretation is usually by assigned group, regardless of whether every participant completed the intended treatment. This is an intention-to-treat interpretation: it answers a treatment-policy question about assignment, not the biological effect among only those who fully adhered.
| Contrast | Clinical question | What it does not isolate |
|---|---|---|
| Durvalumab versus active surveillance | Does assignment to adjuvant durvalumab change DFS compared with surveillance after nephrectomy? | The effect of drug exposure among only patients who complete treatment. |
| Durvalumab plus tremelimumab versus active surveillance | Does assignment to the combination strategy change DFS compared with surveillance? | The effect of either drug separately from the combination strategy. |
| Durvalumab plus tremelimumab versus durvalumab | Does adding tremelimumab within the assigned combination strategy change DFS compared with durvalumab alone? | A universal, context-free effect of tremelimumab in other populations or regimens. |
Active surveillance is a management strategy, not an absence of care. It can include follow-up visits, imaging, and treatment if disease recurs, according to the protocol and clinical practice. Thus, treatment-versus-surveillance contrasts compare strategies with different immediate treatment exposure and monitoring implications. Their clinical value includes benefits and harms beyond the DFS endpoint; DFS alone cannot summarize toxicity, quality of life, or treatment burden.
Reading a Hazard Ratio in Clinical Context
A common time-to-event summary is the hazard ratio (HR). The hazard at time \(t\) is the instantaneous event rate among people who have remained event-free up to \(t\). Under the proportional hazards assumption, the ratio of the treatment-arm hazard to the comparator-arm hazard is constant over time. An HR below 1 indicates a lower estimated event hazard in the treatment arm during follow-up, under this model.
Worked Example: Interpreting a Hypothetical Hazard Ratio
Suppose, purely for illustration, a hypothetical analysis of durvalumab versus active surveillance reports an HR of 0.78 with a 95% confidence interval from 0.61 to 0.99. Assume the endpoint is protocol-defined DFS and the proportional hazards model is reasonable.
The relative hazard difference is \(1-0.78=0.22\), or 22% lower estimated event hazard with assignment to durvalumab than with assignment to surveillance. The interval corresponds to hazard reductions from \(1-0.99=0.01\), or 1%, to \(1-0.61=0.39\), or 39%. Because the interval narrowly excludes 1, it is compatible with a statistically significant difference at the conventional two-sided 5% level, assuming this is the relevant prespecified comparison and no multiplicity adjustment changes the decision.
This result would not mean that 22% more patients remain disease-free, nor would it prove an overall-survival benefit. The HR summarizes relative event rates over follow-up under a model; absolute DFS probabilities at clinically relevant times are needed to show the size of the difference in probability terms.
The proportional hazards assumption matters to the single-number interpretation. If hazards change relative to one another over time, the reported HR may be a weighted summary rather than a constant treatment effect. Kaplan–Meier curves can reveal obvious crossings or changing separation, while model-based checks such as scaled Schoenfeld residuals can assess departures from proportionality. These checks are informative rather than proof that the assumption holds; clinical interpretation should also consider whether early and late effects plausibly differ.
From Relative Effects to Absolute DFS
Absolute survival probabilities are often easier to relate to clinical decisions. Let \(S_0(t)\) denote the probability of remaining free of the defined event through time \(t\) in the comparator arm, and let \(S_1(t)\) be that probability in the treatment arm. Under proportional hazards with constant HR \(\theta\), the treatment-arm survival function is \(S_1(t)=S_0(t)^\theta\). This relationship requires the proportional hazards model; it should not be used as a general conversion from an HR to a risk difference.
Assume \(h_1(t)=\theta h_0(t)\) throughout the time interval of interest.
The cumulative hazard is \(H(t)=\int_0^t h(u)\,du\), and survival is \(S(t)=\exp[-H(t)]\).
Since \(H_1(t)=\theta H_0(t)\), \(S_1(t)=\exp[-\theta H_0(t)]=\{\exp[-H_0(t)]\}^{\theta}=S_0(t)^\theta\).
Worked Example: A Hypothetical Absolute DFS Difference
Suppose at a specified follow-up time the hypothetical surveillance DFS probability is \(S_0(t)=0.80\), and a hypothetical treatment-versus-surveillance HR is \(\theta=0.80\). Under proportional hazards, the estimated treatment DFS probability is
The estimated absolute difference is \(0.8365-0.8000=0.0365\), or about 3.65 percentage points at that time. This is an illustration based on the stated model and hypothetical values, not a RAMPART result. A confidence interval for the HR alone does not fully communicate uncertainty in this absolute difference; uncertainty in the survival estimates and their relationship must also be considered.
The formula shows why an HR does not have one fixed absolute benefit for every baseline risk. The same relative hazard ratio can imply different absolute DFS differences when comparator survival differs. If proportional hazards is doubtful, report arm-specific survival estimates at meaningful times and consider a time-specific contrast or another prespecified summary rather than treating this conversion as valid.
