When Regional Advice Appears to Conflict
Two agencies can review the same model and recommend different next steps without disagreeing about the model’s mathematics. One may be addressing whether the model can support dose selection now; another may be asking whether it is sufficient as the sole basis for a final regional recommendation. They may also be considering different populations, endpoints, or consequences of error. The practical task is not to choose which agency is “right” in the abstract. It is to establish what each piece of advice means for the intended decision, identify the evidence or assumption behind the difference, and decide what response is scientifically and regulatorily workable.
This distinction matters because a superficially unified response can leave a real concern unaddressed. Conversely, generating new evidence in response to every difference can be unnecessary if the agencies are answering different questions or if an existing analysis already resolves the issue. Advice records should be read in context: their wording, status, and scope matter. Scientific feedback is not automatically a binding decision, and an applicant should not present one region’s feedback as if it changed another region’s requirements.
The earlier material on context of use and regulatory meeting records provides the basic tools for stating what a model is intended to do and interpreting the scope of feedback. Here, the focus is how to use those tools when the advice differs. Do not treat a positive comment on a model’s technical performance as agreement that it is adequate for every decision, population, or evidentiary role.
A Structured Reconciliation
Begin with the advice as recorded, not with a paraphrase such as “Agency A accepts the model and Agency B rejects it.” Extract each recommendation, its conditions, and its stated rationale. Then compare the underlying questions. A difference in action may arise because one agency is discussing exploratory or supportive use, while another is discussing whether the model can replace or determine a clinical evidence component. A difference may also reflect regional implementation of a shared guideline, distinct populations or labeling questions, or different judgments about the consequences of model error.
Write down the exact development decision, population, model output, and proposed role of that output. Specify whether the model informs, supports, or is proposed to determine the decision.
For each region, record the requested action, any condition attached to it, and the concern or evidence identified. Preserve uncertainty in the wording; do not turn a request for additional justification into a categorical rejection.
Check whether the advice concerns the same age or disease groups, treatment conditions, endpoint, model version, and acceptable risk. Determine whether each agency has seen the same analyses and assumptions.
Identify whether the disagreement persists under common assumptions. Examine relevant subgroup results, uncertainty, sensitivity analyses, and data gaps; explain what each analysis can and cannot establish.
Use an analysis, clarification, targeted data collection, or region-specific action that addresses the stated concern. Record what remains uncertain and how it could affect the decision.
This is a reasoning sequence, not a formula for resolving every disagreement. Its central safeguard is to avoid comparing conclusions before establishing that they concern the same decision. If they do concern the same decision and remain incompatible after the assumptions and evidence are aligned, the difference should be treated as substantive and managed explicitly.
Worked Applications
Worked Example: Supportive Use Versus Sole Reliance
All values and advice in this example are hypothetical. A sponsor is developing fictional medicine NX-17 and uses a population pharmacokinetic model to select a dose for adolescents. In a written record, Region A says the model may support dose selection if the sponsor provides subgroup diagnostics and a sensitivity analysis. Region B requests prospective pharmacokinetic data in younger adolescents before accepting the model as the sole basis for the regional dose recommendation. The sponsor summarizes this as “A accepts the model; B rejects it.” Is that summary justified, and what response should be considered?
First identify the difference in evidentiary role. Region A permits the model to support dose selection subject to further analyses. Region B’s statement concerns using the model as the sole basis for a recommendation in a specified subgroup. Those positions are not necessarily contradictory: supportive use can be acceptable while sole reliance remains unsupported.
Next identify the concern that would need resolution. Suppose Region B’s record points to sparse observed pharmacokinetic data in younger adolescents, while Region A’s record asks for subgroup diagnostics. The sponsor should provide the diagnostics and sensitivity analysis, but should not assume these substitute for the requested data. It should assess whether the analyses answer the concern about extrapolation to younger adolescents. If they do not, a targeted prospective cohort, or a different proposed evidentiary role for the model, may be needed for that region.
A defensible response would clarify the shared model rationale, show the requested subgroup analyses, and state separately what use is proposed in each region. The sponsor would document that Region A’s advice is conditional and that Region B’s concern about sole reliance remains unresolved unless the requested evidence or an agreed alternative addresses it. The conclusion is not that one region’s advice governs the other; it is that the apparent conflict is partly explained by different claims about the model’s role.
Worked Example: A Pooled Result Can Hide a Subgroup Concern
All numbers are hypothetical. Suppose a sponsor’s prespecified internal criterion for model-supported dosing is that at least 90% of simulated patients in each age stratum have predicted exposure within a target range. The simulation results are 94% of 100 older adolescents and 82% of 50 younger adolescents. One agency highlights a pooled result, while another questions the younger-adolescent result. Calculate the pooled fraction and assess the difference.
The number within range is \(100(0.94)=94\) in the older stratum and \(50(0.82)=41\) in the younger stratum. Thus the pooled fraction is
The pooled value is exactly 90%, but it does not meet the stated criterion for every stratum: the younger stratum is at 82%, below 90%. A pooled summary therefore cannot resolve the subgroup concern. It is also important to describe these as fractions of simulated patients under the model and its assumptions, not as observed clinical response rates or as certainty about future patients.
