Why Within-Subject Trends Matter
A population model can allow patients to have different clearances without implying that any one patient’s clearance changes over time. The distinction matters: a stable difference between people is represented by interindividual variability, whereas induction or a changing disease state can produce a systematic evolution of clearance within a person. Treating one pattern as the other can leave time-ordered residual trends and distort estimates of exposure or dose requirements.
The useful comparison is usually the log-concentration slope, or fractional rate of concentration decline, not the raw concentration slope. Even with constant clearance and volume, the raw slope \(dC/dt\) changes as concentration falls. Under a simple one-compartment elimination model with constant clearance and volume, however, \(d\log C/dt\) is constant. A systematic within-subject change in that log slope can be evidence for time-varying clearance, provided other processes affecting concentration are adequately represented.
Model and Key Derivation
Consider a hypothetical one-compartment model after an intravenous bolus. Let \(C(t)\) be concentration, \(V\) be a constant volume of distribution, and \(CL(t)\) be clearance at time \(t\). Assume no further drug input during the interval, no distribution phase affecting the measured concentration, and elimination described by \(dA/dt=-CL(t)C(t)\), where \(A(t)=VC(t)\). These assumptions give:
The second equality follows by dividing the first differential equation by \(C(t)>0\). Integrating from 0 to \(t\) gives the general concentration profile:
If \(CL(t)=CL\) is constant, this reduces to \(C(t)=C(0)e^{-(CL/V)t}\); the log concentration is linear in time with slope \(-CL/V\). If clearance rises, the log slope becomes more negative; if clearance falls, it becomes less negative. The raw derivative \(dC/dt\), by contrast, is \(-[CL(t)/V]C(t)\), so it may change even when \(CL(t)\) is constant because \(C(t)\) itself changes.
This derivation is a diagnostic guide, not a claim that every observed change in log slope proves clearance change. Ongoing input adds an input-to-concentration term to the log derivative. Changing volume, distribution, absorption, dose timing, assay effects, or modelled covariates can also alter the observed profile. The interpretation is strongest in an adequately characterized elimination phase, with the relevant dose history and other influential processes represented.
Worked Example: Comparing Successive Log Slopes
Suppose a hypothetical patient receives an intravenous bolus and has concentrations of 12.0 mg/L at 0 h, 8.04 mg/L at 4 h, and 3.62 mg/L at 8 h. Assume, for this illustration, a one-compartment elimination phase, constant \(V\), no input after the bolus, and sufficiently accurate measurements.
The estimated log-concentration slope over 0–4 h is:
Over 4–8 h it is:
The fractional decline is therefore faster in the later interval. Under the stated assumptions, the pattern is consistent with an increase in \(CL(t)/V\), rather than a single constant value. For comparison, a constant rate equal to the first interval’s estimate predicts \(C(8)=12.0e^{-0.1001(8)}\approx5.39\) mg/L, above the hypothetical observed 3.62 mg/L. The two intervals are only descriptive estimates: measurement error and model misspecification must be considered before concluding that clearance changed.
The raw slopes also differ: \((8.04-12.0)/4=-0.990\) mg/L/h and \((3.62-8.04)/4=-1.105\) mg/L/h. That difference alone would not distinguish changing from constant clearance, because raw slopes can change as concentration falls. The diagnostic comparison here is specifically between log slopes.
Two Time-Varying Clearance Patterns
Induction and changing disease state can both create evolving clearance, but they imply different time courses. The equations below are candidate descriptions, not mechanisms that can be identified from concentrations alone. Parameter meanings and units should be specified before fitting or interpreting either model.
For a simple induction pattern, let clearance increase exponentially from its initial value:
Here \(CL_0\) is clearance at time zero and \(k_{\mathrm{ind}}\) has units of inverse time. With fixed \(V\), integrating the clearance-to-volume ratio gives:
The instantaneous log slope is \(-CL_0e^{k_{\mathrm{ind}}t}/V\), which becomes increasingly negative. Exponential growth over an unlimited time span is rarely a plausible biological description; this form may be useful over a restricted observation window or when a more mechanistic bounded model is not justified.
A changing disease state may instead move clearance from an initial value toward a new level. One parsimonious model is:
Here \(CL_{\infty}\) is the eventual level and \(k_{\mathrm{rec}}\) controls the speed of transition. If \(CL_0<CL_{\infty}\), clearance rises toward its new level; if \(CL_0>CL_{\infty}\), it falls. The concentration profile follows by integration:
Unlike unbounded exponential induction, this model approaches a finite clearance. A clinically measured time-varying covariate could also be used to describe clearance, for example through \(CL_i(t)=CL_{\mathrm{typ}}\exp(\beta x_i(t)+\eta_{CL,i})\), where \(x_i(t)\) is a time-changing covariate and \(\eta_{CL,i}\) represents a stable individual deviation on the log scale. Such a model requires a defensible covariate and a clear time alignment.
Worked Example: Concentration Under an Induction Model
In a hypothetical one-compartment example, let \(CL_0/V=0.0800\ \text{h}^{-1}\), \(k_{\mathrm{ind}}=0.120\ \text{h}^{-1}\), and \(C(0)=20.0\) mg/L. Assume the induction model applies over the full 12 h and the bolus is the only drug input.
At 6 h, the integrated elimination exponent is:
Thus \(C(6)=20.0e^{-0.7030}\approx9.90\) mg/L. At 12 h:
Using the displayed rounded value, \(C(12)=20.0e^{-2.147}\approx2.34\) mg/L. Under constant clearance fixed at its initial rate, the prediction would instead be \(20.0e^{-0.0800(12)}=20.0e^{-0.960}=7.66\) mg/L. The large difference follows from the assumed increase in clearance; it does not establish that induction occurred in a real patient.
