What One Trough Can and Cannot Identify

Why a Trough Is Informative but Incomplete

A trough concentration can be useful for individualizing a forecast, but it is only one observed value at one point in a dosing interval. How much it says about a patient’s pharmacokinetics depends on the concentration model, the timing and accuracy of the sample, and the prior distribution for individual parameters. In particular, a measured trough does not automatically determine both clearance and volume, even when the population model contains both.

To make the information content precise, consider a hypothetical patient receiving repeated intravenous bolus doses \(D\) every \(\tau\) hours in a one-compartment model with first-order elimination. Assume the doses and sampling time are known, the patient is at steady state, and the model adequately describes the patient. Let \(CL\) be clearance and \(V\) be volume of distribution, so the elimination rate constant is \(k=CL/V\). The trough is measured immediately before the next dose.

Even if the observed trough is measured without error, it supplies one equation involving two unknown patient parameters, \(CL\) and \(V\). In most settings the concentration measurement also has residual error, so it supplies a likelihood rather than an exact equation. The Bayesian forecast combines that likelihood with the population prior; it does not turn one observation into two independent parameter measurements.

What the Trough Constrains

It helps to write the trough on the log scale and define \(x=k\tau=CL\tau/V\):

For fixed \(V\), increasing \(CL\) increases \(x\) and lowers the trough. For fixed \(CL\), increasing \(V\) lowers the initial concentration scale \(D/V\), but also decreases \(x\), slowing the fractional decline during the interval. The net trough response to volume is therefore not described by the \(D/V\) term alone.

A local sensitivity calculation makes the joint constraint visible. Let the parameter coordinates be \(\log CL\) and \(\log V\). Differentiating the log trough gives:

These two derivatives describe the local direction in parameter space to which the trough responds. A single observation can inform movement along that direction; movement in a direction that barely changes the predicted trough remains weakly informed by the measurement. The population prior consequently has substantial influence on how the update is divided between clearance and volume.

A Local Bayesian Update

The earlier material on individual concentration likelihoods and population priors motivates a useful local calculation. It is not a replacement for evaluating a full nonlinear posterior. Here suppose the prior for \((\log CL,\log V)\) is approximately bivariate normal near a reference point, the log concentration residual is approximately normal, and the log trough model can be linearized over the parameter changes being considered. These assumptions permit a transparent first-order update.

Write \(\theta=(\log CL,\log V)^\mathsf{T}\), let \(\theta_0\) be the reference point, let \(g\) be the vector of log-trough sensitivities there, and let \(r=\log C_{\mathrm{obs}}-\log C_{\mathrm{trough}}(\theta_0)\). If the prior is centered at \(\theta_0\), with covariance \(\Sigma\), and the log-scale residual variance is \(s_e^2\), the linearized posterior mean shift from \(\theta_0\) is:

This follows by treating the linearized observation as \(r\approx g^\mathsf{T}\Delta\theta+\epsilon\), where \(\epsilon\) has variance \(s_e^2\). The joint normal update weights the residual by the prior covariance in the sensitivity direction; a larger residual variance reduces the update, while greater prior variance along that direction allows more movement.

Which Forecast Features Are Better Informed?

The observation directly informs concentration near the sampling time under the assumed model. It can also constrain model parameters through the sensitivities above. But a forecast at another time, or an exposure summary such as AUC, depends on how the uncertain parameters combine to produce that quantity. The trough may have limited power to reduce uncertainty in a feature that is weakly related to its sensitivity direction.

In the simple bolus model, interval AUC is \(D/CL\), so it depends on clearance but not directly on volume. If the trough update can be explained by several combinations of \(CL\) and \(V\), posterior uncertainty in clearance—and therefore in AUC—can remain appreciable. Conversely, if the prior already makes volume precise, a trough may primarily update clearance. Which situation applies is determined by the prior covariance, the model sensitivities, and the observation error, not by the word “trough.”

A useful way to separate measurement information from prior influence is to compare the prior and posterior distributions of the actual forecast target, not only the fitted parameters. For example, report how much the posterior uncertainty in trough concentration, AUC, or a specified future concentration changes. A parameter estimate can shift while a clinically relevant prediction remains uncertain; a good fit to the observed trough is not evidence that all forecasts are precise.

Assumptions, Diagnostics, and Alternatives

The calculations above require a known dose and interval, correct sample timing, steady state, the stated one-compartment bolus model, and a suitable residual-error model. A delayed or mistimed sample, missed dose, changing clearance, or a patient whose concentration-time shape differs from the model can make the apparent parameter update misleading. A numerical posterior is conditional on those assumptions; it does not verify them.

Check the recorded dose and sampling time first, then inspect whether the observation is plausible under the prior predictive distribution and whether the posterior is being driven to an implausible region. Compare fits across reasonable prior assumptions and examine posterior uncertainty in the intended forecast quantity. A large discrepancy between the observation and prior prediction can signal informative patient data, but it can also signal a data or model problem; the discrepancy alone does not distinguish these explanations.

There are several practical alternatives, each answering a different question. Holding one parameter fixed gives a conditional estimate, as in the first worked example; it is useful for sensitivity analysis but should not be mistaken for data-based identification of that parameter. A full Bayesian analysis retains uncertainty in both parameters and propagates it into predictions, but its result still depends on the population prior and model. Adding observations at informative times can provide another sensitivity direction, whereas repeating a measurement under effectively the same conditions may mainly reduce uncertainty in the same trough direction. The later material on combining peak and trough measurements addresses the additional information available from distinct sampling phases.

A common mistake is to say, “The observed trough is higher than predicted, so clearance must be lower.” That conclusion is valid only if other relevant parameters are fixed or sufficiently well known. With both \(CL\) and \(V\) uncertain, a higher trough can be explained by multiple parameter changes, as the examples demonstrate. The correction is to state the assumptions that isolate clearance, or to report the posterior joint uncertainty and the resulting uncertainty in the forecast.