From PK Objectives to a Sampling Plan
A sparse schedule is not simply a short list of convenient visit times. It is a plan for distributing samples so that the population data contain information about the pharmacokinetic phases and parameters that matter for the study. The design must also be feasible for participants and robust to plausible departures from the predicted concentration–time profile.
Consider a hypothetical study of an orally administered drug. The planning team has a provisional population PK model and expects absorption, a post-peak period informative about disposition, and a measurable terminal phase within 24 hours after a single dose. The schedule below is an illustration, not a recommendation for a real drug. Its windows must be checked against that drug’s anticipated kinetics, assay performance, dosing conditions, and study objectives.
A workable design begins with explicit targets. For example, the study may need precise estimates of apparent clearance \(CL/F\), apparent central volume \(V_c/F\), and the absorption rate constant \(k_a\). If the study also needs distribution parameters or between-subject variability, those requirements should be stated too: a schedule adequate for one set of targets may be inadequate for another. Sampling times should be selected relative to the planned dose, and the planning model should predict where the important phases occur.
Worked Example: Setting Windows Around the Predicted Profile
Worked Example: Setting Windows Around the Predicted Profile
Suppose the provisional model predicts that samples within about 1.5 hours will be most informative about absorption and the approach to peak concentration, samples around 3 hours will cover the post-peak period, and samples around 7 and 20 hours will help characterize later disposition and the terminal phase. A proposed 24-hour schedule uses five windows:
| Window | Allowable time after dose | Nominal target | Planning purpose |
|---|---|---|---|
| 1 | 0.15–0.35 h | 0.25 h | Early absorption |
| 2 | 0.75–1.25 h | 1.00 h | Absorption and approach to peak |
| 3 | 2.5–3.5 h | 3.00 h | Post-peak profile |
| 4 | 6–8 h | 7.00 h | Later disposition |
| 5 | 18–22 h | 20.00 h | Potential terminal phase |
The five target times are separated enough to span the predicted profile, and the allowable windows make the schedule more feasible than requiring collection at exact instants. The labels are conditional on the planning model: a 20-hour sample is useful for the terminal phase only if the predicted profile is still measurable and terminal at that time.
The width of a window also matters. If concentration changes rapidly during an interval, the actual time within that interval can influence the concentration substantially. In that case, the design should use a narrower window or explicitly include realistic timing variation in its evaluation. The nominal target is a planning value; it does not make every sample in the window equivalent.
Allocate Samples Across Participants
Once windows are chosen, decide which participants contribute samples in each window. Balanced allocation helps avoid a schedule in which one phase is well represented and another is represented by only a few participants. It also makes total participant burden and the number of planned samples transparent.
Worked Example: Balancing Three Samples Across Five Windows
Suppose 60 participants are each assigned three of the five windows. There are \(\binom{5}{3}=10\) possible sets of three windows. Assign six participants to each set, and randomize eligible participants to the sets. Each window then appears in \(\binom{4}{2}=6\) sets, so it has \(6\times6=36\) planned samples.
The total number of planned samples is \(60\times3=180\). Checking by windows gives \(5\times36=180\), so the allocation is internally consistent. Each participant has three planned collection occasions, and each window is represented in 36 participants. A pair of windows appears together in \(\binom{3}{1}=3\) sets, or \(3\times6=18\) participants; this overlap can help provide within-participant information across phases.
This combinatorial allocation balances the five windows, but it does not by itself prove that the design is informative. For example, 36 planned samples in a terminal window are of limited value if concentrations are usually below the assay’s quantification limit there. Nor does equal window representation guarantee adequate precision for every parameter. Those questions require checking the expected profile and evaluating the design under the proposed model.
Allocation should also reflect practical constraints. If early samples require close supervision, if later samples can be collected at home, or if certain participants cannot attend particular times, the assignment scheme must account for those differences. Randomizing participants among feasible schedules helps prevent systematic differences between schedule groups, but the plan should still specify what happens when an assigned visit is missed.
Check Phase Coverage Before Precision
Phase coverage is a first-pass design check: compare planned windows with the concentration–time profile predicted by the planning model. Confirm that early observations can inform absorption, that observations around and after the peak span relevant changes in concentration, and that late observations can inform the terminal decline if terminal kinetics are a study objective. Also check whether the planned dose and follow-up period make those phases observable.
These checks rely on assumptions. The planning model must be a reasonable representation of the anticipated drug profile; the dose and dose time must be known; and the proposed windows must be achievable. The evaluation should also reflect expected measurement error, between-participant variability, and plausible missed or mistimed samples. If these assumptions are doubtful, assess more than one plausible scenario rather than treating a single predicted profile as certain.
Worked Example: Revising a Window That Misses Absorption
Suppose a revised planning model predicts a peak near 0.35 hours, much earlier than the peak used to set the original windows. The original first window, 0.15–0.35 hours, still reaches the predicted peak, but it offers little opportunity to observe the early rise. The second window, 0.75–1.25 hours, is now well after the predicted peak. Thus, the original schedule may have weak coverage of the absorption phase even though it contains two early samples.
