Why Sampling Time Determines What Can Be Learned
An effect-compartment model can describe a delayed response even when plasma concentration changes quickly. But a well-fitting model does not guarantee that its equilibration rate, \(k_{e0}\), is precisely estimated. If every response measurement is taken before the effect site has had time to respond, or only after the response has nearly recovered, the observations may contain little information about the rate of equilibration. Sampling must capture how the system changes, not merely record many observations.
The effect-site concentration \(C_e(t)\) follows the first-order equation \(dC_e/dt=k_{e0}(C_p(t)-C_e(t))\), where \(C_p(t)\) is plasma concentration. This tutorial considers design for estimating \(k_{e0}\) from plasma and response measurements. It assumes the effect-compartment model and a specified concentration-response model are scientifically reasonable. In the basic design comparison below, the plasma concentration trajectory is treated as known; if it must also be estimated, its uncertainty and plasma sampling schedule must be included in the design simulation.
Choose Times That Reveal the Dynamics
A useful first scale for planning is the equilibration half-life, \(t_{1/2,e}=\ln(2)/k_{e0}\). It is the time required for the difference between \(C_e\) and a constant \(C_p\) to halve. It is not a guarantee that the response peaks at that time: when plasma concentration itself changes, the peak also depends on the plasma trajectory and the response model.
Measurements have different roles. Early response samples show the onset of effect. Samples near the anticipated effect peak help locate the transition from rising to falling effect. Samples during recovery reveal how quickly the effect site follows declining plasma concentration. Plasma measurements are needed across the relevant input trajectory—often including its rise, peak or end of infusion, and decline—if that trajectory is not already adequately characterized by the pharmacokinetic model. Response samples without suitable plasma information may not distinguish slow equilibration from an incorrectly specified plasma input.
The value of a response measurement depends on its sensitivity to the parameters. For a response model with predicted mean \(m(t;\boldsymbol{\theta})\), where \(\boldsymbol{\theta}\) includes \(k_{e0}\) and the other estimated parameters, define the sensitivity of the prediction to parameter \(\theta_j\) as \(\partial m(t;\boldsymbol{\theta})/\partial\theta_j\). Large sensitivity to \(k_{e0}\) is useful, but not sufficient: if the same observations are similarly sensitive to another parameter, the two parameters may be difficult to separate.
For example, under a linear response model \(m(t)=E_0+\beta C_e(t)\), with fixed \(C_p(t)\), the sensitivity \(S(t)=\partial C_e(t)/\partial k_{e0}\) obeys a convenient differential equation. Differentiating the effect-compartment equation with respect to \(k_{e0}\) gives:
This result follows from the product rule: the derivative of \(k_{e0}(C_p-C_e)\) is \((C_p-C_e)-k_{e0}(\partial C_e/\partial k_{e0})\), assuming the specified plasma input does not itself depend on \(k_{e0}\). Under the linear response model, the response sensitivity to \(k_{e0}\) is \(\beta S(t)\). Sensitivities can be calculated along with model predictions and used to compare candidate times.
If residual errors are independent, normally distributed, and have known standard deviation \(\sigma\), a local approximation to the information from observations at times \(t_i\) is the sensitivity matrix product \(I=J^\mathsf{T}WJ\). Here \(J_{ij}=\partial m(t_i;\boldsymbol{\theta})/\partial\theta_j\) and \(W\) has diagonal entries \(1/\sigma^2\). A well-conditioned information matrix generally indicates that parameter effects are more distinguishable; its inverse approximates the parameter covariance near the assumed parameter values. This approximation is local and depends on the assumed model, parameter values, and error structure.
Worked Example: Spanning the Equilibration Window
Worked Example: Spanning the Equilibration Window
Suppose a hypothetical compound is expected to have \(k_{e0}=0.70\ \text{h}^{-1}\). A plasma input rises quickly, remains near \(10\ \text{mg/L}\) from time zero through 2 hours, and then declines. Plan a response schedule that includes onset, the expected peak region, and recovery.
First calculate the planning half-life:
A candidate response schedule is 0.25, 0.50, 1.0, 1.5, 2.0, 2.5, 4.0, and 7.0 hours. The first measurements are within roughly one equilibration half-life and can show the onset. The 1.5- and 2.0-hour observations describe the approach to the end of the sustained plasma input. Measurements at 2.5 and 4.0 hours sample the transition and decline, while 7.0 hours checks later recovery.