Comparing All Three Arms Without Overclaiming
Worked Example: A Hypothetical Three-Arm Summary
Imagine a hypothetical report gives estimated three-year DFS probabilities of 70% for active surveillance, 78% for durvalumab, and 81% for durvalumab plus tremelimumab. These figures are invented for teaching and are not results from RAMPART.
The estimated absolute difference for durvalumab versus surveillance is \(78\%-70\%=8\) percentage points. For the combination versus surveillance it is \(81\%-70\%=11\) percentage points. The combination-versus-durvalumab difference is \(81\%-78\%=3\) percentage points.
These point estimates alone do not establish that any contrast is statistically reliable or clinically worthwhile. Confidence intervals, the prespecified analysis, follow-up completeness, multiplicity considerations, and adverse effects matter. In particular, the three-point difference does not prove that adding tremelimumab improves DFS: that conclusion depends on the direct randomized contrast and its uncertainty, not on comparing each arm separately with surveillance.
A common mistake is to choose the arm with the largest observed DFS estimate and call it the best treatment. Random variation can produce that ordering even when the strategies do not differ meaningfully. Another mistake is to infer a treatment-versus-treatment difference because one treatment comparison with surveillance is “significant” while another is “not significant.” The difference between significance decisions is not itself a test of the difference between treatments. The direct comparison, where appropriate, addresses that question.
Assumptions, Diagnostics, and Limits
A sound interpretation depends on more than the estimated HR or survival percentage. Randomization supports comparability at assignment, but follow-up and measurement must still be adequate. Censoring should be independent of the event time, at least conditional on variables used in the analysis; otherwise, the observed participants remaining under follow-up may not represent those censored. If loss to follow-up differs across arms or is related to prognosis, estimates can be biased.
- Check the endpoint. Confirm the time origin, recurrence criteria, death definition, and censoring rules. A change in any of these can change what the DFS estimate means.
- Inspect follow-up. Compare follow-up duration and censoring patterns across arms, and consider whether event ascertainment was similarly complete.
- Assess model fit. Examine survival curves and proportional-hazards diagnostics before presenting one HR as a constant relative effect.
- Keep the assigned strategy in view. An intention-to-treat comparison concerns assignment, including treatment discontinuation and subsequent care as they occur under the trial strategy.
- Separate endpoints. DFS findings do not automatically establish better overall survival, fewer treatment harms, or improved quality of life.
A further limitation is generalizability. The randomized contrast applies most directly to patients who met the trial’s eligibility criteria and received care in settings represented by the trial. It need not predict the same effect in people with different disease risk, tumour characteristics, comorbidities, or access to follow-up. Finally, a composite endpoint can be useful for an adjuvant question, but its overall result may be driven by recurrence, death, or both; interpretation should not claim a component-specific effect without evidence for that component.
Common Mistakes
- “An HR of 0.80 means 20% more patients are disease-free.” It means a 20% lower estimated hazard under proportional hazards. Use survival probabilities to describe absolute DFS differences.
- “Surveillance means no care.” It is an assigned follow-up strategy and may include monitoring and treatment after recurrence.
- “The combination result is the effect of tremelimumab.” The direct contrast estimates the effect of assignment to the combination versus durvalumab alone, within the tested regimen and population.
- “A DFS benefit proves longer life.” DFS and overall survival are distinct endpoints. A DFS finding alone does not establish an overall-survival benefit.
Check Your Understanding
Try each question, then compare your reasoning with the solutions.
- In a post-nephrectomy DFS analysis, what event and time origin should you identify before interpreting an HR?
- A hypothetical HR of 0.90 is reported for durvalumab versus surveillance. Under proportional hazards, what does it describe, and what does it not describe?
- At three years, hypothetical DFS estimates are 74% for surveillance and 79% for a combination strategy. What is the estimated absolute difference, and what additional information is needed before claiming a reliable benefit?
- Why does a combination-versus-durvalumab comparison not establish the effect of tremelimumab in every clinical setting?
Solutions. (1) Identify the protocol-defined recurrence-or-death event and the time origin, typically randomization for the stated definition; confirm both from the protocol. (2) If proportional hazards is reasonable, 0.90 represents a 10% lower estimated event hazard at each time, not a 10-percentage-point increase in DFS probability. (3) The difference is \(79\%-74\%=5\) percentage points at three years. A confidence interval or other prespecified uncertainty assessment, along with endpoint and follow-up checks, is needed; the point estimates alone do not establish a reliable benefit. (4) The contrast estimates assignment to the tested combination strategy versus durvalumab alone in the trial population. It does not isolate a universal component effect separate from regimen, population, and treatment context.