From Model Description to Decision Rationale
A model rationale is not a list of technical features. It is a compact argument for why a particular model, used in a specified way, is adequate to inform a particular decision. A rationale is useful when a reader can follow the chain from the question to the evidence and see both what the model supports and where its support ends.
This workshop uses a fictional drug, RX-17, to make that chain concrete. The examples are hypothetical. They do not describe a real drug, study, threshold, or regulatory decision. The model’s role is deliberately limited: it informs selection of a dose for a planned clinical study; it does not establish clinical benefit or replace evidence from that study.
The links matter as much as the individual components. Data that are adequate for describing observed patients may not support prediction in a different population. A model that predicts concentrations accurately may still be inadequate for a decision about a clinical outcome unless the relationship between concentration and that outcome is supported. Similarly, a limitation is not a decorative disclaimer: it may restrict the population, dose, or decision for which the rationale applies.
Build the Rationale as a Connected Argument
Start with the decision, not the software or model type. Then state the specific quantity the model will estimate and the context in which that estimate will be used. For RX-17, the hypothetical decision is whether to carry a 40 mg once-daily dose forward into a planned study in adults whose kidney function falls within a defined range.
Suppose the planned study uses two prespecified exposure criteria: at least 80% of patients should have a trough concentration, \(C_{\min}\), at or above 1.5 mg/L, and no more than 5% should have a peak concentration, \(C_{\max}\), above 8 mg/L. These are hypothetical decision criteria, not general clinical standards. The model’s question is therefore not merely “What concentrations does RX-17 produce?” It is whether the proposed dose is predicted to meet both exposure criteria in the intended population.
These probabilities describe the predicted proportions of patients meeting the criteria under the model and specified population assumptions. They are not probabilities that the dose is clinically effective or safe. The decision rule connects model outputs to an action, but does not by itself validate the model or establish that the exposure criteria are appropriate.
A useful rationale can be drafted by answering the following questions in order:
Name the action and the milestone. Avoid vague purposes such as “support development.”
State the modeled quantity, population, treatment, and relevant time or setting.
Connect the observed data and model structure to the question, and make important assumptions visible.
Report evaluation results that bear directly on the decision, not merely a favorable fit statistic.
Name the applicability domain, consequential uncertainties, and any evidence still needed.
Why the Links Need to Hold
The rationale is an argument with several dependent steps. The data must contain information relevant to the intended prediction; the model must represent the important features of the system at a suitable level of resolution; evaluation must show that its errors are acceptable for the decision; and the proposed use must stay within the supported population and conditions. If any step is weak, the final action needs to be qualified or deferred.
This is a logical justification, not a guarantee. For example, passing an exposure-based decision rule does not prove a clinical outcome, and successful prediction in one population does not automatically justify prediction in another. The rationale should therefore connect evidence only to the claim it can support.
One practical test is to ask what would happen if a key premise were false. If the dose decision changes when a model assumption, population characteristic, or threshold changes, that uncertainty is decision-relevant and should be addressed explicitly. If it cannot be resolved with available evidence, the rationale should narrow the claim or recommend obtaining more evidence.
Worked Example: Drafting a Complete Dose-Selection Rationale
Suppose a fictional RX-17 population pharmacokinetic model was developed using concentration-time data from 120 adults treated with 20, 40, or 60 mg once daily. The intended use is to inform whether 40 mg should be carried into a planned study in adults with estimated kidney function from 60 to 120 mL/min. A separate cohort of 40 similar adults is available for evaluation. Write a concise rationale using the hypothetical results below.
The evaluation found a median absolute prediction error of 0.18 mg/L and a 90th percentile absolute error of 0.62 mg/L for the concentration predictions used in the decision. In the evaluation cohort, 33 of 40 patients met the trough criterion and 1 of 40 exceeded the peak criterion. The fitted model predicts that, at 40 mg, 84% will meet the trough criterion and 3% will exceed the peak criterion.
