Why Structural Choices Matter
A population pharmacokinetic (PK) model is not determined by its parameter estimates alone. It also specifies a structural account of how drug enters and moves through the body, and a residual-error model for the difference between predicted and observed concentrations. A model can fit the sampled concentrations reasonably well yet predict a different early or late exposure profile from a plausible alternative. That distinction matters when development decisions depend on peak concentration, trough concentration, partial AUC, or the timing of target coverage.
This tutorial focuses on the structural choices most directly relevant to standard clinical concentration-time data: one- versus two-compartment disposition, alternative absorption descriptions, and residual-error structures. For clarity, examples use fixed parameter values and do not include between-participant or between-occasion random effects. A structural comparison is not a claim that one parameter set describes every participant; it is a comparison of alternative model forms under stated assumptions.
The basic assumptions used below are linear, time-invariant PK over the modeled dose range: clearance and distribution parameters do not change with concentration or time; the administered dose and sampling times are known; and there is no unmodeled input, such as a second dose, during the interval being calculated. These assumptions may be reasonable starting points, not facts guaranteed by the data. If they fail, a more complex structural model may be needed, but complexity should be supported by design and evidence.
One- and Two-Compartment Disposition
For an intravenous bolus in a one-compartment model, the amount of drug is assumed to distribute instantaneously and uniformly through an apparent volume \(V\). With linear clearance \(CL\), the elimination rate constant is \(k=CL/V\), and concentration declines exponentially. A two-compartment model instead distinguishes a central compartment, where drug is measured, from a peripheral compartment that exchanges drug with the central compartment. For a linear two-compartment bolus model, the central concentration has two exponential components: a relatively rapid distribution phase and a slower terminal phase.
Here \(C_0\), \(A\), and \(B\) have units of concentration; \(k\), \(\alpha\), and \(\beta\) have units of inverse time. In the two-compartment expression, both exponentials contribute at time zero, so \(C_{\text{2c}}(0)=A+B\). The coefficients and rate constants are linked to the underlying volumes, clearance, and intercompartmental exchange in a mechanistic compartment model; they are not four unrelated biological quantities.
Worked Example: Comparing Disposition Predictions in a Hypothetical Clinical Dataset
Suppose a hypothetical participant receives an intravenous bolus, and concentrations are measured at 0, 4, and 24 hours. The invented observations are 10.0, 2.0, and 0.180 mg/L, respectively. Consider two candidate descriptions: a one-compartment model with \(C_0=10.0\) mg/L and \(k=0.200\) h\(^{-1}\), and a two-compartment model with \(A=7.0\) mg/L, \(B=3.0\) mg/L, \(\alpha=1.20\) h\(^{-1}\), and \(\beta=0.120\) h\(^{-1}\). These values are illustrative candidate parameters, not estimates from a real study.
For the one-compartment model at 4 hours,
At 24 hours, the same model predicts \(10.0e^{-4.8}\approx0.08232\) mg/L. The two-compartment predictions are
The two-compartment predictions are closer to the invented observations at both times. The concentration-time curves imply different exposure as well. Integrating each expression from zero to 4 hours gives
These different partial AUCs illustrate why a model choice can matter for a development decision. They do not establish that the two-compartment structure is correct: the parameters have not been estimated by fitting the full dataset, and additional observations could change the comparison. The assumed models also imply different total AUCs to infinity: \(C_0/k=50.0\) mg·h/L for the one-compartment model, compared with \(A/\alpha+B/\beta=30.83\) mg·h/L for the two-compartment model. A fair empirical comparison estimates each model’s parameters from the same data and evaluates the resulting predictions, rather than treating these illustrative parameter sets as a model-selection test.
A useful justification for the AUC calculation is direct integration. For the two-compartment form, integrate each exponential separately: the integral of \(Ae^{-\alpha t}\) from zero to \(T\) is \(A(1-e^{-\alpha T})/\alpha\), with the analogous result for the second component. Taking \(T\) to infinity gives \(A/\alpha+B/\beta\), provided both rate constants are positive. This calculation also shows why terminal concentrations can matter greatly for total AUC: the slower component contributes an amount proportional to \(1/\beta\).
