Simulation Targets And Scenario Definitions

Start With the Question, Not the Simulation

A simulation can generate a large amount of output without answering a useful question. For example, a table of predicted concentrations is not, by itself, a clear target: it does not say which patients the predictions represent, which regimen was used, when concentrations matter, or how the results will guide a decision. A well-defined target answers those questions before data are generated.

In pharmacometrics, the target should connect a model output to a specific scientific purpose. A prediction target describes an outcome under a specified scenario. A comparison target summarizes a difference between scenarios. A decision target evaluates whether a prespecified criterion is met. These targets may use the same model and simulated profiles, but they are different quantities and can support different conclusions.

Definition: A simulation estimand is the precisely specified quantity that a simulation is intended to calculate. Its specification includes the target population, outcome, time horizon, intervention scenario or scenarios, and summary measure or decision rule. The simulation estimates this quantity using the chosen model and inputs; defining the estimand does not establish that the model or inputs are adequate.

Specify the Target Population and Outcome

The target population is the set of people, patients, or other units about whom the simulation is intended to provide information. State it in operational terms where possible: for example, adults with a particular disease and a stated range of renal function. “Patients like those in the model” is usually too vague unless the relevant features and intended application are also made clear. A target population is not merely the records that happen to be available for a simulation.

The outcome must also be defined. It may be a concentration at a particular time, an exposure summary over an interval, a biomarker value, or an indicator that an outcome crosses a threshold. State its units and how it is calculated from the model output. If the question concerns a summary over patients, specify whether the target is a mean, median, percentile, proportion, or another functional. These summaries answer different questions; the mean concentration and the proportion above a concentration threshold cannot be substituted for one another.

For a finite set of target-population records, suppose record $i$ has weight $w_i$, where the weights are nonnegative and sum to one. Let $Y_i^g(t)$ denote the model-predicted outcome for record $i$, under scenario $g$, at time $t$. A weighted mean prediction at time $t$ is

\mu_g(t)=\sum_{i=1}^{N} w_iY_i^g(t), \qquad \sum_{i=1}^{N}w_i=1.

The weights express how the records contribute to the defined target. Equal weights give each record the same contribution; other weights may be needed if records represent different numbers of people. This formula is a definition of the chosen summary, not a claim that a particular set of records represents every possible target population.

Set the Time Horizon and Intervention Scenarios

A time horizon states when the outcome is evaluated or over what interval it is summarized. “At steady state” or “during treatment” may still be ambiguous: specify the time point, dosing interval, treatment duration, or interval summary that makes the target reproducible. The horizon should match the clinical question. A trough concentration immediately before a dose is not the same outcome as a concentration at the end of the interval.

An intervention scenario is a sufficiently detailed description of what is applied in the simulation. Depending on the question, it can specify dose, route, dosing interval, start time, duration, and other relevant treatment conditions. If scenarios differ in more than the intended intervention, a comparison may not isolate the contrast of interest. State relevant common conditions as well as the feature that changes.

A target is only as interpretable as the model assumptions supporting it. Before calculating the summary, verify that the model represents the relevant dosing events and output, that the parameterization is appropriate for the stated purpose, and that the target-population records and weights match the intended population. Also state any important simplifications, such as assuming the regimen is followed exactly. A precise estimand makes these assumptions visible; it does not remove them.

Worked Example: Predicting a Concentration at a Defined Time

Hypothetical question. For a stated target group of five adults receiving a specified oral regimen, what is the equally weighted mean model-predicted concentration 24 hours after the first dose? Assume the model is suitable for this illustrative prediction, the regimen is represented correctly, and all five records are intended to have equal weight. The simulated concentrations at 24 hours are 1.2, 1.8, 2.0, 2.5, and 3.0 mg/L.

The target is a prediction, not a comparison: the target population is the five equally weighted records; the outcome is concentration in mg/L; the time horizon is 24 hours after the first dose; the scenario is the specified regimen; and the summary is the mean.

