Why Stagger Sampling Times?
In a sparse pharmacokinetic (PK) study, it may be impractical or burdensome to collect many samples from every participant. Yet a schedule that samples everyone at the same few times can leave important parts of the dosing interval poorly observed. Staggering assigns different dose-relative sampling times to different participants. Each participant contributes a limited number of measurements, while the combined schedule can cover a wider range of times.
The design goal is not simply to maximize the number of distinct clock times. Sampling times should be chosen to represent relevant PK phases, such as early absorption, the expected peak region, distribution or early decline, and the later elimination phase. The locations of these phases depend on the drug, route, dose, and prior knowledge. Thus, phase windows used to plan a study are informed design choices, not universal time cutoffs.
For a planned dosing interval of length \(\tau\), a dose-relative time \(t\) is usually considered within the interval when \(0 \leq t \leq \tau\). For example, \(t=2\) h means that the sample is planned two hours after the dose, regardless of the participant’s wall-clock dosing time. A valid design depends on accurately recording actual dose and sample times: a planned time is not a substitute for the time that occurred.
The basic design problem has two linked parts. First, allocate a manageable number of samples to each participant. Second, check whether the combined participants cover the phases needed for the study question. A pooled schedule can be well spread over time while still being weak if few participants contribute to a key phase, or if all observations in a phase come from a small subset of participants.
Plan Patterns, Then Check Coverage
A practical approach is to define a small set of sampling patterns, assign participants to those patterns, and summarize the resulting schedule. The pattern is the planned set of times for a participant; it is not a guarantee that every planned sample will be collected. Before assigning patterns, specify the planned interval, the clinically or scientifically relevant phase windows, and the number of samples each participant can reasonably provide.
For a simple coverage table, divide the interval into nonoverlapping bins. For example, for an illustrative 12-hour interval, one might use \([0,1)\), \([1,3)\), \([3,6)\), and \([6,12]\) hours. These illustrative bins might represent early, peak-region, intermediate-decline, and later sampling, but they are not universal PK boundaries. Assign boundary times consistently: under this convention, a sample at 1 h belongs to \([1,3)\), and one at 3 h belongs to \([3,6)\).
For phase \(k\), let \(C_k\) be the number of planned samples in that phase and \(U_k\) the number of distinct participants scheduled to contribute at least one sample there. Counting both is useful: \(C_k\) summarizes planned measurements, whereas \(U_k\) reveals whether coverage is broadly shared. If one participant has two samples in the same phase, that contributes two to \(C_k\) but only one to \(U_k\).
Here \(N\) is the number of participants, \(m_i\) is the number of planned samples for participant \(i\), \(t_{ij}\) is that participant’s \(j\)th planned dose-relative time, \(B_k\) is phase bin \(k\), and \(I(\cdot)\) equals 1 when its condition is true and 0 otherwise. These counts do not prove that a design will estimate every PK parameter well. They are a transparent first check for concentration of samples and empty or weakly covered phases.
Worked Example: Building Four Complementary Patterns
Suppose a hypothetical study plans one 12-hour dosing interval and can schedule 24 participants for four samples each. For illustration, use the bins \([0,1)\), \([1,3)\), \([3,6)\), and \([6,12]\) h. The design team creates four patterns and assigns six participants to each.
| Pattern | Planned dose-relative times (h) | Participants |
|---|---|---|
| A | 0.25, 1, 4, 10 | 6 |
| B | 0.5, 2, 6, 11 | 6 |
| C | 0.75, 3, 8, 11.5 | 6 |
| D | 1.5, 5, 7, 9 | 6 |
Every participant has four planned measurements, so the schedule contains \(24\times4=96\) planned samples. In the early bin, patterns A, B, and C each contribute one time, giving \(6+6+6=18\) samples from 18 distinct participants. In \([1,3)\), patterns A, B, and D contribute one each, again giving 18 samples from 18 participants. In \([3,6)\), patterns A, C, and D contribute one each, also 18 samples from 18 participants.
