Why Increment Independence Needs Conditions
Sequential logrank monitoring uses information accumulated at successive event-time cutoffs. Its familiar working model treats the score gained between two looks as new information, approximately independent of the score accumulated before those looks. This structure makes the joint behavior of the interim statistics manageable, but it is not a consequence of simply labeling analyses as “interim” and “final.” It depends on how events, risk sets, censoring, and the analysis plan generate the scores.
The important distinction is between uncorrelated increments and independent increments. Under the null hypothesis, the logrank score can be represented as a martingale score process: each new contribution has conditional mean zero given the observed history. Contributions from disjoint time intervals are then orthogonal. Approximate independence follows when a suitable large-sample normal approximation also holds. The earlier material on the canonical joint distribution of sequential logrank statistics gives the resulting working covariance structure; here the focus is on when its independent-increment foundation is reasonable.
The scores are constructed from event contributions weighted by the treatment allocation in the risk set at each event time. The risk set at an event time includes participants still under observation and event-free immediately before that time. Its treatment composition determines the expected treatment-group event count under the null and contributes to the variance, or information. The exact score and information calculations must use a consistent treatment of strata and tied event times.
Conditions Supporting the Approximation
The standard argument is most direct under the null hypothesis \(H_0\) of equal event hazards between randomized groups within the prespecified strata used in the analysis. The score process is then centered: conditional on the history just before an event, the expected score contribution at that event is zero. Its conditional variance is given by the corresponding risk-set variance contribution. Summing those contributions over an interval gives that interval’s information.
Several design and observation conditions support this interpretation:
- Valid risk sets. Participants enter and leave risk sets according to the event and censoring information available by each event time. The risk-set composition used for the score must be determined without using future event outcomes.
- Noninformative censoring for the analysis. Conditional on treatment group and any relevant prespecified strata or covariates, censoring should not selectively remove participants according to their unobserved future event experience in a way that changes the event comparison. Administrative censoring at a planned cutoff is generally compatible with this condition.
- Appropriate randomization and analysis groups. Treatment assignment and the analysis definition should be respected, with the logrank score calculated for the groups and strata specified by the design. Risk-set weighting and variance estimation must match the chosen logrank procedure.
- Regular large-sample behavior. No single event contribution should dominate the total information, and each interval should contain enough information for a normal approximation to be credible. Information should increase across looks; an interval with nearly zero information cannot support a stable standardized increment.
- Planned or predictable analysis timing. Cutoffs should be fixed in advance or determined using information available from the observed history. A decision to continue based on earlier data does not, by itself, make future score contributions dependent on past contributions, but outcome-dependent changes to the analysis require separate justification.
Under these conditions, nonoverlapping event intervals contribute different martingale increments. Their conditional means are zero, and their covariance is zero. For approximate independence, an additional step is needed: a martingale central limit theorem must make the vector of interval increments approximately jointly normal. If its variance structure converges to the planned information increments and no few events dominate, zero covariance in that joint normal limit implies independence.
Why Orthogonality Leads to Approximate Independence
A short derivation shows both the value and the limit of the argument. Write \(\xi_r\) for the score contribution at the \(r\)th event, and let \(\mathcal{F}_{r-1}\) denote the information available just before that event. Under the null and the stated risk-set conditions, \(E(\xi_r\mid\mathcal{F}_{r-1})=0\). For two event contributions \(r<s\), the earlier contribution is already known by the time the later one is evaluated. Therefore:
If interval \(a\) contains event contributions indexed by \(A\), and a later, nonoverlapping interval \(b\) contains those indexed by \(B\), then the covariance of their sums is the sum of the cross-covariances. Each is zero by the same argument, so \(\operatorname{Cov}(\Delta U_a,\Delta U_b)=0\). This is a martingale orthogonality result; by itself it does not prove that the increments are independent. In large samples, if the increments are jointly approximately normal and their covariance matrix has zero off-diagonal entries, then the increments are approximately independent.
A further practical qualification is that the information is often estimated from observed risk sets and is therefore random. The canonical approximation is most persuasive when the accumulated information is adequately represented by its planned or realized values and the variance process does not retain substantial random variation that would make the score distribution a mixture of normals. Standardizing with \(Z_k=U_k/\sqrt{I_k}\) does not automatically remove every such problem.