With $N=5$ and $w_i=1/5$, the calculation is

\mu_g(24)=\frac{1.2+1.8+2.0+2.5+3.0}{5} =\frac{10.5}{5}=2.10\ \text{mg/L}.

Thus, the mean model-predicted concentration for these equally weighted records at the stated time is 2.10 mg/L. This does not say that every person has that concentration, nor does it describe a different time point or population.

Comparison Targets Require a Defined Contrast

A comparison target specifies both the scenarios and the contrast. For example, if the outcome is a mean concentration at 24 hours, a difference target could be the mean under scenario $g_1$ minus the mean under scenario $g_0$, evaluated for the same target population. A ratio, change in a threshold proportion, or other contrast may be more appropriate for another question. The direction of subtraction should be stated so the sign has an unambiguous interpretation.

Definition: For a chosen outcome and time horizon, a model-based mean difference between scenarios $g_1$ and $g_0$ in the specified target population is $$ \Delta(t)=\sum_{i=1}^{N}w_i\left\{Y_i^{g_1}(t)-Y_i^{g_0}(t)\right\}. $$ The scenarios, records, weights, outcome, and time point must be defined for this quantity to be interpretable.

A model-based contrast is not automatically a causal conclusion about patients. It is a comparison of outcomes generated under the model and the specified scenarios. Any clinical causal interpretation requires that the model and assumptions adequately represent the consequences of those interventions for the target population.

Worked Example: Comparing Two Regimens

Hypothetical question. Compare two oral regimens using model-predicted concentration at 24 hours in the same four equally weighted target-population records. Assume the model correctly encodes both regimens, all records are evaluated at the same time, and no feature other than the regimen differs between the scenarios. Under regimen $g_0$, the concentrations are 1.0, 1.5, 2.0, and 2.5 mg/L. Under regimen $g_1$, they are 1.4, 1.8, 2.1, and 2.7 mg/L.

The target is the mean difference $g_1-g_0$, in mg/L. Compute the within-record differences: $0.4$, $0.3$, $0.1$, and $0.2$ mg/L. The equally weighted mean difference is

\Delta(24)=\frac{(1.4-1.0)+(1.8-1.5)+(2.1-2.0)+(2.7-2.5)}{4} =\frac{1.0}{4}=0.250\ \text{mg/L}.

The mean model-predicted concentration at 24 hours is 0.250 mg/L higher under $g_1$ than under $g_0$ for these records. The statement is about this specified model-based contrast; it is not a claim that changing regimens causes the same increase in every patient.

Decision Targets Need a Rule

A decision target goes beyond describing a prediction or contrast: it evaluates a prespecified rule. The rule must identify the outcome, threshold, time horizon, population summary, and action associated with the result. For example, one might examine the proportion of a target population whose predicted concentration is at least a specified threshold at a specified time, and compare that proportion with a decision criterion. The threshold and criterion should come from the scientific or clinical question, not be selected after inspecting the simulation results.

Worked Example: Evaluating a Threshold-Based Decision Rule

Hypothetical question. For a specified regimen, evaluate whether at least 80% of eight equally weighted target-population records have model-predicted trough concentrations of at least 1.0 mg/L at steady state. Assume the trough time is defined as immediately before the next dose, the simulated regimen includes the intended dosing schedule, and the eight records are the prespecified target set. The predicted trough concentrations are 0.7, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, and 1.6 mg/L.

Six of the eight values meet or exceed 1.0 mg/L. The target proportion is therefore

\hat{p}_{\geq 1.0}=\frac{6}{8}=0.750=75.0\%.

The rule requires at least 80%. Since $75.0\% < 80\%$, this simulation does not meet the stated criterion. This is a decision about the defined set of model-predicted outcomes and rule; it is not, by itself, a clinical recommendation or a statement about uncertainty in the population proportion.

Why the Summary Must Match the Purpose

A common justification for using the weighted mean is simply that it is familiar. A short derivation shows what it means and why it answers a particular type of question. Suppose the target contains $N$ records with equal contribution, so $w_i=1/N$. The average of their model-predicted outcomes is, by definition, their sum divided by their count:

\frac{1}{N}\sum_{i=1}^{N}Y_i^g(t) =\sum_{i=1}^{N}\frac{1}{N}Y_i^g(t) =\sum_{i=1}^{N}w_iY_i^g(t).