The later bin contains one time in pattern A and two times in each of B, C, and D. Its count is \(6(1+2+2+2)=42\) planned samples. Those samples come from all 24 participants, because every pattern includes at least one time in that bin. The phase counts sum to \(18+18+18+42=96\), matching the total planned sample count. This reconciliation is a useful arithmetic check.
The schedule has a broad spread of times and gives many participants a later measurement. Whether 18, 18, 18, and 42 are adequate counts depends on the study’s objectives and expected variability; the arithmetic alone cannot establish parameter precision. The team would also check whether the chosen times genuinely represent the phases of interest for this drug and dosing interval.
A Short Justification for Counting Participants
Why not assess coverage only by counting samples? Because repeated measurements from one participant are not equivalent to measurements from several participants. For instance, two observations in a phase from the same person provide temporal information for that person, but do not mean that two different people contributed to that phase. In a population study, both within-participant sampling and representation across participants matter.
The distinction follows directly from the indicators in the coverage definitions. Each scheduled measurement contributes separately to \(C_k\), while the participant contributes at most once to \(U_k\). Thus \(U_k\leq C_k\), with equality only when no participant contributes more than one planned sample to that phase. Reporting both counts prevents a dense cluster of repeat measurements from being mistaken for broad participant coverage.
Worked Example: Finding a Major Gap in a Dense-Looking Schedule
Consider a second hypothetical 12-hour study with 16 participants, four assigned to each of four patterns. Each participant is scheduled for three samples. The initial patterns are A: 0.25, 1, 2 h; B: 0.5, 1.5, 2.5 h; C: 0.75, 2, 3 h; and D: 1, 2.5, 3.5 h. The total is \(16\times3=48\) planned samples.
Although 48 sounds substantial, every sample falls before 4 h. Using the bins above, the later bin \([6,12]\) has \(C_k=0\) and \(U_k=0\). The schedule therefore provides no planned measurements in the later interval, despite its nontrivial total sample count. It also provides no measurements between 3.5 and 6 h.
One possible redesign, retaining three samples per participant and four participants per pattern, is A: 0.25, 2, 8 h; B: 0.5, 3, 10 h; C: 0.75, 4, 11 h; and D: 1, 2.5, 11.5 h. The early bin now has \(3\times4=12\) samples from 12 participants. The \([1,3)\) bin has the 1-hour samples from D and the 2- and 2.5-hour samples from A and D, for \(3\times4=12\) samples from eight distinct participants. The \([3,6)\) bin has 3- and 4-hour samples, giving \(2\times4=8\) samples from eight participants. The later bin has one sample from every participant, so \(C_k=16\) and \(U_k=16\).
The redesign fills all four bins, but it still leaves choices to review. For example, it has fewer participants represented in the peak-region and intermediate bins than in the early and later bins. If those phases are especially important, patterns could be reallocated to increase their representation. A coverage table exposes this trade-off; it does not choose the scientifically appropriate balance automatically.
Evaluate Adequacy, Not Just Occupancy
A phase is not automatically adequately covered because its bin contains one or more observations. Review at least four features:
- Phase occupancy: Are all phases required by the study question represented?
- Participant representation: How many distinct participants contribute in each phase?
- Timing within a phase: Are observations spread usefully across the phase, or concentrated at one time?
- Largest time gaps: Is there a long interval with no planned observation where changing concentrations are expected?
The binning approach is a summary, not a replacement for looking at the actual times. For example, two schedules could have the same \(C_k\) in \([6,12]\), but one could place nearly all samples at 6 h while the other includes observations near 6, 9, and 11.5 h. A simple check is to sort the distinct planned times and inspect successive gaps, paying particular attention to gaps that cross phase boundaries. The acceptable gaps depend on expected concentration changes and the intended analysis.
Adequacy also depends on the design assumptions. Planned times should be relative to the correct dose; the dosing interval and relevant phase windows must be defined; participants assigned to patterns should be sufficiently comparable for the intended pooled analysis; and actual timing deviations and missed samples must be recorded. If participant groups differ systematically in ways associated with concentration, a difference across assigned time patterns could be confused with a time-related difference. Random allocation to patterns, when feasible, can reduce this risk, but does not fix every source of bias.
