Selecting Time Horizons and Temporal Resolution

Why Time Is Part of Model Scope

A model can include the right biological states and still give an incomplete answer if it simulates the wrong period or reports results too coarsely. A treatment decision may depend on an early biomarker change, a delayed clinical response, or an effect that persists after dosing stops. The simulation window and the temporal detail must therefore be chosen to match both the biology being represented and the decision being informed.

The time horizon answers how far the simulation runs. Temporal resolution answers how finely time is divided or results are reported. These are related but distinct choices: extending the horizon does not make short events easier to see, and using fine time steps does not reveal what happens after the simulation ends.

A useful design begins with the decision question and its relevant timeline. Mark when interventions begin, change, or stop; when biological responses are expected to emerge; how long those responses may persist; and when the outcome will be assessed. Include the time needed for the model to represent those events, not merely the dosing period. The introductory material on model output boundaries distinguishes decision-relevant outputs from internal model states; the same principle applies here: simulate long enough to evaluate the specified output at the required time.

Choosing the Horizon

The end time \(T\) should cover the latest decision-relevant event or measurement, with sufficient additional time to represent a plausible delay or persistence that could change the decision. The start time \(t_0\) should likewise precede the earliest event that needs to be represented. If a baseline state must be established before treatment, the simulation may need a pre-intervention period; if the model starts from a defined baseline at treatment, that period may not be necessary.

The horizon is not automatically the full lifetime of every biological process. It is set by the context of use and the effects that could affect the decision. However, a shorter horizon is defensible only if excluded later effects are not relevant to that decision or are represented in another justified way. When the timing of an effect is uncertain, use plausible earlier and later timings to check whether the chosen horizon changes the conclusion.

For a simple illustrative response that approaches a plateau, suppose the fractional response is \(R(t)=1-e^{-t/\tau}\), where \(t\) is time since an intervention and \(\tau\) is a response time scale. This expression is useful for illustrating timing, not a general biological law. The time to reach a fraction \(q\) of the plateau follows by solving \(q=1-e^{-t/\tau}\):

Thus, the time to a chosen response level can help assess whether the horizon is long enough. It does not by itself establish that a real response follows this shape or that the chosen level is the relevant clinical outcome.

Choosing Temporal Resolution

Temporal resolution should be fine enough to represent the timing of events and changes that matter. A coarse output grid can miss a short-lived peak, misplace the apparent time of a threshold crossing, or make an abrupt intervention look gradual. A fine grid, on the other hand, creates more output without necessarily adding useful information. It cannot compensate for uncertain biological timing or unsupported model detail.

A practical planning rule is to choose a reporting interval substantially shorter than the shortest decision-relevant event duration. For a regular grid, if the interval is shorter than an event’s duration, at least one grid point will lie within that event, provided the interval extends across the event’s time range. This is a sampling consideration, not a guarantee that the event’s peak or shape will be accurately estimated. More points may be needed to characterize a peak, a rapid transition, or a threshold crossing. Event times should also be represented explicitly when an intervention changes abruptly.

Keep the reporting schedule distinct from the numerical solver’s internal calculations. Internal solver settings control how the equations are numerically evaluated; requested output times control where results are saved or inspected. A dense reporting schedule does not prove that the numerical solution is accurate, and a solver may calculate between requested output times. For this tutorial’s purpose, the key checks are that important event times are included, outputs are sufficiently frequent for the decision, and conclusions do not depend on an arbitrary reporting interval.

Delayed and Persistent Effects

A simulation can conceal an effect in two distinct ways. It can end before the effect begins or reaches a decision-relevant level, or it can report too sparsely to reveal an effect that occurs within the horizon. Persistence creates a related risk: stopping the simulation when dosing stops can omit a continuing biological response. These are scope problems, not merely plotting problems.

A diagnostic timeline should compare the horizon with plausible onset, peak, recovery, and assessment times. If those times are uncertain, examine how the decision changes across plausible values rather than treating one assumed delay as certain. A longer horizon is useful only when it represents a relevant process and the assumptions supporting that extension are credible.

A Practical Selection and Diagnostic Process

The following sequence makes the time choices explicit and provides checks for failure. It is a planning framework, not a substitute for evidence about biological timing or numerical accuracy.

Several warning signs indicate that the temporal setup may be inadequate: a predicted response is still changing at the simulation end; a decision assessment falls after the end time; an intervention starts or stops between widely spaced output points; or the apparent peak and threshold crossing move substantially when output spacing changes. A flat trace at the final time is not proof that a response has reached its lasting level. Likewise, a smooth plotted curve can hide sparse underlying output or interpolation between reported values.

Alternatives, Trade-offs, and Common Mistakes

A uniform, fine reporting grid is straightforward and useful when relevant changes occur throughout the horizon, but it can produce unnecessary output. A coarser grid with denser reporting around interventions or expected rapid transitions can be more efficient, provided the event times and decision-relevant changes are known well enough to place those points. A short horizon is simpler to inspect but can exclude delayed or persistent outcomes. A longer horizon covers more possible effects but depends on assumptions further from the available evidence. These are trade-offs to justify against the decision, not universally ranked methods.

A common mistake is to end the simulation when treatment stops and reason, “There is no more drug input, so nothing further can change.” The correction is that treatment cessation does not imply that biological responses cease at the same time. If delayed or persistent effects could affect the decision, the horizon must include their relevant period or the limitation must be made explicit. A related mistake is to report daily values and infer that no sub-day event occurred; daily output alone cannot establish that.

Temporal choices have limits. No horizon or output grid can recover an effect whose mechanism is absent from the model, and fine time resolution cannot make uncertain timing assumptions precise. Model users should therefore distinguish what the simulation can resolve from what is known biologically, and should avoid interpreting an unobserved late response as evidence that no late response exists.