From Request to Answer
A regulatory information request may ask for a rationale, a numerical result, an explanation of an analysis decision, or some combination of these. A technically correct result can still fail to answer the request if it addresses a nearby question instead, omits the basis for a calculation, or leaves its practical implications unstated. A focused response makes the path from the exact question to the evidence and conclusion easy to follow.
The aim is not to reproduce an entire population PK report in miniature. It is to supply enough context and traceable evidence for a reviewer to understand what was done, what was found, and what can reasonably be concluded. The response should distinguish a reported analysis result from an interpretation of that result and from any uncertainty that remains.
Start by turning the request into a set of answerable parts. Preserve its scope: if the regulator asks why records were excluded and whether the exclusions affected a result, both parts need an answer. Do not silently substitute a broader question, such as whether the overall model is adequate, for the specific issue raised. If a term, population, time period, or requested comparison is ambiguous, state the interpretation used and seek clarification where needed.
Identify the requested population, analysis, comparison, time period, and form of answer. Separate a multi-part request into explicit questions.
Name the data cutoff or analysis set, relevant model or analysis version, and any decision that governs the result. Explain any departure from the requested or planned approach.
Give the result and enough detail to locate or reproduce it: counts, denominators, units, comparison definitions, and references to the relevant output or analysis.
Explain what the result supports in relation to the request. Identify what it does not establish, including uncertainty that materially affects interpretation.
The analysis basis matters because a numerical answer is only interpretable in context. Relevant details may include which subjects or observations were used, how records were classified, what model version generated an estimate, and whether a result is observed or simulated. Include the details needed to understand the answer; avoid adding unrelated model-development history that obscures it. Earlier material on analysis sets and data cutoffs, data lineage, and reporting parameter estimates provides the standard concepts for naming this basis.
Make the Evidence Reproducible
A reviewer should be able to follow each material claim back to its source. For a count, state both the numerator and the denominator. For a percentage, define the denominator and show the calculation when it helps clarify the result. For an estimate, state its units, the analysis that produced it, and any uncertainty measure being reported. If comparing values, identify the reference value and the direction of the comparison.
This is not a demand to reproduce every computational step in the response. It is a demand to expose the steps that determine the conclusion. For example, “the difference was small” is not self-sufficient. The response should identify the compared quantities, calculate the difference, and say why its magnitude matters—or why the available evidence does not permit a firm judgment about its importance.
Worked Example: Explaining Excluded Concentration Records
Hypothetical request: Explain the disposition of concentration records not used in the primary analysis and whether their exclusion materially changed the clearance estimate. Suppose the defined analysis dataset contains 126 records flagged for review. A documented record-level review classifies 8 as below the quantification limit under the specified handling rule, 4 as having unresolved dose times, and 2 as duplicate records. These categories are mutually exclusive. The primary analysis uses 112 records. A sensitivity analysis after source verification of the four dose times gives a clearance estimate of 5.34 L/h, compared with 5.20 L/h in the primary analysis.
Disposition calculation. The number not used is \(8+4+2=14\), and the number retained is \(126-14=112\), matching the stated primary-analysis count. As percentages of the 126 flagged records, the three categories are:
Together, \(14/126\times 100=11.11\%\) were not used, and \(112/126\times 100=88.89\%\) were retained. The response must make clear that these percentages describe the flagged-record set, not all records collected in the study.
Estimate comparison. Relative to the primary estimate, the sensitivity estimate differs by:
The sensitivity estimate is 0.14 L/h, or 2.692%, higher than the primary estimate. Both estimates are in L/h. A focused response would state that the source-verified-time sensitivity analysis produced a modest numerical increase relative to the primary estimate, then describe whether that change affects the interpretation relevant to the request. The calculation alone does not establish that the difference is immaterial: that judgment needs a stated scientific context. Nor does this sensitivity analysis test the consequences of the below-quantification or duplicate-record handling, so it cannot resolve uncertainty about those decisions.
This example illustrates why a response must not allow a percentage to imply more than it measures. The percentage difference uses the primary estimate as its reference; it is not a probability, a confidence interval, or a general measure of model reliability. If an uncertainty interval or other uncertainty result is relevant and available, report it with its correct interpretation rather than substituting an unqualified word such as “robust.”
Interpret Implications Without Overstatement
A request may ask what a model-based result implies for exposure or for a population subgroup. In that case, report the comparison in terms that match the analysis: identify the groups or scenarios, the summary measure, units, and reference. Keep the distinction between a model-based prediction and an observed concentration result clear. A numerical contrast does not by itself establish clinical importance, a causal effect, or a need for a particular regulatory action.
Worked Example: Reporting a Simulated Exposure Contrast
Hypothetical request: Quantify the predicted exposure difference between two renal-function scenarios. Suppose the same specified simulation approach gives mean AUC values of 42 mg·h/L in the reference scenario and 49 mg·h/L in the lower-renal-function scenario. The requested contrast is relative to the reference scenario.
