Independent Increments for Weighted Logrank Monitoring

Why Increment Behavior Matters

A weighted logrank statistic accumulates evidence from event times. In a group sequential design, that accumulation is examined at several interim looks. To understand how evidence at one look relates to evidence at another, it is not enough to know the variance of the final score: we also need to understand the behavior of the score added between looks.

The central distinction is that, under the null, weighted logrank score increments are typically martingale differences: conditional on the observed history, the next increment has mean zero. They are therefore uncorrelated across nonoverlapping intervals under suitable conditions. This does not usually make them exactly independent in a finite sample. Approximate independent increments arise when a martingale central limit theorem gives a Gaussian process with a nonrandom variance function. That approximation is useful for monitoring, but its assumptions must be checked rather than presumed.

Score Increments and the Null Model

Consider a two-arm randomized time-to-event study. At time \(t\), let \(Y_1(t)\) and \(Y_0(t)\) be the numbers at risk in the treatment and control arms, and let \(Y(t)=Y_1(t)+Y_0(t)\). Let \(N_1(t)\) and \(N_0(t)\) count events by time \(t\), and write \(N(t)=N_1(t)+N_0(t)\). Define \(\pi(t)=Y_1(t)/Y(t)\) whenever \(Y(t)>0\). A weight \(w(t)\) is applied to the logrank score contribution at each event time.

The basic null model is equality of the event hazards between arms throughout the monitoring period: conditional on being at risk immediately before \(t\), each participant in either arm has the same instantaneous event hazard \(\lambda_0(t)\). The risk sets must be correctly defined, and censoring must not selectively remove participants according to their unobserved future event times, conditional on the information used in the analysis. The weight must be known just before the event whose contribution it multiplies; a prespecified time function satisfies this predictability condition.

The standard counting-process formulation also assumes that event processes have no simultaneous events, or that ties are handled with a score and variance calculation appropriate to the tie structure. Randomization supports comparability of treatment assignment, but does not by itself guarantee independent censoring or correct event ascertainment.

Why the Score Has Mean-Zero Increments

Under the equal-hazards null, the event counting processes have conditional intensities \(Y_1(t)\lambda_0(t)\) and \(Y_0(t)\lambda_0(t)\). Define their compensated processes by \(dM_g(t)=dN_g(t)-Y_g(t)\lambda_0(t)dt\), for \(g=0,1\). These are martingale increments under the stated model. Since \(Y_1(t)=\pi(t)Y(t)\), the compensator of \(dN_1(t)-\pi(t)dN(t)\) is zero:

Consequently, \(dU(t)=w(t)\{dM_1(t)-\pi(t)[dM_1(t)+dM_0(t)]\}\). If \(w(t)\) is predictable, multiplying by it preserves the martingale property. Thus, for \(t_a<t_b\), the conditional expected score added after \(t_a\), given the observed history through \(t_a\), is zero.

With no tied events, the predictable variance accumulated over time is

This follows because the treatment and control compensated event processes have conditional variance rates \(Y_1\lambda_0\) and \(Y_0\lambda_0\), respectively, and are orthogonal when simultaneous events have probability zero. In practice, the unknown hazard is replaced by an estimate based on the observed events. In the no-tie case, the event-time contribution to the usual variance estimate is \(w(t)^2\pi(t)\{1-\pi(t)\}\).

From Martingale Differences to Independent Gaussian Increments

Let the monitoring times be \(t_1<\cdots<t_K\), and set \(D_k=U(t_k)-U(t_{k-1})\), with \(t_0=0\). The martingale property implies \(E(D_k\mid\mathcal F_{t_{k-1}})=0\), where \(\mathcal F_{t_{k-1}}\) denotes the observed history through the previous look. For \(j<k\), this also implies zero covariance, provided the increments have finite second moments:

The equality uses the fact that \(D_j\) is part of the history available by \(t_{k-1}\). Zero covariance is weaker than independence: dependent random variables can be uncorrelated. In a finite survival trial, later risk sets depend on earlier event and censoring history, so exact independence of the raw increments generally does not follow.

An approximate independent-increment result is available under additional large-sample conditions. In broad terms, the predictable variance accumulated by each monitoring time must converge to a nonrandom limit, and no single event contribution may dominate the total variation. The latter is a Lindeberg-type condition; bounded weights and an increasing number of modest contributions often support it. Under these conditions, a martingale central limit theorem gives a Gaussian limit for the score process. Its increments over nonoverlapping intervals are independent normal variables, because uncorrelated jointly normal variables are independent.

This result concerns the unstandardized score increments in the Gaussian limit. The cumulative standardized statistics at two looks share the earlier score and are therefore correlated. If the cumulative variance at look \(k\) is \(I_k\), their limiting correlation under the canonical Gaussian approximation is \(\sqrt{I_j/I_k}\) for \(j<k\). This relationship is a consequence of shared accumulated evidence, not a contradiction of independent increments.

Monitoring Assumptions, Diagnostics, and Limits

Independent-increment behavior is most credible when the analysis uses a coherent score process, the null hazards are equal, risk-set membership and event times are recorded correctly, censoring is noninformative for the event process under the analysis assumptions, and the weights are predictable. The large-sample approximation further requires enough events and no unusually influential weighted event contributions. If one arm has very few participants at risk, or if a weight becomes extremely large in a sparse tail, the variance may be dominated by a small number of observations and the Gaussian approximation can be poor.

Diagnostics should focus on the conditions that make the approximation plausible. Review risk-set sizes by arm at each look, the distribution of weighted event contributions, the number of events in each monitoring interval, and whether a few events account for a large fraction of estimated variance. Check that censoring patterns and event ascertainment are understood, and that no data-dependent change to weights or monitoring rules has occurred without being incorporated into the analysis. A martingale residual plot or proportional-hazards diagnostic can help describe departures from a simple hazard model, but it cannot prove independent censoring or exact independence of score increments.

A departure from proportional hazards is not, by itself, a failure of the null martingale argument. Under the null of equal hazards, the score can have mean-zero increments even when the planned alternative involves delayed or changing effects. Under an alternative, however, the score increments can have nonzero conditional means; their null independent-increment calibration does not describe that alternative drift. Likewise, informative censoring, a weight that uses future outcomes, or incorrect tie handling can invalidate the stated mean or variance calculation.

The main alternative to relying on a Gaussian independent-increment approximation is to calibrate the joint null distribution using a method that reflects the actual design and data structure, such as simulation under a fully specified null model or an appropriate randomization-based calculation. Such approaches can be useful when event counts are small, ties are common, or risk-set behavior is irregular, but they require their own assumptions and correct implementation. Neither approach repairs misspecified censoring assumptions or outcome-dependent analysis changes.

Common Mistakes

• Equating zero covariance with independence. Martingale differences are uncorrelated across nonoverlapping intervals under the stated conditions, but this does not establish exact finite-sample independence.
• Calling cumulative standardized statistics independent. Earlier score contributions appear in later cumulative statistics, so the statistics at successive looks are correlated.
• Assuming randomization alone is enough. Correct risk sets, appropriate censoring assumptions, predictable weights, and suitable event handling are also needed.
• Assuming any weight is safe. The weight must be available before the event contribution is observed; extreme weights can also undermine the large-sample approximation.