Why Sparse Designs Need Diagnosis
A sparse pharmacokinetic schedule can be efficient, but a small number of samples does not automatically provide useful information about every parameter. A sample collected after absorption and distribution have largely finished may help describe the terminal decline, yet reveal little about the processes that shaped the earlier concentration curve. The design question is therefore not simply how many samples were collected, but which parts of the curve they represent and whether those observations distinguish the parameters of interest.
A concentration–time plot is a first diagnostic. Mark the planned sampling times against the expected profile, then ask whether there are observations during the phases relevant to the intended parameters. Parameter uncertainty provides a second diagnostic: broad intervals, unstable estimates, or strong dependence on starting values can indicate that the data do not adequately distinguish possible parameter values. These warning signs are evidence to investigate, not automatic proof that a design is unusable.
Relating Sampling Times to the Concentration Curve
For a concrete illustration, consider a hypothetical single-dose, one-compartment model with first-order absorption and elimination. Assume the dose and sampling times are known, concentrations are measured without bias, and the structural model is an adequate description of the profile. Let \(k_a\) be the absorption rate constant and \(k_e\) the elimination rate constant. For \(k_a\ne k_e\), write the concentration as
Here \(F\) is bioavailability, \(D\) is dose, and \(V\) is the apparent volume of distribution. This model is an illustration of how sampling times relate to profile shape, not a claim that every drug follows this model. In more complex models, distribution may create a rapid early decline distinct from the terminal elimination phase.
The absorption phase is represented by the rise toward the peak; the peak itself and nearby observations help constrain its timing and magnitude. A distribution phase, when present, is usually most visible early after dosing. The terminal phase is the later portion of the profile, where the log concentration may decline approximately linearly with time and reflect elimination. A schedule that samples only one region can describe that region while leaving other phases poorly characterized.
Worked Example: Locating the Missing Phase in a Schedule
Suppose hypothetical model predictions follow \(C(t)=6(e^{-0.20t}-e^{-1.2t})\) mg/L, with time in hours. A proposed schedule samples at 0.25, 8, and 12 hours. Calculate selected predicted concentrations and assess phase coverage.
At 0.25 hours, \(C(0.25)=6(e^{-0.05}-e^{-0.3})=6(0.951229-0.740818)=1.26247\) mg/L, or approximately 1.262 mg/L.
At 8 hours, \(C(8)=6(e^{-1.6}-e^{-9.6})=6(0.201897-0.0000677)=1.2110\) mg/L. At 12 hours, \(C(12)=6(e^{-2.4}-e^{-14.4})=6(0.0907180-0.000000556)=0.54430\) mg/L.
The model’s peak time is \(t_{\max}=\ln(k_a/k_e)/(k_a-k_e)=\ln(6)/1.0=1.792\) hours. None of the proposed samples is near that peak. The 8- and 12-hour observations are late relative to the peak and are useful for the terminal decline; the 0.25-hour observation is early but does not by itself map the rise or locate the peak. Thus the schedule has a clear gap around peak formation. It may support some information about elimination while providing weaker evidence about absorption timing and peak concentration.
How Parameter Information Appears in the Data
A useful formal idea is sensitivity: how much a predicted concentration changes when a parameter changes. For parameter \(\theta_j\), the local log-sensitivity at time \(t_i\) can be written \(S_{ij}=\partial\log C(t_i)/\partial\log\theta_j\). A small sensitivity means that a modest relative change in the parameter produces little relative change in the predicted concentration at that time. If the observations are insensitive to a parameter, even precise concentration measurements may not estimate it precisely.
Sensitivity must be interpreted jointly across parameters. If two parameters produce very similar changes in the predictions at all sampled times, their effects can be difficult to separate. This is why observations spread across informative phases can be more useful than the same number of samples clustered in one region. In a population model, between-subject variability and residual error also affect precision; design adequacy is conditional on those features as well as the structural model.
