Independent Increments Across Looks

Why Look at Increments?

At interim look $i$, a cumulative test statistic uses the data collected up to that point. Statistics at two looks are therefore related: the later statistic contains the earlier data as well as newly collected data. Their correlation is useful, but it can obscure a simpler structure. If we express the cumulative evidence as the contribution from the first data block, then the contribution from the next block, and so on, the blocks can be independent under suitable assumptions.

This tutorial defines those information-scaled increments and explains the precise conditions for independence. The key distinction is between a cumulative statistic, which reuses earlier data, and an increment, which represents only the new data since the preceding look. The result is a building block for the canonical joint model, not a claim that all statistics from all monitoring designs have independent increments.

Definitions and Assumptions

Suppose there are $K$ planned looks at strictly increasing, fixed information levels $I_1,\ldots,I_K$, with $I_0=0$. Define the information gained between looks $i-1$ and $i$ by $\Delta I_i=I_i-I_{i-1}$, which is positive. Let $U_i$ denote the cumulative score-scale statistic at look $i$, and set $U_0=0$. The information-standardized cumulative statistic is $Z_i=U_i/\sqrt{I_i}$, or equivalently $U_i=\sqrt{I_i}Z_i$.

Definition: The score increment between looks $i-1$ and $i$ is $D_i=U_i-U_{i-1}$. Its information-standardized increment is $$ W_i=\frac{D_i}{\sqrt{\Delta I_i}} =\frac{\sqrt{I_i}Z_i-\sqrt{I_{i-1}}Z_{i-1}}{\sqrt{I_i-I_{i-1}}}. $$ The increment uses only the score contribution accumulated between those two looks. Under the null model described below, it has mean zero and variance one.

For the exact independent-increments argument, assume that the data arrive in disjoint blocks; observations in different blocks are independent; each block contributes a score based only on its own observations; and the planned information levels are fixed. Under the null hypothesis, suppose each block score has mean zero and variance equal to that block's information. In the canonical normal model, also assume the block scores are jointly normal. Independence of data blocks makes their score contributions independent when each contribution is a function of its own block. More generally, zero covariance alone does not establish independence; joint normality is what allows that step for score increments.

A simple exact example is a normal mean with known standard deviation $\sigma$. For a specified null mean $\mu_0$, each observation $X_j$ contributes $Y_j=(X_j-\mu_0)/\sigma$ to the score for the standardized mean difference $(\mu-\mu_0)/\sigma$. Under $H_0$, independent observations give independent $Y_j\sim N(0,1)$. For $n_i$ observations at look $i$, $U_i=\sum_{j=1}^{n_i}Y_j$, $I_i=n_i$, and $Z_i=U_i/\sqrt{n_i}$. This model makes the meaning of the increments concrete; the canonical formulation expresses the same structure in terms of information rather than sample size alone.

Why Disjoint Data Produce Independent Score Increments

Write the cumulative score at look $i$ as the sum of the contributions from all blocks observed so far. Thus $U_i=D_1+\cdots+D_i$, where $D_i$ is the score contribution from block $i$. If the data blocks are independent and each block's score depends only on that block, then $D_1,\ldots,D_K$ are independent. Under the null model, their means and variances are $E(D_i)=0$ and $\operatorname{Var}(D_i)=\Delta I_i$. Dividing by $\sqrt{\Delta I_i}$ gives independent, mean-zero, unit-variance $W_i$. If the block scores are normal, then $W_i\sim N(0,1)$ exactly.

The same conclusion follows from the canonical joint normal model, stated in cumulative-statistic form. Under $H_0$, assume $E(U_i)=0$, $\operatorname{Var}(U_i)=I_i$, and $\operatorname{Cov}(U_i,U_j)=I_i$ whenever $i\leq j$. For $i<j$, the covariance between the earlier cumulative score and a later increment is

\operatorname{Cov}(U_i,U_j-U_i) =\operatorname{Cov}(U_i,U_j)-\operatorname{Var}(U_i) =I_i-I_i=0.

