Standardizing Cumulative Scores Across Looks

Why Standardize a Cumulative Score?

At an interim analysis in a time-to-event trial, the accumulated logrank score reflects both the observed treatment-control separation and how much information has accrued. The score's scale generally changes as more events and risk-set contributions accumulate. A raw score of a given size is therefore difficult to compare across analyses or interpret without its variance.

Standardization puts the cumulative score on a common, dimensionless scale: the score is divided by the square root of its accumulated variance. This gives an interim statistic whose null distribution is approximately standard normal under the usual large-sample conditions. It is a way to measure the score relative to its uncertainty, not a procedure for adjusting significance levels across repeated looks.

Definition: At look $k$, let $U_k$ be the cumulative logrank score through that look, and let $V_k$ be the corresponding accumulated variance estimate. The standardized cumulative logrank statistic is $$ Z_k=\frac{U_k}{\sqrt{V_k}}, \qquad V_k>0. $$ With the score defined as observed minus expected events in the treatment group, a negative $Z_k$ indicates fewer observed treatment-group events than expected under the null allocation of events within risk sets; a positive value indicates more.

Assumptions and Construction

The score convention matters for interpretation. Here, each risk set contributes the treatment-group observed event count minus its null expected count, and the cumulative score adds those contributions through the analysis cutoff. The variance is accumulated using the corresponding logrank risk-set variance contributions. For tied event times, the analysis must use a consistent tie-handling method for both score and variance.

The standardization is used under the following working conditions. The treatment groups and analysis strata are defined as planned; subjects are followed under the trial's censoring and event-recording procedures; and censoring is independent of the event process conditional on the modeled group and any relevant strata. Under the null hypothesis of equal event hazards, the score contributions have mean zero. For the usual large-sample normal approximation, there must also be enough event information for the score's distribution to be adequately approximated by a normal distribution. A proportional-hazards model is not needed to state the null score test, but a single treatment effect interpretation away from the null is most straightforward when proportional hazards is plausible.

At each look, $U_k$ and $V_k$ must refer to the same data cutoff and the same score definition. In a stratified analysis, add the stratum-specific scores and add their variances; then standardize the resulting totals. Do not standardize each stratum and add the resulting Z statistics.

Why the Standardization Works

The key is that the variance of the cumulative score provides its natural scale. If $U_k$ has mean zero under $H_0$ and variance $V_k$, dividing by $\sqrt{V_k}$ gives a statistic with mean zero and variance one. In practice, the score and variance are estimated from observed risk sets, so the standard normal distribution is an approximation, not an exact finite-sample result.

A short justification can be written in terms of sequential score increments. Write $U_k=\sum_{j=1}^k \Delta U_j$, where $\Delta U_j$ is the score added between looks $j-1$ and $j$. Under the null and the logrank martingale conditions, each increment has conditional mean zero given the prior history, and different increments are uncorrelated. If $v_j$ is the corresponding increment in accumulated variance, then:

\begin{aligned} E_0(U_k)&=0,\\ \operatorname{Var}_0(U_k) &=\operatorname{Var}_0\left(\sum_{j=1}^k\Delta U_j\right) =\sum_{j=1}^k\operatorname{Var}_0(\Delta U_j) \approx \sum_{j=1}^k v_j=V_k,\\ \operatorname{Var}_0\left(\frac{U_k}{\sqrt{V_k}}\right) &\approx \frac{V_k}{V_k}=1. \end{aligned}

The approximation in the final line reflects estimated variance and the large-sample distributional argument. This justification establishes the scale of a single look's statistic. It does not make statistics at different looks independent: they reuse accumulated data. Consequently, a one-sided tail probability calculated from $Z_k$ is a marginal, nominal probability for that look, not a repeated-testing-adjusted probability.

Worked Example: Standardizing One Cumulative Score

Suppose a hypothetical trial reaches an interim cutoff with cumulative logrank score $U_1=-8.0$ and accumulated variance $V_1=25.0$. Assume the score and variance use the same event cutoff and tie method, the censoring assumption is reasonable, and enough events have occurred for the normal approximation to be useful.

The standardized statistic is:

Z_1=\frac{U_1}{\sqrt{V_1}} =\frac{-8.0}{\sqrt{25.0}} =\frac{-8.0}{5.0} =-1.60.

The negative sign means fewer treatment-group events than expected under the null, given the observed risk sets. Under the standard normal approximation, the lower-tail probability is $P(Z\le -1.60)\approx 0.0548$. This is the nominal one-sided tail probability for this look; it is not an adjusted probability for a trial that may examine results repeatedly. The statistic is dimensionless: it measures the score in units of its estimated standard deviation.

Successive Looks Use Cumulative Totals

At a later look, the numerator and variance include everything accumulated since the start, not just the new data since the previous analysis. A useful calculation check is to update both totals by adding the new score and variance increments, and only then divide. Because an increment may have a different sign from the earlier score, the standardized statistic need not move steadily in either direction.

Worked Example: Updating the Standardized Statistic Across Three Looks

Consider a hypothetical trial with the following score and variance increments. The score convention is observed minus expected treatment-group events, and the analysis uses consistent risk-set and tie calculations at each cutoff. Assume the censoring and large-sample conditions described above are adequate.