Worked Example: Checking the Canonical Increment Calculation
Consider a hypothetical two-group trial with two planned looks. Suppose the score increments are \(3.2\) and \(-1.5\), while the corresponding information increments are \(16\) and \(9\). Assume, for this calculation, that the null model and large-sample conditions above are reasonable.
The standardized increment at the first look is \(3.2/\sqrt{16}=0.800\). At the second look it is \(-1.5/\sqrt{9}=-0.500\). Under the null approximation, each has variance one and their covariance is zero, so the standardized increments are approximately independent standard normal variables. The observed values are realizations, not evidence that every trial will have those same increments.
The cumulative score at the second look is \(3.2+(-1.5)=1.7\), and cumulative information is \(16+9=25\). Thus the second cumulative statistic is \(1.7/\sqrt{25}=0.340\). Under the canonical null model, the correlation between the two cumulative statistics is \(\sqrt{16/25}=0.800\): the first cumulative score remains part of the second, even though the new increment is approximately independent of it.
The calculation is justified only if the two score intervals are disjoint and the risk-set, censoring, and approximation conditions are credible. The arithmetic alone cannot establish those conditions.
Departures That Matter for Monitoring
The main diagnostic question is not merely whether event times are disjoint. It is whether the score contribution in each interval remains a properly centered, appropriately standardized increment under the intended null model.
- Informative or treatment-dependent censoring. If censoring depends on unobserved event risk after conditioning on the analysis variables, the observed risk sets can become selected. The resulting score may no longer have the assumed null centering or variance. Compare censoring patterns by group and over time, review reasons for withdrawal, and consider whether the assumed censoring condition is plausible. Similar censoring percentages alone do not establish noninformative censoring.
- Many tied events. With substantial ties, the event-level variance and score depend on how tied failures are handled. A calculation that uses a no-tie variance formula when ties are common can misstate information and therefore distort standardized increments. Use a tie treatment appropriate to the data and apply it consistently at each look.
- Risk-set or information miscalculation. Incorrect entry times, censoring dates, event indicators, or stratum-specific risk sets can produce incorrect information increments. An apparent increment with zero or negative information, or a cumulative information total that decreases, signals a data or calculation problem rather than a usable canonical increment.
- Data-dependent analysis changes. Prespecified monitoring based on accumulated event information is different from changing the score, strata, endpoint, or follow-up rules after inspecting treatment differences. Such changes can invalidate the assumed joint distribution unless their effect is incorporated into the analysis.
- Small or highly uneven increments. If one interval has very few events, or a small number of events supplies a large fraction of its information, the normal approximation can be poor even when the martingale argument is sound. A nominally positive information increment is not sufficient evidence of adequate approximation.
Useful checks include verifying that event and censoring records reproduce the planned risk sets, examining interval event counts and information contributions, and comparing observed information with the design assumptions. If there is a specific concern about covariance, estimate or simulate the joint score distribution under a design-consistent null model. Such checks diagnose the approximation; they do not prove that unmeasured informative censoring is absent.
Worked Example: Informative Censoring Can Change the Comparison
In a hypothetical trial, 100 participants are assigned to each group. Suppose 20 participants in the treatment group leave observation early because their worsening health makes them more likely to experience the event, while 5 participants in the control group are censored for unrelated administrative reasons. A standard logrank analysis treats those censored participants as no longer contributing to later risk sets.
If the treatment-group withdrawals are related to future event risk even after conditioning on the variables used in the analysis, the remaining treatment risk set may be unusually healthy. The observed treatment-group event rate among those still followed can then understate the event experience that would have been observed with complete follow-up. The issue is not that “20 is larger than 5” by itself; it is that the timing and reasons for censoring may be associated with future failure differently across groups.
The ordinary independent-increment calculation relies on the censoring condition being credible. Here, investigators should examine withdrawal timing and reasons, assess whether the association can be explained by observed variables, and consider sensitivity analyses under defensible assumptions. They should not claim that the independent-increment approximation is validated merely because the score intervals do not overlap.