The result extends to any two nonoverlapping increments: their covariance is zero because the shared cumulative portions cancel. Since the cumulative vector is jointly normal, its linear transformations—the increments—are jointly normal as well. For jointly normal random variables, zero covariance implies independence. Finally, $\operatorname{Var}(U_i-U_{i-1})=I_i-I_{i-1}=\Delta I_i$, so each standardized increment has variance one. These steps establish independent standard normal increments under this null model; neither the covariance calculation alone nor the word “increment” is sufficient.

Worked Example: A New Block in a Known-Variance Normal Model

Consider a hypothetical monitoring plan for a normal outcome with known $\sigma=10$ mm and null mean $\mu_0=100$ mm. At the first look, 25 observations have mean $102$ mm. By the second look, 39 additional observations have been collected, and their mean is $101$ mm. Use the standardized score contribution $Y_j=(X_j-\mu_0)/\sigma$, for which the information per observation is one under $H_0$.

The first block contributes a total deviation of $25(102-100)=50$ mm, so its score is $D_1=50/10=5$, and its information is $\Delta I_1=25$. The second block contributes $39(101-100)=39$ mm, so $D_2=39/10=3.9$, with $\Delta I_2=39$. The standardized second-block increment is $$ W_2=\frac{D_2}{\sqrt{\Delta I_2}} =\frac{3.9}{\sqrt{39}} =0.6245. $$

The cumulative scores are $U_1=5$ and $U_2=5+3.9=8.9$, with information levels $I_1=25$ and $I_2=64$. Thus $Z_1=5/\sqrt{25}=1$ and $Z_2=8.9/\sqrt{64}=1.1125$. Substitution into the cumulative-statistic formula gives $(8.9-5)/\sqrt{64-25}=3.9/\sqrt{39}=0.6245$, matching the direct block calculation. The observed increment is not itself a p-value; under the stated null model it is one realized value of a standard normal increment.

Worked Example: Recovering an Increment From Cumulative Z Statistics

Suppose a hypothetical canonical design has $I_1=40$ and $I_2=100$. The observed cumulative statistics are $Z_1=0.5000$ and $Z_2=0.8000$. Find the standardized increment for the data added between the looks.

First convert each $Z$ statistic to its cumulative score scale: $U_1=\sqrt{40}(0.5000)=3.1623$ and $U_2=\sqrt{100}(0.8000)=8.0000$. The information added is $100-40=60$. Therefore, $$ W_2=\frac{U_2-U_1}{\sqrt{60}} =\frac{8.0000-3.1623}{7.7460} \approx 0.6245. $$

This calculation identifies the added-data contribution from the two cumulative summaries and their information levels. It does not establish that the observed increment is independent of the first-look statistic by itself. That property follows from the assumed canonical joint normal model, or from the stronger data-block argument when the observations truly arrive in independent, disjoint blocks.

Worked Example: Zero Covariance and Independence

Take a hypothetical two-look null model with $I_1=25$ and $I_2=100$. The cumulative scores have $\operatorname{Var}(U_1)=25$, $\operatorname{Var}(U_2)=100$, and $\operatorname{Cov}(U_1,U_2)=25$. Let $D_2=U_2-U_1$, the contribution from the additional information.

Its variance and its covariance with the first-look score are $$ \operatorname{Var}(D_2)=100+25-2(25)=75, \qquad \operatorname{Cov}(U_1,D_2)=25-25=0. $$

If $(U_1,U_2)$ is jointly normal, then $(U_1,D_2)$ is jointly normal because it is a linear transformation of that vector. Their zero covariance therefore implies that $U_1$ and $D_2$ are independent. The standardized increment $W_2=D_2/\sqrt{75}$ has variance $75/75=1$, and it is independent of $U_1$, and hence of $Z_1=U_1/5$. If joint normality is absent, the calculation proves only zero covariance; independence would require additional assumptions.