The local information idea can be summarized with a sensitivity matrix \(J\), whose rows correspond to observations and columns to parameters. Under a Gaussian-error model with independent errors of known variances (or the corresponding weighted least-squares local approximation), where \(J\) is the Jacobian of the predicted means with respect to the parameters and \(W\) is the inverse error-variance matrix, the local information matrix has the form \(J^\mathsf{T}WJ\). If a parameter’s sensitivity column is small, or if columns are nearly redundant, the information matrix provides little information for separating that parameter from others. This is a diagnostic approximation, not a guarantee that a fitted model will have well-behaved uncertainty.
Worked Example: Why Late Samples Can Miss Absorption Information
Use the same hypothetical model, now comparing the late samples at 8 and 12 hours with a schedule that also includes 0.25 hours. When \(k_a>k_e\) and the absorption exponential has become negligible, the concentration is approximately proportional to \(e^{-k_e t}\). More exactly, the model can be written \(C(t)=A e^{-k_e t}(1-e^{-k_a t})\), where \(A=FDk_a/[V(k_a-k_e)]\).
For the two late samples, the concentration ratio is \(C(12)/C(8)=e^{-4k_e}(1-e^{-12k_a})/(1-e^{-8k_a})\). With \(k_e=0.20\) and \(k_a=1.2\), the factor involving absorption is nearly one: \(1-e^{-9.6}=0.9999323\) and \(1-e^{-14.4}=0.9999994\). The ratio is therefore approximately \(e^{-0.8}\), and changes in \(k_a\) have almost no effect on this late-sample contrast.
By contrast, the early-to-late ratio \(C(0.25)/C(8)=1.26247/1.2110\approx1.0425\) includes a substantial absorption contribution. To quantify the contrast, the derivative of \(\log C(t)\) with respect to \(\log k_a\), for this model, is \(S_{k_a}(t)=-k_e/(k_a-k_e)+k_at e^{-k_at}/(e^{-k_et}-e^{-k_at})\). At 0.25 hours it is approximately \(0.856\); at 8 hours it is approximately \(-0.197\). The early-versus-late contrast thus has sensitivity about \(0.856-(-0.197)=1.053\) to \(\log k_a\), unlike the nearly absorption-insensitive late contrast.
This calculation assumes the structural model and treats its scale as known for the displayed sensitivity. In an actual analysis, scale parameters such as \(F/V\) are estimated too, so uncertainty and confounding must be assessed jointly. The central design warning remains: if all observations are late, the data can show a terminal slope while offering little direct evidence about the absorption phase.
Using Parameter Uncertainty as a Warning Signal
After fitting a model, inspect uncertainty for the parameters that the design is intended to estimate. Large standard errors or broad confidence intervals indicate limited precision under the fitted model and assumptions. Profile likelihoods or bootstrap distributions can reveal asymmetry, flat regions, or estimates that vary substantially across refits. Boundary estimates, implausible values, and sensitivity to starting values are further reasons to investigate. None of these findings alone identifies the cause: weak design information, model misspecification, influential observations, and optimization problems can produce overlapping symptoms.
The following example illustrates uncertainty for a terminal slope under deliberately simplified assumptions. It does not establish that a real population model’s uncertainty can be calculated from two observations in this way.
Worked Example: A Terminal Slope With Conditional Uncertainty
Suppose hypothetical concentrations are measured at 8 and 12 hours as 1.211 and 0.5443 mg/L. Assume the terminal log-concentration follows a straight line with slope \(-k_e\), the two log measurements have independent errors with known standard deviation \(0.09975\) (approximately a 10% coefficient of variation), and absorption is negligible at both times.
The slope-based estimate is \(\hat{k}_e=[\ln(1.211)-\ln(0.5443)]/(12-8)\approx0.1999253\ \mathrm{h}^{-1}\), or approximately \(0.1999\ \mathrm{h}^{-1}\). Because the two log measurements are independent and have the same known standard deviation, the standard error of their difference divided by four hours is \(SE(\hat{k}_e)=\sqrt{2}(0.09975)/4=0.03527\ \mathrm{h}^{-1}\).
