Why Sampling Time Is Part of the Test
A concentration-time profile can look acceptable at one time and still be wrong in a way that matters. An absorption error may be most visible before or near the peak, whereas a tissue-distribution error may emerge during the redistribution phase. A clearance error may become clearer in the terminal decline. If measurements are collected only where competing predictions happen to be similar, a study can produce precise concentrations without meaningfully testing the model.
Sampling time should therefore be chosen in relation to the prediction being checked. The model and parameter set are frozen for this purpose: candidate error scenarios are compared with the prespecified prediction, rather than being used to adjust the model after observing the verification data. This is a design question, not a guarantee that any selected time will identify the true biological cause of a mismatch.
The central design distinction is between coverage and discrimination. Coverage means measuring across important parts of the profile, such as early exposure, the peak region, distribution, and the terminal phase. Discrimination means choosing times at which a particular plausible error changes the prediction. A useful design generally needs both. A single time chosen because it maximizes one contrast can miss other errors or leave the profile’s shape ambiguous.
A Quantitative Rule for Comparing Candidate Times
Suppose the concentration observation at time \(t_i\) is \(Y_i\), and the frozen prediction is \(C_0(t_i)\). Under a specified alternative scenario, the prediction is \(C_1(t_i)\). If assay and observation error are approximately additive, unbiased, and have known standard deviation \(\sigma_i\) at that concentration, a simple measure of separation is the standardized contrast:
This quantity is measured in standard deviations of observation error. A larger \(d_i\) indicates that the two predictions are less likely to be confused by measurement noise alone under the assumed error model. It is not a probability of detecting an error, and it does not account automatically for parameter uncertainty, between-participant variability, or other model discrepancies.
When error is proportional rather than additive, concentrations are often more appropriately compared on the log scale. If the standard deviation of log concentration is \(\tau_i\), the corresponding contrast is:
The error scale should reflect the planned assay and analysis. It should not be chosen after seeing which scale makes the predictions look most separated. In practice, the uncertainty can vary over time: late concentrations may be near the quantification limit, while early concentrations may be affected by timing uncertainty or a rapid concentration change.
For a set of sampling times, contrasts should be considered together. If the vector of prediction differences is \(\Delta\) and the observation-error covariance matrix is \(\Sigma\), a useful measure of total separation is:
This expression weights differences by their uncertainty and accounts for correlation when \(\Sigma\) is specified. Under a multivariate normal observation model with the same covariance under both predictions, the expected log-likelihood advantage of the correct candidate over the other is \(D^2/2\). To see why, write the log likelihood under a prediction vector \(\mu\) as a constant minus \(\frac{1}{2}(Y-\mu)^\mathsf{T}\Sigma^{-1}(Y-\mu)\). If data are generated with mean \(\mu_0\), the expected difference between the log likelihood at \(\mu_0\) and at \(\mu_1\) is \(\frac{1}{2}(\mu_0-\mu_1)^\mathsf{T}\Sigma^{-1}(\mu_0-\mu_1)\), because the mean-zero error terms cancel. This justifies considering the whole sampling pattern rather than selecting times independently by eye.
The calculation is only as credible as its inputs. It assumes that candidate predictions are meaningful, the error scale is reasonably specified, and the proposed times can be collected accurately. It does not establish that the alternatives exhaust the ways the model could be wrong.
Match Sampling Times to the Error Mechanism
For an oral dose, early samples and samples around the predicted peak can reveal errors in absorption rate, lag time, or input timing. A sample well before the peak can distinguish a delayed rise from a faster rise; a sample near the peak can help reveal a shifted peak or an incorrect peak concentration. A single peak-region observation cannot establish the time of the peak, so neighboring samples are valuable when the peak time itself is in question.
Distribution errors are often most visible after the initial rise, when movement between blood and tissues affects the profile’s curvature. Depending on the compound and model structure, the relevant interval might be shortly after the peak or later during redistribution. A central concentration measurement alone does not directly reveal tissue concentrations; it tests the plasma or blood prediction resulting from the assumed distribution processes.
Terminal samples help assess the late decline and can be informative about clearance, but a terminal concentration is not a direct measurement of clearance. The terminal shape may also depend on distribution, the size of the amount remaining in the body, and other processes. Multiple late samples are more useful for distinguishing a changed terminal slope from a vertical shift in the curve than one late sample alone.
