Why Clearance May Change During a Treatment Course
A concentration profile from the first dose does not necessarily describe disposition several weeks into treatment. Clearance can change over treatment time—for example, as physiology, disease status, or concomitant therapy changes. But a change in the observed decline of concentration is not, by itself, evidence that clearance changed with calendar time. Nonlinear elimination can also make the decline depend on the concentration reached after each dose.
The modeling task is therefore twofold: describe longitudinal concentration data with an explicit clearance model, and test whether a temporal trend explains the observations better than a plausible concentration-dependent alternative. A fitted trend is evidence about the pattern in the data; it does not establish the biological cause.
Specify the Time-Varying Model
Let \(t\) denote time since the start of a dose interval, and let \(\tau\) denote time since treatment began. Keep these clocks distinct: \(t\) determines where a sample falls within a dosing interval, while \(\tau\) indexes possible change across the treatment course. Consider a one-compartment model with known doses, constant volume \(V\), and clearance \(CL(\tau)=CL_0\exp(\beta\tau)\). Here \(CL_0\) is clearance at \(\tau=0\), and \(\beta\) is the rate of change per unit of treatment time. For example, if \(\tau\) is in weeks, \(\beta\) has units of per week.
This specification assumes that the dose history and sample times are known, that the chosen compartment structure adequately describes the data, and that \(V\) is constant over the treatment course. The exponential function ensures positive clearance. A positive \(\beta\) means clearance rises over treatment time; a negative \(\beta\) means it falls. The model does not assume that every individual has the same clearance. In a population analysis, for example, a subject-specific baseline can be represented as \(CL_{0,i}=CL_0\exp(\eta_i)\), where \(\eta_i\) describes between-subject variation.
For a single individual or a population model, concentrations are predicted by solving the model across the actual dosing history. A common observation model for positive concentrations is \(\log C_{\mathrm{obs},j}=\log C_{\mathrm{pred},j}+\epsilon_j\), with independent, mean-zero normal errors of constant variance \(\sigma^2\), conditional on any subject effects. This represents log-normal residual error. The error assumption should be checked: systematically changing residual spread or trends in residuals can indicate that it is inadequate.
What a Concentration Decline Can Identify
A simple derivation clarifies what can be learned from a dose-free interval. If there is no input, \(R_{\mathrm{in}}(t)=0\), then \(dC/dt=-CL(\tau)C/V\). For a short interval during which treatment-time clearance is effectively constant, dividing by \(C\) and integrating gives a log concentration decline. This is a useful diagnostic and approximation; fitting the full dosing history is preferable when clearance changes appreciably during the interval.
Thus the decline estimates \(CL/V\), not clearance alone. Separating \(CL\) from \(V\) requires information that identifies volume, such as sufficiently informative profiles under the specified model. If \(V\) is uncertain or also changes during treatment, an apparent change in \(CL/V\) cannot automatically be attributed to clearance.
Worked Example: Estimating a Treatment-Time Trend
Suppose a hypothetical individual has dose-free concentration measurements at 2 and 10 hours after a dose in treatment week 1: 8.00 and 4.00 mg/L. In week 4, the corresponding measurements are 8.00 and 2.80 mg/L. Assume the one-compartment approximation applies within each interval, and that independently supported volume is \(V=40.0\) L.
The elapsed time is 8 hours in both intervals. In week 1, the estimated elimination rate constant is \(\lambda_1=\log(8.00/4.00)/8=\log(2)/8=0.08664\ \mathrm{h}^{-1}\). The corresponding clearance estimate is \(CL_1=V\lambda_1=(40.0\ \mathrm{L})(0.08664\ \mathrm{h}^{-1})=3.466\ \mathrm{L/h}\).
In week 4, \(\lambda_4=\log(8.00/2.80)/8=1.0498/8=0.1312\ \mathrm{h}^{-1}\), giving \(CL_4=(40.0\ \mathrm{L})(0.1312\ \mathrm{h}^{-1})=5.248\ \mathrm{L/h}\). If week 1 is \(\tau=0\) and week 4 is \(\tau=3\) weeks, the exponential trend estimated from these two clearances is \(\beta=\log(CL_4/CL_1)/3=\log(5.248/3.466)/3=0.1383\ \mathrm{week}^{-1}\).
Under this illustrative model, clearance is higher in week 4, and the fitted value corresponds to an increase of about \(e^{3\beta}-1=51.4\%\) over the three-week interval. This calculation is not a substitute for fitting all observed concentrations: with only two interval estimates, uncertainty and model adequacy cannot be assessed reliably.
Distinguish Time Dependence From Concentration Dependence
A different mechanism is concentration-dependent elimination. For Michaelis–Menten elimination, the amount eliminated per unit time is \(V_{\max}C/(K_m+C)\), so the equivalent clearance is \(V_{\max}/(K_m+C)\). Here \(V_{\max}\) is the maximum elimination rate and \(K_m\) is the concentration at which the rate is half of \(V_{\max}\). Clearance is therefore lower at higher concentrations and rises as concentration falls. Unlike \(CL(\tau)\), this change is driven by concentration, not by time since treatment began.
Worked Example: Recognizing Concentration-Dependent Decline
Consider a hypothetical dose-free profile with \(V=40.0\) L, \(V_{\max}=32.0\) mg/h, and \(K_m=4.00\) mg/L. At \(C=8.00\) mg/L, the instantaneous log-decline rate is \(32.0/[40.0(4.00+8.00)]=0.06667\ \mathrm{h}^{-1}\). At \(C=4.00\) mg/L it is \(32.0/[40.0(4.00+4.00)]=0.1000\ \mathrm{h}^{-1}\). Thus, within the same dose-free profile, the proportional decline speeds up as concentration falls.
