Parallel Endpoint Testing With Alpha Recycling

Why Recycle Alpha Across Parallel Endpoints?

When a trial evaluates several endpoints, a separate test at level \(\alpha\) for each endpoint can inflate the probability of making at least one false rejection. A weighted graph offers a prespecified way to allocate the familywise error rate across endpoint hypotheses. If an endpoint null hypothesis is formally rejected, the graph can transfer some or all of its allocated level to other hypotheses that remain unresolved.

With interim analyses, there are two distinct allocations to track. The graph allocates error across endpoints; a group sequential design allocates error across looks within an endpoint test. Recycling changes the endpoint allocation. It does not mean that an endpoint can be tested repeatedly at its new level without a valid sequential calibration.

Definition: A graphical alpha-recycling procedure specifies an initial local significance level for each endpoint hypothesis and transition weights from each hypothesis to the others. When a hypothesis is formally rejected, its local level is removed and transferred to the remaining hypotheses according to those weights. A rejection is formal only when the endpoint’s prespecified test, including its interim-look rules, supports rejection.

Setting Up the Graph and Sequential Tests

Suppose a family contains \(m\) endpoint null hypotheses \(H_1,\ldots,H_m\), and the familywise significance level is \(\alpha\). Assign nonnegative initial weights \(w_i\) whose sum is at most one. The initial local level for endpoint \(i\) is \(\alpha_i=\alpha w_i\). For each directed edge from endpoint \(i\) to endpoint \(j\), specify a nonnegative transition weight \(g_{ij}\). The outgoing weights from a node must sum to at most one.

If \(H_i\) is rejected, endpoint \(j\)'s local level increases by \(\alpha_i g_{ij}\). The rejected node is removed, and any later transitions follow the prespecified graph update rule. For a simple two-endpoint graph with a single directed edge, this calculation is direct: the edge transfers the specified fraction of the rejected endpoint’s local level to the other endpoint.

For each endpoint, use a valid group sequential test at its current local level. Its planned looks, information times, test statistic, and boundary or alpha-spending calibration must be appropriate for that endpoint. Equivalently, one can compare a valid sequentially adjusted p-value with the endpoint’s current local level. The sequential adjustment accounts for the planned repeated looks; an ordinary p-value from one look generally does not.

The basic assumptions are that the endpoint family and graph are prespecified, each endpoint’s sequential test is valid under its own null at the assigned local level, and the rule for changing allocations during monitoring is specified and valid. Graphical weighted Bonferroni procedures can control familywise error under arbitrary dependence between endpoints when their component tests are valid. Dependence may affect power and the joint distribution of results, but it does not license replacing a valid local test with an unadjusted one.

Key distinction: A graph allocation is a local significance level for an endpoint test; an alpha-spending function describes how that level is used over interim looks. After a graph transition during monitoring, the remaining sequential testing must follow the prespecified valid rule for the updated allocation and the information already observed.

A Short Justification for Alpha Bookkeeping

The graph’s local levels do not grow without bound. If endpoint \(i\) is rejected, its level \(\alpha_i\) is removed. The total level transferred from it is \(\alpha_i\sum_{j\ne i}g_{ij}\), which cannot exceed \(\alpha_i\) because the outgoing weights sum to at most one. Thus, after the update,

\sum_{j\ne i}\left(\alpha_j+\alpha_i g_{ij}\right) = \sum_{j=1}^{m}\alpha_j-\alpha_i +\alpha_i\sum_{j\ne i}g_{ij} \leq \sum_{j=1}^{m}\alpha_j \leq \alpha.

This conservation check helps detect incorrect updates, but it is not by itself a proof of strong familywise error control: the probability of a false rejection also depends on how the individual tests and rejection paths are defined. The closure principle and graphical error-control results provide that justification when their conditions hold. In this setting, the group sequential component must also be calibrated so that its test remains valid at the local level assigned to it.

Worked Example: Recycling Between Two Parallel Endpoints

Consider two endpoint null hypotheses in a hypothetical trial, with familywise \(\alpha=0.05\). The initial weights are \(w_1=0.60\) and \(w_2=0.40\). All outgoing weight from \(H_1\) goes to \(H_2\), and all outgoing weight from \(H_2\) goes to \(H_1\). At the planned final look, valid group sequential analyses give sequentially adjusted p-values \(0.018\) for endpoint 1 and \(0.034\) for endpoint 2.

The initial local levels are \(0.05(0.60)=0.030\) for \(H_1\) and \(0.05(0.40)=0.020\) for \(H_2\). Since \(0.018\leq0.030\), reject \(H_1\). The graph transfers all of its \(0.030\) local level to \(H_2\), so the updated level for \(H_2\) is \(0.020+0.030=0.050\).

Now compare endpoint 2’s sequentially adjusted p-value with its updated level: \(0.034\leq0.050\). The procedure therefore rejects \(H_2\) as well. These conclusions rely on the supplied p-values being valid for the planned sequential tests and on the graph transition being part of the prespecified procedure.

Applying Several Transitions

With more endpoints, each rejection can change several local levels, and the order of formal rejections matters. A hypothesis that has not been rejected retains its current allocation; a rejected hypothesis leaves the set of hypotheses still eligible for rejection. In an actual group sequential design, the procedure must also specify how any transition during monitoring is implemented with the accrued information and remaining looks.

Worked Example: Three Endpoints and Two Transitions

Suppose a hypothetical trial tests \(H_1,H_2,H_3\) at familywise \(\alpha=0.05\). Initial weights are \(0.50,0.30,0.20\), respectively. The graph sends all of \(H_1\)'s level to \(H_2\) and \(H_3\), with transition weights \(0.60\) and \(0.40\). It sends all of \(H_2\)'s level to \(H_3\). Valid sequentially adjusted p-values at the planned analysis are \(0.012,0.024,0.031\).

