Structural Uncertainty in Pharmacometric Models

Why Structure Matters to a Development Decision

A pharmacometric model contains both numerical quantities to estimate and a description of how those quantities relate to exposure and response. Structural uncertainty concerns the latter: for example, whether drug effect is expressed directly on the measured response or indirectly through response turnover, and whether the concentration-response relationship is linear or saturable. Two structures can fit observed data comparably yet imply different responses at a proposed dose, at a later time, or in a less-studied population. Those differences can change a go/no-go decision, a dose choice, or the value of collecting additional evidence.

Structural uncertainty is not the same as uncertainty in a parameter within a chosen model. If a linear model is fixed, uncertainty in its slope is parameter uncertainty; uncertainty about whether the relationship should be linear or Emax-shaped is structural uncertainty. Neither should be confused with variability among patients. In comparing structures, hold the decision question and outcome definition fixed, and make clear which predictions are conditional on each model.

Definition: Structural uncertainty is uncertainty about the form of the model used to represent a system, such as its response dynamics or concentration-effect function. A structural alternative is a specified, scientifically plausible model form. Decision-relevant structural uncertainty is present when credible alternatives yield materially different predictions or actions for the same development question.

The goal is not to list every imaginable equation. It is to identify alternatives that are biologically and statistically plausible, assess whether available data distinguish them, and show what their implications are for the decision. A model that fits existing observations well is not thereby guaranteed to predict well outside the observed exposure and time range.

Direct and Indirect Response Structures

In a direct-response model, drug concentration is linked directly to the measured response. For a simple linear effect, one possible form is $R(t)=R_0+\beta C(t)$, where $R(t)$ is response, $R_0$ is baseline response, $C(t)$ is concentration, and $\beta$ is the response change per concentration unit. Such a model can represent an effect that follows concentration rapidly relative to the sampling schedule. A delay in observed response does not automatically disprove a direct effect: delayed absorption, distribution, sampling, or other processes may also affect the observed time course.

An indirect-response model instead represents response as a quantity that is produced and removed over time. Drug may alter the production rate, the removal rate, or both. The response therefore changes dynamically, rather than being set immediately to a concentration-specific value. The model below assumes drug stimulates production while the removal rate remains constant.

Model definition: Let $R(t)$ be response, $k_{\mathrm{in}}$ the zero-drug production rate, and $k_{\mathrm{out}}$ the first-order removal-rate constant. With drug stimulation of production and constant concentration $C$, write $$ \frac{dR(t)}{dt}=k_{\mathrm{in}}\left(1+\frac{S_{\max}C}{SC_{50}+C}\right)-k_{\mathrm{out}}R(t). $$ Here $S_{\max}$ is the maximum fractional stimulation and $SC_{50}$ is the concentration producing half of that maximum stimulation. In the absence of drug, the steady-state response is $R_0=k_{\mathrm{in}}/k_{\mathrm{out}}$.

This form assumes constant concentration over the interval considered, positive $k_{\mathrm{out}}$, and fixed model parameters. It also assumes a well-mixed response pool with first-order removal and the specified concentration effect on production. These are modeling assumptions, not facts established by writing the equation. If concentration changes over time, the production term must use $C(t)$, and the corresponding time-varying differential equation must be solved or simulated.

Justifying the Turnover Time Course

For constant $C$, define the drug-stimulated production rate as $K=k_{\mathrm{in}}\left(1+\frac{S_{\max}C}{SC_{50}+C}\right)$. The differential equation is then $dR/dt=K-k_{\mathrm{out}}R$. Its equilibrium is $R_{\mathrm{ss}}=K/k_{\mathrm{out}}$. The following integrating-factor derivation gives the transient response from any initial value $R(0)$, including the case where the initial response already equals the new equilibrium.

\begin{aligned} \frac{dR}{dt}+k_{\mathrm{out}}R&=K,\\ e^{k_{\mathrm{out}}t}\frac{dR}{dt} +k_{\mathrm{out}}e^{k_{\mathrm{out}}t}R &=Ke^{k_{\mathrm{out}}t},\\ \frac{d}{dt}\left(e^{k_{\mathrm{out}}t}R(t)\right) &=Ke^{k_{\mathrm{out}}t},\\ e^{k_{\mathrm{out}}t}R(t)-R(0) &=\frac{K}{k_{\mathrm{out}}} \left(e^{k_{\mathrm{out}}t}-1\right),\\ R(t)&=R_{\mathrm{ss}}+ \bigl(R(0)-R_{\mathrm{ss}}\bigr)e^{-k_{\mathrm{out}}t}. \end{aligned}

No division by $R(t)-R_{\mathrm{ss}}$ is needed in this derivation. If $R(0)=R_{\mathrm{ss}}$, the formula gives $R(t)=R_{\mathrm{ss}}$ at every time, as the differential equation requires. Otherwise, the response approaches the new equilibrium exponentially, with time scale governed by $k_{\mathrm{out}}$. This explicit solution depends on the constant-concentration assumption; it should not be used unchanged when concentration varies.