Non-Proportional Hazards and the Null Approximation
Non-proportional hazards require a careful distinction between the null distribution and the behavior under an alternative. Under \(H_0\), equal hazards over time support the centered logrank score increments; the independent-increment approximation can still be useful. Under an alternative in which the hazard ratio changes over time, however, the mean score contribution can change from one interval to another. The increments can retain approximately independent random noise while having different, time-dependent means. A single constant drift per unit information is then generally not adequate.
Worked Example: Independent Noise With Changing Drift
Suppose a hypothetical trial has 40 units of information in an early interval and 60 in a later interval. Assume the log hazard ratio is approximately \(\beta_1=\log(0.70)=-0.3567\) early and \(\beta_2=\log(1.20)=0.1823\) later. For a local-alternative illustration, approximate each interval’s expected score by its information multiplied by its interval-specific log hazard ratio.
The approximate early mean score is \(40(-0.3567)=-14.27\), and the approximate later mean score is \(60(0.1823)=10.94\). The total approximate mean is \(-14.27+10.94=-3.33\). The early and late increments may have approximately independent score noise under the regularity conditions, but their means have different signs. Pooling them into one constant drift would hide the early benefit and later harm represented by these assumed hazard ratios.
This is an approximation for illustrating the distinction, not an exact finite-sample expectation formula for every logrank implementation. It shows why non-proportional hazards chiefly complicate interpretation and power under alternatives; they do not, by themselves, establish that the null increments are dependent.
Alternatives and Limits of the Working Model
The canonical independent-increment approximation is attractive because it gives a compact description of the joint null behavior. It is not the only option when its assumptions are doubtful. A direct joint covariance calculation can represent correlations estimated from the actual score construction, including a design’s particular stratification or tie handling. Null simulation can preserve event accrual, censoring, and planned analysis rules, and can reveal non-normal behavior that a covariance matrix alone cannot show. These approaches depend on correctly specifying the covariance calculation or simulation mechanism; they are not automatic fixes for unknown informative censoring or unplanned design changes.
| Approach | What it provides | Important limitation |
|---|---|---|
| Canonical independent increments | A simple large-sample null model based on successive information increments | Relies on valid score construction, regularity, and an adequate information approximation |
| Direct joint covariance model | Explicit covariance estimates for the statistics at the planned looks | Correct covariance does not guarantee a good normal approximation or valid censoring assumptions |
| Design-consistent null simulation | Can represent discrete events, accrual, censoring, and analysis rules | Requires a credible data-generating model and sufficient simulation precision |
The approximation is also specific to the score process and analysis plan used. It does not imply that all statistics computed at successive looks are independent, that cumulative statistics are independent, or that arbitrary changes to the endpoint and risk-set definition preserve the result. Cumulative statistics share earlier data and are therefore correlated. The independent-increment statement applies to the new score contributions between nonoverlapping looks, under the stated conditions.
Common Mistake
A common error is to argue, “The interim and final analyses use different event times, so their logrank statistics are independent.” That reasoning confuses disjoint new contributions with cumulative statistics: the final cumulative score includes the interim score and is not independent of it. Even disjoint score increments are only approximately independent after the martingale and large-sample conditions justify a joint normal approximation. Check the data-generating and analysis conditions first, and distinguish each interval’s increment from the cumulative statistic at its endpoint.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why does zero covariance between two score increments not, by itself, establish their independence?
- A trial has a first information increment of 25 and a second of 75. Under the canonical null approximation, what is the correlation between the first and second cumulative statistics?
- What feature of censoring threatens the logrank increment argument: a higher censoring percentage in one group, or censoring related to future event risk after conditioning on analysis variables? Explain.
- Does a time-varying hazard ratio necessarily invalidate independent increments under the null? Distinguish its effect on the null approximation from its effect under an alternative.
Solutions. (1) Zero covariance gives orthogonality, but dependent variables can have zero covariance. Approximate independence follows here only when a joint normal approximation is also justified. (2) Total information is \(25+75=100\), so the correlation is \(\sqrt{25/100}=0.500\). (3) Censoring related to future event risk after conditioning is the direct threat because it can select the observed risk sets and undermine score centering or variance. Unequal percentages warrant investigation but do not alone prove informative censoring. (4) No. Under the null, the equal-hazard score can retain its centered martingale structure. Under a time-varying alternative, the increments may have different means, so a constant-drift interpretation can fail even when their random noise is approximately independent.