Absorption Ambiguity From Sparse Oral Samples

Why sparse oral samples can leave absorption unresolved

A small number of oral concentrations can be compatible with more than one account of absorption. A relatively rapid absorption rate with a longer lag may produce a concentration similar to that from slower absorption with a shorter lag. Even when the dose time is known, sparse samples may not reveal whether a concentration was observed while the profile was rising, near its peak, or on its decline. The data can therefore support distinct combinations of absorption rate, lag, and phase rather than one well-determined absorption history.

This is a question of posterior ambiguity, not simply whether an optimizer returns a numerical estimate. A single maximum a posteriori estimate (MAP) can sit on one plausible explanation while concealing another. The practical task is to diagnose whether the individual data meaningfully distinguish these explanations, and to communicate the remaining uncertainty before using an absorption estimate for an individual forecast.

A model for reasoning about absorption shape

Consider one oral dose \(D\) at a known time, with linear one-compartment disposition, first-order absorption, and a lag \(\tau\) before absorption begins. Let \(k_a\) be the absorption rate constant, \(k\) the elimination rate constant, and \(A=FD/V\), where \(F\) is bioavailability and \(V\) is volume of distribution. For elapsed time \(s=t-\tau>0\), the concentration predicted by this model is

For \(t\leq\tau\), the predicted concentration from this dose is zero. This expression assumes a known dose and dose time, positive \(k\), \(k_a\), and \(\tau\geq 0\); a known value of \(A\); and the stated linear, first-order model. In practice \(A\) or \(k\) may also be uncertain, but allowing them to vary adds sources of ambiguity. Here, they are treated as known so the focus remains on absorption. The formula also requires care when \(k_a=k\): its continuous limit is \(C(t)=Aks e^{-ks}\).

The likelihood connects those explanations to the observations. For example, under independent additive normal errors with known standard deviation \(\sigma\), observations \(C_i^{\mathrm{obs}}\) at known times \(t_i\) have likelihood proportional to

This likelihood assumption is used for the examples below: conditional on the model predictions, residuals are independent, normally distributed, mean zero, and have the stated common standard deviation. A different residual error model changes the likelihood and should be used when it better matches the measurement process. The individual posterior is proportional to this likelihood times the prior density for the parameters. Thus similar likelihoods do not by themselves imply similar posterior support if the prior assigns very different densities to the competing regions.

Why rate and lag can trade off

The concentration curve has a peak when its time derivative is zero. For \(s>0\), differentiating the model gives a factor proportional to \(-ke^{-ks}+k_ae^{-k_as}\). Setting this equal to zero and rearranging yields the elapsed time to peak:

The limit as \(k_a\) approaches \(k\) is \(s_{\max}=1/k\). The peak time therefore depends on both the absorption rate and lag. A sample observed at a given clock time may be on the rising limb for one parameter combination and on the declining limb for another. If there are few observations around the peak, the data may constrain observed concentrations without sharply constraining the peak time or the shape that connects them.

A related source of ambiguity is that changing \(k_a\) changes the curve’s shape while changing \(\tau\) shifts it in time. With only one concentration, many combinations can give nearly the same prediction. Additional observations help only to the extent that their times and measurement precision reveal differences between those curves.

Quantifying support for competing rates

A useful diagnostic is to compare the likelihood or posterior support across plausible parameter regions, rather than reporting only the best-fitting point. The following hypothetical model evaluation illustrates a comparison of two absorption rates while all other settings, including lag and sample times, are held common. Suppose three observations are \(0.420\), \(0.570\), and \(0.280\ \mathrm{mg/L}\), with independent normal errors having known \(\sigma=0.100\ \mathrm{mg/L}\). At the three sample times, the model predictions are \((0.246,0.561,0.318)\ \mathrm{mg/L}\) for \(k_a=0.60\ \mathrm{h}^{-1}\), and \((0.409,0.564,0.290)\ \mathrm{mg/L}\) for \(k_a=1.20\ \mathrm{h}^{-1}\).

Diagnosing ambiguity in an individual posterior

Inspect the joint posterior for \(k_a\) and \(\tau\), as well as the implied concentration curves. Posterior correlation may show a rate–lag trade-off, while separated clusters can indicate distinct supported explanations. Neither a correlation coefficient nor marginal intervals alone reveal the full geometry: a curved ridge can have modest linear correlation, and similar marginal summaries can conceal different joint configurations. The introductory material on posterior correlation and identifiability provides the broader interpretation; here the diagnostic question is whether the supported combinations imply meaningfully different absorption timing or profile shape.

A practical assessment can combine complementary views:

• Joint plots and conditional summaries: plot posterior draws or density contours for \(k_a\) and \(\tau\), and inspect curves generated by draws from each region.
• Likelihood or posterior profiles: vary one parameter while optimizing or integrating over the others. A flat or irregular profile signals weak determination or competing regions, though a profile alone does not represent all posterior uncertainty.
• Sampling diagnostics: with MCMC, check whether chains visit all apparent regions and mix between them. Agreement in a single chain’s summaries is not reassuring if it never explores a second mode. With Laplace approximation, check whether a single local Gaussian approximation is appropriate; it can miss a remote mode or poorly represent a long curved ridge.
• Prior sensitivity: compare whether the locations and relative support of plausible absorption explanations change materially under reasonable alternative priors. Strong prior influence is important to report, not evidence that the data distinguish the alternatives.

These checks rely on correctly implemented computation as well as a plausible model. If chains do not explore the posterior, a reported single mode may be a computational failure rather than a feature of the data. Conversely, a well-mixed sampler cannot make sparse data informative: it can accurately represent a broad or multimodal posterior.

Common pitfalls and limits

• Reporting only the MAP estimate. A precise-looking optimum does not show whether other regions also fit. Inspect joint support and the resulting curves.
• Calling a posterior correlation proof of non-identifiability. Correlation is a clue, not a complete diagnosis. Check for curved ridges, distinct modes, and consequences for predicted profiles.
• Reading a best-fitting rate as a data-only conclusion. Posterior support combines likelihood and prior. When observations are sparse, prior sensitivity may be substantial.
• Assuming samples are informative because they are numerous enough to fit a model. Several samples clustered on a similar part of the curve may contribute little information about lag or peak timing. Sampling placement matters.
• Confusing an absorption explanation with a verified mechanism. The conclusions are conditional on the dose record, sampling times, residual model, and assumed PK structure. Incorrect timing, misspecified absorption, or unmodeled variability can produce misleading posterior structure.

Profile likelihoods, a local Laplace approximation, and posterior simulation serve different purposes. A profile can expose weakly constrained directions; Laplace approximation can be efficient near a well-behaved single mode; MCMC or importance sampling can represent more complex posterior shapes when diagnostics support their use. None replaces checking the model and the sampling information. In particular, the present assessment holds disposition and amplitude fixed. If they are also uncertain, the absorption assessment must account for those additional dependencies rather than treating this simplified comparison as a complete individual analysis.