Worked Example: Clearance Recovering Toward a New Level
Consider hypothetical clearance-to-volume rates \(CL_0/V=0.120\ \text{h}^{-1}\) and \(CL_{\infty}/V=0.0600\ \text{h}^{-1}\), with \(k_{\mathrm{rec}}=0.200\ \text{h}^{-1}\). Clearance falls toward its new level. Assume the same one-compartment, no-input conditions and \(C(0)=10.0\) mg/L.
At 5 h, the integrated rate is:
Therefore \(C(5)=10.0e^{-0.4896}=6.13\) mg/L. If clearance had remained at its initial level, the concentration would be \(10.0e^{-0.120(5)}=5.49\) mg/L; if it had been at the eventual level throughout, it would be \(10.0e^{-0.0600(5)}=7.41\) mg/L. The transitional model gives a value between these predictions, as expected for a rate declining from 0.120 to 0.0600 \(\text{h}^{-1}\).
Distinguishing Time Variation From Stable Variability
A common population model allows each individual to have a stable clearance, such as \(CL_i=CL_{\mathrm{typ}}\exp(\eta_{CL,i})\), where \(\eta_{CL,i}\) is an individual random effect. In the simple elimination setting, that individual’s log-concentration slope is constant at \(-CL_i/V_i\) if both clearance and volume are stable. Different people may have different slopes, but stable between-subject variability does not itself produce a systematic change in one person’s log slope over time.
Evidence for time variation is more persuasive when the within-subject trend recurs across individuals in a way aligned with a plausible induction or disease time course, and when a constant-clearance model leaves ordered residuals. For instance, a model that underpredicts late concentrations after overpredicting early concentrations may be missing a decline in clearance, although other misspecifications can produce similar patterns. Residuals should be examined against time since dose, study time, and relevant changing covariates, with dose timing and input represented correctly.
The principal alternatives answer different questions. A stable random effect describes persistent differences among individuals. A time-varying clearance model describes evolution within an individual. A time-varying covariate model ties that evolution to measured state information, while a smooth time function may describe a trend without specifying its cause. These approaches can be compared by examining fit, parameter plausibility, predictive performance, and whether the added complexity is supported by the sampling design. A lower objective function alone does not prove the biological mechanism.
| Approach | What it represents | Key limitation |
|---|---|---|
| Stable interindividual clearance variation | Persistent differences in clearance between people | Cannot explain a systematic within-person change in log slope when other assumptions hold |
| Time-varying clearance function | Clearance evolving with time, such as exponential increase or relaxation toward a new level | May fit a trend without identifying its cause; shape can be weakly supported by sparse data |
| Time-varying covariate model | Clearance linked to a measured changing state | Requires reliable covariate measurements, appropriate timing, and adequate separation from other time trends |
Diagnostics, Limitations, and a Common Pitfall
Before interpreting a changing log slope as evolving clearance, verify that the concentration observations are in a phase where the assumed elimination equation applies. Confirm dose and sample times, account for continued input where relevant, and consider whether changes in volume, distribution, or disease-related processes could explain the pattern. In population data, inspect individual profiles as well as population-level residuals: a population trend can arise from differences in sampling times or patient composition rather than within-person evolution.
Time-varying models can be difficult to identify. Sparse sampling may estimate an overall exposure change but not distinguish the initial clearance, eventual clearance, and transition rate. If all observations occur early, \(CL_{\infty}\) may be poorly determined; if they occur late, the initial value and rate may be weakly informed. Strong parameter correlations, wide uncertainty, implausible estimates, or materially different estimates from reasonable starting values are warnings. Predictions should also be checked over the time range where the model will be used, especially for an unbounded induction function.
Common mistake: “The raw concentration slope became more negative, so clearance must have increased.” This is not sufficient. With constant clearance and volume, \(dC/dt=-(CL/V)C(t)\) changes as concentration changes. The appropriate simple-model comparison is \(d\log C/dt\), which remains \(-CL/V\) under constant clearance and no input. Even a changing log slope is evidence to investigate, not proof of a clearance mechanism, because the assumptions and alternative explanations still matter.
The central limitation is that concentration profiles identify the combined consequences of drug input, distribution, and elimination, not clearance in isolation under every design. A time-varying clearance function is useful when the data and biological context support it, but a more complex model should not be used merely to absorb unexplained residual patterns. The interpretation should remain conditional on the model structure, observation window, and available covariate information.
Check Your Understanding
Answer each question, then compare with the concise solutions below.
- With constant \(CL\) and \(V\), what is constant during a no-input elimination phase: \(dC/dt\) or \(d\log C/dt\)?
- A patient’s log-concentration slopes over two successive intervals are \(-0.05\ \text{h}^{-1}\) and \(-0.09\ \text{h}^{-1}\). Under the one-compartment assumptions, what direction of change in \(CL/V\) is consistent with this pattern?
- In the recovery model, if \(CL_0>CL_{\infty}\), does clearance rise or fall over time?
- Name two reasons, other than changing clearance, that could produce a changing observed log-concentration slope.
Solutions. (1) \(d\log C/dt=-CL/V\) is constant; the raw derivative generally changes with concentration. (2) The slope becomes more negative, consistent with increasing \(CL/V\). (3) Clearance falls toward \(CL_{\infty}\), since \(e^{-k_{\mathrm{rec}}t}\) decreases from 1 toward 0. (4) Examples include ongoing drug input, changing volume, distribution effects, incorrect dose or sample times, and measurement error.