One possible revision is to use an early window of 0.10–0.25 hours and a near-peak window of 0.30–0.60 hours, then reassess the remaining windows against the predicted post-peak and terminal profile. The change is justified by the revised predicted peak time, not by a general rule that every oral drug should be sampled at those times. Before adopting it, check whether these early collections are feasible and whether their timing can be recorded accurately enough for the intended analysis.
Estimate Expected Precision Under Explicit Assumptions
Phase coverage is necessary but not sufficient. A design can include early and late observations yet still provide weak estimates of a parameter, particularly when parameters have similar effects on the predicted profile. Expected precision is therefore assessed using the proposed model, parameter values, sampling schedule, measurement-error assumptions, and plausible between-participant variability. A typical design evaluation simulates hypothetical data under those assumptions and fits the planned model to each simulated data set.
For a parameter estimate \(\hat{\theta}\) with estimated standard error \(SE(\hat{\theta})\), the relative standard error is
RSE is a relative measure, not a guarantee of accuracy. For positive parameters, design evaluations often work on the log scale. If \(\log(\hat{\theta})\) has standard error \(SE_{\log}\), an approximate 95% interval is
This interval is a model-based description of uncertainty under the stated approximation. The design team should set precision criteria according to the study’s objectives; no single RSE cutoff is suitable for every parameter or application.
Worked Example: Interpreting Predicted Precision
In a hypothetical design simulation, the planned model estimates \(CL/F\) at 4.00 L/h with a predicted log-scale standard error of 0.13. The approximate RSE is \(100\times0.13=13\%\). The approximate interval is
The arithmetic uses \(1.96\times0.13=0.2548\); exponentiating gives a multiplicative factor of about 1.290. Thus the interval extends from \(4.00/1.290=3.10\) to \(4.00\times1.290=5.16\) L/h. This is a prediction of uncertainty under the simulation assumptions, not a result observed in an actual study.
For \(k_a\), suppose the design simulation predicts an estimate of 1.20 h\(^{-1}\) and a log-scale standard error of 0.27. The approximate RSE is 27%, and the interval is
The larger relative uncertainty for \(k_a\) than for \(CL/F\) signals that the proposed schedule is less informative about absorption under this model. The team should compare both results with its prespecified precision goals and consider changing early windows or increasing the number of appropriately allocated participants.
Use Design Diagnostics to Guide Revisions
A design evaluation is only as useful as its assumptions and diagnostics. Check more than one summary of precision and inspect whether estimation failures cluster around particular parameters or simulated data sets. Warning signs include frequent convergence problems, large or unstable standard errors, strong parameter correlations, systematic bias in estimates, and samples predicted to fall below the assay’s quantification limit. A schedule can have a respectable average precision while performing poorly in a substantial fraction of simulated studies.
When results are weak, revise the part of the schedule connected to the weakness. For example, limited information about \(k_a\) may motivate earlier, more discriminating windows; weak terminal-phase information may motivate later measurable samples; and high correlation between two parameter estimates may require windows where their predicted effects differ. Then repeat the evaluation under the revised schedule and under plausible alternative assumptions, including missed visits or timing deviations where relevant.
The main alternatives serve different purposes. Collecting rich profiles from every participant can provide broader within-participant information but increases burden and may be infeasible. A sparse balanced schedule reduces per-participant collections while relying on information distributed across the population. Opportunistic or convenience sampling may be easier to implement, but without planned phase coverage it can leave important periods underrepresented. The choice is not simply “rich” versus “sparse”: it depends on objectives, feasibility, and the information each candidate design is predicted to deliver.
A key limitation is that precision predictions depend on the assumed model and parameter values. If the true absorption rate, variability, residual error, or concentration range differs materially from the planning assumptions, actual precision may differ from the prediction. Robust design work therefore checks plausible scenarios and treats design outputs as conditional forecasts, not promises.
Common Mistake
Another mistake is treating an allowable window as if all times within it were interchangeable. If concentrations change sharply within a window, the actual collection time matters. Define feasible windows, choose nominal targets, and include realistic timing variation in the design assessment.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a 48-participant study, each participant is assigned two of four sampling windows. If every possible pair is assigned to the same number of participants, how many participants are assigned to each pair, and how many planned samples fall in each window?
- A planning model predicts the peak at 0.4 hours, but the schedule’s earliest window is 1–2 hours. What phase-coverage concern should be raised?
- A hypothetical simulation predicts an estimate of \(CL/F=5.0\) L/h with log-scale standard error 0.20. Calculate the approximate RSE and 95% interval.
- Why is equal sample count per window not sufficient evidence that the schedule meets the study’s precision objectives?
Solutions. (1) There are \(\binom{4}{2}=6\) pairs, so each pair receives \(48/6=8\) participants. Each window occurs in three of the six pairs, giving \(3\times8=24\) planned samples per window; the total is \(4\times24=96=48\times2\). (2) The schedule starts after the predicted peak and may miss the absorption rise and peak region. Consider earlier feasible windows, then reassess the schedule under the model. (3) The approximate RSE is \(100\times0.20=20\%\). The interval is \([5.0\exp(-0.392),5.0\exp(0.392)]\), approximately \([3.38,7.40]\) L/h. (4) Equal counts describe allocation, not information about each parameter. Precision depends on phase placement, model assumptions, variability, measurement error, and parameter relationships, so it must be evaluated for the parameters that matter.