Why Spend Alpha Across Interim Looks

Why a Trial Needs an Alpha Budget

A group sequential trial may examine accumulating evidence before the planned final analysis. If the treatment appears effective early, the trial may stop and declare efficacy; otherwise, it may continue to another look. This can be valuable, but it changes the error-control question. The relevant type I error is not the probability of a false rejection at one specified analysis. It is the probability of a false efficacy conclusion at any analysis in the trial.

Let \(H_0\) denote the null hypothesis of no treatment effect. Under \(H_0\), a one-sided efficacy rule rejects when a test statistic crosses its efficacy boundary. With \(K\) planned looks, denote by \(R_k\) the event that the rule rejects at look \(k\). The trial’s overall type I error is \(P_{H_0}(R_1\cup\cdots\cup R_K)\). The events can be dependent because interim and later statistics are calculated from overlapping data. Therefore, the overall error is not generally the sum of the nominal significance levels at the looks, nor is it generally equal to the significance level of one look.

The central idea of alpha spending is to treat \(\alpha\) as a trial-level budget. As information accumulates, the design permits some of this budget to be used for the possibility of an early false efficacy conclusion. The rule must also preserve enough control for analyses that have not yet occurred. The budget is about the probability of a false rejection over the entire monitoring process, not a quantity that is literally subtracted from the observed data.

How Repeated Testing Uses the Budget

The union of rejection events can be partitioned according to the first look at which rejection occurs. Let \(E_k\) be the event that the procedure first rejects \(H_0\) at look \(k\). These events are mutually exclusive: a trial that first rejects at look 1 cannot first reject again at look 2. They also cover every way the trial can reject. Consequently,

This decomposition gives a useful interpretation of spending. Each look adds a chance of a first false crossing, and the sum of those chances is the overall type I error. The probability of crossing at a later look must be evaluated among paths that have not already crossed; it is not simply the unconditional significance level assigned to a single test. The earlier material on null-path derivations uses this same partition to calculate overall alpha for a specified boundary design.

A short calculation shows why repeatedly using a single-look level can be problematic. Suppose, only for illustration, that two tests under \(H_0\) were independent and each had type I error 0.05. The probability of no false rejection at either test would be \(0.95\times0.95\). Thus the probability of at least one false rejection would be

This is 9.75%, not 5%. Actual interim statistics are usually dependent, so this independence calculation is not a group sequential design formula. It isolates the basic problem: two opportunities to reject can produce a larger probability of at least one false rejection. Dependence changes the size of the increase, but does not make repeated testing automatically safe. A union bound, \(P_{H_0}(\cup_k R_k)\leq\sum_k P_{H_0}(R_k)\), gives a general conservative control principle; a calibrated sequential design can use the joint null distribution more efficiently.

From Per-Look Testing to Cumulative Spending

A fixed per-look rule might say, for example, “test at 0.01 at each interim analysis.” That instruction does not by itself describe how much of the trial-level error budget has been used after a particular amount of information has accrued. Nor does it naturally adapt if the timing or number of analyses changes. A spending rule instead describes how much cumulative type I error the design allows by a given point in the trial.

The information fraction is the information available at an analysis divided by the information planned at the maximum analysis. In a simple design with fixed variance and sample size proportional to information, it may be approximated by the fraction of the maximum sample size accrued. Information time is not necessarily the same as calendar time: an analysis scheduled after six months is not automatically at the same information fraction in every trial.

A cumulative rule gives meaning to an analysis at an unplanned information time. If a look moves from one information fraction to another, the cumulative allowance can be read at the information actually accrued, rather than being tied only to the ordinal label “look 2.” This addresses a key weakness of a rule that assigns a separate fixed allowance to each planned look. It does not mean that one may change the design arbitrarily after seeing results: the boundary procedure, monitoring plan, and statistical assumptions still matter.

Why Index the Rule to Information Time?

Interim schedules can change for operational reasons. Recruitment may be slower or faster than anticipated, an analysis may be delayed, or the team may decide that a planned look should occur after more information has accrued. A rule defined only as an allowance “at look 1, look 2, and look 3” is tied to the original schedule. If a look is added, removed, or moved, the design team must determine how the error budget is to be reassigned and whether the original boundaries still control overall alpha.

An information-time rule makes the intended cumulative allowance a function of information accrued rather than just the number of completed looks. The following simple function is illustrative only; it is not a recommendation for a particular trial. Suppose the target is \(\alpha=0.05\), and an example rule is \(A(t)=0.05t^2\). At information fractions 0.25, 0.50, and 1.00, the cumulative allowances are 0.003125, 0.0125, and 0.05, respectively. If operational circumstances instead place analyses at 0.30, 0.70, and 1.00, the corresponding values from the same illustrative rule are 0.0045, 0.0245, and 0.05. The rule can therefore be evaluated at the actual information fractions.

These values are cumulative allowances, not the nominal significance levels for separate tests, and not test-statistic cutoffs. Constructing boundaries requires accounting for the joint distribution of the sequential statistics and the probability of having crossed earlier boundaries. The later material on cumulative spending and boundary construction develops those steps.

Alternatives, Diagnostics, and Limits

A design with a fixed, pre-specified schedule can use classical calibrated boundaries: the cutoffs are chosen jointly so the overall null rejection probability is the desired \(\alpha\) for that schedule. This can be effective when the planned looks and assumptions remain appropriate. If the schedule changes, the original calibration may no longer apply, and a new analysis of the design may be needed. A conservative alternative is to use a multiplicity adjustment such as Bonferroni, limiting the sum of per-look error probabilities to \(\alpha\). This can provide control without relying on a detailed dependence calculation, but may spend the budget inefficiently. Spending rules offer a way to express cumulative error control over information time; they still require a valid procedure for converting that rule into boundaries.

Several checks help identify when the planned error control may be in danger:

• Compare the actual and planned information times. Record information accrued at each analysis, not only its calendar date or look number.
• Count every efficacy opportunity. An extra unplanned look can increase the chance of a false crossing if it is not included in the error-control procedure.
• Check what triggered the analysis. A schedule change based on external operations is different from one triggered by emerging treatment results. Data-dependent timing may require additional justification; a spending curve does not automatically validate any stopping or analysis rule.
• Separate error control from power. A rule that controls overall type I error does not, by itself, ensure adequate power, a desired expected sample size, or acceptable operating characteristics under clinically relevant alternatives.
• Revisit the assumptions and calibration. Changes in the statistical model, information estimates, or analysis method can affect the joint null distribution used to calibrate boundaries.

The main limitation is that “spend alpha over time” is a design principle, not a complete testing algorithm. The spending rule states how the cumulative error allowance is indexed; it does not specify the test statistic, the boundary at each look, or the consequences of every possible operational deviation. Those must be handled by the full sequential design and analysis plan.

A Common Mistake

Wrong reasoning: “The trial has a 5% alpha level, so each interim look can use a 5% test; the 5% applies separately every time.” This confuses a single-analysis error rate with the probability of at least one false rejection over the whole trial. Repeated testing creates multiple opportunities to cross an efficacy boundary.

Correction: Define the trial-level rejection event as crossing at any look, then control its probability under \(H_0\) at the chosen \(\alpha\). Calibrated fixed boundaries or a suitable spending-based procedure can do this. Do not infer overall control from the nominal level at one look, and do not assume that simply dividing \(\alpha\) equally is necessary or optimal.