Why sampling schedules need to be compared by simulation
A pharmacokinetic (PK) sampling schedule determines which parts of a concentration-time profile are observed. A schedule that collects several early samples may describe initial concentrations well but provide little information about the terminal decline. A schedule with later samples may better distinguish slow from fast elimination, but late concentrations can be close to the assay’s lower limit of quantification (LLOQ). These differences can affect estimates of clearance and exposure, and can carry forward into exposure classification or treatment decisions.
The useful design is therefore not necessarily the one with the most samples or the smallest standard error for one PK parameter. It is the feasible schedule that performs acceptably on prespecified targets, under plausible assumptions about patients, dosing, residual error, and analysis. The introductory material on clinical trial simulation provides the general framework for comparing designs; here, the simulation is focused on when concentration samples are collected and what those samples support.
Before simulating, specify the design constraints and targets. Constraints may include the total number of samples per participant, clinic visit windows, assay capacity, and burden on participants. Targets may include precision and bias for clearance or other PK parameters, accuracy of classifying participants above an exposure threshold, and the frequency with which the planned treatment analysis makes a suitable decision. Each target should be defined before comparing schedules; otherwise a design can appear preferable simply because the criterion was chosen after seeing the results.
Model assumptions and a simple information calculation
A simple one-compartment model after an intravenous bolus illustrates why sample times matter. Assume dose \(D\) enters the central compartment immediately, volume \(V\) is constant, and elimination is first-order with clearance \(CL\). Define \(k=CL/V\). The predicted concentration at time \(t\) is \(C(t)=(D/V)\exp(-kt)\). For this illustration, assume measured concentrations are positive and have multiplicative lognormal error: the log concentration equals the log prediction plus independent, mean-zero normal error with standard deviation \(\sigma\). The calculation does not include absorption, multiple compartments, or concentrations censored below the LLOQ; those features must be added when they are relevant to the intended study.
Taking logs gives a straight-line model, \(\log C(t)=a-kt+\epsilon\), where \(a=\log(D/V)\) and \(\epsilon\) has variance \(\sigma^2\). Under these assumptions, with a common \(\sigma\) and independently observed samples at times \(t_i\), the least-squares estimate of the slope has standard error \(\sigma/\sqrt{S_{tt}}\), where \(S_{tt}=\sum_i(t_i-\bar t)^2\). This follows because the slope estimate is the sum of \((t_i-\bar t)\log C(t_i)\), divided by \(S_{tt}\); the independent errors contribute variance \(\sigma^2\sum_i(t_i-\bar t)^2/S_{tt}^2=\sigma^2/S_{tt}\). Wider time separation can therefore improve precision for \(k\), provided the model remains appropriate and late samples are measurable.
Worked Example: Comparing time spread for an elimination-rate estimate
Suppose a hypothetical participant receives a 100 mg intravenous bolus. For four measurements, compare an early schedule at 0.5, 1, 2, and 4 hours with a schedule at 1, 4, 8, and 12 hours. Assume log-scale residual standard deviation \(\sigma=0.15\), independent errors, and that all concentrations are quantifiable. These assumptions let us use the straight-line calculation above; they do not establish that the model is adequate for a real drug.
For the early schedule, \(\bar t=(0.5+1+2+4)/4=1.875\) hours, and \(S_{tt}=(-1.375)^2+(-0.875)^2+(0.125)^2+(2.125)^2=7.1875\) hours squared. Thus \(SE(\hat k)=0.15/\sqrt{7.1875}=0.0560\) per hour.
For the schedule spread across 1 to 12 hours, \(\bar t=(1+4+8+12)/4=6.25\) hours, and \(S_{tt}=(-5.25)^2+(-2.25)^2+(1.75)^2+(5.75)^2=68.75\) hours squared. Its standard error is \(0.15/\sqrt{68.75}=0.0181\) per hour. Under this simplified model, the later-spread schedule gives more information about \(k\). In a real design, that apparent advantage must be checked against quantification limits, assay error at low concentrations, and possible model misspecification.
This calculation is a justification for examining time spread, not a substitute for the full design simulation. In population PK, participants can differ in \(CL\) and \(V\); schedules can involve different sample counts or staggered times; and measured data may be censored, noisy, or collected outside the nominal visit time. A simulation can represent these features and assess the analysis that will actually be used.
