Rules for Triggering a Redesign

Why a Redesign Trigger Must Be Explicit

An adaptive design needs a rule for when to reconsider its current plan. Without one, the decision can drift: a team may redesign after an inconvenient result, or continue with a plan that no longer fits its purpose or constraints. A redesign trigger makes the decision reproducible by specifying what will be monitored, the threshold that prompts action, and what action follows.

A trigger is a design-management rule, not a statistical test of a scientific hypothesis. It does not establish that a model is true, that an effect exists, or that a parameter has crossed a clinically meaningful boundary. Instead, it identifies conditions under which the expected value or feasibility of reconsidering the design is judged sufficient. As in sequential design and interim updating, the trigger should be evaluated using the information actually available at that interim.

A trigger may monitor posterior uncertainty about a parameter, predicted information gain from a candidate redesign, or a change in operational conditions. These measures answer different questions. Posterior uncertainty asks whether current knowledge is precise enough; predicted information gain asks whether changing the plan is expected to add useful information; an operational trigger asks whether the current plan remains feasible. They should not be treated as interchangeable.

Specify the Rule Before Observing the Interim Data

Write the rule in terms of a quantity that can be calculated or recorded when the decision is due. For example, let \(\theta\) be a target model parameter and \(p(\theta\mid D_t)\) its posterior distribution at interim \(t\). A simple uncertainty rule is to trigger redesign when the posterior standard deviation exceeds a prespecified maximum:

Here \(\tau_U\) is expressed in the units of \(\theta\). If the parameter is on a log scale, the threshold is in log units; it should not be presented as though it were on the original scale. For a parameter vector, a rule could monitor one prespecified component or use scaled uncertainty measures, but the selected summary and scaling must be stated. A multivariate posterior can be uncertain in directions that a single marginal standard deviation does not reveal.

A trigger based on predicted information gain uses a different quantity. Let \(d_1\) be a candidate redesign and \(d_0\) the current remaining plan. The expected information gain can be defined as the expected reduction in posterior uncertainty from the observations planned under a design. A direct comparison is the incremental expected information gain of \(d_1\) over \(d_0\). A rule might trigger reconsideration when that increment exceeds a prespecified threshold. The threshold must be interpreted in the information measure's units, such as nats for entropy measured with natural logarithms.

An operational rule should specify a measurable change, such as a reduction in available sample slots, a change to allowable sampling windows, or loss of access to a required assay. A statement such as “redesign if operations become difficult” is not reproducible. State the baseline, the monitored operational quantity, the threshold, and the consequence of crossing it.

Why Expected Information Gain Has a Useful Gaussian Form

A simple derivation illustrates how a model can translate measurement precision into a trigger quantity. Suppose the current posterior for a scalar parameter is normal with variance \(v\). Under a candidate design, suppose one new observation has a normal measurement model with known variance \(r\), independent of the current data conditional on the parameter. Assume the observation mean is the parameter itself, and that both variances are known. Normal-normal updating gives posterior variance \(v_{\text{new}}=(1/v+1/r)^{-1}\). Because these variances do not depend on the observed value, the expected entropy reduction equals the entropy reduction for any observation.

The first equality follows from the entropy of a normal distribution with variance \(w\), which is \(\frac{1}{2}\log(2\pi e w)\). The final expression is nonnegative because \(v>0\) and \(r>0\). It grows when current uncertainty \(v\) is larger or the planned observation is more precise (smaller \(r\)). This calculation depends on the stated Gaussian assumptions. In a nonlinear pharmacometric model, with nuisance parameters, non-Gaussian posteriors, or parameter-dependent observation variance, expected information gain generally requires averaging over possible future data under the model rather than using this shortcut.

Compare the Candidate Plan With the Current Plan

A threshold on the information from a candidate plan alone can be misleading: even a useful candidate may add little beyond what the current plan would already provide. A more decision-relevant rule compares plans on the same basis. For example, trigger reconsideration if the candidate's predicted information gain exceeds that of the current plan by at least \(\tau_I\):

The comparison must use the same posterior at interim \(t\), parameter target, information measure, and horizon. Otherwise, the difference may reflect inconsistent calculations rather than the redesign. Also specify whether the threshold accounts for costs or constraints. Information gain is not automatically equivalent to clinical value: it quantifies learning under a model, not the utility of a decision.

Operational Triggers and Their Consequences

Operational triggers are especially useful because a plan can become infeasible even if the posterior is sufficiently precise and a redesign offers little expected information gain. Predefine how the operational change will be measured. An absolute threshold, such as a loss of at least two sampling slots, may be appropriate when every slot has a similar role. A relative threshold, such as a loss of at least 15% of planned capacity, is more comparable across study sizes. If the consequence is loss of a required sampling window or assay, a binary trigger may be more suitable than a percentage.

Crossing an operational threshold should prompt a specified action, commonly a feasibility check and constrained re-optimization of the remaining plan. It does not guarantee that a suitable alternative exists. The team should record the revised feasible set, what must be sacrificed, and whether the design still meets its prespecified aims. If no feasible plan remains, the trigger has identified a problem to escalate; it has not solved it.

Choose Thresholds by Considering Their Consequences

A low uncertainty threshold or a small information-gain threshold makes redesign more likely. This may improve the chance of responding to inadequate precision or a genuinely better plan, but it can also consume time and resources on changes with little practical benefit. A high threshold reduces unnecessary redesigns but can leave the study with substantial uncertainty or a missed opportunity. Operational thresholds have analogous trade-offs: an overly sensitive rule may respond to minor fluctuations, whereas an overly permissive one may allow feasibility problems to accumulate.

Thresholds should be selected before examining the interim result that will determine action. A practical assessment is to apply candidate rules across plausible parameter values, data patterns, and operational scenarios, then examine how often they trigger, how many resources they use, and how the resulting designs perform. This connects the trigger to robust design considerations without pretending that one threshold is universally optimal. The threshold and its consequences depend on the objective, remaining budget, model assumptions, and the cost of changing course.

A key diagnostic is to check whether the monitored quantity is stable and defined consistently at each interim. For an uncertainty trigger, confirm the parameter scale, posterior summary, and treatment of correlations or nuisance parameters. For an information-gain trigger, verify that the current and candidate plans use the same posterior and future-data assumptions. For an operational trigger, verify the baseline and count what is actually available. If the rule's inputs cannot be calculated reproducibly, its apparent precision is misleading.

Common mistake: “The candidate plan has positive expected information gain, so we should redesign.” This reasoning ignores what the current plan would have learned and the cost of changing plans. The correction is to compare candidate and current plans on a common basis, use a prespecified threshold for the incremental benefit, and separately assess constraints and practical value. A trigger is a decision rule, not a claim that every positive gain warrants action.