Separating Clearance and Volume With Sparse Data

Why Clearance and Volume Can Be Hard to Separate

In sparse pharmacokinetic data, a concentration at one time is the combined consequence of several processes. For a known intravenous bolus dose, a low concentration might reflect a large distribution volume, rapid elimination, or both. The parameters are not interchangeable: clearance \(CL\) governs elimination, while volume \(V\) sets the concentration scale. But with too little information, many pairs of values can explain the same observation.

This tutorial uses a one-compartment model with first-order elimination to show what the data can and cannot determine. The model is deliberately simple: the point is to understand the geometry of the inference, not to treat this model as universally adequate. Oral absorption ambiguity is a separate issue; here we assume the dose enters the systemic circulation immediately.

The elimination rate constant \(k\) has units of inverse time. Taking logarithms gives a straight-line representation:

The slope is \(-k\); the intercept is \(\log(D/V)\). If both slope and intercept are informed by data, \(k\) and the initial concentration can be estimated, after which \(V=D/C(0)\) and \(CL=kV\). Sparse observations may fail to inform one or both of these quantities adequately.

A Single Concentration Defines a Ridge

For one exact concentration \(C\) measured at time \(t\), the model implies \(C=(D/V)\exp[-(CL/V)t]\). This is one equation in two unknown individual parameters. For each chosen positive value of \(k\), the matching volume and clearance are \(V=D\exp(-kt)/C\) and \(CL=kV\). Thus a continuum of parameter pairs can reproduce the same observation.

This nonidentifiability can also be seen locally. Write the parameter vector as \(\eta=(\log V,\log CL)\). The predicted log concentration is \(\log C=\log D-\log V-(CL/V)t\), and its gradient is \(\left( -1+kt,\,-kt \right)\). With one observation, the local information contribution is proportional to the outer product of this one gradient with itself. Such a matrix has rank at most one, so it cannot supply independent curvature in two parameter directions. The likelihood is locally flat, or nearly flat, along a direction that trades off \(CL\) and \(V\).

What Additional Times Contribute

With exact concentrations at two distinct times \(t_1\) and \(t_2\), the log-concentration difference eliminates the intercept:

Once \(k\) is obtained, the model gives \(C(0)=C(t_1)\exp(kt_1)\), then \(V=D/C(0)\) and \(CL=kV\). This algebra establishes structural identification under the idealized model. It does not guarantee precise estimates from noisy data: when times are close, small concentration errors can produce substantial uncertainty in the slope.

Sampling times matter because they determine how much the concentrations change relative to their measurement error. Two samples taken almost together can behave much like one sample for estimating a slope. A sample late enough to show meaningful decay can add information about \(k\), but a very late sample may be close to the assay's quantification limit or dominated by residual error. No single timing rule applies without regard to the error model and the expected PK profile.

Reading Posterior Geometry

In Bayesian individualization, the posterior combines the likelihood from the individual's concentrations with the prior distribution for the PK parameters. A prior can make an otherwise unbounded ridge yield a proper, finite posterior. That is regularization, not evidence that the individual observations independently determined both parameters. When the likelihood is weak in one direction, posterior estimates in that direction can be substantially influenced by the prior.

A posterior plot of \(\log V\) against \(\log CL\) helps reveal this structure. A long, narrow contour indicates that some combinations are much better determined than others. The orientation shows how the parameters trade off; the breadth along the contour shows the remaining uncertainty. Posterior correlation is a useful summary, but a correlation close to zero does not by itself establish good identification: nonlinear or curved relationships can be poorly represented by one correlation coefficient.

A local curvature calculation gives another diagnostic. If a fitted log-likelihood or log-posterior is approximated near its mode by a quadratic, its curvature matrix describes local information. For example, an illustrative positive-definite local information matrix in two log-parameter coordinates might be

Its determinant is \(5(145)-25^2=100\), and its trace is \(150\). Its eigenvalues are \((150\pm\sqrt{150^2-4(100)})/2\), approximately \(149.3\) and \(0.6697\). The much smaller eigenvalue corresponds to a direction with far less local curvature than the other direction; the eigenvalue ratio is about \(223.0\). This is an example of an ill-conditioned geometry, not a universal threshold for acceptable estimation. The matrix must be interpreted in the parameterization and scaling used, and posterior curvature also reflects prior information.

Alternatives, Failure Modes, and Limits

The main data-based alternative to relying on one concentration is to obtain samples at distinct, informative times and estimate the log-concentration slope and intercept. A nonlinear model fit to all observations is generally preferable when residual errors are specified on the concentration scale or when observations do not follow a straight line after logging. A Bayesian fit adds prior information and represents uncertainty jointly; the earlier material on MAP estimates, posterior means, posterior correlation, and prior sensitivity provides relevant context. These approaches differ in estimation and uncertainty representation, but none can recover information absent from the observations and assumptions.

• One sample treated as two pieces of information. A concentration depends on both \(CL\) and \(V\); it does not separately measure them. Report the joint uncertainty rather than interpreting two precise marginal point estimates as data-driven.
• Close samples treated as a reliable slope. The slope divides a concentration difference by elapsed time. If the time interval is small relative to measurement noise, slope uncertainty can be large.
• Prior precision mistaken for individual identification. A narrow posterior may result from a strong prior. Compare with the likelihood geometry and examine sensitivity to plausible prior choices.
• Incorrect model or records. Missed doses, incorrect sample times, distribution phases, nonlinear elimination, or an unsuitable error model can distort the apparent slope and intercept. A tidy posterior under a misspecified model does not ensure valid parameter estimates.

There are important limits to the algebraic two-time result. It assumes a single exponential decline after an IV bolus, with no changing clearance or volume and no additional disposition phase. Measurement noise means the exact two-point solution is random; additional samples can help assess fit and reduce uncertainty, but only if their information is not undermined by poor timing, high error, or model failure. A population prior may stabilize sparse individual estimates, yet the resulting answer should be described as combining individual data and population information.