Worked Example With Early Response and Overall Survival

Why Pair an Early Response With Overall Survival?

A multi-arm, multi-stage design can use a response measured at a fixed time to decide which experimental treatments should continue. The attraction is practical: response may be observed earlier than enough deaths have accrued for an overall-survival comparison. The statistical challenge is that the early outcome is a selection tool, while overall survival remains the final endpoint. A treatment that looks promising on response is not thereby shown to improve survival.

This tutorial considers three randomized arms: experimental treatments A and B, and shared control C. At a prespecified time \(\tau\), the trial applies a response rule to the experimental arms. Arms that meet the rule continue to an overall-survival comparison with control. The rule, the analysis population, the handling of missing response assessments, and the final inferential procedure must be specified before outcomes are examined.

For example, a protocol might define response as a confirmed tumour response by week 12, with death or progression before week 12 classified as nonresponse. That is only one possible definition: the endpoint and the rule for unevaluable assessments must match the clinical question and be fixed in advance. The survival endpoint should also specify its time origin, event definition, and censoring rules.

The usual response-rate calculation treats randomized participants as the analysis population and requires outcome classification to be sufficiently complete and comparable across arms. If response is missing, silently excluding those participants can bias the comparison; a prespecified missing-data rule or sensitivity analysis is needed. For an ordinary logrank analysis of overall survival, participants should be randomized comparably, follow-up and event ascertainment should be reliable, and censoring should be independent of the event time within the analysis groups, at least conditional on variables used in the analysis. A hazard-ratio interpretation also relies on proportional hazards; the logrank test itself can be less informative when treatment effects vary over time.

Specify the Rule Before Seeing the Data

A simple rule compares each experimental arm's observed response rate with the control response rate. For instance, retain arm \(i\) if \(\hat p_i-\hat p_C\geq \delta\), where \(\delta\) is a clinically chosen minimum difference. This is a decision threshold, not automatically a hypothesis test: unless the protocol says otherwise, it does not require a response p-value to cross a significance boundary. A rule that selects the top-ranked experimental arm is another option, but it can retain an arm even when both experimental response rates are poor.

At selection, a dropped arm generally stops recruiting, while already randomized participants continue to be followed as specified. The final comparison should use the participants and control data allowed by the protocol; excluding participants merely because their arm was dropped can distort the planned analysis. Selection also changes the distribution of the final test statistic. The material on familywise error with intermediate-based selection and operating characteristics of intermediate-based rules is directly relevant: final error rates and power must reflect the complete selection-and-testing procedure, not just an isolated survival test.

Continue Follow-Up and Analyze Overall Survival

For a standard two-group logrank comparison of selected arm A with control C, each observed death contributes to the observed event count and to an expected count under the null hypothesis of equal survival distributions. Let \(O_A\) be the observed number of deaths in A, \(E_A\) the number expected in A under the null, and \(V_A\) the variance of the observed-minus-expected count, calculated from the risk sets at event times. The approximate standardized logrank statistic is

With the sign convention above, a negative value means fewer observed deaths in A than expected under the null, consistent with a possible survival advantage for A. Under the null and usual large-sample conditions, \(Z\) is approximately standard normal. This supplies a nominal p-value for the comparison; in a design that selected A using response, the final decision rule may need a selection-adjusted boundary or calibration from the full design.

Justification and Design Checks

The logrank standardization follows from the null distribution of its score. At each death time, the risk set determines how many deaths would be expected in A if the hazards were equal. Summing the observed-minus-expected contributions gives \(O_A-E_A\); the corresponding variance contributions sum to \(V_A\). Under the null, this score has mean approximately zero and variance approximately \(V_A\). Standardizing therefore gives an approximately unit-variance statistic:

This approximation is most dependable when there is adequate event information and no severe violation of the assumptions behind the planned analysis. For interpreting the logrank comparison as a constant hazard ratio, proportional hazards is an additional substantive assumption. If survival curves cross or effects emerge only after a delay, a single hazard ratio can be difficult to interpret and the ordinary logrank test can have reduced power.

A useful operational sequence is:

Threshold Sensitivity, Alternatives, and Failure Modes

A threshold is a design choice with consequences. In the example, if the response-difference threshold were 0.05 instead of 0.08, B's observed difference of 0.060 would also pass. The final stage would then include two experimental-versus-control comparisons, changing resource use and the multiplicity problem. The choice of threshold should be motivated clinically and evaluated by simulation or another valid operating-characteristic calculation under plausible relationships between response and survival. The earlier material on sensitivity analysis for the intermediate-final relationship addresses why that relationship cannot be assumed to be stable or perfectly predictive.

The fixed-difference rule can be compared with two common alternatives. A top-ranked rule always chooses the apparent leader, even if the difference from control is negligible; a minimum-difference rule can drop all experimental arms if none meets its threshold. A confidence-bound or test-based response rule accounts more explicitly for sampling variability, but it is not interchangeable with a point-estimate rule and must be specified and calibrated in advance. No response-based rule eliminates the need to assess overall survival.

For the survival stage, the logrank test is a conventional unadjusted comparison that uses event-time risk sets. A Cox model can additionally provide a hazard-ratio estimate and may adjust for prespecified covariates; under proportional hazards, that ratio has a relatively direct interpretation. Neither approach automatically fixes selection-induced error inflation. A valid final decision procedure must account for the treatment-selection process and any planned multiple comparisons, for example through a design-specific adjustment or calibration. The nominal p-value in the worked example is therefore illustrative, not a claim that a selected-arm test always controls the overall type I error at its nominal level.

Diagnostics should be considered before interpreting the final result. Check response assessment completeness and timing by arm, confirm that the selection rule was applied without post hoc changes, and examine event and censoring follow-up by arm. For survival, inspect whether risk sets become sparse, whether censoring patterns differ materially, and whether the proportional-hazards interpretation is plausible. A discrepancy between response and survival is not a computational error by itself: intermediate and final effects can disagree. Conversely, a favorable response difference cannot be used to declare a survival benefit when the survival analysis does not support that conclusion.