Using a normal 95% interval under these assumptions gives \(0.1999253\pm1.96(0.03527)=0.1999253\pm0.06913\), or approximately \(0.1308\) to \(0.2691\ \mathrm{h}^{-1}\). This interval is fairly broad relative to the estimate, even for this simplified calculation. It concerns the terminal rate constant only; it does not establish that the schedule identifies \(k_a\), distribution parameters, or population variability. In a real analysis, the error variance is usually estimated and model uncertainty may make the interval wider or otherwise different.
Diagnostics, Alternatives, and Limitations
Use the plot and parameter uncertainty together. Overlay planned sample times on a plausible concentration–time profile, preferably including uncertainty in the expected profile rather than relying on one predicted curve. Check whether the early rise, expected distribution period, peak region, and terminal decline have observations when those phases matter to the objectives. Then examine whether fitted parameters have plausible estimates and reasonably stable uncertainty under the intended model. A plot can expose a phase gap that a single summary statistic hides; uncertainty can show that the gap has practical consequences for estimation.
For a multi-compartment model, a schedule containing only late samples may be unable to distinguish rapid distribution from other early-profile behavior because the fast component has largely disappeared by the time sampling begins. Conversely, samples concentrated near the peak may characterize peak exposure while providing a short or absent terminal segment. Sampling time alone is not the whole story: assay precision, concentration limits, dose history, and between-subject variability influence the information available at each time.
When a design appears weak, the main alternatives depend on the question. Rich sampling across the profile can provide direct phase coverage within individuals. Balanced or staggered sparse sampling can spread observations across the interval across participants, though it relies on the model to combine those observations. A design comparison using predicted parameter precision can help distinguish candidate schedules, but predictions depend on assumed parameter values, residual error, and model structure. If collecting more data is not feasible, narrowing the estimation target or using external information may be considered, while explicitly stating that conclusions then rely more heavily on those assumptions.
A common mistake is to treat a large sample count as proof that each parameter is well informed: “There are many late samples, so absorption must be estimated precisely.” The number of observations does not fix a phase-coverage gap. Many similar late concentrations can sharpen a terminal trend while remaining nearly insensitive to absorption timing. The correction is to ask whether the observed portions of the curve change enough, and differently enough, when the target parameters change.
These checks have limits. A sparse design is not inherently inadequate if the estimand is narrow and the available information addresses it. Conversely, a visually attractive spread of times does not guarantee reliable estimates under a misspecified structural model or an inappropriate residual-error model. Treat plots, sensitivity analyses, and uncertainty summaries as complementary evidence, not as interchangeable pass/fail tests.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions below.
- A schedule samples only at 10, 16, and 24 hours after a dose. What phase may be represented, and which phases could be weakly represented?
- Why can many terminal samples still provide little information about an absorption rate constant?
- In the terminal-slope example, what does the interval from 0.1308 to 0.2691 per hour describe, and what does it not describe?
- How would adding an observation near the predicted peak help diagnose a schedule with an early sample and two late samples?
- Name two reasons, other than a weak sampling design, that could produce unstable or implausible parameter estimates.
Solutions. (1) The late observations may represent terminal elimination; absorption and any early distribution phase may be poorly represented. (2) Once absorption is essentially complete, changes in \(k_a\) have little effect on late concentrations, so repeated late observations mainly refine the terminal trend. (3) Under the stated simplified error and terminal-phase assumptions, it is a 95% interval for \(k_e\), in \(\mathrm{h}^{-1}\); it is not an interval for \(k_a\), distribution parameters, or population variability. (4) A near-peak observation would add evidence about the profile’s rise and peak region, potentially making absorption timing easier to distinguish from late-phase behavior. (5) Examples include structural-model misspecification, influential data, inappropriate residual-error assumptions, or optimization difficulties.