Where the Measurements Go Matters
A population pharmacokinetic study does not distribute its concentration measurements in only one way. It may collect many measurements from each participant, a few measurements from many participants, or combine the two approaches. These choices change the balance between learning about concentration patterns within individuals and representing differences across individuals.
The total number of concentration records alone does not describe that balance. For example, 120 records could come from 15 participants measured eight times each, 60 participants measured twice each, or a mixture. Those plans have the same record count but differ in how many people they represent and how much repeated information they contain for each person.
These are descriptions of the distribution of measurements, not guarantees about the quality of a study. A design is informative only in relation to its scientific aims, the population model, the variability in the data, and whether the measurements are usable. Rich sampling is not automatically best, and increasing the number of participants is not automatically enough.
Two Sources of Information
In a population pharmacokinetic model, measured concentrations are related to time, dosing, model parameters, and unexplained variability. A general schematic representation is
Here \(C_{ij}\) is the concentration measured for participant \(i\) at observation \(j\), \(t_{ij}\) is its measurement time, \(\theta_i\) denotes that participant's model parameters, and \(\theta\) denotes population-typical parameters. The term \(\eta_i\) represents participant-level deviations from typical values; \(\epsilon_{ij}\) represents residual variation not explained by the model. This is a conceptual framework: the precise forms of \(f\), \(g\), and the error terms depend on the model.
Repeated observations within a participant can help describe that participant's concentration profile and can help distinguish a pattern that recurs within that person from variation among measurements. Measurements from more participants can better represent population diversity and provide more individuals from whom to estimate the distribution of participant-level parameters. These contributions are related, but they are not interchangeable. Many measurements on a small number of participants do not create the same population coverage as measurements on many different participants.
The value of either source depends on the question. If the study needs detailed profiles for model development, within-participant density can matter. If the study needs to characterize variability across a target population, the number and range of participants matter. A good design identifies the intended balance rather than treating all records as equivalent units of information.
Worked Example: Equal Records, Different Allocations
Suppose three hypothetical study plans each collect 120 concentration measurements. Compare their allocation across participants.
Rich plan: 15 participants are measured 8 times each, so \(15\times 8=120\) records. There are 8 observations per participant.
Sparse plan: 60 participants are measured 2 times each, so \(60\times 2=120\) records. There are 2 observations per participant.
Hybrid plan: 30 participants are measured 4 times each, so \(30\times 4=120\) records. There are 4 observations per participant.
The arithmetic confirms that each plan has 120 records, but the rich plan has the fewest participants and the most repeated observations per person; the sparse plan has the most participants and the fewest repeated observations per person. The hybrid plan lies between them on both counts. The plans therefore offer different balances of within-participant detail and participant coverage. The record count by itself cannot establish which plan best serves a particular analysis.
A Simple Variance Decomposition
A small derivation shows why repeated measurements do not always compensate for having few participants. Consider a deliberately simplified setting in which each measurement targets the same population quantity. Let
where \(\mu\) is the population mean, \(b_i\) is a participant-specific deviation with mean zero and variance \(\tau^2\), and \(\epsilon_{ij}\) is a residual deviation with mean zero and variance \(\sigma^2\). Assume participant deviations are independent across participants, residuals are independent across observations and participants, and the two sources are independent. Suppose each of \(N\) participants has \(M\) measurements. The overall average is
Independence makes the variance of a sum the sum of its variances. The participant-deviation term averages \(N\) independent values, while the residual term averages \(NM\) independent values. Thus
The first component decreases as the number of participants increases. The second decreases as the total number of measurements increases. This illustrates a central trade-off: adding repeated observations can reduce residual contribution, but it does not add independent participants to reduce uncertainty arising from participant-to-participant differences.
This derivation is not a formula for the precision of every population PK parameter. Real concentration profiles vary with time, and observations from the same participant are generally related through the participant's parameters. The simplified model isolates the allocation principle under explicit assumptions; PK design assessment must account for the actual model and schedule.
Worked Example: Precision Under a Simplified Model
Consider three hypothetical balanced designs with 96 measurements each: 12 participants measured 8 times, 24 measured 4 times, or 48 measured 2 times. For illustration, suppose the shared participant-level standard deviation is \(\tau=0.30\) and the residual standard deviation is \(\sigma=0.20\), on a log-concentration scale. The assumptions of the preceding random-intercept model apply. In particular, this calculation treats every observation as targeting the same quantity; it is not a PK profile analysis.