Look	New score increment	New variance increment
1	$-4.0$	$9.0$
2	$-3.0$	$7.0$
3	$+1.0$	$8.0$

At look 1, the cumulative totals are $U_1=-4.0$ and $V_1=9.0$, so $Z_1=-4.0/\sqrt{9.0}=-1.33$. At look 2, add the second increments: $U_2=-4.0-3.0=-7.0$ and $V_2=9.0+7.0=16.0$. Thus $Z_2=-7.0/\sqrt{16.0}=-1.75$.

At look 3, the new positive score increment partially offsets the earlier negative score: $U_3=-7.0+1.0=-6.0$, while $V_3=16.0+8.0=24.0$. Therefore $Z_3=-6.0/\sqrt{24.0}\approx -1.22$. All three statistics are on the same standardized scale, but their change reflects both new score contributions and increasing variance. A less extreme statistic at look 3 does not imply that the earlier calculation was wrong.

Combining Strata Before Standardizing

For a stratified logrank analysis, stratum-specific scores contribute to one overall score, and their variances contribute to its variance. This preserves the weighting implied by the score's uncertainty. The calculation is not an unweighted average of separate standardized statistics.

Worked Example: Standardizing a Stratified Cumulative Score

At a hypothetical interim cutoff, suppose two prespecified strata have cumulative logrank scores and variances as follows. Each stratum uses the same treatment coding, cutoff, censoring assumptions, and tie method; assume the strata are independent for this standard variance addition.

Stratum A has $U_A=-5.0$ and $V_A=16.0$. Stratum B has $U_B=+1.0$ and $V_B=9.0$. The combined totals are:

U=-5.0+1.0=-4.0,\qquad V=16.0+9.0=25.0, \qquad Z=\frac{-4.0}{\sqrt{25.0}}=-0.800.

Thus the combined score is $0.800$ estimated standard deviations below zero. For comparison, the separate standardized values are $Z_A=-5.0/4.0=-1.25$ and $Z_B=1.0/3.0=0.333$. Adding those Z statistics would give $-0.917$, which is not the combined logrank statistic. The correct result comes from adding the scores and variances first.

Interpretation, Diagnostics, and Limits

A standardized cumulative score answers a focused question: how far from zero is the cumulative score, measured in its estimated standard-deviation units? It makes results from looks with different accumulated variances readable on a common scale. It does not supply a stopping boundary, correct for repeated testing, or establish that a treatment effect is clinically important. Sequential boundaries and error control require a separate design-level procedure.

Several checks help identify problems before interpreting $Z_k$:

Check the cutoff alignment. The score and variance must use the same included event times, risk sets, strata, and tie convention.
Check the variance. It must be positive and plausibly accumulated from the corresponding risk-set contributions. A zero or implausibly small value makes standardization undefined or unstable.
Check the sign convention. Under the convention here, negative values indicate fewer observed treatment events than expected. Reversing the group coding reverses the sign.
Check approximation quality. Sparse events, very small risk sets, or heavy ties can weaken the normal approximation; a displayed Z value alone does not diagnose this.
Check censoring and strata. Informative censoring or inconsistent stratum handling can invalidate the score's null calibration or the variance calculation.

The main alternatives serve different purposes. The raw cumulative score retains the direction and magnitude of the observed-minus-expected contrast but is not comparable across different variance levels. The standardized statistic expresses that score relative to its uncertainty and supports a familiar marginal normal reference distribution. A sequentially adjusted boundary or a prespecified combination procedure addresses repeated looks; replacing one with the other changes the inferential procedure. In particular, it is a mistake to treat each $Z_k$ as an independent standard normal observation: the cumulative data overlap, so the statistics across looks are dependent.

Key takeaway: At each interim look, compute the cumulative score and its matching accumulated variance, then form $Z_k=U_k/\sqrt{V_k}$. Interpret its sign using the stated coding and its magnitude on the marginal standard-normal scale, while handling repeated testing separately.

Check Your Understanding

Try each question, then compare your reasoning with the solutions.

At a look, $U_k=6.0$ and $V_k=36.0$. What is $Z_k$, and what does its sign mean under the convention used here?
Between two looks, a new score increment is positive. Must the later cumulative Z statistic be more positive? Explain.
Two independent strata have scores $-3.0$ and $-4.0$, and variances $4.0$ and $12.0$. Calculate the combined standardized statistic.
Why is a marginal normal tail probability at an interim look not automatically a repeated-testing-adjusted probability?

Solutions. (1) $Z_k=6.0/\sqrt{36.0}=1.00$; the positive sign means more observed treatment-group events than expected under the null. (2) No. The cumulative score increases, but accumulated variance also increases; the ratio can move either way. (3) Add first: $U=-3.0-4.0=-7.0$, $V=4.0+12.0=16.0$, so $Z=-7.0/\sqrt{16.0}=-1.75$. (4) Statistics at successive looks use overlapping cumulative data and are dependent, so controlling the overall error rate requires a sequential procedure rather than a single-look tail probability.

Look	New score increment	New variance increment
1	\(-4.0\)	\(9.0\)
2	\(-3.0\)	\(7.0\)
3	\(+1.0\)	\(8.0\)