Build the simulation around the decision
A defensible comparison states both how data are generated and how they are analyzed. For example, a population model might draw each participant’s \(CL\) and \(V\) from distributions representing expected between-participant variability, calculate concentrations at the assigned sampling times, add residual error, and apply the assay’s LLOQ rule. The analysis then fits the planned PK model to the simulated observations. If the proposed analysis estimates population parameters, the simulation should include the intended number of participants and the planned distribution of sampling times across them. A schedule that gives each person only a few samples may still be informative if times are distributed across participants, but the allocation must be represented explicitly.
For each replicate, record the outcomes tied to the design question. Parameter outcomes can include relative bias, standard error or interval width, convergence rate, and interval coverage. Exposure outcomes can include accuracy and sensitivity for a prespecified threshold. Decision outcomes can include how often a predefined treatment rule selects the preferred regimen. Summaries across replicates estimate these operating characteristics; Monte Carlo uncertainty should be checked, particularly when a decision is rare or the schedules have similar performance. The earlier material on estimating operating characteristics and quantifying Monte Carlo error describes these general simulation summaries.
Specify candidate sample times, sample counts, participant allocation, visit windows, and resource limits. Keep these constraints visible when interpreting performance.
Use stated PK parameters, between-participant variability, residual error, dosing assumptions, and assay behavior. Include plausible deviations from the primary model in sensitivity scenarios.
Fit each replicate using the analysis intended for the trial, including its handling of below-quantification observations, missing samples, and parameter constraints.
Compare precision and bias, exposure classifications, and treatment decisions, while checking failures and Monte Carlo uncertainty.
Worked simulation comparison
Worked Example: Comparing two sparse schedules across simulated trials
Consider a hypothetical trial with 240 participants receiving a 100 mg intravenous bolus. Under the generating model, typical \(V=20\) L and typical \(CL=4\) L/h, so typical \(k=CL/V=0.20\) per hour and typical AUC is \(D/CL=25\) mg·h/L. Participants vary in clearance and volume, concentration measurements have lognormal residual error with \(\sigma=0.15\), and the assay LLOQ is 0.10 mg/L. Compare Schedule A, four samples per participant at 0.5, 1, 2, and 4 hours, with Schedule B, four samples at 1, 4, 8, and 12 hours. Assume the analysis fits the same population PK model for both. These values and the results below are illustrative simulation outputs, not findings from a real study.
Suppose 1,000 simulated trials are analyzed per schedule. The simulation calculates relative bias in typical \(CL\) as \(100(\text{mean estimate}-4)/4\), relative standard error as \(100\times SE(\widehat{CL})/CL\), and empirical 95% interval coverage as the proportion of intervals containing the generating typical \(CL\). For a clinical target, it classifies each participant as above or below 30 mg·h/L AUC and compares that label with the label based on generated AUC. A prespecified downstream rule selects the regimen if its estimated probability of meeting a clinically meaningful response threshold exceeds 0.80. Assume that, in this hypothetical setting, one regimen is the preferred regimen under the generating model. The table reports illustrative results from these definitions.
| Performance measure | Schedule A: 0.5, 1, 2, 4 h | Schedule B: 1, 4, 8, 12 h |
|---|---|---|
| Relative bias in typical \(CL\) | +2.1% | +0.8% |
| Median relative standard error for typical \(CL\) | 18% | 11% |
| Empirical coverage of 95% intervals for typical \(CL\) | 92.6% | 94.7% |
| Correct AUC-threshold classification | 82.4% | 91.3% |
| Correct preferred-regimen selection by the rule | 74.8% | 84.1% |
The results favor Schedule B on each displayed target, but the decision is not simply “choose B.” First, check feasibility: late samples may require additional visits and may be more vulnerable to missingness. Second, examine diagnostics and sensitivity scenarios before interpreting the numerical differences as reliable. Third, compare the advantage with its operational cost. For example, if Schedule B’s late samples are frequently lost in a plausible dropout scenario, its apparent advantage may shrink. A schedule with slightly less precision could be preferable if it is more reliably implemented.
Exposure classification and downstream decisions
An exposure classification is only as useful as its definition and purpose. If the threshold is 30 mg·h/L AUC, a participant with true AUC 29.8 mg·h/L and one with true AUC 30.2 mg·h/L are likely to be harder to distinguish than participants far from the threshold. Report overall accuracy together with measures that describe errors in each direction: sensitivity is the proportion of truly above-threshold participants classified above threshold, and specificity is the proportion of truly below-threshold participants classified below threshold. Also show results by distance from the threshold or across clinically relevant thresholds when appropriate. Overall accuracy alone can look high when one class is common.
