Why Endpoint Alpha Allocation Changes Power
A multi-endpoint trial can have one overall type I error budget but different clinical priorities across endpoints. Giving one endpoint a larger share of that budget generally makes its rejection boundary less stringent and its marginal power higher; the tradeoff is that less alpha remains available to other endpoints. In a group sequential design, this allocation interacts with the interim boundaries: each endpoint’s local error budget must cover its opportunities to reject across the planned looks.
The aim is not to maximize every endpoint’s power independently. Rather, it is to choose and document an allocation that reflects the clinical objectives while preserving familywise error control. The calculations below first use a single-look normal test, which makes the marginal power tradeoff transparent. The same principle applies to a sequential test, but its crossing probability must account for all the looks and the specified boundary rule.
For strong familywise error control, the probability of at least one false rejection must be no greater than \(\alpha\) under every configuration of true and false endpoint null hypotheses. One direct approach is to use valid local tests with levels \(\alpha_j\) and ensure that the sum of the local error budgets used for testing is no greater than \(\alpha\). The union bound then gives familywise error at most \(\sum_j\alpha_j\leq\alpha\), without requiring independence among endpoints.
In a sequential design, “local level” refers to the probability of crossing that endpoint’s boundary at any permitted look under its null, not the nominal level at each individual look. A group sequential boundary or spending rule must be calibrated to the endpoint’s local budget. If an allocation rule later transfers alpha after a rejection, the complete transition rule—not merely the initial weights—must also be valid for strong familywise error control.
From Local Alpha to Marginal Power
Suppose a single-look, one-sided test uses a standardized statistic \(Z_j\). Under the null, assume \(Z_j\sim N(0,1)\); under a specified alternative, assume \(Z_j\sim N(\delta_j,1)\), where \(\delta_j>0\) is the standardized effect in the direction of benefit. A level-\(a_j\) test rejects when \(Z_j>z_{1-a_j}\), where \(z_q\) is the \(q\)-quantile of the standard normal distribution. Its power is therefore
This formula shows the tradeoff directly. Increasing \(a_j\) lowers the critical value \(z_{1-a_j}\), and hence raises power for a fixed positive \(\delta_j\). It does not make the endpoint’s treatment effect larger; it changes the rejection rule. For a group sequential test, the corresponding marginal power is the probability of crossing at least one of the endpoint’s sequential boundaries. That probability depends on the complete boundary sequence and the joint distribution of the endpoint’s statistics across looks.
Worked Example: Allocating the Initial Error Budget
Consider a hypothetical two-endpoint trial with familywise level \(\alpha=0.05\). The team assigns initial weights \(w_1=0.60\) and \(w_2=0.40\). Find the local budgets and check the basic error-budget condition.
The budgets are \(\alpha_1=w_1\alpha=0.60(0.05)=0.030\) and \(\alpha_2=w_2\alpha=0.40(0.05)=0.020\). The weights sum to \(0.60+0.40=1.00\), so the local budgets sum to \(0.030+0.020=0.050\).
If each endpoint’s test—including all its planned sequential looks—is valid at its assigned local level, the union bound gives familywise error no greater than \(0.030+0.020=0.050\). This check does not establish the validity of any particular interim boundary or recycling rule; those must be specified and justified separately.
Worked Example: Comparing Marginal Power Under Two Allocations
Suppose, hypothetically, that two one-sided endpoint statistics have distributions \(Z_1\sim N(2.50,1)\) and \(Z_2\sim N(2.30,1)\) under their respective alternatives. Compare the allocation \(0.030,0.020\) with an equal allocation of \(0.025,0.025\). These are fixed-look calculations, not claims about a multi-look boundary.
With local level \(0.030\), the critical value is \(z_{0.97}\approx1.8808\). Endpoint 1 has power \( \Phi(2.50-1.8808)=\Phi(0.6192)\approx0.7321\). At local level \(0.020\), \(z_{0.98}\approx2.0537\), so endpoint 2 has power \( \Phi(2.30-2.0537)=\Phi(0.2463)\approx0.5973\).