This is a starting design, not proof of optimality. The expected response peak depends on the actual plasma profile and response model. Plasma samples should be scheduled to characterize the changing input as required by the PK model; response samples should not be concentrated only around the presumed maximum. If the plausible \(k_{e0}\) range is broad, evaluate the schedule across that range rather than relying on one half-life estimate.
Compare Designs by Simulating Their Precision
A simulation-based comparison asks what parameter estimates a design would produce repeatedly if the assumed model and error process were true. For each candidate schedule, simulate data at the planned times, fit the same model that will be used in the study, and summarize the resulting estimates. Useful summaries include bias, standard deviation or root mean squared error for \(k_{e0}\), and coverage of a stated confidence interval. A design should not be chosen on the basis of a small standard error alone if it also produces substantial bias or poor interval coverage.
Set plausible values or distributions for \(k_{e0}\), response parameters, the plasma trajectory, and residual variability. State which quantities are fixed and which are estimated.
For each candidate schedule, calculate model predictions at its plasma and response times, then simulate observations using the assumed error model.
Fit the planned analysis model to every replicate. Record failed fits as well as successful estimates, and summarize bias, precision, and interval coverage.
Repeat across plausible parameter values, residual variability, and plasma-input uncertainty. Prefer a schedule that performs adequately across credible scenarios, not just at one nominal setting.
Worked Example: Comparing Three Hypothetical Schedules
Consider an illustrative simulation with true \(k_{e0}=0.70\ \text{h}^{-1}\), a known plasma trajectory, and response mean \(m(t)=100+3C_e(t)\), where response is in hypothetical response units and concentration is in \(\text{mg/L}\). Each response observation has independent normal error with standard deviation \(0.50\) response units. The fitted model estimates \(k_{e0}\), baseline \(E_0\), and slope \(\beta\); it treats the plasma trajectory and residual standard deviation as known. Thus this comparison isolates response-time effects and does not account for uncertainty in the PK input.
Three eight-sample schedules are compared: an early schedule at 0.15, 0.30, 0.45, 0.60, 0.75, 0.90, 1.05, and 1.20 hours; a peak-region schedule at 1.8, 2.0, 2.2, 2.4, 2.6, 2.8, 3.0, and 3.2 hours; and a spread schedule at 0.25, 0.50, 1.0, 1.5, 2.0, 2.5, 4.0, and 7.0 hours. For this hypothetical setup, assume 2,000 replicate data sets were generated and fitted for each schedule, using the same estimation procedure and Wald intervals throughout.
| Schedule | Mean estimated \(k_{e0}\) (\(\text{h}^{-1}\)) | Monte Carlo SD (\(\text{h}^{-1}\)) | 95% interval coverage |
|---|---|---|---|
| Early | 0.704 | 0.064 | 93% |
| Peak-region | 0.712 | 0.047 | 92% |
| Spread | 0.701 | 0.022 | 95% |
For example, the spread schedule’s estimated bias is \(0.701-0.700=0.001\ \text{h}^{-1}\), while its Monte Carlo standard deviation is \(0.022\ \text{h}^{-1}\). In these assumed simulations, the spread schedule had the smallest variability and the closest-to-nominal interval coverage. The result does not establish that spreading samples is universally optimal: a different PK profile, error level, response function, or parameter range could change the ranking. It does show why comparing schedules through repeated fits is more informative than counting observations alone.
Worked Example: Interpreting a Sensitivity-Based Warning
Worked Example: Interpreting a Sensitivity-Based Warning
Suppose the same linear response model is planned, and a proposed schedule puts all response measurements very early, when \(C_p(t)\) is approximately constant at \(10\ \text{mg/L}\). For a constant input beginning at time zero and \(C_e(0)=0\), the effect-site solution is \(C_e(t)=10(1-e^{-k_{e0}t})\). Its sensitivity is:
At \(k_{e0}=0.70\ \text{h}^{-1}\) and \(t=0.10\ \text{h}\), this equals \(10(0.10)e^{-0.07}=0.9324\ \text{mg/L per }(\text{h}^{-1})\). The response sensitivity is therefore \(3(0.9324)=2.797\) response units per \(\text{h}^{-1}\), given \(\beta=3\). But at early times, \(1-e^{-k_{e0}t}\) is approximately \(k_{e0}t\). Consequently, \(m(t)\) is approximately \(E_0+\beta(10k_{e0}t)\), so a change in \(k_{e0}\) can resemble a change in \(\beta\). Early observations may be individually sensitive yet still fail to distinguish these parameters well.