The relevant comparison is now clear: an overall summary may support a population-average argument, while the other agency’s concern addresses the youngest subgroup. The sponsor should examine why the subgroup prediction differs and whether it is robust to plausible assumptions. If the model’s intended use explicitly requires adequate performance in each stratum, the stated internal criterion is not met for younger adolescents. If the intended decision is different, the criterion and its rationale must be reconsidered transparently rather than silently replaced by the pooled result.
Justifying a Proportionate Response
A response should be linked to the uncertainty that drives the difference. For example, a request for clearer context of use may be addressed by a more precise decision statement; a question about a subgroup may call for stratified diagnostics or additional data; and a concern about sensitivity to an assumption may call for a prespecified sensitivity analysis. These actions are not interchangeable. A new analysis is useful only if its result could change the assessment of the concern, and new data are not automatically necessary if the question can be answered adequately with existing evidence.
The logic can be stated as a simple decision check. Let \(C\) denote the proposed decision criterion and \(E\) the evidence under the assumptions relevant to that decision. The model supports the decision only if the evidence, including relevant uncertainty, is adequate for \(C\); a favorable overall estimate alone is not sufficient when the criterion applies to particular subgroups. If a material assumption changes the conclusion, the evidence does not robustly support the decision under that assumption set. This is not a universal regulatory threshold: the criterion and the required degree of support depend on the decision and must be justified in context.
Worked Example: Recording a Sensitivity Result and Remaining Risk
All figures are hypothetical. In the younger-adolescent example, the modeled fraction within the target range is 82% under the primary assumptions. A plausible alternative assumption about maturation raises it to 88%, still below the sponsor’s prespecified 90% criterion. Show how to interpret this sensitivity result and what should be recorded.
The change is \(88\%-82\%=6\) percentage points. The direction is favorable, but the criterion remains unmet because \(88\%<90\%\). Therefore, the sensitivity analysis shows that the estimate depends on the maturation assumption, yet it does not remove the identified concern under the stated criterion. It would be inaccurate to call the conclusion robust merely because the alternative is more favorable.
The record should identify the primary and alternative assumptions, the corresponding results (82% and 88%), the rationale for considering the alternative plausible, and the consequence for the proposed dose decision. It should state that uncertainty about maturation remains material, and specify a plan proportionate to the risk—for example, obtaining targeted pharmacokinetic data or limiting the model’s role for that subgroup pending adequate evidence. The record should also distinguish actions that are agreed from those that are proposed or still require agency discussion.
A useful written reconciliation makes the reasoning auditable. For each region, record the advice and its status, the decision and model role under discussion, the evidence considered, the concern identified, the sponsor’s response, and any outstanding question. If a common analysis addresses both regions, explain that shared response; if regional differences remain, explain the basis for separate actions. State residual risk in terms of the decision it could affect, not merely as “uncertainty remains.”
Diagnostics, Alternatives, and Limits
Several warning signs indicate that reconciliation is going wrong. A sponsor may paraphrase conditional feedback as approval, compare a population-average result with a subgroup requirement, or claim that two agencies agree because they used similar words while addressing different decisions. Other warning signs include analyses that do not test the stated concern, unexplained changes to a decision criterion after seeing results, and a response plan that omits who must act or what evidence would close the issue.
The main alternatives are to seek clarification, generate targeted evidence, revise the model’s proposed role, or maintain distinct regional plans. Clarification is appropriate when the advice is ambiguous or the underlying question is unclear. Additional analysis or data are appropriate when a specific evidentiary gap remains. A narrower supportive role may be reasonable when the evidence informs but does not independently determine the decision. Separate regional actions may be necessary where advice reflects distinct requirements or unresolved judgments. None is a shortcut: each requires an explanation of why it addresses the particular concern.
Reconciliation has limits. A sponsor cannot infer that silence means agreement, that informal scientific advice is binding, or that one region’s position determines another region’s decision. Nor can a document eliminate scientific uncertainty. The goal is to make the decision logic, commitments, and remaining uncertainty explicit, so that development choices do not depend on an overstated impression of consensus.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- One region says a model may support dose selection after sensitivity analysis; another requests clinical data before sole reliance. Why is “one accepts and one rejects” an incomplete summary?
- A model predicts 96% of a larger subgroup and 84% of a smaller subgroup within a target range. Why might a pooled result be insufficient?
- A sensitivity analysis changes a result from 86% to 89%, while the stated criterion is at least 90%. What can and cannot be concluded?
- Name four items that should be included in a record of the sponsor’s response to divergent advice.
Solutions. (1) The regions are discussing different evidentiary roles: supportive use versus sole reliance, and may also have different conditions. (2) If the decision criterion applies to each subgroup, a pooled fraction can conceal that one subgroup falls short; subgroup results and their uncertainty must be assessed directly. (3) The alternative assumption improves the result by 3 percentage points, but 89% remains below 90%, so the criterion is still unmet under that analysis. (4) Appropriate items include the decision and context of use, the advice and its status in each region, the evidence and concern, the response and rationale, planned actions, and residual risk with its possible decision consequence.