Diagnostics, Alternatives, and Limitations

In application, check the structure that makes the result credible before using independent increments in a group sequential calculation. Confirm that information is increasing, that each $\Delta I_i$ is positive, and that the score at each look is cumulative with the earlier score included exactly once. For an independent-block design, verify that new observations do not overlap with earlier observations and that each block's score is computed consistently. If information is estimated from the same data, or if future sample size depends on interim outcomes, the fixed-information argument may not apply as stated.

A useful diagnostic is to compare the covariance implied by the planned model with the covariance arising from the actual data structure. In the canonical model, $\operatorname{Cov}(U_i,U_j)=I_i$ for $i\leq j$, so the corresponding cumulative $Z$ statistics have correlation $\sqrt{I_i/I_j}$. If shared observations, clustering, repeated measurements, changing score definitions, or shared nuisance-parameter estimates alter that covariance, the independent-increments representation may fail or may need a more careful approximation. A numerical observation that sample increments appear uncorrelated is not proof of population independence.

The main alternative is to work directly with the joint distribution of the cumulative statistics, retaining their covariance rather than transforming to increments. This is natural when the covariance is known and the increments do not simplify the problem. The increment representation is especially convenient when new blocks are independent, because it separates the contribution of each interval in information units. Neither representation repairs a misspecified covariance model. In particular, nonnormal dependent data, data reuse, information chosen adaptively, or score contributions that share estimated quantities can invalidate exact independence; approximation or a model tailored to that design may be needed.

Common Mistakes

Calling nested cumulative statistics independent. They share data and are generally correlated. It is the disjoint score increments, not the cumulative statistics, that are independent under the stated conditions.
Equating zero covariance with independence in every model. Zero covariance is enough only with an appropriate joint normal assumption (or another justification of independence).
Using sample-size differences when information differs from sample size. Standardize by $\sqrt{\Delta I_i}$, not automatically by the square root of the number of newly enrolled participants.
Forgetting to recover the cumulative score first. The difference $Z_i-Z_{i-1}$ is generally not the score increment. Use $\sqrt{I_i}Z_i-\sqrt{I_{i-1}}Z_{i-1}$.

Key takeaway: Under fixed increasing information, disjoint independent data blocks give independent score increments when each block contributes its own score. In the canonical joint normal model, the same conclusion follows from the covariance structure and joint normality. Standardize each increment by the square root of the information gained in that interval.

Check Your Understanding

Answer each question, then compare with the solutions below.

For $I_1=16$, $I_2=64$, $Z_1=0.5000$, and $Z_2=0.7500$, calculate $W_2$.
Why are cumulative statistics at nested looks usually correlated even when increments are independent?
Under a jointly normal canonical model, what does zero covariance between two increments establish? What would it establish without joint normality?
A design has information levels $I_1=30$ and $I_2=30$. Can the standardized increment formula be used for the second look as written? Explain.
Name one feature of a study or analysis that could make independent increments questionable.

Solutions. (1) $U_1=\sqrt{16}(0.5000)=2$ and $U_2=\sqrt{64}(0.7500)=6$. Thus $W_2=(6-2)/\sqrt{64-16}=4/\sqrt{48}=0.5774$. (2) Each later cumulative statistic contains the earlier observations, so it shares data with the earlier statistic; independent new contributions do not remove that overlap in the cumulative totals. (3) Joint normality and zero covariance imply independence. Without joint normality, zero covariance alone establishes only lack of linear covariance, not independence. (4) No. Here $\Delta I_2=0$, so the denominator is zero and no positive information was added; the definition requires strictly increasing information. (5) Examples include overlapping or reused observations, dependence across data blocks, data-dependent information levels, or score contributions that share estimated nuisance quantities; any may require a different joint model or additional justification.