The initial local levels are \(0.05(0.50)=0.025\), \(0.05(0.30)=0.015\), and \(0.05(0.20)=0.010\). For \(H_1\), \(0.012\leq0.025\), so reject it. The transfers are \(0.025(0.60)=0.015\) to \(H_2\) and \(0.025(0.40)=0.010\) to \(H_3\). Their updated levels are \(0.015+0.015=0.030\) and \(0.010+0.010=0.020\).

For \(H_2\), \(0.024\leq0.030\), so reject it. Its full local level of \(0.030\) transfers to \(H_3\), whose updated level becomes \(0.020+0.030=0.050\). Since \(0.031\leq0.050\), reject \(H_3\). The total level remains \(0.050\) throughout the updates: after the first rejection, \(0.030+0.020=0.050\), and after the second, \(0.050\) remains for \(H_3\).

Interim Looks: What a Transition Does and Does Not Do

A graph transition may occur when an endpoint is rejected at an interim look, rather than at the final analysis. The rejection must meet the endpoint’s valid group sequential rule at the level currently assigned to it. Once the transition occurs, later testing of the remaining endpoints must use the prespecified procedure for the updated local levels and the observed data history. That rule may use a recalibrated spending schedule or another valid sequential construction.

The graph does not, on its own, tell the trial team how to spend newly allocated level over the remaining looks. Nor does a transition automatically make the entire updated level available as though no interim data had been observed. The endpoint-level allocation and the within-endpoint sequential calibration must work together. This distinction is essential when the transition happens before all planned looks are complete.

Worked Example: Checking an Interim Transition

In a hypothetical two-endpoint trial with \(\alpha=0.05\), each endpoint initially receives \(0.025\), and a rejection of \(H_1\) transfers its full level to \(H_2\). At an interim look, the prespecified sequential test formally rejects \(H_1\). The graph therefore changes \(H_2\)'s local level from \(0.025\) to \(0.050\).

Suppose the design’s prespecified rule for \(H_2\), after this transition and given the interim history, reports a valid sequentially adjusted p-value of \(0.041\). Since \(0.041\leq0.050\), the procedure rejects \(H_2\) under that rule. The conclusion is justified by both the graph update and the stated sequential calibration.

By contrast, an ordinary interim p-value of \(0.041\) would not be enough information to justify that conclusion. It may not account for repeated looks, and it may not be calibrated for the transition history. The graph determines the new local level; it does not validate an otherwise unadjusted interim test.

Diagnostics, Alternatives, and Limitations

Before interpreting results, check the following aspects of the implementation:

Initial allocation: confirm that all weights are nonnegative and sum to at most one, so the initial local levels sum to at most \(\alpha\).
Transition arithmetic: verify that the rejected endpoint’s level is removed and that each recipient gets the specified fraction, not the whole level unless its edge weight is one.
Formal rejection: confirm that a transition follows rejection under the planned endpoint test, not a promising trend, nominal interim p-value, or futility finding.
Sequential validity: check that local tests account for all planned looks, the endpoint’s information pattern, and the updated allocation when a transition occurs during monitoring.
Recorded decisions: document the look, local level, test result, and graph update after each formal rejection. This makes it possible to reproduce the path through the graph.

A common mistake is to treat recycled alpha as a fresh, unrestricted allowance: “\(H_1\) was rejected, so \(H_2\) can now be tested at 0.05 with its ordinary interim p-value.” The correction is to use the updated local level only within the prespecified valid sequential test for \(H_2\). A second mistake is to transfer alpha after a failure to reject because the endpoint appears unlikely to succeed. Failure to reject does not establish the null hypothesis, so it does not trigger the graph’s rejection-based transition.

Compared with a fixed allocation that never recycles alpha, a graph can give more testing level to endpoints that remain after a rejection, potentially increasing the opportunity to reject them. In contrast, a hierarchical strategy makes claims in a specified order and may restrict testing later endpoints until earlier conditions are met. A graph is useful when the intended priorities are better expressed through initial weights and transfer paths. Its limitations are equally important: the graph encodes a chosen testing policy, does not guarantee high power for every endpoint, and cannot repair invalid local tests or an unsuitable interim implementation.

Key takeaway: Alpha recycling is a prespecified update to endpoint-level significance allocations following a formal rejection. In a group sequential design, every endpoint test—including any test after a graph transition—must remain valid across its interim looks.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

At familywise \(\alpha=0.05\), two endpoints start with weights \(0.70\) and \(0.30\). If the first endpoint is rejected and transfers all its level to the second, what is the second endpoint’s updated local level?
Three hypotheses have local levels \(0.020,0.015,0.010\). The first is rejected and transfers 25% of its level to the second and 75% to the third. What are the updated levels of the two unrejected hypotheses?
Why is a nominal interim p-value not necessarily sufficient to trigger alpha recycling?
If an endpoint is not rejected at an interim look, may its graph level automatically be transferred to another endpoint? Explain.

Solutions. (1) The first endpoint starts at \(0.05(0.70)=0.035\), and the second at \(0.05(0.30)=0.015\). After the full transfer, the second has \(0.015+0.035=0.050\). (2) The second receives \(0.020(0.25)=0.005\), giving \(0.015+0.005=0.020\). The third receives \(0.020(0.75)=0.015\), giving \(0.010+0.015=0.025\). (3) A nominal interim p-value may not account for repeated looks or the rule for testing after a graph transition. A formal rejection must follow the valid sequential procedure. (4) No. The described transitions are triggered by formal rejection, not by nonrejection or futility; any different rule would need to be specified and justified separately.