Absorption Structures for Oral Data
For an oral dose, the observed concentration reflects both absorption into the systemic circulation and disposition after absorption. In a one-compartment model with first-order absorption, the absorption rate constant is \(k_a\), and a lag time \(T_{\text{lag}}\) delays the start of absorption. With dose \(D\), bioavailability \(F\), and apparent volume \(V\), concentration after the lag is obtained by integrating the input over time while drug is eliminated.
This expression assumes a single, instantaneous oral dose, first-order absorption, linear one-compartment disposition, and no presystemic or time-varying input beyond the factor \(F\). It also requires \(k_a\ne k\); when the rates are equal, the limiting form is used. The equation follows from the input-response convolution: each amount entering at time \(u\) contributes to concentration at time \(t\) after elimination for \(t-u\). Integrating the first-order input \(FDk_a e^{-k_a u}\) against the disposition factor \(e^{-k(t-u)}/V\) from \(u=0\) to \(u=t\) yields the difference of exponentials above.
Worked Example: First-Order Absorption With and Without a Lag
Consider hypothetical oral-dose predictions described by \(C(t)=12.0(e^{-0.200t}-e^{-1.20t})\) mg/L, where time is in hours. The coefficient is consistent with \(FD/V=10.0\) mg/L, because \(FDk_a/[V(k_a-k)]=10.0(1.20)/(1.20-0.200)=12.0\). Compare this model with the same input delayed by a one-hour lag.
The AUC from zero to 4 hours for the no-lag model is found by integrating each exponential:
With \(T_{\text{lag}}=1\) hour, there is no concentration from this dose during the first hour. The interval from zero to 4 hours therefore contains three hours of post-lag exposure:
The partial AUCs differ by about \(23.1226-17.3445=5.7781\), or 5.78 mg·h/L. In contrast, the AUC to infinity is \(FD/CL=50.0\) mg·h/L in both models: a lag shifts the concentration-time profile but does not change the total absorbed amount under these assumptions. Thus, a lag can be important for early exposure or onset even when total AUC is unchanged.
First-order absorption is not the only plausible description. A transit-compartment model represents input as delayed and spread over time through successive transit stages; a zero-order input assumes a constant absorption rate over a specified duration. These alternatives can produce similar curves over a narrow sampling window but different peak times and early exposure. Choosing among them requires sampling informative about the absorption phase. A very rapid absorption rate may be difficult to estimate if the first sample occurs after the peak; a slow terminal decline can also be confused with slow absorption, a phenomenon often called flip-flop kinetics.
Residual Error Is a Separate Modeling Choice
Even when the structural prediction \(f_i\) is appropriate, observed concentration \(y_i\) will generally differ from it. The residual-error model describes this difference. An additive model assumes a constant absolute standard deviation; a proportional model makes variability increase with the predicted concentration; and a combined model includes both components. A commonly used combined variance is the sum of the additive and proportional variances.
These normal-error forms assume residuals are conditionally independent, centered on the structural prediction, and adequately described by their stated variance. The simple proportional form can assign positive probability to negative observations; for positive concentrations, a lognormal error model is another option. In a lognormal specification, the model is defined on the log scale, so its central prediction must be interpreted accordingly rather than automatically as the arithmetic mean.
Worked Example: How Error Structure Changes Residual Scaling
Suppose a hypothetical concentration is observed as 10.0 mg/L and the structural model predicts 8.0 mg/L. For an additive error model with standard deviation \(\sigma_a=1.00\) mg/L, the residual is 2.00 mg/L and its standardized value is \(2.00/1.00=2.00\).
Under a proportional model with \(\sigma_p=0.200\), the residual standard deviation at the prediction is \(0.200(8.0)=1.60\) mg/L. The standardized residual is therefore \(2.00/1.60=1.25\). Under a combined model with \(\sigma_a=0.500\) mg/L and \(\sigma_p=0.150\), the standard deviation is
The corresponding standardized residual is \(2.00/1.300=1.54\). The same observed-minus-predicted difference thus has different relative weight under the three error assumptions. The example is arithmetic, not a basis for selecting an error model from one observation; selection should use the pattern across concentrations and participants, together with the fitting method’s likelihood assumptions.