Worked Example: Comparing Immediate and Turnover Responses

Suppose, hypothetically, that baseline response is 100 response units, $k_{\mathrm{in}}=20$ response units per day, and $k_{\mathrm{out}}=0.20$ per day. A constant concentration of $2$ mg/L is maintained, with $S_{\max}=1$ and $SC_{50}=2$ mg/L. Compare the indirect-response prediction after two days with a direct model whose immediate effect is linear with slope 25 response units per mg/L.

First, the drug-stimulation fraction is $S_{\max}C/(SC_{50}+C)=(1)(2)/(2+2)=0.50$. The stimulated production rate is $K=20(1+0.50)=30$ response units per day, so the new equilibrium is $R_{\mathrm{ss}}=30/0.20=150$ response units. The initial response is $R(0)=100$, consistent with the drug-free equilibrium $20/0.20=100$.

Using the derived time course, $$ R(2)=150+(100-150)e^{-(0.20)(2)} =150-50e^{-0.40} \approx 116.5\ \text{response units}. $$ The direct linear model predicts an immediate response of $100+(25)(2)=150$ response units at this concentration. Thus the two hypothetical structures predict 116.5 versus 150 response units at two days. This comparison illustrates the possible decision relevance of delayed turnover; it does not establish that either structure is correct. It assumes the stated parameters and constant concentration throughout the two days.

Linear and Emax Concentration-Response Structures

A linear concentration-effect model assumes that each additional concentration unit produces the same additional response over the modeled range. An Emax model allows the effect to rise toward a maximum, so the increment from increasing concentration becomes smaller at higher concentrations. Both can be plausible over a restricted range: an Emax curve can look approximately linear at low concentrations, while a linear model may be an adequate local approximation when saturation is not observed.

Model definitions: For a baseline-adjusted response change $\Delta R(C)$, a linear model is $\Delta R(C)=\beta C$, where $\beta$ is the slope. An Emax model is $\Delta R(C)=E_{\max}C/(EC_{50}+C)$, where $E_{\max}$ is the maximum effect and $EC_{50}$ is the concentration producing half of $E_{\max}$. These forms assume nonnegative concentration and, for the examples here, positive parameters and an increasing effect.

Worked Example: A Target Exposure Under Linear and Emax Effects

Consider a hypothetical development rule requiring a response increase of at least 30 units. A linear model has slope $\beta=12$ units per mg/L. An alternative Emax model has $E_{\max}=60$ units and $EC_{50}=4$ mg/L. Assume both models describe the same response increase, and compare the concentration needed to meet the rule.

Under the linear model, solve $30=12C$, giving $C=30/12=2.50$ mg/L. Under the Emax model, solve $30=60C/(4+C)$. Multiplying by $4+C$ gives $120+30C=60C$, hence $C=4.00$ mg/L. At $C=3$ mg/L, the linear prediction is $12(3)=36$ units, which meets the rule. The Emax prediction is $60(3)/(4+3)=180/7\approx25.71$ units, which does not. Thus, under the stated deterministic rule, the structures imply different target-exposure requirements and different conclusions at 3 mg/L.

These calculations assume the models apply at the concentrations being evaluated and ignore parameter uncertainty for clarity. If observed data cover only low concentrations, the apparent linear relationship may not determine whether saturation occurs at higher exposure. Extrapolating either structure beyond its evidential range should therefore be reported as model-dependent.

Comparing Structural Alternatives for a Decision

A useful comparison starts with a common endpoint, time horizon, target population, and decision rule. For each plausible structure, calculate the prediction relevant to that same decision. If uncertainty in parameters has already been propagated, report the resulting probability or interval conditional on each structure; do not silently treat the structures as if they were parameter draws from one model. A probability assigned to each structure requires a justified model-weighting approach. Without such weights, model-conditional results should remain separate.

Worked Example: Conditional Go/No-Go Probabilities

Suppose a hypothetical program proceeds if the probability that the response increase exceeds 30 units is at least 0.60. Under one plausible structural model, the predictive response increase is $X_A\sim N(34,5^2)$ units. Under another, it is $X_B\sim N(27,4^2)$ units. Treat each stated normal distribution as a predictive distribution conditional on its model, incorporating parameter uncertainty but not assigning probability to either model structure.

