Independent Increments Across Trial Stages

Why Stagewise Independence Requires a Data-Structure Argument

In a seamless trial, evidence from an earlier stage may help determine which treatment or dose proceeds. Stagewise statistics are often treated as independent under a common null, but that property does not follow merely from labeling observations “stage 1” and “stage 2.” It follows from assumptions about how the observations were generated and whether the stages share data.

Here, the target is the relationship between stage-specific treatment-effect statistics under the null hypothesis that the treatment effect for the estimand being tested is zero. The issue is not whether the observed statistics happen to look unrelated. The issue is whether their joint sampling distribution has the claimed dependence structure. Disjoint participant cohorts with independent outcomes can support independence; shared participants or shared outcomes generally cannot.

Definition: Stagewise statistics have independent increments under \(H_0\) when the evidence contributions from successive stages are independent under that null. In the two-stage setting, this means the stage 1 and stage 2 statistics are independent random variables under \(H_0\). For a sequence of more than two stages, each new increment must be independent of the evidence accumulated before it.

A common null means the stages address the same null hypothesis for the same estimand—for example, no treatment effect in either stage—not that the observed effects must be equal or that all stage-specific estimates are numerically identical. Independence is a property of the null sampling distribution.

A Basic Derivation for Disjoint Cohorts

Suppose stage \(k\) compares a treatment with control using a difference in sample means, \(\hat{\theta}_k=\bar{X}_{Tk}-\bar{X}_{Ck}\). Assume participants in the two arms and across stages are independent, the outcome variance in each arm is the same known value \(\sigma^2\), and the sample sizes are fixed. Under \(H_0:\theta=0\), each stage estimate has mean zero. Its variance is \(\sigma^2/n_{Tk}+\sigma^2/n_{Ck}\), so the standardized statistic is

Z_k=\frac{\hat{\theta}_k}{SE_k}, \qquad SE_k=\sqrt{\frac{\sigma^2}{n_{Tk}}+\frac{\sigma^2}{n_{Ck}}}.

For normally distributed outcomes, each \(Z_k\) is exactly standard normal under the null. With sufficiently large samples, the same standard-normal approximation may be justified by a central limit theorem under appropriate regularity conditions. The stage estimates are functions of disjoint sets of independent observations, so they are independent. Standardizing each by a fixed constant preserves that independence.

The same reasoning can be expressed with stagewise score contributions. Let \(U_k\) be the score contribution from stage \(k\), with null variance, or information, \(I_k\). If the stage data are independent, then \(U_1\) and \(U_2\) are independent under \(H_0\). With fixed information, \(Z_k=U_k/\sqrt{I_k}\) are also independent. This score formulation applies beyond differences in means, provided the likelihood or estimating procedure has the stated independent-data structure and the null calibration is valid.

Key condition: Disjoint participant lists are not enough on their own. Independence also requires that the relevant outcomes or score contributions are independent across stages, and that the statistic is correctly calibrated under the common null. Shared households, sites, batches, or other clusters can create dependence even when no individual appears in both stages.

Worked Example: Two Independent Stage Cohorts

Suppose a hypothetical trial measures a continuous endpoint. In stage 1, 50 participants are assigned to treatment and 50 to control; in stage 2, a new 80 are assigned to each arm. Assume independent participants, normal outcomes with known variance \(\sigma^2=100\) in both arms, fixed sample sizes, and the same estimand in both stages. The hypothetical observed treatment-minus-control differences are 4.0 units in stage 1 and \(-1.0\) unit in stage 2.

For stage 1, the standard error and statistic are

SE_1=\sqrt{\frac{100}{50}+\frac{100}{50}}=2, \qquad Z_1=\frac{4.0}{2}=2.00.

For stage 2,

SE_2=\sqrt{\frac{100}{80}+\frac{100}{80}} =\sqrt{2.5}\approx1.581, \qquad Z_2=\frac{-1.0}{\sqrt{2.5}}\approx-0.6325.

Under \(H_0\), each statistic is standard normal in this specified model. Since stage 2 uses new participants whose outcomes are independent of stage 1, \(Z_1\) and \(Z_2\) are independent under \(H_0\). The observed values 2.00 and \(-0.6325\) illustrate the calculations; their numerical contrast is not the reason for the independence conclusion. That conclusion comes from the sampling assumptions and cohort structure.

