Start With the Question the Model Must Answer
A model is useful only in relation to a stated purpose. A compact curve can summarize how a biomarker changed over observed time, yet that same curve may be unreliable for forecasting beyond the observed period or for estimating what would happen if a dose were changed. The relevant choice is not simply which model can fit the data; it is which model structure represents the relationships needed for the intended use.
The earlier material on description, prediction, and extrapolation distinguishes these uses. Here, the focus is the broad class of model that may suit each one. Empirical models describe relationships using patterns in data; mechanistic models represent specified processes in the system; and hybrid models combine both. These are useful categories, not guarantees of validity. A mechanistic label does not make a model causal or correct, and an empirical model is not inherently unsuitable for prediction.
These classes describe how a model represents the system, not a ranking from weak to strong. An empirical curve may be the clearest choice for describing a pattern within observed conditions. A mechanistic model may be useful when the question depends on how a process changes under a new condition. A hybrid can represent a known process while allowing the data to characterize a less understood component. In all cases, the intended use determines which parts of the system must be represented credibly.
What Each Model Class Can Contribute
An empirical model might express an outcome as a function of an observed predictor, such as time, dose, or concentration. For example, a linear model summarizes an average change per unit time. Other empirical forms include polynomial curves, splines, and flexible regression functions. Their usefulness depends on whether the estimated relationship is an adequate representation for the question at hand. A good description within the observed range does not by itself establish reliable behavior outside it.
A mechanistic model encodes assumptions about processes. A pharmacokinetic model, for example, may represent drug entry, distribution, and elimination. Its equations connect quantities through process-based parameters, such as clearance or volume. The structure can make predictions under changed conditions more interpretable, but only if its assumptions are reasonable for those conditions. A simplified biological mechanism is still a model, not a complete account of reality.
A hybrid model uses process-based structure where it is useful and empirical relationships elsewhere. One common strategy is to represent drug exposure with a pharmacokinetic model and relate exposure to response with a flexible empirical function. Hybrid structure can be practical when some components are well understood but others are not. It also creates a need to distinguish which conclusions come from process assumptions and which rely on empirical patterns.
| Model class | Typical strength | Question to ask | Important limitation |
|---|---|---|---|
| Empirical | Summarizes a pattern in the data | Does this form represent the relationship needed for the intended use? | Behavior may not carry over to new conditions, especially beyond the observed range. |
| Mechanistic | Represents specified processes and their connections | Are the represented processes and assumptions suitable for this use? | Incorrect or incomplete process assumptions can yield misleading outputs. |
| Hybrid | Combines process structure and data-based relationships | Are the boundaries between process-based and empirical components clear? | Uncertainty in either component can affect the combined result. |
Worked Examples of Matching Class to Use
Worked Example: Describing an Observed Pattern
Suppose a hypothetical early study records a biomarker at three scheduled times after treatment. The observed group means are 8 units at 1 hour, 10 units at 2 hours, and 12 units at 3 hours. The stated aim is to summarize the pattern over these observed times, not to predict a later response or estimate the effect of changing dose.
An empirical straight-line summary is reasonable as a first description. Let \(t\) be time in hours and let \(\hat{y}(t)\) be the fitted mean biomarker value in units. The line through these three means has slope \((12-8)/(3-1)=2\) units per hour. Its intercept is \(8-2(1)=6\) units, so
At 2 hours, the predicted value is \(6+2(2)=10\) units, matching the observed mean. The line is a concise description of this hypothetical pattern. Its arithmetic does not establish a biological mechanism, and the three points alone do not justify extending the line to later times. If the purpose changed to forecasting or evaluating a changed treatment, the model choice and supporting evidence would need to be reconsidered.
Worked Example: Forecasting With a Process-Based Model
Consider a hypothetical intravenous bolus dose of 100 mg. Suppose a one-compartment model is appropriate for the intended forecast, with instantaneous mixing, first-order elimination, and constant clearance and volume over the period of interest. Let volume \(V=20\) L and clearance \(CL=4\) L/hour. Under these assumptions, concentration after time \(t\) is
The initial concentration is \(D/V=100\text{ mg}/20\text{ L}=5\) mg/L. The elimination rate constant is \(CL/V=4/20=0.2\) per hour. At 3 hours,
The result is a model-based forecast of 2.744 mg/L under the stated assumptions. This process-based form is useful because it represents elimination explicitly. The calculation does not show that the one-compartment assumptions are appropriate for every drug or patient; if distribution is not adequately represented by one compartment, the forecast may be systematically wrong.
Worked Example: Combining Process and Empirical Components
Suppose a hypothetical development question asks how response would differ between two dose regimens. A hybrid model uses a pharmacokinetic component to predict exposure and an empirical maximum-effect curve to relate exposure to mean response. Assume the PK component predicts concentrations of 2 and 6 mg/L for the two regimens, and the response component is
where response is measured in arbitrary units, the baseline is 10 units, the maximum modeled increase is 40 units, and the concentration giving half that increase is 4 mg/L. For the first regimen,
For the second regimen,
The modeled difference is \(34.00-23.33=10.67\) response units. The hybrid form separates the exposure prediction from the exposure-response pattern. To interpret this contrast as a causal response to changing regimen, the model must be supported by an appropriate design and assumptions about the treatment comparison; the hybrid structure alone does not establish that interpretation.