For model A, $$ P(X_A>30)=1-\Phi\left(\frac{30-34}{5}\right) =\Phi(0.80)\approx0.7881. $$ Since $0.7881$ exceeds the decision threshold $0.60$, model A supports proceeding under this rule.

For model B, $$ P(X_B>30)=1-\Phi\left(\frac{30-27}{4}\right) =1-\Phi(0.75)\approx0.2266. $$ Since $0.2266$ is below $0.60$, model B does not support proceeding under the same rule. The result is a structure-dependent decision, not an overall probability that the program succeeds. No weighted average of 0.7881 and 0.2266 is warranted unless model weights are specified and justified.

When structures imply different actions, the disagreement itself is decision-relevant. It can motivate checking whether the data contain information that distinguishes the structures, examining the consequences of each conditional prediction, or considering whether additional evidence is worth obtaining. It does not by itself show which structure is true or dictate a particular development action.

Diagnostics, Comparisons, and Limitations

Structural alternatives should be evaluated against both observed data and the prediction task. Useful checks include whether residual patterns change systematically over time or concentration, whether the model reproduces the observed response trajectory, and whether predictions remain plausible at the intended dose and time. A direct model may miss a reproducible delay between concentration and response. A turnover model may fit a delay but estimate production or removal parameters poorly if the sampling schedule is too sparse to distinguish its time course. A linear effect may predict implausibly large changes when extrapolated far beyond observed concentrations; an Emax model may be weakly identified if the data never approach saturation.

These checks diagnose possible inadequacy; they do not mechanically select a biological truth. Good fit to existing data is not enough if the competing structures make different predictions in the decision region. Conversely, a more elaborate model is not automatically preferable simply because it can represent more patterns. It may require information the study did not collect and can leave key predictions highly uncertain.

Structural choice	What differs	Decision-relevant question	Common limitation
Direct versus indirect response	Immediate concentration-response link versus turnover dynamics	Does the predicted timing of response affect the endpoint or dosing decision?	Sampling may not resolve the delay or turnover rate.
Linear versus Emax effect	Constant slope versus a saturating concentration-effect curve	Does the target exposure or response depend on saturation?	Observed concentrations may not distinguish the curve shapes.
Separate model-conditional predictions	Each structure is evaluated on its own	Do plausible structures support the same action?	Does not produce one combined probability without justified model weights.

A common mistake is to choose the structure with the best in-sample fit and then report its decision prediction as if structural uncertainty had disappeared. The correction is to check how strongly the data distinguish the alternatives and to show whether the decision changes across the plausible structures. Another mistake is to average their predictions equally merely because two models were compared. Equal weighting is a substantive assumption, not a neutral default. Model discrepancy—the possibility that all candidate structures miss important features of the system—is a separate issue and should not be hidden by comparing structures alone.

Key takeaway: Define structural alternatives explicitly, compare them on the same decision-relevant outcome, and report model-conditional implications. If credible structures lead to different actions, communicate that dependence rather than presenting one structure's result as uniquely determined.

Check Your Understanding

Answer each question, then check the solutions below.

In one sentence, distinguish structural uncertainty from parameter uncertainty in an Emax model.
For a turnover model with constant concentration, what happens to $R(t)$ if $R(0)=R_{\mathrm{ss}}$? Why?
A linear model predicts 24 response units at a concentration of 2 mg/L. What is its slope, and what response increase does it predict at 3 mg/L?
Why is it not automatically appropriate to average two model-conditional probabilities of meeting a target with equal weights?
Name one data pattern that could make an immediate direct-response structure questionable, and one reason that pattern might still fail to establish an indirect-response mechanism.

Solutions. (1) Parameter uncertainty concerns values such as $E_{\max}$ or $EC_{50}$ within the Emax form; structural uncertainty concerns whether the Emax form is appropriate rather than, for example, a linear form. (2) The response remains constant at $R_{\mathrm{ss}}$, because the solution gives $R(t)=R_{\mathrm{ss}}+(R_{\mathrm{ss}}-R_{\mathrm{ss}})e^{-k_{\mathrm{out}}t}=R_{\mathrm{ss}}$, and the differential equation has zero net change there. (3) The slope is $24/2=12$ response units per mg/L, and at 3 mg/L the predicted increase is $12(3)=36$ units. (4) Equal weights assert a particular model-averaging assumption; they are not implied by having two alternatives, and a combined probability needs justified weights. (5) A consistent delay between concentration change and response could challenge an immediate direct-effect model. However, the delay could also reflect processes such as absorption or distribution, so that pattern alone does not prove a turnover mechanism.