Model-Based Bioequivalence With Sparse Sampling

Why Sparse Sampling Needs a Design Check

Sparse sampling can reduce participant burden, but fewer concentration measurements do not affect every inference in the same way. A schedule may estimate an exposure contrast adequately while leaving individual absorption-rate estimates imprecise. Conversely, a model can converge and produce a precise-looking contrast even when the sampling times provide little information about the curve’s shape. The key question is therefore not simply how many samples are collected: it is whether the proposed times, model, and study size support the parameters and exposure metrics needed for the bioequivalence estimand.

The earlier material on structural models and model-derived exposure establishes how concentration-time curves can be fitted and how quantities such as AUC are obtained from model predictions. Here the focus is on evaluating a sampling design before relying on those estimates. A design simulation repeatedly generates data under explicit assumptions, fits the intended model, and measures how well it recovers parameters and exposure contrasts.

What the Simulation Must Represent

For illustration, consider a one-compartment oral concentration model with first-order absorption and elimination:

Here \(D\) is dose, \(F\) is bioavailability, \(V\) is apparent volume, \(k_a\) is the absorption-rate constant, and \(k_e\) is the elimination-rate constant. A design assessment must specify the structural model, the assumed parameter distributions or fixed values, the residual-error model, the sampling times, and the estimator that will be used on the simulated data. For a formulation comparison, it must also specify how formulation effects are represented and which model-derived contrast is the target.

A simulation is informative only to the extent that its assumptions are credible. If the fitted model is the same as the model used to generate the data, the exercise evaluates information under correct specification; it does not show that the model will remain reliable when the real concentration curve differs from that assumed. A useful basic assessment states its assumptions explicitly. For example, it may assume the structural model above is correct, residual errors are independent and normally distributed on the log-concentration scale with standard deviation \(\sigma\), all planned samples are quantifiable, and the estimation algorithm converges to the intended solution. More elaborate assessments can include subject variability and realistic operational complications, but they should not hide those assumptions.

Sparse schedules are particularly sensitive to where samples fall on the curve. Measurements concentrated in one phase may provide little information about another. Early samples can help describe absorption and the peak region; later samples can help describe elimination. Whether a particular time is informative depends on the model, parameter values, residual error, and the metric of interest. A schedule should therefore be evaluated under a range of plausible parameter values, not just one convenient curve.

Why Sampling Time Affects Precision

A short information calculation makes the principle concrete. Suppose, temporarily, that the curve’s shape is known and only its multiplicative scale \(S=DF/V\) is unknown. On the log scale, an observation can be written as \(\log C(t_i)=\log S+g(t_i)+\varepsilon_i\), where \(g(t_i)\) is known and \(\varepsilon_i\) has variance \(\sigma^2\). The derivative with respect to \(\log S\) is 1 at every sampling time. Thus, with \(m\) independent observations, the information about \(\log S\) is \(m/\sigma^2\), and its standard error is \(\sigma/\sqrt{m}\). When the shape parameters are also unknown, the information is a matrix: derivatives of predicted concentrations with respect to different parameters can be similar at the chosen times, making their effects difficult to separate.

This calculation is conditional on a known curve shape; it is not a claim that every extra sample reduces all parameter standard errors by exactly the same factor. With unknown shape parameters, the fitted estimates depend on the joint information for scale, absorption, and elimination. If a schedule observes only a narrow time region, changes in \(k_a\) and \(k_e\) may produce similar predictions there. The estimates can then be highly correlated, and the uncertainty in a derived quantity may differ substantially from the uncertainty in any one parameter.

A practical justification for using design simulation follows in four steps:

Worked Design Simulation

The following hypothetical simulation compares a sparse schedule with a richer schedule. The goal is to assess both fitted rate parameters and a model-derived log AUC ratio. For each of 1,000 simulated crossover studies, 24 subjects receive both formulations. The analysis uses the same structural model for both formulations, includes shared subject-level variation, and estimates a within-subject formulation contrast. Residual errors are generated independently on the log-concentration scale with standard deviation 0.10. The sparse schedule samples at 1, 4, and 12 hours; the richer schedule samples at 0.5, 1, 2, 4, 8, and 12 hours. These are illustrative assumptions, not observations from a real study. Period and sequence effects are not part of this design comparison.

For each simulation \(b\), the log AUC-ratio error is \(\widehat{\theta}_b-\theta\), where \(\theta\) is the true simulated log ratio. Across \(B\) simulations, bias is the mean error, the standard deviation describes variability of the estimates, and root mean squared error (RMSE) combines both:

From Metric Uncertainty to a Bioequivalence Contrast

For an AUC comparison, a common target is the log geometric mean ratio, denoted here by \(\theta\); its exponentiation gives the geometric mean ratio. If two log exposure estimates are independent, the variance of their difference is the sum of their variances. In a crossover analysis, estimates may be correlated within subject, so the variance of the contrast must reflect that covariance; treating paired estimates as independent can give the wrong standard error. The following calculation deliberately uses independent estimates to show the variance addition and interval transformation.

Diagnostics, Alternatives, and Limits

A simulation should be examined for signs that sparse data are not supporting the intended fit. Relevant diagnostics include the proportion of fits that fail or reach parameter boundaries, large correlations among parameter estimates, unstable standard errors, and systematic prediction errors concentrated at particular sampling times. Plotting simulated predictions and residuals by time can reveal that a schedule fits the observed sampling points while missing the curve between them. A high rate of failed fits is itself a design result, not a reason to silently discard those replicates.

Also examine whether conclusions change over plausible values of \(k_a\), \(k_e\), between-subject variability, and residual error. An apparently precise design under one parameter set may be weak when absorption is faster or slower, or when residual variability is higher. A useful report distinguishes Monte Carlo uncertainty in the simulation summaries from variability across simulated studies. Increasing the number of simulation replicates reduces Monte Carlo uncertainty in the summaries; it does not make an individual study more informative.

The main alternatives address different limitations. Richer sampling can add information about curve shape, but increases collection burden. Increasing the number of subjects can improve precision for population contrasts, but may not resolve a structural ambiguity caused by uninformative sampling times. External information or stronger population-level sharing can stabilize estimation, but conclusions then depend more on those modeling assumptions. A simulation-based assessment compares these trade-offs for the intended estimand rather than assuming that one remedy works in every setting.

Finally, simulation is conditional on the data-generating scenarios and analysis being considered. A correctly specified simulation can show that a schedule is informative under its assumptions; it cannot prove that the assumed model is correct for future data. Sparse schedules may be especially vulnerable to misspecification because there are fewer observations with which to detect an incorrect curve shape. Design assessment should therefore be interpreted as evidence about expected performance under stated scenarios, not as a guarantee of reliable estimation in every study.

Common Pitfalls

• Counting samples without considering their times. The same number of samples can provide very different information if one schedule captures both early and late phases while another clusters measurements in one region.
• Assuming precise metrics require precise parameters. A particular combination of parameters may yield a stable AUC even when individual rate estimates are uncertain. Assess both parameters and the target metric.
• Ignoring covariance when forming a contrast. The variance of a paired log ratio depends on the covariance between formulation estimates; adding two marginal variances is appropriate only when the estimates are independent.
• Treating convergence as proof of identifiability. An optimizer can return estimates even when the data weakly distinguish parameters. Check uncertainty, parameter correlations, boundaries, and sensitivity to starting values.
• Reporting only average error. Small bias can coexist with large variability. Report a spread measure and a combined measure such as RMSE, and inspect failure rates.