1. State the question and use. The question is whether 40 mg once daily is predicted to satisfy both prespecified exposure criteria in adults with kidney function from 60 to 120 mL/min. The model informs dose selection for a planned study; it does not decide whether the dose is effective.
2. Connect the data and model to the question. The development data include repeated concentration-time observations across doses that include 40 mg. A population pharmacokinetic model is suited to describing concentration profiles and predicting patient-to-patient exposure variation. Because the model is being applied to adults within the stated kidney-function range, the planned use stays within the population described in this hypothetical evidence package.
3. Show the decision arithmetic. The predicted trough proportion is 84%, which exceeds the required 80% by 4 percentage points. The predicted proportion exceeding the peak limit is 3%, which is below the maximum of 5% by 2 percentage points. Thus, the model outputs satisfy both parts of the stated rule:
4. Relate evaluation to the intended use. In the separate cohort, \(33/40 = 0.825\), or 82.5%, met the trough criterion, and \(1/40 = 0.025\), or 2.5%, exceeded the peak criterion. Those observed proportions are close to the corresponding model predictions of 84% and 3%. The prediction-error summaries also provide information about concentration-level accuracy. This supports, but does not prove, adequate prediction for this use.
5. State a bounded conclusion. A defensible rationale is: “In adults with kidney function from 60 to 120 mL/min, the population pharmacokinetic model predicts that 40 mg once daily meets the prespecified trough and peak exposure criteria. Predictions were evaluated in a separate, similar cohort, where the observed criterion proportions were close to model predictions and concentration prediction errors were summarized. On this evidence, 40 mg is supported for consideration in the planned study. The model does not establish clinical benefit, and its use is not justified here for patients outside the evaluated kidney-function range.”
The evaluation cohort contains only 40 patients, so the observed proportions are imprecise estimates of population proportions. The close values are supportive evidence, not proof that the model’s long-run prediction is exactly 84% or 3%. The rationale keeps that limitation visible rather than treating the observed proportions as definitive.
Evaluation Evidence, Alternatives, and Boundaries
Evaluation evidence should match the claim. If the claim concerns concentration prediction, concentration-level predictive checks and prediction errors are relevant. If the decision concerns proportions above or below thresholds, the rationale should also examine those threshold-based predictions. A model may have small average prediction errors yet still misclassify patients near a decision threshold. Conversely, an overall fit summary may look imperfect while predictions near the decision boundary remain adequate. The relevant evidence depends on the use.
The choice of model should also be compared with plausible alternatives. A simpler empirical summary may describe observed concentrations but be unable to predict a new dose or account for differences between patients. A more complex mechanistic model might represent additional processes, but complexity is useful only if the data identify those processes and the added detail changes the decision. The rationale need not catalogue every possible model; it should explain why the selected approach is adequate relative to realistic alternatives for the stated use.
Diagnostics and failure modes should be considered before a positive conclusion. Warning signs include systematic underprediction or overprediction in an important subgroup, poor predictions near a threshold, sparse data in the dose range of interest, and predictions that depend strongly on unsupported assumptions. A discrepancy should be investigated in context: it may indicate a model limitation, a data problem, or a population mismatch. Do not hide it in a broad statement that the model “fit well.”
Worked Example: A Predicted Result Outside the Supported Domain
Suppose the same model is used to consider 40 mg in adults with kidney function from 35 to 59 mL/min. The model predicts that 91% meet the trough criterion and 7% exceed the peak criterion. The peak criterion is not met because \(7\% > 5\%\). More importantly, the intended population in the model rationale was kidney function from 60 to 120 mL/min, and the evaluation cohort did not include the lower range.
The calculation is straightforward:
The model output therefore fails the stated two-part rule even if its predictions were otherwise trusted. But the output should not be presented as a reliable estimate for this lower kidney-function range: the population lies outside the stated applicability domain. A sound conclusion is that the current rationale does not support selecting 40 mg for this group. Further data, a justified model extension, or a different decision approach would be needed. Merely producing a numerical prediction does not establish that the prediction is fit for use.
