Why the Sampling Design Matters
A population PK model can use observations collected at different times from different people. That flexibility makes sparse sampling practical, but it does not make every sparse design informative. If observations cluster at times when concentration changes little, or if a subgroup is represented by very few participants, some parameters can remain weakly estimated even when the total number of observations looks large.
The key question is not simply how many concentrations were measured. It is whether the data distinguish the parameters and covariate effects that matter for the development decision. Sampling time, number of observations per participant, allocation across time windows, and subgroup representation all contribute. Simulation can make these design consequences concrete before interpreting a fitted model.
From Sampling Times to Information
A useful local approximation connects parameter sensitivity to information. Let \(\boldsymbol{\theta}\) be the vector of model parameters, and let \(f_j(\boldsymbol{\theta})\) be the model-predicted response for observation \(j\). Near a specified parameter value, the sensitivity matrix \(J\) contains derivatives of those predictions with respect to the parameters. If observation errors are independent with known variance \(\sigma^2\), the approximate information matrix is \(J^\mathsf{T}J/\sigma^2\). The inverse of this matrix approximates the parameter covariance matrix when the model is locally identifiable and the approximation is suitable.
This is a local approximation, not a universal formula for the standard errors from a nonlinear mixed-effects model. Such models account for random effects and the likelihood of the observed data; their information depends on the full model and design. The approximation nevertheless explains a central design principle: if two parameters affect predictions in nearly the same way at the sampled times, their sensitivities are hard to separate and their estimates can be highly correlated. If the information matrix is singular, at least one parameter combination is not identified by that design under the stated model.
Worked Example: Why Repeated Samples at One Time Can Be Weak
Consider a hypothetical single-dose IV-bolus model with concentration \(C(t)=D/V\exp(-CLt/V)\). Assume dose \(D\) is known and independent errors on log concentration have standard deviation \(0.10\). At the illustrative values \(CL/V=0.20\) per hour, the log-concentration sensitivities to \(\log V\) and \(\log CL\) are \(-1+0.20t\) and \(-0.20t\), respectively.
Suppose three observations are collected at 1, 2, and 4 hours. The sensitivity rows are \((-0.8,-0.2)\), \((-0.6,-0.4)\), and \((-0.2,-0.8)\). Thus \(J^\mathsf{T}J\) is:
Because the error variance is \(0.10^2=0.01\), the approximate covariance matrix is \(0.01(J^\mathsf{T}J)^{-1}\). The approximate standard errors on the log scale are \(\sqrt{0.01(0.84/0.56)}=0.1225\) for \(\log V\) and \(\sqrt{0.01(1.04/0.56)}=0.1363\) for \(\log CL\).
Now place all three observations at 2 hours. Every sensitivity row is \((-0.6,-0.4)\). The information matrix has rank one: its determinant is zero. In this local approximation the data cannot distinguish changes in \(\log V\) from changes in \(\log CL\), however many repeated observations are taken at that same time. This example assumes the specified model, known dose, and independent log-scale error; it illustrates design geometry rather than providing a general standard-error calculation for population PK models.
Compare Designs by Simulation
When a planned design is sparse or complicated, a practical assessment is to simulate datasets under plausible population models and fit the intended analysis model to each dataset. The simulation must state its assumptions: for example, the structural model, parameter values, variability, residual error, covariate distribution, sampling schedule, and planned estimation method. Results are conditional on those choices. If important plausible scenarios are excluded, the exercise may give reassuring but misleading precision estimates.
Name the parameters or covariate effects whose precision matters, and define what would count as adequate precision for the decision.
Represent participants, doses, sampling times, and subgroup allocations as they are actually planned, including intended missing observations if relevant.
Generate data under each plausible scenario and apply the planned model and estimation procedure, rather than evaluating a more favorable model than the one intended for analysis.
Review bias, variability or relative standard error, interval coverage, convergence, and whether estimates are stable across scenarios. Examine target parameters individually; a good overall fit does not guarantee precise estimation of every parameter.
Worked Example: Comparing Two Sparse Schedules
Consider a hypothetical development study of 120 participants after a single oral dose. The assumed data-generating model is one-compartment PK with first-order absorption and elimination, interindividual variability in clearance and volume, and proportional residual error. The same model and estimation procedure are used to fit data generated under both candidate schedules.
In Schedule A, every participant is sampled at 1 and 6 hours. In Schedule B, participants are divided equally among three two-sample schedules: 0.5 and 2 hours, 1 and 6 hours, or 4 and 12 hours. Each design therefore has 240 planned concentrations. The simulation uses 500 hypothetical replicate datasets per design under the specified model and parameter values. Suppose the resulting illustrative summaries are:
| Summary | Schedule A | Schedule B |
|---|---|---|
| Median relative standard error for clearance | 24% | 15% |
| Median relative standard error for volume | 18% | 13% |
| Replicates meeting the convergence criterion | 96% | 99% |
| Coverage of nominal 95% clearance intervals | 88% | 94% |
For clearance, the relative reduction in median relative standard error is \((24\%-15\%)/24\%=0.375\), or 37.5%. The simulations suggest that spreading observations across early, intermediate, and later times may improve the separation of absorption and disposition features for this assumed model, despite an unchanged total sample count. The 88% interval coverage under Schedule A also signals that its nominal 95% intervals may not achieve their intended coverage in this scenario.
