Independent-Increment Designs For Interim Arm Selection

Why the Interim–Final Relationship Matters

At an interim analysis, a trial may retain one treatment arm and stop enrolling patients to the others. The selected arm is then evaluated using additional observations. A central design question is whether the final statistic uses only those additional observations or combines them with the interim data. The distinction matters: an analysis that reuses interim data is statistically dependent on the selection decision, even when the observations collected after interim are new.

The independent-increment model makes this dependence explicit. It gives the joint distribution of interim and final statistics, and therefore allows the probability of a final rejection to be evaluated conditional on the interim results and selection. It does not, by itself, adjust for selection or guarantee error control. Those depend on the selection rule and the testing procedure.

The Canonical Independent-Increment Model

Consider a treatment contrast with a stable effect \(\theta\), measured on a scale for which the information accumulated by analysis time \(k\) is \(I_k\). Write \(U_k\) for its cumulative score statistic and define the standardized statistic \(Z_k=U_k/\sqrt{I_k}\). In the canonical model, the score increment between analyses has mean equal to the effect times the increment in information, and variance equal to that information increment.

Here \(\theta=0\) represents the null hypothesis for this contrast. The normal model is exact when the score increments are normal with known variance; it is also a common large-sample approximation. Its use requires a correctly specified information scale, a stable treatment effect over the stages, and independent increments. In practice, independence is plausible when the stages use disjoint patient observations and the score process is correctly formed. Repeated measurements, shared data across stages, or a change in the estimand can break the assumed model.

Deriving the Joint and Conditional Distributions

The final statistic can be written as a weighted sum of the interim statistic and the standardized new increment. Let \(\Delta I=I_2-I_1\) and \(Z_{\Delta}=\Delta U/\sqrt{\Delta I}\). The independent-increment assumptions give \(Z_1\sim N(\theta\sqrt{I_1},1)\), \(Z_{\Delta}\sim N(\theta\sqrt{\Delta I},1)\), and independence between them. Therefore:

The weights have squared sum one. Set \(\rho=\sqrt{I_1/I_2}\), so that \(\sqrt{\Delta I/I_2}=\sqrt{1-\rho^2}\). Since \(Z_1\) and \(Z_{\Delta}\) are independent with unit variances, \(Z_2\) has variance one. Its covariance with \(Z_1\) is \(\rho\), because the only shared random component is the interim score. The resulting joint distribution is:

For a bivariate normal distribution, conditioning on \(Z_1=z_1\) gives a normal distribution whose mean is \(\mu_2+\rho(z_1-\mu_1)\) and whose variance is \(1-\rho^2\). Substituting \(\mu_1=\theta\sqrt{I_1}\), \(\mu_2=\theta\sqrt{I_2}\), and \(\rho=\sqrt{I_1/I_2}\) yields:

Under the null hypothesis \(\theta=0\), this simplifies to \(Z_2\mid Z_1=z_1\sim N(\rho z_1,1-\rho^2)\). Thus a large interim statistic shifts the conditional distribution of the cumulative final statistic upward under the null. The conditional probability of exceeding a one-sided final boundary \(c\) is:

where \(\Phi\) is the standard normal cumulative distribution function. This is a conditional probability for the specified interim result, not the unconditional type I error rate of the whole adaptive design.

Selection Among Arms and Conditional Testing

With several candidate treatments, let \(Z_{j1}\) and \(Z_{j2}\) denote the interim and final statistics for arm \(j\), and let \(\theta_j\) be its effect relative to control. If each arm has the same information schedule and the vector of score increments is jointly normal with stable correlation matrix \(R\), the same argument applies arm by arm. The cross-time covariance is \(\operatorname{Cov}(Z_{j1},Z_{\ell2})=\rho R_{j\ell}\), while the within-time correlation between arms is \(R_{j\ell}\). The stage-one and stage-two score vectors must be independent; correlations across arms within a stage need not be zero.

If the rule selects \(J\) using interim data, the conditional distribution for the selected arm must reflect that choice. Given the full interim vector, and under the model above, its final statistic has conditional mean \(\rho Z_{J1}+\theta_J(I_2-I_1)/\sqrt{I_2}\) and variance \(1-\rho^2\). In particular, under a null effect for the selected contrast, the conditional rejection probability at boundary \(c\) is the upper-tail formula derived above. Selection can make the selected interim statistic unusually large, so the conditional probability can be much greater than the nominal one-sided level.

Cumulative Data Versus New-Increment Testing

A cumulative final statistic includes the selected arm's interim data, so it is correlated with the selection event. By contrast, \(Z_{\Delta}\), calculated only from the new score increment, is independent of the interim score under the model. If selection depends only on interim data, then under a null effect \(Z_{\Delta}\sim N(0,1)\) even conditional on that selection. A test based only on this increment can therefore avoid the specific dependence created by reusing the interim observations.

That independence does not make the two tests interchangeable. The increment-only test discards information and generally has less power than a valid design that uses all data while correctly accounting for selection. It also requires that the new-stage statistic and its variance be properly specified; post-interim changes to the estimand, analysis, or data collection can undermine the argument.

Assumptions, Diagnostics, and Alternatives

Before using these calculations, check that the analysis follows the design's information scale and that the interim and later score increments correspond to disjoint, appropriately analyzed data. Also check that the contrast and effect scale remain the same across stages. For multiple arms, specify the correlation structure induced by shared control and allocation; treating arm statistics as independent when they share control can give incorrect joint probabilities.

Useful diagnostics include verifying the information fractions from the actual allocation and variance assumptions, checking that interim and post-interim records do not overlap, and comparing the assumed cross-time correlation with the value implied by the information ratio. If information is estimated, outcomes are non-normal, or the treatment effect changes over time, the canonical normal model may be an approximation rather than an exact description. Simulation under the planned data-generating model can then assess the approximation, but it does not repair a misspecified analysis or selection rule.

The principal alternative is confirmation using an independent validation sample, which avoids reusing the data that drove selection. Another option is to use only the independent post-interim increment for the selected contrast. Cumulative-data procedures retain more information but must account for the dependence induced by selection and continued testing. In all cases, the selection rule, multiplicity plan, and final test must be evaluated together. Independent increments describe the joint data distribution; they are not, on their own, a multiplicity adjustment.