Solving the First Interim Boundary

From First-Look Spending to a Boundary

At an interim analysis, an efficacy boundary translates an alpha-spending target into a decision rule for the observed test statistic. The key calculation at the first look is direct: under the null hypothesis, find the upper-tail cutoff whose probability equals the amount of alpha scheduled to have been spent by that look. The information fraction matters because it determines the spending target; once that target is known, the first boundary follows from the null distribution of the statistic.

This tutorial considers a one-sided efficacy test, with the statistic oriented so that larger positive values support the treatment effect of interest. It focuses on the first interim look only. There has been no earlier opportunity to cross an efficacy boundary, so the probability of crossing at the first look is also the cumulative probability of crossing by that look.

The canonical statistic \(Z_1\) has a standard normal null distribution under the assumed model: \(Z_1\sim N(0,1)\) under \(H_0\). In a large-sample analysis this is typically an exact or asymptotic standardization, depending on the trial model. The calculation requires that the statistic and its direction match the design used to specify the spending target.

Write \(\Phi(z)=P(Z\le z)\) for the standard normal cumulative distribution function when \(Z\sim N(0,1)\). Because the desired event is an upper-tail crossing, \(P(Z_1\ge b_1)=1-\Phi(b_1)\). Equating that probability to the target and inverting gives the boundary.

Thus, the first-look boundary is not chosen by applying the full planned alpha at the first look. It is obtained by using the cumulative amount scheduled to be spent by the actual first information fraction. If \(A(t_1)=0\), the boundary is \(+\infty\): no finite statistic can cross a zero-spending efficacy threshold. In ordinary designs, \(0<A(t_1)<1\), so the boundary is a finite positive quantile when \(A(t_1)<0.5\).

Why the First-Look Formula Works

The justification is a one-step null probability calculation. At the first look the design has no previous test statistics or stopping events to condition on. Therefore the null probability of crossing by the first look is simply the marginal null probability of \(Z_1\) being at or above \(b_1\). Setting this probability equal to the cumulative spending target calibrates the first boundary to that target.

This calibration relies on several conditions. The spending function must be specified before the interim result is used; its value at \(t_1\) must be a valid cumulative spending target for the design; and the null distribution used for the quantile must describe the actual test statistic at that look. The statistic must also be oriented so that the efficacy direction is the upper tail. If evidence for benefit corresponds to smaller values, the statistic or tail must be reoriented before applying the formula.

For a continuous standard normal statistic, the probability of equality to the boundary is zero, so using \(Z_1\ge b_1\) or \(Z_1>b_1\) gives the same null probability. With a discrete or materially nonnormal statistic, that equality and the normal quantile calculation may not apply exactly; the boundary then needs calibration using the appropriate null distribution or a justified approximation.

Using the Boundary in a Decision

Once \(b_1\) has been calculated, the interim decision is a comparison on the same statistic scale. A value below the boundary does not meet the prespecified first-look efficacy criterion, even if it appears favorable or would be statistically significant under a different, nonsequential threshold. Conversely, a value at or above the boundary meets this efficacy crossing rule; it does not by itself guarantee that every other design condition for stopping has been satisfied.

Checks, Comparisons, and Limitations

A useful diagnostic is to calculate the tail probability implied by the proposed boundary and compare it with the spending target. If the boundary is \(b_1\), verify that \(1-\Phi(b_1)\) is approximately \(A(t_1)\). This check catches common errors such as using \(\Phi^{-1}(a_1)\), which gives a lower-tail quantile, or using \(\Phi^{-1}(1-\alpha)\) when the first-look target is \(A(t_1)\), not the full planned alpha.

The first boundary also depends on the statistic's null calibration. If its standard error is incorrect, if the statistic has been standardized using a different model than the one assumed in the design, or if the planned information fraction does not match the information actually available, the calculated cutoff may not deliver the intended null crossing probability. A striking discrepancy between the boundary's implied tail probability and the target is a reason to check the tail direction, the spending-function evaluation, the quantile calculation, and the null distribution assumption.

A nominal fixed-look threshold and a spending-based threshold answer different design questions. For example, using \(z_{1-\alpha}=\Phi^{-1}(1-\alpha)\) at the first interim look allocates the full one-sided alpha to that look. When \(A(t_1)<\alpha\), this is less stringent than the spending-based boundary and spends more alpha early than planned. Classical group-sequential boundaries instead use a prespecified set of cutoffs for a planned look schedule; a spending approach calculates the target at the information fraction and calibrates boundaries accordingly. At the first look, the probability inversion is simple, but the first boundary still must use the design's actual target.

The formula is limited to the one-sided upper efficacy boundary and the stated null distribution. It does not calculate a lower futility boundary, represent two-sided spending, or determine a boundary at a later look. Nor does it make an unplanned change in information or analysis timing harmless: the spending target must be evaluated consistently with the design's rules. The introductory material on alpha spending explains why cumulative spending is controlled across looks; here the important first-look point is that, with no earlier looks, the first cumulative target is calibrated by a single null tail probability.