Thus the weighted formula reduces to the ordinary arithmetic mean when all records have equal weight. This derivation establishes the calculation, not that a mean is the right scientific target. If the question asks how many records exceed a threshold, the relevant summary is instead the proportion of indicators:

\hat{p}_{\geq c}=\sum_{i=1}^{N}w_i I\{Y_i^g(t)\geq c\},

where $I\{\cdot\}$ is 1 when the condition in braces is true and 0 otherwise. A mean can conceal a subgroup below a clinically important threshold, while a threshold proportion does not report the magnitude of concentrations above that threshold.

Prediction, comparison, and decision targets are therefore alternatives for different questions, not interchangeable reporting styles. A prediction may be the right target for describing expected outcomes under one regimen. A contrast is needed to quantify a difference between regimens. A decision target is needed when an explicit rule determines an action. Reporting several summaries can be useful, but each should have a stated role rather than being treated as a substitute for the prespecified primary target.

Diagnostics, Pitfalls, and Limitations

Before running a simulation, check whether another analyst could identify the target from its specification alone. In particular, look for an unnamed population, a missing time point, an incompletely described regimen, or a summary that does not match the question. After simulation, verify that the outputs have the expected units and time indexing, that all intended records and weights were included, and that the scenario labels and contrast direction were implemented correctly. A plausible-looking result can still answer the wrong question if one of these definitions was lost in translation.

Vague population. “Typical patients” does not identify who the result is intended to represent. State the relevant population and how the records or weights correspond to it.
Unspecified horizon. “Concentration during the interval” could mean a trough, peak, or interval summary. Name the time or calculation.
Scenario mismatch. If the scenarios differ in dose, schedule, and duration at once, the contrast cannot be attributed to one of those changes without a more specific target and interpretation.
Wrong summary. A mean cannot answer a question about the fraction meeting a threshold. Define the statistic that corresponds to the decision rule.
Overclaiming. A model-generated contrast is not automatically evidence of a causal treatment effect. Its interpretation depends on the adequacy of the model and assumptions.

A frequent mistake is to generate profiles first and then choose whichever summary appears most favorable. The correction is to define the population, outcome, horizon, scenario, and summary before examining results. This reduces the risk of selective reporting and makes the intended interpretation testable. A second limitation remains: even a perfectly specified target can be estimated poorly if the model is unsuitable, the input records do not represent the target, or the simulated events do not match the scenario.

Key takeaway: A simulation target is not merely a model output. Specify whose outcomes matter, what outcome is measured, when it is measured, which scenario applies, and how the outcomes are summarized or used in a decision.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

A simulation reports a mean concentration at 12 hours but does not describe the population or regimen. Name two essential omissions that prevent a complete target specification.
Four equally weighted records have predicted outcomes of 2, 3, 5, and 6 mg/L. What is their mean prediction?
In a paired comparison, scenario A gives 1.0 and 2.0 mg/L; scenario B gives 1.5 and 2.2 mg/L for the same two equally weighted records. What is the mean difference $B-A$?
Five records have troughs of 0.8, 1.0, 1.1, 1.2, and 1.4 mg/L. What proportion meets a threshold of at least 1.0 mg/L?
Why does a model-based difference between regimens not automatically establish a causal effect in patients?

Solutions. (1) It must define the target population and specify the intervention scenario, including the regimen; it should also state the outcome and time horizon clearly. (2) $(2+3+5+6)/4=16/4=4.00$ mg/L. (3) The paired differences are $0.5$ and $0.2$ mg/L, so their mean is $(0.5+0.2)/2=0.350$ mg/L. (4) Four of five meet the threshold, so the proportion is $4/5=0.800$, or 80.0%. (5) The result describes outcomes generated by the specified model and assumptions. A causal interpretation additionally depends on whether the model and assumptions adequately represent the effects of those interventions in the target population.