For 12 participants with 8 measurements each, \(N=12\) and \(NM=96\):
The standard error is \(\sqrt{0.007917}=0.0890\) log units. For 24 participants with 4 measurements each:
For 48 participants with 2 measurements each:
Under these assumptions, the design with 48 participants has the smallest standard error for the population mean because the participant-level component falls as \(N\) increases, while the total number of measurements—and therefore the residual component—stays fixed. This result does not show that the sparsest plan is universally superior: it addresses only this simplified mean-estimation problem, not the information available for describing individual profiles or estimating a full PK model.
What a Hybrid Design Can Contribute
A hybrid plan assigns different measurement burdens to different participants. A richer subgroup can provide detailed repeated observations, while a larger sparse group can extend coverage to more individuals. This can be attractive when intensive participation is burdensome or costly, or when the study has distinct goals requiring both profile detail and broader representation.
Worked Example: Counting a Hybrid Plan
A hypothetical study enrolls 24 participants. Four participants each provide 6 concentrations, and the other 20 each provide 2. The total is
measurements. The rich subgroup is \(4/24=16.7\%\) of participants but contributes \(24/64=37.5\%\) of measurements. The sparse subgroup contributes the remaining 40 measurements from 20 participants. Relative to a plan with all 24 participants measured twice, the 16 extra measurements are concentrated in four people. Relative to a plan that spent all 64 records on 8 participants measured 8 times each, the hybrid plan covers 16 more people.
The four richer profiles can support examination of repeated patterns within participants, while the 20 additional participants contribute broader population coverage. Whether this allocation is useful depends on whether the richer subgroup and the sparse group are suitable for the intended joint analysis—for example, whether selection into the richer schedule creates an important difference between the groups. The count alone does not answer that question.
Comparing the Main Approaches
| Approach | Relative strength | Important limitation |
|---|---|---|
| Rich sampling | More repeated observations per participant; supports detailed within-person profiles. | For a fixed measurement budget, often covers fewer participants and may represent population diversity less broadly. |
| Sparse sampling | Can distribute measurements across more participants and broaden population coverage. | Few observations per person provide less within-person detail; a population pattern must be learned across participants. |
| Hybrid sampling | Combines a rich subset with a broader sparse group. | Requires attention to how the subgroups are selected and whether their data can inform the same population model. |
These approaches should be compared against the same objectives, not ranked in isolation. A fixed budget may force a choice between more people and more measurements per person, but the scientifically relevant comparison is the information each candidate allocation provides for the parameters and population features of interest. A rich design may be valuable for profile detail even when it has fewer participants; a sparse design may be valuable for population coverage even when individual profiles are incomplete; a hybrid may be a practical compromise.
Diagnostics, Failure Modes, and Limits
A design can appear balanced on paper yet provide weak information in practice. Review the allocation by participant, not just the grand total. Check whether the planned rich subgroup is large enough to serve its purpose, whether the broader group actually adds participant coverage, and whether missing or unusable samples could leave a few individuals carrying most of the repeated information. Also consider whether participants assigned to different schedules differ systematically in ways relevant to the analysis.
A key limitation is that the simple variance calculation assumes every measurement targets the same quantity and has the same independent residual variance. In a PK study, concentrations change over time, observations within a participant share participant-level features, and parameters may be informed differently by different parts of a profile. Therefore, the formula demonstrates the distinction between participant-level and residual variation; it does not replace evaluation under the intended PK model.
- Counting records as independent participants. Repeated observations from one person do not supply the same population coverage as observations from additional people.
- Assuming sparse data automatically represent the population. A larger participant count helps only if the enrolled participants and available data support the intended population description.
- Assuming a rich subset represents everyone. If participants receiving rich schedules are selected differently, their profiles may not alone characterize the full study population.
- Choosing a hybrid only by its total count. The allocation between rich and sparse participants, and the analysis connecting them, matters as much as the number of records.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Two hypothetical plans each collect 80 concentrations. Plan A measures 10 participants 8 times each; Plan B measures 40 participants twice each. What differs between them?
- In the simplified variance model, what happens to the participant-level term when the number of measurements per participant rises but the number of participants stays fixed?
- A hybrid plan enrolls 18 participants: 3 provide 5 measurements each and 15 provide 2 each. How many measurements are collected?
- Why does the variance formula in this tutorial not establish which design is best for estimating every PK parameter?
Solutions. (1) Both have \(80\) records, since \(10\times8=80\) and \(40\times2=80\). Plan A has more repeated observations per person; Plan B covers more participants. (2) In \(\operatorname{Var}(\bar{Y})=\tau^2/N+\sigma^2/(NM)\), the participant-level term \(\tau^2/N\) is unchanged if \(N\) stays fixed; increasing \(M\) reduces only the residual term in this simplified model. (3) \(3\times5+15\times2=15+30=45\) measurements. (4) The derivation assumes a simplified common-quantity random-intercept model with independent residuals; actual PK profiles vary with time, and different schedules can contribute differently to different model parameters.