For a downstream treatment-effect analysis, the simulation must use the actual planned exposure-response model or decision rule. Better PK parameter precision does not guarantee a better clinical decision: exposure estimates may remain uncertain at the individual level, the exposure range may be narrow, or the treatment decision may depend on patients near a response threshold. Conversely, a schedule may be adequate for choosing between regimens even if individual clearance estimates are imprecise. Match each design criterion to the decision it is intended to support, and report decision performance under both beneficial and null or unfavorable scenarios where relevant.
A useful short derivation clarifies why individual classification can remain difficult even with an unbiased population estimate. Suppose an individual’s estimated exposure is \(\widehat A=A+e\), where \(A\) is true exposure and \(e\) is estimation error. Classification at threshold \(T\) is correct if \(A+e>T\) when \(A>T\), or \(A+e\leq T\) when \(A\leq T\). When \(A\) is close to \(T\), a small error can reverse the label; reducing population-level bias alone does not eliminate that probability. The simulation should therefore evaluate participant-level classification, not infer it from the precision of a typical population parameter.
Diagnostics, alternatives, and limitations
Look for signs that a schedule’s apparent performance is being driven by an analysis problem rather than useful information. Check the proportion of concentrations below the LLOQ by time and participant, parameter estimates at imposed bounds, convergence and failure rates, residual patterns over time, and whether estimates are unstable across initial values. Compare estimates with generating values and inspect intervals, not just point estimates. The earlier material on numerical convergence, simulation validation, and model misspecification provides relevant checks; here, those checks should be applied separately for each schedule.
Also test assumptions that could change the ranking: alternative values of between-participant variability, residual error, assay LLOQ, dose, model structure, missing samples, and visit-time deviations. If one schedule wins only under a narrow set of assumptions, that dependence is itself a design result. A model that omits absorption or distribution phases may make a schedule appear informative for the wrong parameter. Simulated conclusions are conditional on the generating scenarios; they do not establish that the assumed model describes future participants.
| Approach | Strength | Main limitation |
|---|---|---|
| Rich sampling in every participant | Provides detailed individual profiles and can support several PK questions. | Can increase burden, cost, and missed visits; may not be feasible at scale. |
| Sparse sampling with simulation | Directly compares feasible schedules using the intended analysis and decision criteria. | Results depend on the simulated PK, error, missingness, and analysis assumptions. |
| Staggered or population sampling | Distributes times across participants and may cover a broad profile with few samples per person. | Requires suitable allocation and an analysis that can use the resulting information; individual estimates may remain imprecise. |
| Prior-informed or optimal-design calculations | Can efficiently identify candidate times for a specified model and information target. | May be sensitive to prior values or model assumptions, and an information criterion need not optimize a clinical decision. |
A simulation comparison does not remove the need for scientific judgment. The optimal schedule for estimating typical clearance may differ from the schedule that best classifies individual exposure or supports a treatment decision. A schedule may also be infeasible despite good statistical performance. State these limitations, present the trade-offs, and avoid treating one summary metric as a universal design ranking.
Check Your Understanding
Answer each question, then compare with the solutions.
- Why can samples spread over time improve estimation of an elimination rate in a one-compartment model?
- A schedule has low bias for typical clearance but poor sensitivity for identifying participants above an exposure threshold. Is it necessarily adequate for an exposure-guided decision? Explain.
- In a simulation, 90% of participants are truly below an exposure threshold. Why can overall classification accuracy be misleading?
- Name two diagnostics that could reveal a schedule is performing poorly because of assay or fitting problems.
- What should be done if the preferred schedule changes when the assumed LLOQ or residual error changes?
Solutions. (1) Under the log-linear model with independent equal-variance errors, the slope standard error is \(\sigma/\sqrt{\sum_i(t_i-\bar t)^2}\); wider time spread increases the denominator, if the model and observations remain valid. (2) No. Poor sensitivity means many truly above-threshold participants are missed, which may undermine that decision even when the population clearance estimate is nearly unbiased. (3) A classifier that labels everyone below threshold is already correct for 90% of participants, so high accuracy could conceal failure to identify the smaller above-threshold group. Report sensitivity and specificity as well as accuracy. (4) Examples include the fraction below LLOQ by time, convergence and boundary rates, residual patterns, and instability across initial estimates. (5) Treat the ranking as assumption-sensitive, investigate which scenario is plausible, and report the trade-off; do not present one schedule as uniformly superior without qualification.