The time to fall from \(C_0\) to \(C_1\) follows by separating variables: \(dt=-(V/V_{\max})(K_m/C+1)dC\). For \(8.00\) to \(4.00\) mg/L, this gives \(t=(40.0/32.0)[4.00\log(8.00/4.00)+(8.00-4.00)]\) \(=1.25[2.7726+4.00]=8.466\) h. The average log-decline rate is consequently \(\log(8.00/4.00)/8.466=0.08191\ \mathrm{h}^{-1}\), between the two instantaneous rates, as expected.
A time-varying clearance model instead predicts a change tied to treatment time. With clearance held fixed over a short interval, it predicts a constant log-decline rate within that interval. Repeated profiles at comparable concentration ranges, collected at different treatment times, help assess temporal change; profiles spanning a wide concentration range within the same treatment period help assess concentration dependence.
Fit and Compare Candidate Models
Fit candidate models to the same concentration observations, dose records, and observation model. At minimum, consider constant clearance, time-varying clearance, and a concentration-dependent elimination model when the design and data support them. Estimate parameters by maximizing the likelihood, or with the corresponding population-model approach, and inspect uncertainty as well as goodness of fit. A model comparison is meaningful only when the data, likelihood, and error assumptions are held consistent.
Worked Example: Comparing Three Fits
Suppose a hypothetical analysis uses the same 48 concentration observations and log-normal residual model for three candidate fits. Assume the reported objective is \(-2\log L\), the number \(k\) counts all estimated parameters in each model, and the fits are maximum-likelihood estimates.
| Model | \(-2\log L\) | \(k\) | AIC |
|---|---|---|---|
| Constant clearance | 214.6 | 3 | 220.6 |
| Exponential time-varying clearance | 201.8 | 4 | 209.8 |
| Michaelis–Menten elimination | 205.1 | 4 | 213.1 |
Using \(\mathrm{AIC}=-2\log L+2k\), the time-varying fit gives \(201.8+2(4)=209.8\); the other two calculations are \(214.6+2(3)=220.6\) and \(205.1+2(4)=213.1\). The lower AIC favors the time-varying model among these candidates, while not proving that its mechanism is correct.
The constant-clearance model is nested in the exponential trend model when \(\beta=0\). The likelihood-ratio statistic is \(214.6-201.8=12.8\). If the regularity conditions for the usual one-degree-of-freedom chi-square approximation hold, this gives \(p\approx0.00035\). This is evidence against constant clearance in favor of the added temporal-trend parameter under the specified models. The Michaelis–Menten model is not being tested against the time-varying model by that likelihood-ratio calculation; AIC provides a relative comparison here, not a significance test.
A good comparison also includes visual and numerical diagnostics. Plot observations and predictions against both time since dose and time since treatment began. Examine residuals against predicted concentration, observed concentration, treatment time, dose, and sampling time; check whether residual spread changes and whether the model systematically misses early or late profiles. Compare individual trajectories where data allow. A lower objective alone does not rescue a model with structured residuals or implausible parameter estimates.
Assumptions, Failure Modes, and Interpretation
The time-varying model relies on correct dosing and sampling records, an adequate structural model, and enough information across treatment time to estimate \(\beta\). The exponential trend imposes a particular smooth pattern; it may be a useful approximation over the observed course without being credible over a much longer extrapolation. A change in measured concentrations can also reflect changes in bioavailability, adherence, volume, assay behavior, or other time-varying factors. If such factors are omitted, the estimated \(\beta\) can absorb their effects.
- Confusing the time scales. A trend against time since dose is not automatically a trend across the treatment course. Label both clocks in the data and model.
- Attributing all changing slopes to time. Concentration-dependent elimination can alter the decline within a single profile. Check residual patterns against concentration and consider a nonlinear elimination model where justified.
- Over-interpreting clearance. Dose-free slopes identify \(CL/V\) under the simple model. Without adequate information on \(V\), changing \(CL/V\) does not uniquely identify changing clearance.
- Reading a fitted association as a cause. A treatment-time trend describes changing disposition in the fitted model. It does not show which biological process caused that change.
Alternatives include a constant-clearance model when there is no meaningful temporal signal, a concentration-dependent model when elimination changes with concentration, and a prespecified piecewise or covariate model when the scientific hypothesis supports a different pattern. More flexible time functions can fit data better but may be poorly identified, especially with sparse sampling. The design should include informative samples across the treatment course and, where possible, comparable concentration ranges at different treatment times.
Check Your Understanding
Work through these short problems. Solutions follow each question.
- If \(CL(\tau)=CL_0\exp(\beta\tau)\), what does \(\beta<0\) mean, and what units does \(\beta\) have when \(\tau\) is measured in days?
- In a dose-free interval, concentration falls from 6.00 to 3.00 mg/L over 5.00 hours. Calculate \(CL/V\) under the constant-clearance approximation.
- At a fixed treatment time, how does Michaelis–Menten elimination affect the proportional decline as concentration falls?
- A model with time-varying clearance has lower AIC than a constant-clearance model. Does this establish the biological cause of the trend?
Solutions. (1) Clearance decreases as treatment time increases; \(\beta\) has units of \(\mathrm{day}^{-1}\), so that \(\beta\tau\) is dimensionless. (2) \(CL/V=\log(6.00/3.00)/5.00=\log(2)/5.00=0.1386\ \mathrm{h}^{-1}\). The result is a ratio, not clearance alone. (3) For Michaelis–Menten elimination, the log-decline rate is \(V_{\max}/[V(K_m+C)]\), which increases as \(C\) decreases. (4) No. Lower AIC favors that model among the candidates compared, but does not prove a mechanism or rule out omitted time-varying factors.