Defining the Decision Question and Analysis Population

From a Dosing Decision to a Prediction Target

An NLME analysis can produce precise parameter estimates and still fail to answer the question that matters. A prediction of exposure in the trial participants, for example, is not automatically a prediction of exposure in the patients who will receive the treatment. Before choosing an outcome to predict, specify which decision the prediction will inform, whose outcomes matter, and under what dosing conditions.

This specification is not a substitute for fitting or checking a model. It sets the target against which those activities should be judged. A useful decision question names a population, an outcome and its time point, the dosing context, and a prediction target that can be connected to a decision rule. If any one of these is vague, different analyses may appear to answer the same question while actually targeting different quantities.

Specify the Population, Outcome, and Dosing Context

A population description should be concrete enough to determine who is included and which characteristics matter to the prediction. “Patients with the disease” may be insufficient if age, renal function, disease severity, or prior treatment differs across the intended use group. A useful specification states the clinically relevant eligibility characteristics and, where relevant, the distribution of important covariates. It should also make clear whether the intended target is a broad population or a defined subgroup.

The analysis population and target population need not be identical. Data may come from a trial with restrictive eligibility criteria, while the decision concerns a broader treated population. That difference is not automatically a flaw, but it requires a reasoned basis for applying the model to people unlike those observed. A model-based prediction cannot supply information about a target group that is absent from, or poorly represented in, the data without additional assumptions.

Define the outcome in measurable terms. “Exposure” could mean area under the concentration-time curve over a dosing interval, peak concentration, or trough concentration. State the units and the time window. For a repeated-dose question, clarify whether the outcome is at the first dose, at steady state, or over a specified treatment period. For a dose-selection question, identify the target range or threshold and whether the decision concerns an average, an individual prediction, or the fraction of the population expected to meet the target.

The dosing context is part of the question, not a detail to fill in after prediction. Specify dose amount, route, interval, duration, and any dose adjustment rule that defines the regimen. If the prediction is conditional on covariates or adherence, say so. A prediction under a fixed dose does not answer a question about a regimen that adjusts dose according to renal function unless that adjustment is represented in the prediction.

Why the Prediction Target Must Be Explicit

Consider a standard NLME representation in which an individual's response at time \(t\) under dose \(D\) depends on individual parameters \(\theta_i\), and those parameters vary across individuals. Write the structural prediction as \(f(t,D,\theta_i)\), with \(\theta_i\) determined in part by covariates \(X_i\) and random effects \(b_i\). For a target population with covariate distribution \(G\), a population mean prediction is an average over both covariates and random effects.

This expression assumes the model adequately represents the response and parameter variation for the target population, the covariate distribution \(G\) is specified appropriately, and the residual error has mean zero for the outcome scale being averaged. The expectation over \(b\) reflects the random-effects distribution. If the target is a proportion above a threshold, the target is instead a probability, such as \(P_{\mathrm{target}}\{Y(t,D)\geq c\}\), and requires the relevant outcome variability as well.

The distinction follows directly from the definition of an expectation: first compute the outcome for each combination of covariates and random effects, then average those outcomes using the target distributions. In general, this is not the same as evaluating the function at average inputs. For example, if clearance is lognormal, exposure under a fixed dose is proportional to the reciprocal of clearance. The mean reciprocal clearance is not the reciprocal of mean clearance.

Worked Decisions Need a Defined Target

For an exposure comparison, specify whether the decision concerns a difference or ratio in mean exposure, a difference in a quantile, or a change in the proportion of patients meeting a target. For dose selection, define the target and the required level of attainment before calculating a dose. These choices are not interchangeable: two regimens can have similar mean exposure but different variability or different proportions above a safety threshold.

Assumptions, Diagnostics, and Alternatives

A clearly worded decision question makes assumptions visible, but does not make them true. Before interpreting a prediction, check whether the target population is represented by the data in relevant covariates, whether the dose and measurement conditions match the intended use, and whether the model supports the requested outcome and time horizon. Extrapolation beyond observed covariate or dosing ranges is a warning that the prediction depends more heavily on model assumptions. A prediction for a subgroup with little or no supporting data should be identified as uncertain rather than presented as directly established by the analysis.

Useful warning signs include a target population described only as “typical patients,” a stated exposure goal without a time window or units, and a dose recommendation with no specified attainment criterion. Another warning sign is reporting a prediction for the observed analysis sample as if it applied to a different target population. Check subgroup representation and compare the covariate mix in the data with the intended population. If a target group is not represented, state what model-based extrapolation is being made.

The main alternatives answer different questions. A prediction at the population typical parameter is simple and can describe a representative model profile, but it does not by itself describe the distribution of patient outcomes. A summary of observed participants describes that sample, but may not transport to a differently composed target population. A population prediction that integrates over random effects and target covariates can address a population-level decision, but relies on the model and target covariate distribution. Individual-specific predictions may be relevant to a patient-level decision, but are not the same as a population-average result. Choose among these approaches by matching the quantity to the decision, not by choosing whichever summary is easiest to calculate.

The limitations should be stated in the same terms as the target. A model prediction is conditional on the structural model, variability assumptions, covariate relationships, dosing conditions, and target-population definition. It does not prove that a regimen causes a clinical benefit, and a predicted target-attainment proportion does not summarize unmodeled harms. When plausible alternative assumptions could change the decision, that sensitivity belongs in the interpretation. The decision question should therefore be precise enough to guide analysis while remaining honest about what the model can support.

Common Pitfall: Letting the Data Define the Decision

• Wrong reasoning: “The trial enrolled these participants, so the analysis population is automatically the population for the dose recommendation.”
• Correction: The trial population identifies whose data were analyzed. The target population identifies whose outcomes the decision concerns. State and justify any difference between them.
• Wrong reasoning: “The predicted value at typical clearance is the average exposure.”
• Correction: With nonlinear functions of variable parameters, a prediction at typical values need not equal an average over the population. Name the prediction target and calculate it accordingly.
• Wrong reasoning: “The regimen with the larger mean exposure must be the better choice.”
• Correction: The decision may depend on target attainment, a quantile, toxicity, or another defined outcome. A mean alone does not determine those quantities.