Why residuals can follow the concentration profile
A model can describe the overall concentration scale and terminal decline reasonably well yet place concentrations at the wrong times within an oral-dose interval. One clue is a sequence of residuals whose signs change in an orderly way as concentrations rise toward a peak and then fall toward a trough. This is different from isolated positive and negative deviations: it suggests that the model may be getting the timing or shape of the profile wrong.
For this diagnosis, define a concentration residual as observed concentration minus the model prediction at the recorded sampling time. A positive residual means the observation is above the prediction; a negative residual means it is below. If residuals are standardized or weighted, their scale changes, but the sign has the same interpretation. Compare residuals calculated on the same prediction basis; do not mix, for example, population predictions and individual conditional predictions in one pattern.
A useful starting model is a one-compartment oral dose with first-order absorption and elimination, no lag, and linear pharmacokinetics. For a single dose, its concentration at time \(t\) after dosing is
Here \(D\) is dose, \(F\) is bioavailability, \(V\) is apparent volume, \(k_a\) is the first-order absorption rate constant, and \(k_e\) is the first-order elimination rate constant. In this model, \(F D/V\) has concentration units and both rate constants have units of inverse time. The formula assumes instantaneous dose administration into an absorption process, constant parameters, and no lag; it is a diagnostic reference, not a claim that all oral profiles follow this structure.
The absorption rate affects the rise and peak, while elimination also shapes the declining profile. If the fitted \(k_a\) is too large relative to the true rate, the model absorbs the dose too quickly: it may predict concentrations that are too high early and place the peak too early. If fitted \(k_a\) is too small, the early prediction may be too low and the peak too late. These directions are conditional on the other structural and individual parameters being reasonably represented. Parameter compensation, dose timing errors, and sparse observations can change what is visible in residuals.
Why the peak shifts when absorption changes
The peak time follows from the maximum of the concentration curve. Differentiating the single-dose expression with respect to \(t\), setting the derivative to zero, and solving gives
The result shows why an absorption-rate error can look like a phase error: changing \(k_a\) changes the time at which the curve reaches its maximum, as well as its shape. The formula is for the stated single-dose model. With repeated dosing, the concentration is the superposition of contributions from prior doses, so the peak time within an interval is not generally given by this single-dose expression.
Nor does an incorrect absorption rate guarantee a particular number of residual sign changes. The exact residual curve depends on which parameters were fitted, the observation times, and measurement variability. The useful diagnostic is a reproducible temporal ordering that aligns with rising and falling limbs, not a rule that every early residual must have one sign and every late residual another.
Worked Example: A Slower True Absorption Rate Than the Fitted Rate
Consider a hypothetical single oral dose with \(F D/V=10\ \mathrm{mg/L}\) and \(k_e=0.10\ \mathrm{h}^{-1}\). Suppose the true absorption rate is \(k_a=0.70\ \mathrm{h}^{-1}\), but the fitted model uses \(k_a=1.40\ \mathrm{h}^{-1}\). For this illustration, compare the two model mean curves without adding measurement noise.
The true curve is
The fitted curve is
At \(t=2\ \mathrm{h}\), the true concentration is \(11.6667(e^{-0.2}-e^{-1.4})=11.6667(0.818731-0.246597)=6.675\ \mathrm{mg/L}\). The fitted prediction is \(10.7692(e^{-0.2}-e^{-2.8})=10.7692(0.818731-0.060810)=8.162\ \mathrm{mg/L}\). Thus the residual, using true minus fitted concentration, is \(6.675-8.162=-1.487\ \mathrm{mg/L}\): the faster fitted absorption overpredicts this early concentration.
At \(t=5\ \mathrm{h}\), the true concentration is \(11.6667(e^{-0.5}-e^{-3.5})=11.6667(0.606531-0.030197)=6.724\ \mathrm{mg/L}\). The fitted prediction is \(10.7692(e^{-0.5}-e^{-7})=10.7692(0.606531-0.000912)=6.522\ \mathrm{mg/L}\). The residual is \(6.724-6.522=+0.202\ \mathrm{mg/L}\), now in the opposite direction.