With equal local levels \(0.025\), the critical value for each test is \(z_{0.975}\approx1.9600\). Endpoint 1’s power is \( \Phi(2.50-1.9600)=\Phi(0.5400)\approx0.7054\); endpoint 2’s power is \( \Phi(2.30-1.9600)=\Phi(0.3400)\approx0.6331\).
Thus the \(0.030,0.020\) allocation raises endpoint 1’s marginal power relative to equal allocation, while lowering endpoint 2’s. The familywise budget is unchanged. The conclusion concerns these specified standardized effects and one-look tests; another effect profile or sequential boundary could yield different power differences.
Transition Rules and Power After Rejection
Initial weights are not the only way to prioritize endpoints. A prespecified transition rule can allow alpha associated with a rejected hypothesis to be used for another endpoint. This can improve the latter endpoint’s chance of rejection on paths where the transfer is available. Its marginal power is then an average over paths both with and without a transfer. Consequently, the power gain is not generally equal to the gain from simply testing at the larger level on every path.
A transition rule must be assessed as a whole. The testing thresholds before and after each possible rejection need to obey a procedure that controls familywise error under every configuration of true nulls. Alpha cannot be spent twice: a budget used for one rejection can be passed on only as allowed by the prespecified rule, and a nominal threshold at a particular look is not by itself proof of validity.
Worked Example: A Prespecified Transfer to a Secondary Endpoint
Consider a hypothetical single-look illustration with \(\alpha=0.05\). Initially, endpoint 1 is tested at \(0.030\) and endpoint 2 at \(0.020\). If endpoint 1 is rejected, the rule permits endpoint 2 to be tested at \(0.050\); otherwise endpoint 2 remains at \(0.020\). Assume independent endpoint p-values for this calculation, and suppose the endpoint alternatives give the powers in the preceding example: endpoint 1 has power \(0.7321\), while endpoint 2 has power \(0.5973\) at \(0.020\) and \(0.7438\) at \(0.050\).
Endpoint 2 is rejected either when its p-value is at most \(0.020\), or when it is between \(0.020\) and \(0.050\) and endpoint 1 is rejected. By independence, its marginal power is
The transfer therefore increases endpoint 2’s power above its initial-level value of \(0.5973\), but not to \(0.7438\), because the expanded threshold is available only when endpoint 1 rejects. In this illustration the rule also preserves familywise control: if endpoint 1 is true and endpoint 2 is false, the only possible false rejection is endpoint 1, whose probability is at most \(0.030\). If endpoint 1 is false and endpoint 2 is true, endpoint 2’s rejection event is contained in the event that its p-value is at most \(0.050\), so its probability is at most \(0.050\). When both nulls are true, endpoint 2 can be falsely rejected after a transfer only on a path where endpoint 1 is already falsely rejected; the familywise event is contained in the union of rejection of endpoint 1 at \(0.030\) and rejection of endpoint 2 at its initial \(0.020\), also bounded by \(0.050\). These arguments use valid null p-values; the power arithmetic additionally used independence.
Planning and Interpretation in a Group Sequential Design
For endpoint \(j\), let \(Z_{jk}\) be its test statistic at look \(k\), \(k=1,\ldots,K\), and let \(b_{jk}\) be its prespecified efficacy boundary. The endpoint is rejected if it crosses an applicable boundary. Its marginal power under an alternative is
The boundary sequence must be calibrated so that the probability of any crossing under the endpoint null is no greater than its available local error budget. For an overall allocation with no recycling, this means calibrating each endpoint to its assigned \(\alpha_j\). With recycling, the available budget and threshold can depend on earlier rejections, so marginal power must average over those possible paths under the assumed alternative configuration.
The standard planning model should be stated before interpreting these probabilities. In a common large-sample formulation, endpoint statistics are standardized so that their null marginal distributions are approximately standard normal, and their means under specified alternatives encode the effect and information at each look. The interim schedule, endpoint estimands, boundary rules, and alpha allocation are fixed in advance. Local sequential tests must be valid under their nulls; any calculation that assumes independence between endpoints or between a rejection path and another statistic must state and justify that assumption. In realistic designs, simulation or numerical integration may be needed to obtain the crossing probabilities.