Worked Example: Exposing a Slower Absorption Scenario
Consider a hypothetical oral-dose design with predictions in mg/L from a frozen PBPK model and a slower-absorption alternative. Assume, for this illustration, additive observation error with standard deviation \(0.20\) mg/L at each listed time.
| Time (h) | Frozen prediction (mg/L) | Slower-absorption prediction (mg/L) | Absolute contrast (mg/L) |
|---|---|---|---|
| 0.25 | 0.80 | 0.30 | 0.50 |
| 0.50 | 2.00 | 1.10 | 0.90 |
| 1.00 | 3.20 | 2.60 | 0.60 |
| 2.00 | 2.50 | 2.60 | 0.10 |
| 4.00 | 1.40 | 1.50 | 0.10 |
At 0.50 h, the standardized contrast is \(|1.10-2.00|/0.20=0.90/0.20=4.50\). At 1.00 h it is \(|2.60-3.20|/0.20=0.60/0.20=3.00\), and at 2.00 h it is \(0.10/0.20=0.50\). Thus 0.50 h is the most discriminating listed time for this particular pair of predictions under the stated error assumption. The model’s predicted peak is near 1 h, so a 1 h sample still provides useful peak-region coverage; it should not be discarded just because 0.50 h has the largest single-time contrast.
This design targets absorption timing, not every aspect of the profile. If the purpose also includes estimating peak location, samples on both sides of the expected peak are more informative than a single observation at the apparent maximum. The table’s values are hypothetical and do not imply that all slower-absorption alternatives have this pattern.
Worked Example: Choosing Between Two Terminal-Phase Times
For a hypothetical terminal-phase comparison, a frozen prediction is \(1.20\) mg/L at 8 h and \(0.12\) mg/L at 24 h. A specified slower-terminal alternative in the candidate set predicts \(1.60\) mg/L at 8 h and \(0.35\) mg/L at 24 h. Assume an additive observation-error standard deviation of \(0.20\) mg/L at both times.
At 8 h, the absolute contrast is \(|1.60-1.20|=0.40\) mg/L, giving \(0.40/0.20=2.00\) standard deviations. At 24 h, it is \(|0.35-0.12|=0.23\) mg/L, giving \(0.23/0.20=1.15\) standard deviations. Therefore, 8 h has the larger separation from the frozen prediction among these two candidate times. A 24 h sample may still be useful for late-time coverage, but it is less separated for this comparison.
The candidate’s label describes the full simulated alternative, not a mechanism that can be proven from either concentration alone. These two contrasts compare predicted concentrations; they do not by themselves identify clearance or establish a terminal slope. To assess a slope difference, plan multiple appropriately placed late samples and check that the predicted concentrations remain measurable with adequate precision.
Worked Example: Targeting a Distribution-Phase Difference
Suppose two hypothetical predictions differ because of a tissue-distribution assumption. Their plasma concentrations and an assumed additive observation-error standard deviation of \(0.15\) mg/L are:
| Time (h) | Frozen prediction (mg/L) | Distribution alternative (mg/L) | Absolute contrast (mg/L) |
|---|---|---|---|
| 0.25 | 3.00 | 2.70 | 0.30 |
| 1 | 2.80 | 2.00 | 0.80 |
| 4 | 1.50 | 1.00 | 0.50 |
| 8 | 0.80 | 0.65 | 0.15 |
| 24 | 0.20 | 0.18 | 0.02 |
At 1 h, the standardized contrast is \(0.80/0.15=5.33\). At 4 h it is \(0.50/0.15=3.33\), while at 24 h it is \(0.02/0.15=0.133\). The 1 h sample is the strongest single-time discriminator in this table, and the 4 h sample adds information about how the curves diverge through the distribution phase.
A large contrast is evidence that the predictions differ, not proof that tissue distribution is the cause if an observed value disagrees with the frozen model. Absorption, clearance, or another structural assumption could produce a similar plasma difference. Including times from more than one phase helps reveal whether the pattern is consistent with the prespecified distribution scenario.
Designing a Useful Sampling Pattern
A practical design begins with a small set of scientifically plausible error scenarios, such as slower absorption, altered tissue partitioning, or changed clearance. Simulate each scenario under the intended dose and sampling conditions, then compare its predictions with the frozen model on a justified error scale. Candidate times should be feasible and chosen before data are observed. A concise selection process is:
State the prediction and the particular process or assumption that the alternative changes. Do not treat a generic “wrong model” as a single scenario.
