Why Combine Stagewise Evidence?
Under nonproportional hazards, a treatment effect may vary over event time. A weighted logrank test can emphasize event times that are important for the anticipated effect, but evidence gathered in one stage need not have the same pattern as evidence gathered in another. A prespecified combination rule lets a trial combine stagewise evidence without requiring one constant treatment effect or one unchanged weighting function to describe every stage.
The central distinction is between the weight used to construct a stagewise test and the coefficient used to combine stagewise tests. The first determines which event times contribute most to a stage’s weighted logrank statistic. The second determines how much each stage’s standardized evidence contributes to the final statistic. Neither choice should be made after examining the observed stagewise results.
This method is useful when stages contribute different amounts of information or when the design prespecifies different weighted logrank tests for different stages. Its validity depends on having properly calibrated, independent stagewise evidence under the null. In particular, separate tests calculated from overlapping cumulative data are not automatically independent.
Stagewise Statistics and the Combination Rule
The weight function used to calculate \(U_k\) must be specified for that stage. It may be the same across stages or may differ according to a prespecified design. The score and its variance must be calculated consistently with that choice. Under the null, standard weighted logrank theory gives approximately standard normal \(Z_k\) when the usual large-sample and censoring conditions hold.
where the coefficients \(c_k\) are fixed in advance. Larger values of \(C\) support the prespecified treatment-benefit direction. For a one-sided test with final significance level \(\alpha\), reject the global null when \(C>z_{1-\alpha}\), where \(z_{1-\alpha}=\Phi^{-1}(1-\alpha)\). The corresponding combined p-value is \(1-\Phi(C)\).
The normalization \(\sum c_k^2=1\) is essential to the standard normal calibration when the stagewise \(Z_k\) are independent standard normal variables under the null. The coefficients are commonly nonnegative so that evidence favoring treatment in any stage supports the same directional alternative. Other choices require a clear, prespecified interpretation and appropriate calibration.
Assumptions and a Short Justification
The combination rule relies on assumptions about the stagewise tests, not on proportional hazards. In particular, the global null is that there is no treatment effect over the event times relevant to the tests. Under that null, each stagewise statistic must be correctly standardized, and the stages’ contributions must be independent or adequately approximated by independent increments. Stagewise p-values should then be approximately uniform under the null.
For weighted logrank statistics, the usual large-sample justification also requires appropriate risk-set and censoring conditions. A common sufficient framework includes randomized treatment assignment, independent censoring conditional on the modeled factors, prespecified weights that do not use future outcomes, and enough events for the score statistics to have a useful normal approximation. The exact conditions depend on the trial design and analysis.
The null calibration follows directly from the independent standard normal model. If \(Z_k\) are independent \(N(0,1)\) variables, then a fixed linear combination is normal, with mean zero and variance equal to the sum of squared coefficients:
Therefore \(C\sim N(0,1)\) under the null, and comparing it with \(z_{1-\alpha}\) gives a test with size \(\alpha\) under this model. With large-sample weighted logrank statistics, the result is generally an asymptotic justification rather than an exact finite-sample guarantee.
Worked Example: Combining Two Independent Stages
Suppose a hypothetical two-stage trial prespecifies a one-sided test for treatment benefit and uses the same weighted logrank construction in both stages. The stages use nonoverlapping event increments. In stage 1, \(U_1=12\) and \(V_1=100\); in stage 2, \(U_2=30\) and \(V_2=225\). Use prespecified combination coefficients \(c_1=\sqrt{0.4}\) and \(c_2=\sqrt{0.6}\). Assume the normal approximation and independent-increments model are adequate.
First standardize each stage’s score:
The corresponding one-sided p-values are approximately \(p_1=1-\Phi(1.200)=0.1151\) and \(p_2=1-\Phi(2.000)=0.02275\). The coefficients have squared sum \(0.4+0.6=1\), so the combined statistic is
The combined one-sided p-value is \(1-\Phi(2.308)\approx0.0105\). At a prespecified final level of \(0.025\), the critical value is approximately \(1.960\), so the combined statistic crosses the threshold. Under the stated model, the result is evidence against the global null in the treatment-benefit direction. It does not establish that the treatment effect was constant across stages.
Choosing Coefficients When Information Changes
Standardization puts stagewise evidence on a common null scale, but it does not mean every stage has the same precision or expected sensitivity to a particular alternative. A common design choice is to set \(c_k\) proportional to the square root of the planned information in stage \(k\), then normalize the coefficients so their squares sum to one. This gives greater combination weight to stages expected to carry more information. The choice is a design decision, not a requirement for null validity: any fixed coefficients satisfying the normalization give the stated null calibration under the independence assumptions.
When the anticipated treatment effect and test construction are similar across stages, information-proportional coefficients can be a sensible choice. Under nonproportional hazards, however, stagewise expected effects can differ as well as information. A stage using weights that emphasize early events may respond differently from one emphasizing later events. Information alone therefore does not guarantee that a particular coefficient choice is most powerful for the anticipated alternative.
Worked Example: Planned Coefficients for Unequal Information
Consider a hypothetical trial with two independent stages, planned null information \(I_1=25\) and \(I_2=100\), and prespecified coefficients proportional to the square roots of those information amounts. Suppose the stagewise weighted logrank scores and variances are \(U_1=4,V_1=100\) and \(U_2=36,V_2=400\). Assume both stages meet the same null-calibration and censoring conditions.
