Robustness to Misspecified Delay and Weight Functions

Why Misspecification Matters

A weighted logrank design is often chosen because a particular pattern of treatment benefit is considered plausible. For example, a treatment may be expected to have little effect initially and a later benefit. A weight that downweights early events and emphasizes later events can be more sensitive to that pattern than the ordinary logrank test. But the assumed delay and the chosen weight are design assumptions, not guarantees about the true hazard pattern.

If the benefit begins earlier or later than expected, the weighted statistic may give relatively more influence to events carrying little treatment signal and less influence to events carrying more signal. This changes power and can change the chance of crossing an interim efficacy boundary. It does not, by itself, invalidate a test whose weights and boundaries were prespecified and calibrated under the null. The key is to examine operating characteristics under more than one plausible alternative.

How the True Hazard Pattern Enters the Statistic

Let \(t\) denote event time, and let \(w(t)\) be a fixed, prespecified weight function. For a two-arm comparison, the weighted logrank score accumulates weighted differences between observed and expected treatment-arm events. With a sign convention in which a treatment benefit contributes positively, its null variance is estimated by the corresponding weighted event-information sum. Write the cumulative score and its variance as \(U_w(t)\) and \(V_w(t)\), and the standardized statistic as \(Z_w(t)=U_w(t)/\sqrt{V_w(t)}\).

For interpretation under an alternative, let \(q(t)=-\log\{HR(t)\}\), where \(HR(t)\) is the treatment-to-control hazard ratio. Thus \(q(t)>0\) represents a beneficial treatment effect. Under the usual large-sample, local-alternative approximation, the expected weighted score is approximately the weight- and information-weighted treatment signal. In continuous time, this gives the following useful expression for the standardized mean:

Here \(dI(t)\) denotes the incremental logrank information at time \(t\), and \(\tau\) is the analysis time. This expression is an approximation, not a replacement for the trial’s full design calculations: allocation, censoring, event-time distributions, and the risk sets all affect the information. It nevertheless shows the core robustness issue. The numerator rewards weight placed where the true treatment signal is present, whereas the denominator reflects the variance created by weighting events.

A Short Derivation and the Role of Sequential Monitoring

Divide follow-up into \(J\) time intervals and suppose, for illustration, that the unweighted score increments are approximately independent with variances \(I_j\) and means \(q_j I_j\). Applying a constant weight \(w_j\) within interval \(j\) gives a total weighted score with approximate mean \(\sum_j w_jq_jI_j\) and variance \(\sum_jw_j^2I_j\). Standardizing yields:

This derivation makes a useful point about a delayed effect. If \(q_j=0\) early and positive later, a weight that is small early and larger later can increase the signal relative to its standard deviation. If the effect instead begins early, that same weight may discard useful early signal. If the effect begins later than anticipated, weight concentrated in the originally expected benefit period may again be poorly aligned.

In a group sequential design, let looks occur at cumulative information values \(V_1,\ldots,V_K\). Under the standard canonical joint-normal approximation under the null, the cumulative standardized statistics have correlation \(\operatorname{Corr}(Z_i,Z_j)\approx\sqrt{V_i/V_j}\) for \(i\leq j\). Under an alternative, their means depend on the true hazard pattern and the prespecified weights. Therefore, an overall power calculation must account for the joint distribution and the stopping boundaries; final-look power alone does not determine the probability of stopping early or the total probability of rejecting the null.

Worked Sensitivity Analyses

A practical robustness analysis keeps the trial design fixed—weights, information targets, looks, and boundaries—and varies the data-generating hazard pattern. Scenarios should be chosen before examining unblinded interim treatment comparisons. Useful changes include moving the onset of benefit earlier or later, changing the size of the post-onset effect, or allowing a gradual rather than abrupt onset. For each scenario, calculate the operating characteristics using the same design and analysis rule.

Interpreting Type I Error, Diagnostics, and Alternatives

A crucial distinction is between misspecification under the alternative and misspecification under the null. If a weight function is fixed in advance and the group sequential boundaries are correctly calibrated for that statistic, a different-than-expected treatment effect generally changes power and stopping behavior, not the null Type I error. In contrast, selecting or changing weights using unblinded interim treatment contrasts can change the null distribution of the final procedure. Such adaptation requires an analysis and calibration that account for the selection rule; one cannot assume the original boundaries still control the error rate.

Review the sensitivity analysis for warning signs such as a steep drop in power for a small shift in onset, most of the expected power coming from one narrow scenario, or substantial changes in early stopping probabilities across plausible patterns. Also check that the modeled event information and risk-set behavior are credible for each scenario. A calculation based on identical information accrual across scenarios may be misleading if changing the hazard pattern would materially change event timing or censoring.

The main alternatives are not interchangeable. An ordinary logrank analysis uses constant weights and can be less sensitive to a particular delayed pattern, but it may be less vulnerable to choosing the wrong time emphasis. A prespecified weighted logrank test may gain power for its target pattern, at the cost of potentially losing power when the pattern differs. A design that combines evidence across multiple prespecified weights can reduce dependence on a single weight, but its statistic and sequential boundaries must be calibrated as a combined procedure. Any alternative should be compared using the same plausible scenarios, error target, and stopping rules.

Robustness is not a guarantee against every possible hazard pattern. Scenario analyses depend on assumptions about treatment effects, event accrual, censoring, and the information process. They also do not turn a weighted logrank result into a universal summary of treatment benefit: the statistic answers a design-specific weighted comparison. The appropriate conclusion is therefore conditional—how the planned procedure behaves over the scenarios considered—not that it is optimal for every possible true effect.

Common Mistake

A common mistake is to conclude that a prespecified weight is “robust” because the trial controls Type I error under the null. Null calibration and power robustness are different properties. A procedure may control its false-positive rate at the target level and still have poor power, or very different early-stopping probabilities, when the true treatment effect begins at a time other than anticipated. Conversely, poor power under one alternative does not by itself show that null Type I error is inflated.