Worked Comparison of Population PK Sampling Plans

Why Population Sampling Plans Need a Population-Level Comparison

An individual PK schedule is often judged by how well it estimates parameters for one person. A population sampling plan has a broader task: data collected across many people must support estimates of both typical PK parameters and differences among people. These are not interchangeable goals. A plan may estimate typical clearance accurately while providing weak information about the amount of interindividual variability, or it may estimate variability adequately while leaving the typical clearance imprecise.

A simulation comparison makes this distinction explicit. It generates many hypothetical datasets under a stated population model, fits each dataset using the same analysis model, and summarizes how the competing plans recover parameters of interest. The comparison is meaningful only when the model, simulated population, fitting procedure, and performance measures are specified clearly.

Model and Assumptions for the Worked Comparison

Consider a hypothetical single-dose, one-compartment IV bolus study. The dose is \(D\), the volume of distribution \(V\) is known and fixed at 50 L, and individual clearance varies in the population. For individual \(i\), let \(CL_i=\theta_{CL}\exp(\eta_i)\), where \(\eta_i\sim N(0,\omega^2)\). Thus, \(\theta_{CL}\) is the population median (typical) clearance, while \(\omega\) is the standard deviation of log clearance. For observation \(j\) on individual \(i\), the model is:

The assumptions for this illustration are that individuals are independent; the \(\eta_i\) values are independent draws from the stated normal distribution; residual errors are independent of each other and of the \(\eta_i\); doses and sampling times are recorded correctly; and the model used for fitting matches the model used to generate data. The log-scale residual model implies multiplicative error on the concentration scale. We also assume no informative dropout, no covariates, and no model misspecification. These simplifying assumptions make the design comparison interpretable; they do not guarantee that the model is appropriate for a real study.

For a lognormal parameterization, the coefficient of variation of individual clearance is \(CV_{CL}=\sqrt{\exp(\omega^2)-1}\). For example, if the true \(\omega\) is 0.29, then:

So the corresponding clearance CV is about 29.62%. This is a population spread on the original clearance scale; it is not the same quantity as \(\omega\), which is measured on the log scale.

What to Measure Across Replicates

For each simulated dataset, fit the population model and retain estimates of \(\theta_{CL}\), \(\omega\), and any other prespecified parameters. Across \(R\) successful replicates, the bias of an estimate \(\hat{\psi}\) for a true value \(\psi\) is the average estimate minus that truth. RMSE combines bias and replicate-to-replicate variation:

Relative bias, \(100\times\operatorname{Bias}(\hat{\psi})/\psi\), can help compare parameters on different scales when \(\psi\ne0\). It should not replace absolute bias or RMSE, and it is unsuitable when the true value is zero. For variability parameters, report the scale being evaluated: bias in \(\omega\) is not bias in the clearance CV. If CV is scientifically important, transform each replicate’s \(\hat{\omega}_r\) to \(\sqrt{\exp(\hat{\omega}_r^2)-1}\), then summarize those transformed estimates against the true CV.

Fixed-effect recovery and variability recovery answer different questions. In this example, \(\theta_{CL}\) describes typical clearance; \(\omega\) describes dispersion across individuals. A complete comparison reports both, along with uncertainty summaries such as the replicate distribution or interval coverage when relevant. Coverage is the fraction of simulated intervals that contain the fixed true parameter under the simulation conditions; it is not the probability that a particular interval contains the truth.

Why Within-Person Information Helps Separate Variability

The following simplified derivation explains why sparse sampling can make between-person variability difficult to estimate. Suppose, only as an approximate diagnostic representation, that each person’s log-clearance information can be expressed as \(Y_i=\mu+\eta_i+u_i\). Here \(\mu\) is the population mean on the log scale, \(\eta_i\) has variance \(\omega^2\), and \(u_i\) represents estimation noise with variance \(\tau^2\). Assume these terms are independent, have finite variances, and the same \(\tau^2\) applies across people. Then:

The sample variance across people therefore reflects both true population heterogeneity and noise in each person’s information. More observations at informative times can reduce the uncertainty in each person’s clearance information, reducing the contribution analogous to \(\tau^2\). If that contribution is large, observed person-to-person variation can be hard to attribute to \(\omega^2\) rather than measurement and estimation noise. This decomposition is an explanatory approximation, not a substitute for fitting the nonlinear population model: in practice, the likelihood uses the full concentration model and the planned observation times.