The normalized coefficients are
The stagewise statistics are \(Z_1=4/\sqrt{100}=0.4000\) and \(Z_2=36/\sqrt{400}=1.800\). Thus,
The one-sided combined p-value is approximately \(1-\Phi(1.789)=0.0368\). The result does not cross a one-sided \(0.025\) threshold of \(1.960\). The coefficients give the more informative planned stage greater influence, but this does not imply that the stage-specific treatment effects are identical. Nor should the coefficients be switched to equal weights after seeing the stagewise results.
Stage-Specific Weights Under Nonproportional Hazards
A prespecified combination rule can combine stagewise tests that use different event-time weights. For example, a design might give one stage greater sensitivity to an anticipated early effect and another stage greater sensitivity to a later effect. Each stage’s score, variance, and p-value must correspond to its own prespecified test. The combination statistic then aggregates the resulting standardized evidence; it does not create a single weighted logrank test or a single treatment-effect estimate.
This flexibility is especially relevant when the expected benefit changes over time. It does not make the analysis immune to a misspecified effect pattern: a chosen weight may be inefficient or may produce little evidence if it emphasizes event times where the effect is weak. The combination remains interpretable as a test of the global no-effect null when the component tests are valid under that null, but its power and the meaning of its alternative depend on the prespecified components.
Worked Example: Combining Prespecified Early- and Late-Sensitive Tests
Suppose a hypothetical trial prespecifies a stage 1 weighted logrank test that emphasizes earlier events and a stage 2 test that emphasizes later events. The stages use independent event increments, and their weights are fixed before outcomes are examined. Assume \(U_1=15,V_1=100\) and \(U_2=10,V_2=100\), with coefficients \(c_1=0.6\) and \(c_2=0.8\).
The coefficients are normalized because \(0.6^2+0.8^2=1\). The stagewise statistics are \(Z_1=15/\sqrt{100}=1.500\) and \(Z_2=10/\sqrt{100}=1.000\). Combining them gives
The combined one-sided p-value is approximately \(1-\Phi(1.700)=0.0446\). This is evidence in the prespecified treatment-benefit direction, but it does not cross a one-sided \(0.025\) threshold. The combined statistic is not an estimate of a hazard ratio, nor does it say that the early and late treatment effects were equal. It summarizes the evidence from the two prespecified tests.
Diagnostics, Alternatives, and Limitations
Before interpreting a combined result, check whether the design assumptions match how the data were analyzed. Review the stage definitions, the event-time weight functions, the sign convention, and the calculation of \(U_k\) and \(V_k\). Confirm that the test used at each stage was specified in advance and that the stagewise evidence does not reuse observations in a way that breaks independence. Also assess whether the event counts and censoring patterns support the normal approximation. These checks diagnose potential problems; they do not repair a data-dependent analysis rule.
The independent-increments material on weighted logrank monitoring motivates why disjoint stagewise contributions can be treated as approximately independent under the null. By contrast, separately calculating weighted logrank statistics on all data accrued up to successive looks generally produces correlated cumulative statistics. Treating those cumulative statistics as independent p-values and applying the formula above can miscalibrate the final test.
An information-weighted cumulative statistic is a natural alternative when the same score construction applies across stages and the design is based on its independent increments. An inverse-normal combination test is more flexible about combining prespecified component tests, including different stage-specific weights, but requires valid stagewise p-values and appropriate independence or conditional calibration. A group sequential boundary design is another distinct approach when decisions are made at interim looks: repeated opportunities to stop for efficacy require boundaries calibrated to control the overall type I error. A final combination test alone does not supply those interim boundaries.
A further limitation is adaptation. If the combination coefficients, weights, or stagewise analyses are selected in response to observed interim outcomes, the fixed-coefficient derivation no longer automatically applies. Some adaptive designs can preserve error control using conditional-error or other formally calibrated methods, but simply choosing the most favorable combination after seeing the data is not one of them. The introductory material on independent increments and prespecified weighted logrank monitoring provides the relevant design context; the combination rule must still be specified and justified for the particular analysis.
Practical Sequence for a Prespecified Combination Test
Define the stages, direction of benefit, stage-specific weighted logrank tests, event-time weights, and score-variance calculations before reviewing stagewise results.
State why each statistic is calibrated under the global null and why the stages are independent, or why the independent-increments approximation is adequate.
Choose coefficients based on the design objective, such as planned information, and verify that their squared sum is one.
Standardize each stage’s score, calculate \(C\), and compare it with the prespecified critical value. Interpret the result as evidence from the specified component tests, not as proof of a constant effect.
Check Your Understanding
Answer the questions, then check the short solutions below.
- Two independent stages have one-sided stagewise statistics \(Z_1=0.50\) and \(Z_2=1.50\), with coefficients \(c_1=0.6\) and \(c_2=0.8\). Calculate the combination statistic and its approximate one-sided p-value.
- Why is it generally invalid to combine cumulative statistics from successive looks as though they were independent?
- Can stage-specific weighted logrank functions differ while the combination test remains valid under the global null? State the key conditions.
- What does a significant inverse-normal combination test establish, and what does it not establish about treatment effects over time?
Solutions. (1) \(C=(0.6)(0.50)+(0.8)(1.50)=1.50\), giving \(1-\Phi(1.50)\approx0.0668\). (2) The later cumulative statistic reuses earlier observations, so the statistics are correlated; the independent-stage null variance calculation no longer applies. (3) Yes, if each stage’s test and variance are prespecified and valid under the null, and the stagewise evidence is independent or has the required conditional calibration. (4) It provides evidence against the global no-effect null in the prespecified direction at the calibrated level. It does not establish a constant treatment effect, estimate one common hazard ratio, or show that every stage has a beneficial effect.