Worked Example: Checking Minimum Participant Representation
A hypothetical team plans 20 participants, four assigned to each of five patterns, with three samples per participant. The patterns are A: 0.25, 2, 8 h; B: 0.5, 3, 9 h; C: 0.75, 4, 10 h; D: 1, 5, 11 h; and E: 1.5, 6, 11.5 h. The team uses the same four illustrative bins and sets a preliminary planning target of at least eight distinct participants in each bin. This target is a design criterion for this example, not a universal standard.
For \([0,1)\), patterns A, B, and C contribute one sample each: \(U_k=3\times4=12\) participants. In \([1,3)\), patterns A, D, and E contribute one sample each: \(U_k=12\). In \([3,6)\), patterns B, C, and D contribute one sample each: \(U_k=12\). In \([6,12]\), every pattern contributes at least one time, so \(U_k=5\times4=20\). The schedule meets the stated minimum of eight participants in every bin.
There are 60 planned samples in total. The bin counts are 12 early, 12 in \([1,3)\), 12 in \([3,6)\), and 24 in \([6,12]\), which sum to 60. In the later bin, pattern E contributes two samples per participant, but those repeated samples do not raise its distinct-participant count above 20. The team would next inspect the actual times and ask whether the bins and eight-participant target are appropriate for its scientific objectives.
How Staggered Sampling Compares With Alternatives
| Approach | What participants contribute | Main design trade-off |
|---|---|---|
| Staggered sparse sampling | Different limited subsets of times across participant groups | Broad pooled time coverage with fewer samples per person, but depends on adequate representation and comparability across groups |
| Balanced sparse sampling | The same planned time pattern for each participant | Consistent within-participant timing, but the shared pattern may cover fewer distinct times |
| Rich sampling | Many measurements across the interval for each participant | Detailed within-participant time courses, generally with greater participant burden |
Staggering is most useful when the study needs broad coverage but cannot collect a dense profile from everyone. Balanced sparse sampling may be preferable when each participant should have comparable sampling times or when the relevant phases can be represented by a small shared set of times. Rich sampling may be needed when detailed individual profiles are central and burden and logistics permit it. These approaches are not interchangeable by sample count alone: they distribute information differently across participants and time.
Common Pitfalls and Limitations
- Equating total sample count with adequate coverage. A schedule can contain many measurements but none in a critical phase. Inspect the time distribution, not just the total.
- Counting repeated samples as distinct participants. Report both \(C_k\) and \(U_k\); multiple samples from one person do not increase the number of participants represented.
- Treating phase boundaries as universal. Illustrative time bins must be adapted to the drug and study question. A peak may occur outside a planned “peak-region” bin.
- Ignoring long gaps within populated bins. A nonzero bin count can conceal that all samples occur near one edge. Review the actual planned times and successive gaps.
- Assuming a schedule guarantees successful analysis. Coverage is only an initial design check. It does not by itself establish identifiability, precision, or robustness to missing data.
Staggered schedules also have a practical limitation: a participant’s few samples may not show that individual’s full concentration-time profile. The design relies on combining observations across participants under the planned analysis assumptions. If participants differ substantially, or if attendance and actual timing depart from the plan, the pooled schedule may not provide the intended evidence. Record actual times and evaluate the collected data against the planned coverage rather than treating the schedule table as a record of what was observed.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a 10-hour interval, what makes a schedule “staggered sparse” rather than rich sampling?
- In one phase, six participants contribute one sample each and one participant contributes three. What are \(C_k\) and \(U_k\)?
- A plan has 80 samples, all scheduled before 2 h of a 12-hour interval. What can be concluded about later-phase coverage?
- Why should a design review inspect exact sampling times even after a phase-coverage table has been completed?
Solutions. (1) Participants are assigned different planned time subsets, each contributing a limited number of measurements, while the pooled schedule covers more times than any one subset. (2) \(C_k=6+3=9\) planned samples and \(U_k=6+1=7\) distinct participants. (3) The total count does not establish adequate coverage; the plan has no scheduled samples after 2 h, so later phases are unrepresented. (4) Counts by phase can hide samples clustered at one time or large gaps within a phase; the exact times reveal those limitations.