Simulate the frozen and alternative profiles on a sufficiently fine time grid, then identify early, peak-region, distribution, and terminal intervals where their predictions diverge.
Use additive, proportional, or other justified observation-error assumptions. Consider assay performance, timing precision, and correlations rather than relying only on raw concentration differences.
Select times that separate key alternatives while retaining samples across relevant phases. Check feasibility, participant burden, sample handling, and quantification limits.
Confirm that predictions at the proposed times are not artifacts of solver error, dose-event handling, or interpolation, and that the protocol can collect samples at those times accurately.
The last check matters because an apparent prediction contrast smaller than numerical error is not a reliable design signal. Earlier material on solver convergence and extracting predictions at observation times provides the relevant computational checks. Sampling should be scheduled with respect to actual dose and infusion events, not merely convenient clock times.
Alternatives, Diagnostics, and Limitations
Evenly spaced sampling is simple and can provide broad coverage, but it may spend many samples where candidate predictions are nearly identical. Log-spaced schedules provide denser coverage at early times and can be useful for profiles spanning a wide duration, but they still need to be checked against the specific error scenarios. Choosing only the time with the largest standardized contrast is more targeted, but it risks poor phase coverage and can be fragile if the assumed error scale or candidate alternative is wrong. A compromise is to include a small number of high-contrast times for each important error class plus samples that anchor the overall profile.
Several warning signs indicate that the planned design may not answer its intended question:
- Candidate curves nearly overlap at all chosen times. The design has little power to distinguish those scenarios, even if the number of samples is large.
- Only one phase is sampled. A mismatch may be detected, but its timing pattern may not help distinguish absorption, distribution, and clearance explanations.
- Late samples are expected below quantification limits. Nominal separation is not useful if the assay cannot reliably measure the concentrations.
- Predictions change sharply near a planned time. Small dose-time or blood-draw errors may then create substantial differences unrelated to the intended mechanism.
- Errors are correlated or concentration-dependent. Treating every observation as having the same independent additive error can exaggerate the value of a dense cluster of samples.
Sampling time alone cannot solve structural non-identifiability. Different mechanisms may produce similar plasma concentration-time curves, and a plasma sample does not directly observe tissue concentrations or clearance. Additional biomarkers, study conditions, or independent evidence may be needed for mechanistic attribution. The design also depends on the intended population, dose, formulation, route, and protocol; a schedule that distinguishes predictions in one setting may not do so in another.
A common mistake is to choose the latest time because it “must be best for clearance,” or to select the time with the largest raw concentration difference without considering measurement error. The correction is to compare the relevant predictions at feasible times using an appropriate uncertainty scale, then retain enough samples to cover the profile and distinguish curve shape from a simple vertical displacement. For example, in the terminal-phase calculation above, 8 h—not 24 h—has the greater standardized separation, while the 24 h observation can still contribute late-time coverage.
Check Your Understanding
Answer each question using the stated assumptions; the solutions follow.
- Two predictions at 0.5 h are 2.0 and 1.4 mg/L. If the additive observation-error standard deviation is 0.20 mg/L, calculate the standardized contrast.
- At a candidate time, two predictions differ by 0.30 mg/L. At another time they differ by 0.45 mg/L, but the observation-error standard deviations are 0.10 and 0.30 mg/L, respectively. Which time has the larger standardized contrast?
- Why is one late sample insufficient to establish that a mismatch is caused by clearance?
- A study has a strong 1 h contrast for an absorption alternative, but all its other samples are between 0.5 and 2 h. Name one limitation and one design improvement.
Solutions. 1. \(d=|2.0-1.4|/0.20=0.60/0.20=3.00\), or three observation-error standard deviations. 2. The first time has \(0.30/0.10=3.00\); the second has \(0.45/0.30=1.50\). The first time has the larger standardized contrast despite its smaller raw difference. 3. A single concentration can reflect both the terminal curve’s level and its shape, and distribution or other processes can also affect late concentrations. Multiple late samples are needed to examine the decline pattern, and even then the result is not a direct measurement of clearance. 4. The study has poor coverage of the distribution and terminal phases, so it may not expose those error classes. Add feasible samples in relevant later phases, selected by comparing the corresponding candidate predictions and accounting for assay precision.