At \(t=10\ \mathrm{h}\), the true concentration is \(11.6667(e^{-1}-e^{-7})=4.281\ \mathrm{mg/L}\). The fitted prediction is \(10.7692308(e^{-1}-e^{-14})=3.962\ \mathrm{mg/L}\). The residual is approximately \(+0.31952\ \mathrm{mg/L}\), which rounds to \(+0.320\ \mathrm{mg/L}\) using the unrounded model values. The fitted rate predicts an earlier peak: the single-dose peak times are \(\ln(7)/0.60=3.243\ \mathrm{h}\) for the true curve and \(\ln(14)/1.30=2.030\ \mathrm{h}\) for the fitted curve. This hypothetical comparison illustrates an early negative residual followed by positive residuals; it does not establish that absorption rate is the cause in real data.
Reading a recurring pattern across oral-dose intervals
In repeated-dose data, inspect residuals against time since the most recent dose as well as against absolute study time. A wave that recurs at similar positions within successive intervals is more consistent with a repeatable profile-shape mismatch than a sequence tied only to one visit or one clock time. For a model with absorption that is too fast, a common pattern is overprediction during the observed rising phase, followed by underprediction later in the interval. If absorption is too slow, the signs may reverse. The relevant evidence is the observed profile and its sampling times, not merely the nominal label “peak” or “trough.”
A trough is a concentration near the end of a dosing interval, not necessarily a point where the derivative of concentration is zero. Likewise, a sample labelled “peak” may not have been collected at the actual maximum. Use actual dose and sample times where available. When a model’s residuals are plotted by time after dose, assess whether the pattern repeats across individuals or occasions, and whether the same individuals have enough observations on both the rising and declining portions to make that comparison informative.
Worked Example: Residual Signs Across Two Hypothetical Intervals
A hypothetical patient receives an oral dose every 12 hours. The table gives observed concentrations and fitted predictions in two intervals. The values are illustrative; the task is to calculate residuals and assess whether the signs follow a repeated temporal pattern.
| Interval | Time after dose (h) | Profile position | Observed (mg/L) | Predicted (mg/L) |
|---|---|---|---|---|
| 1 | 1 | Rising | 4.8 | 5.6 |
| 1 | 3 | Near peak | 7.1 | 7.5 |
| 1 | 6 | Falling | 5.4 | 5.0 |
| 1 | 11 | Near trough | 2.4 | 2.1 |
| 2 | 1 | Rising | 4.9 | 5.7 |
| 2 | 3 | Near peak | 7.0 | 7.4 |
| 2 | 6 | Falling | 5.3 | 5.0 |
| 2 | 11 | Near trough | 2.3 | 2.0 |
For interval 1, the residuals are \(4.8-5.6=-0.8\), \(7.1-7.5=-0.4\), \(5.4-5.0=+0.4\), and \(2.4-2.1=+0.3\ \mathrm{mg/L}\). For interval 2, they are \(4.9-5.7=-0.8\), \(7.0-7.4=-0.4\), \(5.3-5.0=+0.3\), and \(2.3-2.0=+0.3\ \mathrm{mg/L}\). In both intervals, the fitted model is high early and low later, a repeated sign sequence compatible with absorption that is too fast or another phase mismatch.
The repeated pattern strengthens the case for checking the absorption-time description, but it does not identify the cause uniquely. These eight observations alone cannot rule out systematically early recorded sample times, a shared dose-time error, an incorrect lag representation, or other structural discrepancies. Nor do they quantify uncertainty; replicate residuals from one hypothetical patient are not independent evidence of a population-wide effect.
Distinguishing absorption rate from timing and sampling
Several explanations can produce residuals that appear to move with the concentration profile. They imply different checks and should not be collapsed into a single conclusion.
| Candidate explanation | Expected clue | Useful check |
|---|---|---|
| Incorrect absorption rate | Profile shape or peak timing is systematically wrong; signs may recur by time after dose. | Check the model’s rising and falling limbs against actual times, and assess whether the pattern recurs across intervals or individuals. |
| Lag or delayed input | Input appears shifted in time relative to the recorded dose. | Verify dose administration and sample timestamps first; consider whether a time shift is plausible. A shift and a rate change are not interchangeable in every part of the curve. |
| Sampling-time error or sparse sampling | Some samples may be displaced in time, or the actual peak may lie between widely spaced samples. | Check source records and actual collection times; examine whether sampling covers both sides of the peak. |
| Other model or data problem | Residuals may also vary with concentration, occasion, or individual rather than recurring by time after dose. | Review prediction basis, data handling, and other residual displays before attributing the pattern to absorption. |
A timing error can matter most on a steep part of a curve. For a small timing displacement \(\delta t\), a first-order approximation is \(C(t+\delta t)-C(t)\approx C'(t)\delta t\). Near a peak, \(C'(t)\) is close to zero, so a small timing shift has little first-order effect on concentration there; on a steep rising limb, the same shift can have a larger effect. This is a local approximation, not a way to infer a timestamp correction without checking records.