Why Shared Participants Create Dependence

A shared participant can contribute random variation to both stage estimates. A shared control group is a particularly clear example: its sample mean is subtracted in both treatment-versus-control contrasts. The resulting estimates may each be unbiased under \(H_0\), yet they are correlated because they contain the same random term.

For two estimates \(D_1=\bar{X}_{T1}-\bar{X}_{C}\) and \(D_2=\bar{X}_{T2}-\bar{X}_{C}\), suppose the two treatment groups and the control group are mutually independent, while the same control participants are used in both comparisons. Then

\operatorname{Cov}(D_1,D_2) =\operatorname{Cov}(\bar{X}_{T1}-\bar{X}_{C},\bar{X}_{T2}-\bar{X}_{C}) =\operatorname{Var}(\bar{X}_{C}).

The positive sign matters: both contrasts subtract the same control mean, so a control mean that happens to be high tends to push both treatment differences downward. In a jointly normal model, the standardized contrasts are jointly normal. They are independent only if their covariance is zero; here it is positive, so they are not independent.

Worked Example: Two Comparisons With a Shared Control

In a hypothetical design, 50 participants receive control and are used in both stage comparisons. Separate groups of 50 receive treatment A and treatment B. Assume all outcomes are independent across these three groups, normal, and have variance 100. Suppose the observed means are 100 for control, 103 for A, and 102 for B.

The two estimated effects are \(D_A=103-100=3\) units and \(D_B=102-100=2\) units. Each has variance \(100/50+100/50=4\), so each standard error is 2 units. The corresponding statistics are

Z_A=\frac{3}{2}=1.50, \qquad Z_B=\frac{2}{2}=1.00.

Their covariance is the variance of the shared control mean:

\operatorname{Cov}(D_A,D_B)=\operatorname{Var}(\bar{X}_{C}) =\frac{100}{50}=2.

Standardizing both estimates by 2 gives \(\operatorname{Cov}(Z_A,Z_B)=2/(2\cdot2)=0.50\). Each statistic has variance 1, so their correlation is 0.50. The statistics are therefore not independent, even though the treatment groups are distinct and each comparison is individually calibrated under its null. Treating them as independent would ignore the shared control’s contribution to both.

Selection Does Not Automatically Destroy Fresh-Stage Independence

Stage 1 may select a treatment using stage 1 data. That makes the selected stage 1 evidence data-dependent and can make it unusually favorable. It does not, by itself, make a statistic based on entirely fresh stage 2 data dependent on the stage 1 observations. To see why, suppose stage 1 provides a vector of statistics for candidate treatments, and stage 2 uses new independent participants. If the stage 2 statistics are independent of the entire stage 1 vector under the common null, then selecting an index from stage 1 leaves the corresponding fresh stage 2 statistic null-calibrated, provided the relevant stage 2 null distribution is the same for each selectable treatment.

Worked Example: Choosing a Candidate Before a Fresh Stage

Consider two candidate treatments A and B. Under the hypothetical common null, their stage 1 statistics are \(Z_{1A}=1.20\) and \(Z_{1B}=1.80\), and the prespecified rule advances whichever has the larger stage 1 statistic. Stage 2 enrolls new, independent participants. Assume the stage 2 statistic for either candidate is standard normal under its null and that the fresh stage 2 data are independent of all stage 1 data.

The rule selects B because \(1.80>1.20\). Suppose its hypothetical stage 2 statistic is \(Z_{2B}=-0.40\). Conditional on all stage 1 results, including the fact that B was selected, the stage 2 statistic still has the null distribution \(N(0,1)\). In particular, for any real \(c\),

P(Z_{2B}\le c\mid \text{stage 1 data and B selected})=\Phi(c),

where \(\Phi\) is the standard normal cumulative distribution function. For example, at \(c=0\), this conditional probability is 0.50. This conclusion concerns the fresh stage 2 statistic; it does not assert that the selected stage 1 statistic is standard normal after selection. It also depends on the common-null calibration and the independence of the new observations. Shared controls, unequal or misspecified null distributions across candidates, or an invalid adaptive analysis can invalidate the argument.