The correction is not simply to avoid early samples. They help establish onset and baseline behavior. Instead, add observations at times when the curvature and recovery of the response provide information that differs from the effect of changing the response slope. Then check parameter correlations, uncertainty, and fit failures in the simulation. Sensitivity to one parameter should always be assessed alongside sensitivity to the others.
Diagnostics, Alternatives, and Limitations
A design may be failing to identify \(k_{e0}\) if simulated fits often reach parameter bounds, estimates vary widely between replicates, the fitted \(k_{e0}\) is strongly correlated with response parameters, or confidence intervals are wide or have poor coverage. Frequent convergence failures are also important design results, not records to discard. In study data, a flat profile likelihood or strong dependence of estimates on starting values can similarly signal weak information; the earlier material on profile likelihood provides a complementary way to examine uncertainty after fitting.
The main alternatives to simulation-based design evaluation are useful but answer narrower questions. Evenly spaced sampling is simple to implement, but it does not automatically allocate observations to informative phases. Sensitivity or Fisher-information calculations are computationally efficient and help explain why times are informative, but they approximate uncertainty around specified parameter values and may be misleading for nonlinear, bounded, or weakly identified problems. Simulation directly evaluates the planned fitting procedure under stated assumptions, but it can be computationally expensive and only evaluates the scenarios that were simulated.
Practical constraints also matter. Blood-volume limits, assay logistics, clinic hours, and missed visits can make a mathematically attractive schedule infeasible. A robust design may include a few early measurements, observations around the anticipated transition, and later recovery samples rather than placing nearly all sampling at a theoretical optimum. If plasma concentrations are uncertain, simulate that uncertainty too; treating the input as known can overstate the precision of \(k_{e0}\). Finally, design performance does not validate the effect-compartment model or prove a physiological interpretation of its estimated rate.
- Pitfall: maximizing the number of samples near the predicted response peak. Closely clustered observations may add little new information about the time course and can leave onset or recovery poorly characterized. Compare schedules with the same sampling burden and include measurements from distinct phases.
- Pitfall: choosing times from the nominal half-life alone. The half-life is a useful scale, not a complete sampling plan. Consider uncertainty in \(k_{e0}\), the plasma input, response parameters, and operational constraints.
- Pitfall: treating a small simulated standard deviation as proof of a good design. Check bias, interval coverage, failed fits, and robustness to plausible assumptions as well.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If \(k_{e0}=0.35\ \text{h}^{-1}\), what is the equilibration half-life? Give the calculation and units.
- Why can a schedule containing only very early response measurements confound \(k_{e0}\) with the slope \(\beta\) in a linear response model?
- A proposed design has a small Monte Carlo standard deviation for \(k_{e0}\), but its confidence intervals cover the true value in only 70% of simulations when nominal coverage is 95%. Is this sufficient evidence to prefer the design? Explain.
- In a simulation comparison, the plasma trajectory is treated as known exactly, although the planned study will estimate it from sparse plasma samples. What limitation does this introduce?
Solutions. 1. \(t_{1/2,e}=0.6931/(0.35\ \text{h}^{-1})=1.980\ \text{h}\). 2. At early times, \(1-e^{-k_{e0}t}\approx k_{e0}t\), so \(E_0+\beta C_e(t)\approx E_0+\beta k_{e0}C_p t\); the data may mainly inform the product \(\beta k_{e0}\), not the two parameters separately. 3. No. The poor coverage indicates that the reported intervals are not reliably representing uncertainty under that simulation scenario; compare bias, fit behavior, and interval construction before preferring the design. 4. It can overstate precision because the simulated estimates omit uncertainty in the plasma input. Reevaluate the design with plausible PK estimation uncertainty or joint plasma-response data generation and fitting.