Worked Example: Checking Whether a Timing Shift Is Plausible
Use the fitted curve from the first example, \(C_{\mathrm{fit}}(t)=10.7692(e^{-0.10t}-e^{-1.40t})\ \mathrm{mg/L}\). At \(t=1\ \mathrm{h}\), its slope is
For a small shift of \(\delta t=-0.25\ \mathrm{h}\), the local predicted concentration change is approximately \(2.74348(-0.25)=-0.68587\ \mathrm{mg/L}\), or \(-0.686\ \mathrm{mg/L}\) to three decimal places. Thus, an observation collected 15 minutes earlier than its recorded time could be lower than the prediction by roughly \(0.686\ \mathrm{mg/L}\) on this steep rising limb. This is only a first-order approximation.
At \(t=2\ \mathrm{h}\), the same model’s slope is \(10.7692(-0.10e^{-0.2}+1.40e^{-2.8})=10.7692(-0.081873+0.085134)=0.0351\ \mathrm{mg/L/h}\). A 15-minute shift would change the local prediction by only about \(0.0351(-0.25)=-0.0088\ \mathrm{mg/L}\). The contrast shows why timing errors can have strongly time-dependent effects. It does not show that timing error explains a recurring pattern; records and the full profile still need examination.
Diagnostics, limitations, and a common mistake
Before modifying a model, plot residuals against time since dose, with dose intervals aligned where appropriate, and inspect observed and predicted profiles on the same time scale. Look for recurrence across intervals and individuals, not just one striking sequence. Check whether the pattern persists under the same prediction basis and whether it is concentrated in the rising or falling phase. Review actual dose and sample timestamps, dose history, and the placement of samples relative to the expected peak and trough. Sparse data may make the peak time poorly determined; a few observations near trough cannot establish the shape of the absorption phase.
A residual wave is a diagnostic clue, not a formal test or a unique mechanism. Individual variability and measurement error can obscure or mimic it. Also, changing \(k_a\) may trade off against other parameters, so an apparent improvement in one profile region can worsen another. Compare candidate explanations using the whole data pattern and appropriate model evaluation, rather than selecting a mechanism from a single residual plot. This tutorial uses a simple first-order absorption curve to explain the direction of a possible mismatch; it does not establish more complex absorption structures.
Common mistake: “Residuals are negative on the rising limb and positive later, so absorption rate is definitely too high.” The sign pattern is compatible with a fitted absorption rate that is too high under the simple model, but it is not conclusive. Incorrect sample times, a shifted input, sparse sampling, or other model misspecification can produce related patterns. The correction is to treat the pattern as a hypothesis, check data timing and sampling support, and see whether the shape discrepancy recurs in the appropriate time coordinates.
A second mistake is to call any point near the highest observed concentration “the peak” and infer a precise peak-time error from it. The observed maximum depends on when samples were collected. With sparse sampling, the true maximum may occur between samples, and the highest recorded point need not estimate its time accurately. Report what the sampling schedule supports, not more precision than the data contain.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If residuals are observed minus predicted, what does a negative residual mean?
- In the stated single-dose model, what peak-time expression follows by setting the concentration derivative to zero?
- A fitted model overpredicts concentrations on the rising limb and underpredicts them later in two successive intervals. Give one compatible absorption-rate explanation and one alternative that should be checked.
- Why can a modest timing error have a larger concentration effect on the rising limb than near the peak?
- Does a recurring peak–trough residual pattern uniquely prove absorption-rate misspecification? Explain briefly.
Solutions. (1) The observed concentration is below the prediction at that sample time. (2) \(t_{\max}=\ln(k_a/k_e)/(k_a-k_e)\), for the stated single-dose model with \(k_a\ne k_e\). (3) A fitted \(k_a\) that is too large is compatible with that pattern; actual sample-time or dose-time errors are alternatives to check. (4) The local concentration change from a small shift is approximately \(C'(t)\delta t\); the slope is usually steeper on the rising limb and near zero at the peak. (5) No. The pattern is suggestive but can also reflect timing errors, lag, sparse sampling, or other misspecification, so it requires corroborating checks.