Marginal power is not the probability of rejecting every required endpoint, nor the probability of making at least one clinically useful claim. Those are joint or family-level operating characteristics. Likewise, increasing an endpoint’s local alpha does not guarantee a particular increase in marginal power: the effect size, information, number and timing of looks, boundary shape, and transition paths all matter.
Identify which endpoint claims are essential, which are supportive, and what power is important under clinically relevant alternative configurations.
Record the initial weights, local budgets, and every condition under which alpha can be transferred. Confirm that the complete rule is valid for strong familywise error control.
Use the planned sequential boundaries and alternative model to calculate each endpoint’s crossing probability. Report the assumptions and distinguish marginal from joint power.
Compare clinically plausible allocations and alternatives, including scenarios where a hoped-for transfer is unavailable because an earlier endpoint does not reject.
Diagnostics, Alternatives, and Limitations
Power calculations are especially sensitive to errors in the design specification. Check that endpoint weights are nonnegative and sum to no more than one, each boundary is calibrated to the correct local budget, and a transferred budget is not counted both at its source and destination. For sequential calculations, verify that the number and timing of looks in the calculation match the design. A useful diagnostic is to evaluate power across several plausible effect configurations, not just the configuration in which every endpoint has its target effect.
A fixed allocation is transparent and easy to communicate, but it may spend alpha on an endpoint that is less clinically important. A hierarchical or gatekeeping strategy can reserve access to a secondary claim until a primary claim succeeds. A graphical recycling strategy can allow rejected hypotheses’ budgets to flow to other endpoints. These alternatives can all preserve familywise error control when correctly specified, but they produce different rejection paths and marginal powers. The most appropriate procedure depends on the claims the trial is intended to support, not solely on which option gives the largest number for one endpoint.
Common limitations include reliance on a particular effect-size model, approximate normality, and assumed dependence structures for joint power calculations. Further, a nominally higher power for one endpoint may come at a clinically unacceptable reduction elsewhere. Weight selection is therefore a design decision that should be justified clinically and documented before unblinded interim results are available; it should not be adjusted informally in response to observed trends.
- Mistake: “Endpoint 2 has a \(0.05\) threshold after recycling, so its marginal power is its power at level \(0.05\).” Correction: That threshold applies only on paths where the transfer condition holds; marginal power averages over both transferred and untransferred paths.
- Mistake: “Each endpoint’s interim nominal alpha can be compared directly with the overall \(\alpha\).” Correction: For a sequential test, control concerns the probability of crossing any boundary under the null, and the full boundary sequence must be calibrated.
- Mistake: “A valid initial allocation automatically makes any later reallocation valid.” Correction: Strong familywise control depends on the entire transition rule and its behavior under every configuration of true nulls.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- With \(\alpha=0.05\) and weights \(0.70\) and \(0.30\), what are the two initial local budgets, and do they exhaust the familywise budget?
- For a one-sided normal test with standardized alternative mean \(\delta=2.0\), compare power at local levels \(0.05\) and \(0.01\). Use \(z_{0.95}=1.6449\) and \(z_{0.99}=2.3263\).
- Why is power at the expanded threshold not generally the marginal power under a recycling rule?
- What must be calibrated in a group sequential endpoint test: the nominal level at each look, or the probability of crossing any boundary under the null?
Solutions. (1) The budgets are \(0.70(0.05)=0.035\) and \(0.30(0.05)=0.015\); they sum to \(0.050\). (2) At \(0.05\), power is \(\Phi(2.0-1.6449)=\Phi(0.3551)\approx0.6388\). At \(0.01\), power is \(\Phi(2.0-2.3263)=\Phi(-0.3263)\approx0.3721\). The stricter level has lower power. (3) The expanded threshold is available only on paths satisfying the transfer condition, so the marginal probability averages over paths with and without transfer. (4) Calibrate the probability of crossing any boundary under the null to the endpoint’s available local error budget.