Sparse-Sampling BE With Population PK

Why Sparse Sampling Requires a Population Model

In a sparse-sampling study, each participant contributes only a small number of concentration measurements per treatment period. Those measurements may be insufficient to estimate that participant’s pharmacokinetic parameters reliably on their own. A population pharmacokinetic (PK) model addresses this by estimating shared population features from all participants while allowing individual parameters to vary. It can then estimate individual exposure metrics using both the participant’s observations and information from the population.

That borrowing of information is useful, but it has an important consequence: an individual estimate is not a direct measurement. It can be pulled toward the population average, especially when the person’s data are sparse or noisy. A bioequivalence (BE) analysis must therefore account for uncertainty in individual exposure estimates rather than treating each estimate as known without error. The earlier material on model-based BE using individual AUC estimates provides the context; here the focus is how sparse data and population PK affect those estimates and the resulting inference.

For example, a model may describe the typical apparent clearance and volume of distribution, while representing each participant’s clearance as a typical value multiplied by an individual deviation. Observed concentrations are then related to the participant’s PK parameters through a structural concentration-time function, with residual error accounting for assay variation and other unexplained differences. This analysis relies on assumptions: the structural model is adequate over the observed dose and time range; the variability distributions are appropriate; the residual error model is reasonable; the sampling schedule is informative about the parameters of interest; and missing or below-quantification observations are handled appropriately.

From Sparse Concentrations to Exposure

A simple one-compartment oral model illustrates how a structural model connects concentrations to AUC. For a single dose with first-order absorption and elimination, one expression for concentration is:

Here \(D\) is the dose, \(k_a\) is the absorption rate constant, \(CL/F\) is apparent clearance, and \(V/F\) is apparent volume. Under this model, assuming complete integration to infinity and linear PK, \(AUC_{0-\infty}=D/(CL/F)\). For oral data, \(F\), \(CL\), and \(V\) are not generally identified separately; apparent parameters are used. For comparing formulations at the same dose, the model can estimate relative exposure even though absolute bioavailability is not separately identified.

Why Individual Estimates Shrink

A concise normal model shows the source of shrinkage. Let \(y_i\) be a noisy estimate of participant \(i\)’s log AUC, and suppose \(y_i\mid\theta_i\sim N(\theta_i,s_i^2)\), where \(\theta_i\) is the participant’s true log AUC and \(s_i^2\) is the observation variance. Suppose the population distribution is \(\theta_i\sim N(\mu,\tau^2)\), where \(\mu\) and \(\tau^2\) are the population mean and between-subject variance. If these population quantities are treated as known, the conditional distribution of \(\theta_i\) given \(y_i\) is normal.

The weight \(\tau^2/(\tau^2+s_i^2)\) is between 0 and 1. When an individual estimate is noisy, \(s_i^2\) is large and the conditional estimate moves toward \(\mu\). When the individual data are more informative, \(s_i^2\) is smaller and the estimate stays closer to \(y_i\). This conditional estimate is often described as an empirical Bayes estimate when the population quantities are estimated from the study data. Plugging in estimated population parameters adds uncertainty beyond the conditional variance shown above.

This derivation also explains a diagnostic warning: a set of individual estimates that looks unusually close to the population mean may reflect extensive shrinkage rather than genuine homogeneity. Shrinkage is not automatically a model failure; it is a consequence of limited individual information. But substantial shrinkage means individual estimates should not be interpreted as precise subject-specific measurements, and it can make analyses that treat them as error-free misleading.

Carry Uncertainty Into the BE Contrast

For a two-formulation comparison, the target is typically the population geometric mean exposure ratio, test relative to reference, for a prespecified exposure endpoint such as AUC. In a crossover design, the model should preserve the within-participant pairing and account for relevant design effects, such as treatment and period. The ratio is most naturally estimated on the log scale: the treatment contrast is the difference in mean log exposure, and exponentiating that contrast gives the geometric mean ratio.

The uncertainty calculation must reflect the fact that the exposure metrics came from a fitted population model. One suitable strategy is joint estimation of the PK model and the treatment contrast. Another is a participant-level bootstrap that resamples participants and refits the full model at each replicate. Either approach should carry forward uncertainty in population PK parameters, individual exposure estimates, and the treatment contrast. Simply calculating a standard error from the shrunken point estimates as if they were observed without error generally understates uncertainty.

A frequentist 90% confidence interval does not mean there is a 90% probability that the fixed ratio lies in this particular observed interval. Its 90% coverage refers to the long-run performance of the procedure under its assumptions. By contrast, a 90% Bayesian posterior interval has 90% posterior probability conditional on the specified model, prior, and data that the parameter lies in the interval. These interpretations should not be interchanged.

Assumptions, Checks, and Alternative Approaches

A sparse-sampling BE conclusion is only as credible as the model and design that support it. Before interpreting the interval, examine whether the sampling schedule is informative for the exposure metric and whether the model reproduces the observed concentration data. Useful checks include observed-versus-predicted concentration plots, residual patterns over time and concentration, prediction-corrected visual predictive checks, parameter plausibility, and stability of the treatment contrast under reasonable model variations. A model can fit the central concentration pattern yet still give unstable AUC estimates if terminal elimination or absorption is poorly informed.

Investigate practical failure modes directly. Sparse samples clustered at similar times may not distinguish absorption from elimination. Strong parameter correlations, poor convergence, boundary estimates, or wide profile likelihoods can signal weak identifiability. A few influential concentrations, incorrect handling of below-quantification observations, or unmodeled formulation-specific absorption can distort estimates. Extensive shrinkage is a reason to be cautious about individual estimates, while instability of the population treatment contrast is a reason to question the BE conclusion itself. Design simulations before the study can reveal whether the planned sampling times and participant count are likely to estimate the contrast with useful precision.

Dense-sampling noncompartmental analysis estimates exposure from each participant’s observed concentration-time profile with fewer structural assumptions, but it requires sufficiently informative sampling. A two-stage model can first estimate individual parameters and then compare derived exposures; it is vulnerable to noisy individual estimates and to failing to carry their uncertainty into the second stage. A population PK approach can make use of sparse data and borrow strength, but its conclusions depend more directly on structural and distributional assumptions. The choice is not simply about which method produces a narrower interval: the design, model adequacy, and estimand must support the method used.