Why an Empirical Relationship Is Not Automatically Causal
A central distinction is between describing an observed relationship and estimating the response to an intervention. An empirical model may estimate an association between dose and outcome. That association equals a causal dose effect only under additional conditions, such as an appropriate treatment assignment or adequate control of common causes of dose and outcome. Model class does not remove the need for those conditions.
A simple derivation shows the issue. Suppose the outcome follows \(Y=\beta_0+\beta_A A+\gamma U+\varepsilon\), where \(A\) is dose, \(U\) is a factor related to both dose and outcome, and the error \(\varepsilon\) has mean zero and is uncorrelated with \(A\). If a linear regression of \(Y\) on \(A\) omits \(U\), its population slope is \(\operatorname{Cov}(A,Y)/\operatorname{Var}(A)\). Substituting the outcome equation gives
Unless \(\gamma=0\) or \(\operatorname{Cov}(A,U)=0\), the observed regression slope differs from \(\beta_A\), the treatment coefficient in the stated outcome model. A mechanistic model may make assumptions about causal pathways explicit, but it does not prove those assumptions. A hybrid model has the same requirement: its causal interpretation depends on the design and assumptions supporting the intervention contrast.
A Practical Choice and Its Checks
Choose a candidate model class by asking what structure the intended use requires. The following sequence is a way to make that rationale explicit; it is not a substitute for assessing data adequacy or validating a model.
Specify whether the model must summarize an observed pattern, forecast an outcome, or support a response-to-intervention question.
Ask which relationships can be represented as data-based patterns and which depend on specified biological or pharmacological processes.
Use empirical structure for a data-based summary when that is sufficient; use mechanistic structure when explicit processes are needed; combine them when different components warrant different representations.
Look for systematic residual patterns, implausible predictions, or sensitivity to assumptions that matter to the intended output. For causal uses, also examine whether the design and causal assumptions support the intervention contrast.
Diagnostics depend on what the model is meant to do. For description, systematic departures from the fitted pattern suggest the chosen form may be an inadequate summary. For forecasting, compare predictions with relevant observed outcomes when possible, and check whether model assumptions remain credible under forecast conditions. For a causal response question, good fit alone is insufficient: confounding, treatment assignment, and the meaning of the intervention contrast remain central. These checks indicate possible failure; passing them does not prove that a model is correct.
The main alternative to a single class is often a hybrid, but combining structures is not automatically an improvement. A hybrid can make a useful division between well-specified processes and relationships better represented empirically. It can also inherit limitations from both components and make the source of a prediction less transparent. The appropriate choice is the simplest class that represents the relationships the intended use requires, with assumptions made explicit.
Common Mistake and Limitations
Common mistake: “The mechanistic model predicts the outcome, so its estimate of the dose effect is causal.” That reasoning confuses a model’s structure with evidence for its assumptions. A mechanistic equation can encode a plausible pathway, but causal interpretation still depends on whether the pathway and relevant assumptions are supported for the question. The correction is to state the intervention contrast, identify the assumptions that connect the model to that contrast, and assess whether the study design and evidence support them.
No class is best for every purpose. Empirical models may behave poorly when asked to represent conditions unlike those informing the fitted pattern. Mechanistic models can be misleading when important processes are omitted or parameters are poorly supported. Hybrid models can be difficult to interpret if their component roles are unclear. Selecting a class is therefore a reasoned starting point, not a declaration that the model is fit for use. Data, model resolution, validation, and decision consequences also matter, but they are distinct parts of a complete fit-for-purpose assessment.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A team needs a compact summary of a biomarker’s pattern over the observed visits. Which model class might be suitable, and what would that choice not establish?
- A pharmacokinetic forecast uses a one-compartment model with \(D=80\) mg, \(V=16\) L, and \(CL=3.2\) L/hour. Under the stated bolus and first-order assumptions, calculate the predicted concentration at 2 hours.
- A dose-response regression shows an association between dose and outcome. Does selecting a mechanistic or hybrid model, by itself, make the estimated dose contrast causal? Explain.
- Give one reason to use a hybrid model and one limitation it may retain.
Solutions. (1) An empirical model may be suitable for summarizing the observed pattern. That choice does not establish a biological mechanism or justify forecasting beyond the observed conditions. (2) The initial concentration is \(80/16=5\) mg/L and the elimination rate constant is \(3.2/16=0.2\) per hour. Thus \(C(2)=5\exp(-0.2\times2)=5\exp(-0.4)\approx5(0.6703)=3.352\) mg/L. This calculation relies on the stated one-compartment, bolus, and first-order assumptions. (3) No. Causal interpretation requires a design and assumptions that support the intervention contrast; model structure alone does not establish them. (4) A hybrid can represent a well-specified pharmacokinetic process while using an empirical exposure-response relationship. It may still be limited by unsupported assumptions in either component, including the empirical component’s behavior under new conditions.