Comparison, Diagnostics, and Limitations
| Choice | What it can represent | Useful evidence and caution |
|---|---|---|
| One- versus two-compartment disposition | Single exponential decline versus an additional distribution phase and terminal component | Look for systematic early-to-late pattern in residuals and adequate samples across phases; sparse or late-only data may not identify both phases. |
| First-order absorption versus lag, transit, or zero-order input | Different onset, absorption delay, peak timing, and input duration | Early samples help distinguish input patterns; an apparent lag or slow absorption can be poorly identified when samples miss the absorption phase. |
| Additive, proportional, or combined residual error | Constant absolute, concentration-dependent, or mixed residual spread | Assess residual spread across the prediction range; avoid using a variance form that systematically underweights one concentration range. |
A practical structural review starts by plotting observed concentrations and candidate predictions on both linear and logarithmic concentration scales. A one-compartment curve that repeatedly overpredicts early concentrations but underpredicts the terminal phase may be missing distribution structure; a curve that misses the observed peak timing may have an inadequate input model. Residuals should be examined against time and predicted concentration: a fan-shaped spread can signal proportional variability, while a roughly constant absolute spread can support additive error. These patterns are clues, not proof, because sampling, measurement error, data anomalies, or omitted variability can produce similar appearances.
Compare candidate models using the same records and a compatible estimation and error framework. A more complex model can improve fit simply by having additional parameters; that improvement does not alone show that the extra compartment or input process is biologically necessary or that its parameters are estimable. Check parameter plausibility, uncertainty, convergence, and whether the data cover the portions of the profile relevant to the intended decision. Where a model is used to predict beyond the sampled times or doses, the extrapolation warrants particular scrutiny.
The principal limitation is identifiability: several structures may describe the observed samples similarly while implying meaningfully different unobserved profiles. This is especially likely when sampling misses distribution or absorption phases, or when concentrations are near the assay’s quantification limit. Structural choices should therefore be treated as a source of prediction uncertainty. A good fit is necessary but not sufficient; the model must also be supported by the design, have interpretable parameters, and predict the exposure summaries needed for the development question.
Choose disposition, absorption, and residual-error forms that correspond to the dosing route, design, and clinical question.
Ask whether samples cover the expected distribution, absorption, peak, and terminal phases, and whether the observations provide information about the parameters being added.
Review residual patterns and parameter stability, then compare the exposure summaries that matter for the development decision.
Check Your Understanding
Try each question, then compare your reasoning with the solutions.
- In a two-compartment bolus model, what does the slower exponential component contribute to the terminal profile?
- A one-hour lag is added to the oral model in the worked example. Does it change AUC to infinity under the stated assumptions? Explain.
- An additive residual model has a standard deviation of 0.80 mg/L. What is the standardized residual when the observation is 5.0 mg/L and the prediction is 3.4 mg/L?
- Why may a two-compartment model be hard to distinguish from a one-compartment model when early samples are absent?
- Residual spread increases as predicted concentration increases. Which error structures could be considered, and what does the pattern not prove by itself?
Solutions. (1) The slower component, with rate \(\beta\), governs the long-time decline and contributes \(B/\beta\) to AUC to infinity. (2) No. A pure lag shifts the same absorption profile later, so its total integral remains \(FD/CL\); the partial AUC over a fixed early interval can nevertheless decrease. (3) The residual is \(5.0-3.4=1.6\) mg/L, so the standardized residual is \(1.6/0.80=2.00\). (4) Without early samples, the rapid distribution phase may be missed, leaving little information to estimate its contribution separately from a single exponential decline. (5) Proportional or combined error may be considered because they allow spread to grow with concentration. The pattern alone does not prove that either error structure is correct; structural misspecification, design, and data quality must also be considered.