The fixed typical clearance can also be poorly recovered when the sampling schedule gives little leverage on the concentration-time slope. Thus, a plan’s number of samples alone does not determine its performance. Their timing, the number of participants, residual error, and the distribution of individual clearances all matter.

Checking Estimation and Comparing Plans Fairly

A plan comparison should use the same generating assumptions, fitting model, parameter definitions, and replicate count for each plan. When possible, shared simulated population draws can make paired comparisons less noisy; the earlier material on common random numbers and paired design differences describes that approach. The sampling plans must still be evaluated using their own observation times and resulting datasets.

Convergence is part of the evidence, not a housekeeping detail. Report the number and proportion of attempted fits that converged, and inspect whether failures cluster in particular parameter regions or datasets. When a fit fails, parameter estimates may be unavailable or unreliable. Performance summaries calculated only among successful fits are therefore conditional on success; they do not describe all attempted replicates. Report failure rates alongside conditional bias, RMSE, and interval performance, and avoid treating a failed estimate as though it were an ordinary parameter value.

• Check estimates at boundaries. Estimates of \(\omega\) close to zero can indicate weak evidence for heterogeneity, limited information, or estimation difficulty. Review the simulated population and model fit rather than interpreting every boundary estimate as proof of no variability.
• Inspect replicate distributions. A mean bias can conceal skewness, occasional extreme estimates, or a cluster of estimates near a boundary. The earlier material on performance distributions across replicates is useful for reading these features.
• Check model diagnostics. Systematic residual patterns over time, implausible individual estimates, or strong dependence between parameter estimates can reveal poor information or model-data conflict. Convergence alone does not establish a reliable fit.
• Separate parameter goals. A plan that performs well for \(\theta_{CL}\) need not perform well for \(\omega\). State which parameters matter and compare their recovery explicitly.

The main alternatives are not different labels for the same performance measure. Comparing only standard errors emphasizes precision under the fitted model but can miss bias. Comparing only bias misses replicate variability. Comparing only overall likelihood or fit statistics does not directly establish recovery of the population parameters. A useful design assessment therefore reports several complementary measures, and relates them to the study’s prespecified goals. The earlier material on uncertainty in design rankings and Pareto fronts can help when no plan is best on every measure; here the essential task is to provide sound population-level performance summaries for the competing plans.

Limitations and Common Pitfalls

A simulation comparison is conditional on its generating scenarios. If the simulated variability, residual error, participant count, or model structure is unrealistic, a plan may appear favorable for the wrong reason. Evaluate plausible ranges of assumptions rather than relying on one convenient scenario, and interpret the result as design evidence under those scenarios—not as a guarantee of performance in future participants.

A frequent mistake is to conclude that a plan with a lower standard error for typical clearance is automatically the better population plan. That reasoning overlooks \(\omega\), potential bias, and estimation failures. The correction is to define the target parameters first, report fixed-effect and variability recovery separately, and include convergence and uncertainty in the comparison. Another mistake is to read \(\hat{\omega}\) as a percentage coefficient of variation. It is a log-scale standard deviation; convert it using \(\sqrt{\exp(\hat{\omega}^2)-1}\) when a clearance-scale CV is wanted.

Finally, a simulated advantage does not by itself establish clinical utility, and a design with more observations may have a different operational cost. A population plan should be judged against its scientific objectives and practical constraints. This tutorial’s comparison is about simulated recovery of population PK parameters, not adaptive allocation or the clinical consequences of dosing decisions.