These numbers are hypothetical simulation outputs, not a general prediction that Schedule B is superior. The conclusion depends on the assumed model and parameter values, error structure, and implementation of the designs. A different plausible absorption rate, for example, could change which early sampling times are most useful. Repeat the comparison under alternative plausible scenarios before treating a design ranking as robust.
Unequal Sampling and Subgroup Imbalance
An overall sample count can conceal uneven information across time or subgroups. If most samples occur during one part of the concentration-time profile, the model may estimate some parameters precisely and others poorly. Similarly, a covariate effect is informed by contrast between covariate values, not by the total number of participants alone. A large study with few participants in one subgroup can have substantially weaker precision for a subgroup contrast than a smaller, balanced study.
The following simplified calculation shows the role of subgroup allocation. Suppose the log-clearance outcome follows a linear model with a binary covariate \(G\), where \(G=1\) denotes a subgroup, and independent normal residuals with known standard deviation \(\sigma\). In that model, the standard error for the estimated difference in mean log clearance is \(\sigma\sqrt{1/n_0+1/n_1}\), where \(n_0\) and \(n_1\) are subgroup sample sizes. A full population PK analysis is more complex; this calculation isolates the allocation effect under stated assumptions.
Worked Example: Precision Under Subgroup Imbalance
Suppose a hypothetical study has 100 participants and a residual standard deviation of \(0.20\) on the log-clearance scale. Compare a balanced allocation, 50 participants in each group, with an imbalanced allocation of 90 and 10. Assume the design otherwise supports the same outcome measurement and the simplified model above applies.
For the balanced allocation, the standard error of the subgroup contrast is \(0.20\sqrt{1/50+1/50}=0.0400\). For the imbalanced allocation, it is \(0.20\sqrt{1/90+1/10}=0.06667\). Thus imbalance increases the standard error by \(0.06667/0.0400=1.667\), or about 66.7%.
For illustration, suppose the estimated log-clearance contrast is \(0.20\) in either design. The approximate 95% interval is the estimate plus or minus \(1.96\) standard errors. With balanced allocation, this is \(0.20\pm1.96(0.0400)=(0.1216,0.2784)\). Exponentiating gives a clearance ratio of approximately \((1.129,1.321)\). With imbalanced allocation, it is \(0.20\pm1.96(0.06667)=(0.0693,0.3307)\), or a ratio of approximately \((1.072,1.392)\). These intervals describe uncertainty in a ratio under this simplified model; they do not establish that subgroup membership causes a clearance difference.
In an actual population PK design, subgroup imbalance can be more serious if it coincides with unequal sampling times, dose, or other covariates. Simulation can preserve those joint allocations and show whether a covariate effect remains recoverable under realistic assumptions.
Diagnostics, Alternatives, and Limitations
Look beyond whether model fitting completes. For each target, inspect its relative standard error, interval width and coverage across simulated replicates, and the frequency of extreme or implausible estimates. Strong parameter correlations, a poorly conditioned information matrix, or frequent estimation failures can signal that the design does not separate important effects. Review simulation-based predictive checks and goodness-of-fit diagnostics as well: a design can yield apparently precise estimates when the fitted model is systematically wrong. The earlier material on simulation-based diagnostics and predictive checks and on parameter uncertainty provides relevant context for these checks.
The main alternatives are complementary. Rich sampling can provide more information per participant about individual concentration-time profiles, but may be burdensome or infeasible. A planned sparse design can reduce participant burden while distributing observations across informative times and covariate groups. Local information calculations can compare designs efficiently around chosen parameter values, but they are sensitive to those values and rely on a local approximation. Simulation is more flexible for nonlinear population models and realistic variability, but it is only as useful as its scenarios and fitting procedure.
No design metric repairs a lack of relevant data. Post hoc adjustment cannot reliably separate a subgroup effect from sampling-time differences when the groups have little overlap in time or other important characteristics. Likewise, a low relative standard error under one assumed model does not establish robustness to model misspecification. Simulation is a way to expose design weaknesses and uncertainty, not a guarantee that the eventual estimate is unbiased or decision-worthy.
Check Your Understanding
Try each question, then compare your reasoning with the solutions below.
- In the worked IV-bolus example, why did three observations at 2 hours fail to distinguish \(\log V\) and \(\log CL\)?
- A design simulation reports 20% median relative standard error for a parameter under one design and 16% under another. What is the relative reduction?
- With a fixed total of 100 participants and \(\sigma=0.20\), which allocation gives greater precision for a binary subgroup contrast: 50/50 or 90/10? Explain.
- Name two reasons simulation results from one design scenario may not generalize to the actual study.
Solutions. (1) At 2 hours all sensitivity rows are the same, \((-0.6,-0.4)\), so the information matrix has rank one and cannot identify both parameter directions separately. (2) The reduction is \((20\%-16\%)/20\%=20\%\). (3) The 50/50 allocation is more precise: its standard error is \(0.20\sqrt{1/50+1/50}=0.0400\), compared with \(0.20\sqrt{1/90+1/10}=0.06667\) for 90/10. (4) Results depend on assumptions such as the structural model, parameter values, variability, residual error, and the simulated allocation; the actual study may differ in any of these respects.