Bioavailability and Clearance Confounding in Individual Posteriors

Why Oral Data Can Leave Two Parameters Entangled

For an individual receiving an oral drug, a low measured concentration may reflect low bioavailability, high clearance, or a combination of both. This ambiguity matters in individualization: a posterior can predict exposure reasonably well while remaining uncertain about the separate values of bioavailability and clearance. If those parameters are then interpreted or used separately, the apparent precision of a point estimate can be misleading.

Consider linear pharmacokinetics with absolute oral bioavailability \(F\), dose \(D\), and clearance \(CL\). For a dose at steady state every \(\tau\) hours, the interval AUC is \(F D/CL\), so average concentration is \(F D/(CL\tau)\). Thus, if the data primarily inform AUC or average exposure, they primarily inform the ratio \(F/CL\). This is not merely a correlation caused by a particular estimation algorithm: an AUC-only likelihood has an exact ridge of equally supported parameter combinations.

The distinction is important. More precise measurements of the same exposure summary can narrow the ratio estimate without separating its components. Additional information that changes the concentration-time shape, or a defensible source of prior information, may help distinguish them. Neither guarantees that individual \(F\) and \(CL\) are well determined.

What the Likelihood Can Identify

Suppose the observations are summarized by interval AUC under linear, time-invariant pharmacokinetics, with the administered dose and interval known. Assume complete dose records and that the AUC summary is appropriate for the dosing interval. A simplified observation model is \(AUC_\tau = FD/CL + \epsilon\), where \(\epsilon\) represents measurement or model error. The likelihood for these data depends on \(F\) and \(CL\) through their ratio.

The implication follows because \(D\) is fixed and known. Consequently, any two feasible pairs with the same \(F/CL\) predict the same AUC in this model. If the residual model also depends only on the predicted AUC, the likelihood is constant along that set of pairs. The data alone cannot choose one point on the ridge over another.

A concentration-time profile can contain more information than AUC alone. For example, after absorption is effectively complete in a one-compartment model, concentration declines exponentially with rate constant \(k=CL/V\). The slope over time can inform \(k\), while the profile’s scale depends on bioavailability and volume. If \(V\) is known, that shape information can help estimate \(CL\), after which the scale can inform \(F\). If \(V\) is not known, the oral profile may still leave important parameter combinations unresolved. Sampling times, absorption, measurement error, and the model’s adequacy all affect how much separation is possible.

How Sampling Can Weaken or Break the Ambiguity

The following calculation illustrates what shape information can add under restrictive conditions. Assume one-compartment disposition, first-order elimination, a known dose, known volume \(V\), and samples taken after absorption is effectively complete. Then \(C(t)=FD/V\exp(-CLt/V)\). With two post-absorption samples, the ratio cancels the concentration scale, allowing the elimination rate to be estimated. The absolute scale can then inform \(F\), conditional on the stated known \(V\) and model assumptions.

A useful diagnostic is to ask which features of the observations change when \(F\) varies and when \(CL\) varies. AUC and average exposure may change similarly under compensating adjustments. A well-observed decline over time can supply a distinct shape feature, but only when the model and sampling support its interpretation. The posterior should be inspected for a narrow direction corresponding to a stable exposure ratio and a broader compensating direction in \(F\) and \(CL\). A high posterior correlation is a warning of dependence, not proof by itself that the likelihood contains no separate information.

Parameterization and Prior Information

Parameterization can make the informed and weakly informed directions easier to see. Let \(f=\log(F/F_{\mathrm{ref}})\) and \(c=\log(CL/CL_{\mathrm{ref}})\), where \(F_{\mathrm{ref}}\) and \(CL_{\mathrm{ref}}\) are fixed positive reference values. For AUC information, the log ratio is represented by \(q=f-c\), up to a fixed reference constant. A parameterization in terms of \(q\) and \(c\) makes the structure explicit: the data may inform \(q\), while \(c\), and therefore the separate values of \(F\) and \(CL\), remains sensitive to prior information.

A change of coordinates does not add information to the likelihood. It can improve computation or clarify the ridge, but it cannot make the data identify a parameter combination that the likelihood does not distinguish. Priors can regularize that weak direction. Their effect should be described accurately: they may produce a proper, concentrated posterior even where the likelihood alone does not, but the resulting separation then depends substantially on the prior.

The distinction between likelihood information and prior contribution can be assessed by comparing the likelihood’s support with the posterior, and by repeating the analysis under scientifically defensible alternative priors. Prior distributions must also respect the parameter’s scale and constraints. Absolute bioavailability is bounded by zero and one; an unconstrained normal prior directly on \(F\) would assign probability to impossible values. A logit-scale prior or a suitably bounded distribution avoids that problem. On log scales, prior distributions also require careful interpretation because they concern multiplicative rather than additive changes.

A further caution concerns mode estimates. A posterior mode is the location of the highest density, and its numerical value can depend on the coordinate system because densities transform with a Jacobian. Posterior probabilities and properly transformed distributions do not acquire information merely from a reparameterization. When comparing parameterizations, transform the prior and posterior distributions consistently; do not compare coordinate-dependent density modes as if they were invariant quantities.

Diagnostics, Alternatives, and Limits

Several checks help determine whether an individual separation is supported by the observations or imposed by assumptions:

• Inspect posterior plots on interpretable scales, including \(F\), \(CL\), and their ratio. A narrow ratio with a long compensating ridge is evidence that exposure is better identified than its components.
• Compare likelihood profiles or fits at fixed feasible values of \(F\), allowing \(CL\) to adjust. Similar fits across a wide range suggest weak separation by the data.
• Check whether the sampling times include information about concentration-time shape, and whether absorption and residual-error assumptions are credible for those times.
• Repeat the analysis under alternative plausible priors and parameterizations. Large changes in the separate estimates, alongside stable predicted concentrations or ratios, indicate prior-sensitive separation.
• Check model predictions against the observed concentrations. A narrow posterior produced by a misspecified model is not reliable evidence of identification.

The main alternatives are not interchangeable solutions to the same problem. A richer concentration-time design can add likelihood information if it samples features that distinguish clearance from an amplitude effect. External information, such as a defensible prior informed by other data, can regularize the individual estimate but makes the inference conditional on that source. Reparameterizing into an informed ratio and a weakly informed component improves transparency and may aid computation, but does not resolve the scientific ambiguity. A population model may also constrain an individual using population information; such borrowing should be acknowledged as a modeling assumption, not attributed to that individual’s oral observations alone.

These conclusions rely on the assumed dose history, linearity, and adequacy of the structural and residual-error models. Incorrect dose records, time-varying clearance, unmodeled adherence, or misspecified absorption can distort the apparent ratio and its uncertainty. Even with an appropriate model, a single individual’s limited data may not support precise separate estimates. In that situation, report the uncertainty and the better-supported exposure quantity rather than presenting a prior-driven point estimate as if it were directly observed.