Assumptions, Diagnostics, and Alternatives

Before asserting independent increments, check the actual data flow rather than relying on the stage labels. Important questions include:

Are participants and outcome records used in stage 1 excluded from the stage 2 statistic?
Do the stages share controls, repeated measurements, sites, laboratories, or other sources of random variation?
Are the stagewise statistics calibrated for the same estimand and the stated common null?
Are sample sizes, variances, and test statistics handled correctly if stage 2 design features are adapted using stage 1 results?

A design diagram or participant-level data audit can reveal overlapping records. For shared controls, compute the covariance implied by the common control contribution, as in the preceding example. For clustered observations or repeated measures, the analysis must account for within-cluster or within-person dependence; treating all observations as independent can understate uncertainty and produce an incorrect stagewise correlation. When sample sizes or other design features change adaptively, verify that the stage 2 statistic retains its claimed null distribution conditional on the adaptation. Independence cannot repair invalid calibration.

Zero covariance and independence are not interchangeable in general. The disjoint-data argument establishes independence directly. Alternatively, if stagewise statistics are jointly normal, zero covariance implies independence; without joint normality, zero covariance alone is insufficient. Conversely, a nonzero covariance rules out independence when the relevant second moments exist.

The main alternative to assuming independent increments is to model the actual joint distribution, including overlap or clustering, and use a test procedure calibrated for that dependence. Depending on the design and endpoint, this may involve a covariance model, cluster-aware variance estimation, or a randomization-based analysis that reproduces the allocation and reuse structure. A design can also avoid the issue prospectively by using nonoverlapping stage cohorts. These options are not interchangeable shortcuts: the method must match the source of dependence and the inferential target.

Common Pitfall and Practical Limits

Wrong reasoning: “Stage 1 and stage 2 have different labels and separate treatment groups, so their statistics are independent.” This overlooks shared controls, repeated outcomes, and common clusters. Two contrasts can use different treatment participants and still be dependent because they subtract the same control mean.

Correction: Identify every random observation or score contribution used in each statistic. Establish independence from disjoint, independent contributions, or derive the covariance and joint distribution when contributions overlap. Also distinguish an unconditional statement under \(H_0\) from a conditional statement after an adaptation: the latter requires valid conditional calibration for the stage 2 statistic.

Independent increments are an assumption about a specified statistical model and design, not a guarantee for every seamless trial. Departures from normality may make zero covariance inadequate to establish independence; shared participants can induce dependence even when each stage estimate is unbiased; and changes to the estimand or endpoint can make the two stage statistics incomparable. When these limitations apply, the dependence structure must be represented explicitly rather than assumed away.

Key takeaway: Independent increments follow most directly when stagewise evidence uses disjoint, independent data and each statistic is valid under the same null. Shared participants or shared random components can create dependence; treatment selection alone does not do so when the later statistic is based on fresh, independently sampled data with valid conditional calibration.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions.

Two normally distributed stage estimates use disjoint independent cohorts and have fixed standard errors. What additional null-distribution fact allows zero covariance to imply independence if you establish the joint normal model?
Stagewise treatment-control contrasts use separate treatment groups but the same control participants. Which shared random quantity creates their covariance?
A candidate is selected using stage 1 statistics and tested with new stage 2 participants. State the conditions needed for the selected stage 2 statistic to retain its null distribution conditional on stage 1.
Can zero covariance by itself establish independence for arbitrary stagewise statistics? Explain.
Name one design feature, other than a participant appearing in both stages, that can undermine independence across stages.

Solutions. (1) Joint normality: jointly normal variables with zero covariance are independent. (2) The shared control sample mean; it enters both contrasts with a minus sign, giving covariance equal to its variance under independent treatment groups. (3) The stage 2 observations must be fresh and independent of stage 1, and the stage 2 statistic must be correctly calibrated under the common null for the selected treatment, conditional on the adaptation. (4) No. Zero covariance does not generally imply independence outside an appropriate joint normal model. (5) Examples include shared sites or clusters, repeated measurements, or common batches that induce correlated outcomes.