Deriving the Inverse Normal Null Distribution

Why a Normal Reference Distribution Appears

In a multi-stage trial, each stage can provide a p-value for the same null hypothesis. A combination procedure can transform those p-values to a common scale and combine the resulting scores. The distribution of that combined quantity under the null is only as reliable as the assumptions behind the transformation and the relationship among stages.

This tutorial derives one specific reference result: when each stagewise p-value is exactly uniform under the null, the p-values are independent, and the fixed weights have squared sum one, the weighted sum of their upper-tail normal scores has a standard normal distribution. The result is about the null distribution; it does not by itself specify a complete testing procedure or establish that the assumptions hold in a particular trial.

The use of \(1-P_i\), rather than \(P_i\), fixes the direction of the score: for example, \(P_i=0.05\) maps to approximately \(1.645\). With the alternative convention \(\Phi^{-1}(P_i)\), the score has the same standard normal null distribution but the sign is reversed. A consistent direction is essential when scores are combined.

Assumptions for the Result

The exact distributional result uses a specific set of assumptions. Under the null hypothesis, each stagewise p-value must be continuous and exactly uniform on \((0,1)\). The p-values from different stages must be mutually independent. The weights \(w_1,\ldots,w_K\) are fixed constants, not choices made after observing the stagewise p-values, and satisfy \(\sum_{i=1}^K w_i^2=1\). Weights may be negative mathematically, although trial designs often use nonnegative weights.

Exact uniformity is stronger than validity of a p-value. A valid p-value under a composite null is often super-uniform: for every \(t\) between 0 and 1, \(P(P_i\leq t)\leq t\). Such a p-value need not be exactly uniform, and its transformed score need not be standard normal. The derivation below therefore applies directly to exact continuous null p-values; it should not be silently extended to discrete or merely conservative p-values.

Independence is a substantive condition, not just an algebraic convenience. It can be justified in settings where stagewise evidence is based on independent increments, provided the stagewise p-values are constructed so that they preserve the required null properties. If stages reuse information, or the data-dependent design creates dependence between their p-values, independence needs separate justification. Earlier material on independent stage increments explains why this issue matters for combination procedures.

Deriving the Null Distribution

The first step is the probability integral transform. If \(P_i\) is uniform on \((0,1)\), then \(1-P_i\) is also uniform on \((0,1)\). Applying the standard normal quantile function to a uniform random variable produces a standard normal random variable. Thus \(Z_i\sim N(0,1)\) under the null.

If the \(P_i\) are mutually independent, applying a fixed function separately to each preserves independence. The scores \(Z_1,\ldots,Z_K\) are therefore independent standard normal random variables. Define the weighted sum \(T=\sum_{i=1}^K w_iZ_i\). A linear combination of independent normal random variables is normal. Its mean and variance can be calculated directly:

The variance formula uses zero covariances between different scores. Independence guarantees those covariances vanish; pairwise uncorrelatedness is sufficient for this variance calculation. However, uncorrelatedness alone does not in general establish that a weighted sum has a normal distribution. Here, independence gives the stronger joint normal structure needed for the distributional conclusion.

The unit squared-sum condition is what sets the variance to one. Without it, the same assumptions give \(T\sim N(0,\sum_i w_i^2)\), not necessarily a standard normal distribution. This distinction is the reason the condition concerns squared weights rather than the weights’ ordinary sum.

What Changes When Assumptions Fail?

Dependence changes the variance through covariance terms. For two scores with unit variances and correlation \(\rho\), $$ \operatorname{Var}(w_1Z_1+w_2Z_2) =w_1^2+w_2^2+2w_1w_2\rho. $$ Even when the squared weights sum to one, a nonzero covariance generally changes the variance. In particular, normal marginal distributions do not alone ensure that their weighted sum has the standard normal reference distribution; the joint distribution also matters.

If p-values are discrete or merely super-uniform, the normal-score transform can also depart from the standard normal distribution, even if the p-values are independent. A conservative p-value tends to be larger than an exact uniform p-value in the lower tail; after the upper-tail transform, this can make the score distribution differ from \(N(0,1)\). Likewise, if weights are selected using the observed scores, treating them as fixed in the derivation is not justified without additional conditions. Each of these departures requires a suitable calibration argument rather than an unexamined appeal to the theorem.

Diagnostics, Comparisons, and Limits

The derivation is a reference result, not a diagnostic test for the assumptions. In an actual adaptive trial, examine how each stagewise p-value was constructed, whether the relevant null distribution is exact or conservative, whether the stage data are independent or otherwise yield independent evidence, and whether the weights were fixed as assumed. Design documentation and the analysis plan should make these features explicit. Empirical plots of simulated null scores can reveal a mismatch in a specified design, but a simulation does not replace a proof that the simulated data-generating process represents the trial’s null behavior.

The main alternative at this point is to combine p-values using a different function rather than normal scores. The inverse normal transformation places evidence on a normal scale, so its weighted sum has a particularly simple reference distribution under the stated assumptions. Other combination functions can have different distributional requirements and may respond differently to patterns of stagewise evidence. The normal-score result does not show that this approach is universally preferable; it establishes the null distribution in the case where its assumptions hold.

The result also has limits. It does not establish the power of a design, the appropriateness of a particular set of weights, or the validity of stagewise p-values under every adaptive procedure. Nor does the theorem itself handle discreteness, dependence, or data-driven weights. Those are separate design and calibration questions. Its value is that it makes clear exactly why a standard normal reference is available in the ideal continuous, independent, fixed-weight setting.

Common Mistake

• Wrong reasoning: “The weights have squared sum one, so the weighted sum is standard normal.” This omits the distributional assumptions on the scores.
• Correction: Unit squared-sum weights give variance one only when the scores have the specified variances and covariance structure. The standard normal conclusion additionally follows here because the transformed scores are independent standard normals. If scores are dependent, pairwise uncorrelatedness can suffice for the variance calculation, but does not by itself establish normality.