The absolute difference is \(49-42=7\) mg·h/L. The relative increase is:
A precise answer is that the simulated mean AUC is 7 mg·h/L, or 16.67%, higher in the lower-renal-function scenario than in the reference scenario under the stated simulation conditions. The calculation assumes that both values refer to the same AUC definition, dosing conditions, model basis, and summary statistic. If those conditions differ, this contrast is not interpretable as a like-for-like comparison.
The result describes a model-based scenario comparison. It does not, on its own, establish a clinical consequence or prove that renal function caused the difference. A response should state the relevant model and simulation limitations, and should not call the increase clinically meaningful unless the request’s scientific context and an appropriate basis support that conclusion.
Separate Changes in Data From Changes in Analysis
When a request concerns a changed estimate, show which changes could explain it. A comparison between two final outputs may combine changes in the data, analysis set, model specification, or computational settings. Unless the comparison holds other factors fixed, it cannot isolate the contribution of any one change. A staged analysis can be informative, but its attribution depends on the sequence in which factors are changed.
Worked Example: Explaining a Changed Clearance Estimate
Hypothetical request: Explain why the clearance estimate changed from an earlier analysis to the submitted analysis. Suppose the earlier estimate was 5.10 L/h. A staged refit gives 5.02 L/h after updating the data while holding the earlier model specification fixed, and 4.92 L/h after applying the final model specification to the updated data.
The total change is \(4.92-5.10=-0.18\) L/h. Relative to the earlier estimate:
Along the stated staged path, the data update changes the estimate by \(5.02-5.10=-0.08\) L/h, or \((-0.08/5.10)\times 100=-1.569\%\) relative to the earlier estimate. The subsequent model-specification change changes it by \(4.92-5.02=-0.10\) L/h, or \((-0.10/5.02)\times 100=-1.992\%\) relative to the intermediate estimate. The component changes sum to the total change: \(-0.08+(-0.10)=-0.18\) L/h.
The response can therefore say that the final estimate is 0.18 L/h, or 3.529%, lower, and that the staged refits associate part of this change with the data update and part with the model-specification update. It should not present these components as a unique causal decomposition: changing the order of refits could produce different components, especially if data and specification interact. State the staged comparison and its limitation.
Checks, Alternatives, and Common Pitfalls
Before finalizing a response, check that the evidence answers the full request and that its numerical statements can be reconciled with the stated analysis basis. In particular, verify the following:
- Scope: Each sub-question in the request has a direct answer, or the response clearly explains why it cannot be answered from the available evidence.
- Definitions: Terms such as “record,” “subject,” “analysis set,” and “change” are used consistently and are not left ambiguous where they affect a result.
- Arithmetic: Counts reconcile; percentages have defined denominators; differences have a named reference; units and rounding are consistent.
- Traceability: The data cutoff, relevant analysis or model, and source outputs are identified sufficiently to locate the evidence.
- Interpretation: The conclusion is no stronger than the analysis supports, and material uncertainty or limitations are visible rather than buried.
A concise response organized by each request question is often easier to review than a general narrative. A table can help when there are many sub-questions, scenarios, or record categories, provided it includes denominators and units. A narrative is usually clearer for explaining a decision or the limits of an interpretation. Either format is appropriate if the evidence remains traceable and the answer is explicit.
A common mistake is to treat a result that is numerically close to a reference value as proof that an issue has no consequence. For example, a 2.692% difference in a sensitivity estimate does not automatically mean the difference is negligible. The correction is to state exactly what was compared, explain the scientific context for judging its magnitude, and disclose what the sensitivity analysis did not test. Similarly, listing a result without the requested rationale, or supplying a rationale without the requested result, leaves a multi-part request only partly answered.
A focused response has limits. It cannot remove uncertainty that is inherent in the analysis, replace missing information, or support a claim beyond the design and evidence available. If the request cannot be answered as phrased, say what is missing, what can be answered, and what additional analysis or clarification would be needed. That is more informative than implying certainty from an incomplete result.
Check Your Understanding
Answer each question, then compare with the solutions.
- A request asks for the percentage of 80 flagged records that were excluded. If 12 were excluded, what percentage is this, and what denominator should the response name?
- A sensitivity estimate is 8.4 units and the primary estimate is 8.0 units. Calculate the relative difference using the primary estimate as reference. What does that calculation establish?
- Why is it insufficient to say that a model-based exposure increase is “not clinically meaningful” without further explanation?
- A final estimate changed after both a data update and a model change. What kind of analysis can help describe the contributions, and what limitation should be reported?
Solutions. (1) \(12/80\times 100=15.00\%\). The denominator is the 80 flagged records; the result should not be described as a percentage of all study records unless 80 is that total. (2) \((8.4-8.0)/8.0\times 100=5.000\%\). It establishes that the sensitivity estimate is 5.000% higher than the primary estimate, not that the difference is unimportant or statistically precise. (3) Clinical importance requires context and an appropriate scientific basis; the numerical result alone does not establish it. (4) Staged refits that change one factor at a time can describe contributions along that sequence. The decomposition may depend on the order and may not uniquely separate interacting changes.