Worked Example: Improving a Weak Rationale
Consider this draft: “The model fits the data well and predicts an 84% response, so 40 mg is a safe and effective dose for all adults.” Identify the reasoning problems and replace it with a defensible statement, using the fictional RX-17 evidence from the first example.
Problem 1: “Fits the data well” is not decision-specific. A favorable fit alone does not show that threshold predictions are accurate. Replace it with the evaluation evidence relevant to the use: in the separate 40-person cohort, 82.5% met the trough criterion and 2.5% exceeded the peak criterion, close to model predictions of 84% and 3%; concentration prediction errors were also summarized.
Problem 2: “Response” changes the modeled quantity. The model predicts exposure proportions, not clinical response. Replace “84% response” with “84% are predicted to have \(C_{\min}\) at or above 1.5 mg/L.”
Problem 3: “Safe and effective” overstates what the model establishes. Meeting the hypothetical peak criterion does not prove safety, and meeting the trough criterion does not prove effectiveness. Replace those claims with the limited statement that both prespecified exposure criteria are met in the modeled population.
Problem 4: “All adults” exceeds the supported population. The described use is limited to adults with kidney function from 60 to 120 mL/min. The corrected conclusion is: “For adults with kidney function from 60 to 120 mL/min, the model predicts that 40 mg once daily meets the prespecified trough and peak exposure criteria. Evaluation in a separate, similar cohort provides supportive predictive evidence. These findings support considering 40 mg for the planned study, but do not establish clinical benefit or justify use in populations outside the stated range.”
Common Pitfalls and a Final Quality Check
A concise rationale can still be incomplete if it leaves a crucial link implicit. Before finalizing one, check whether a reader can answer all of these questions without guessing:
- Wrong reasoning: “The model fits, therefore it is suitable.” Correction: explain what evidence evaluates predictions relevant to the intended decision.
- Wrong reasoning: “The model returns a number, therefore it applies to this population.” Correction: state the applicability domain and flag predictions outside it.
- Wrong reasoning: “The exposure target is met, therefore the treatment works.” Correction: limit the conclusion to the exposure quantity and decision the model actually addresses.
- Wrong reasoning: “A limitation is harmless because the main result is favorable.” Correction: ask whether the limitation could change the decision; qualify or defer the action if it could.
- Wrong reasoning: “More model complexity means more reliable decisions.” Correction: justify complexity by its relevance to the question and the information available in the data.
The final rationale should be short enough to read as one argument, but specific enough to audit. Name the decision, population, model output, evaluation evidence, key limitation, and bounded action. If a number is important, include its units and explain how it relates to the decision rule. If the evidence supports only consideration of a dose in a planned study, do not phrase the conclusion as a general treatment recommendation.
Check Your Understanding
For each question, identify the reasoning needed before deciding what a model rationale can claim. Solutions follow the questions.
- A model predicts that 86% of patients meet a prespecified exposure target, where the decision threshold is 80%. What additional information is needed before concluding that the dose should be carried forward?
- A model was evaluated in adults with kidney function from 60 to 120 mL/min. What is the main rationale problem if its prediction is used for adults with kidney function from 30 to 45 mL/min?
- In a separate evaluation group of 50 patients, 42 meet the target. Calculate the observed proportion and explain how it should be compared with a model prediction of 86%.
- Why is “the model proves the dose is effective” an inappropriate conclusion when the model predicts only drug concentrations?
Solutions. (1) Check that the model predicts the quantity and population relevant to the decision, that evaluation supports prediction near the threshold, that any other decision criterion is also met, and that important uncertainty does not reverse the conclusion. The 86% result alone is insufficient. (2) The proposed population is outside the stated applicability domain. The numerical prediction is not justified for that use without further support, so the rationale must be narrowed or additional evidence obtained. (3) The observed proportion is \(42/50 = 0.84\), or 84%. This is close to the 86% prediction and is supportive evidence, but sampling variability means it does not prove the model’s true predictive proportion is exactly 84% or 86%. (4) Concentration prediction does not by itself establish a clinical outcome. The conclusion must stay with the modeled exposure and the decision it was intended to inform.