Why Look at Residuals Across Time After a Dose?
A model can have residuals that are centered near zero overall and still miss an important part of the concentration-time profile. For example, it may predict the later decline reasonably well but place the peak too late, or match early concentrations while predicting a terminal decline that is too shallow. Summarizing residuals without their post-dose timing can conceal these phase-specific discrepancies.
An IWRES or CWRES plotted against time after dose adds that ordering. The main question is whether residuals remain scattered around zero throughout the profile, or instead show a sustained sign, drift, or change in pattern in a particular phase. The plot can point to where model predictions are systematically high or low. It does not, by itself, prove which parameter or biological mechanism caused the mismatch.
For a simple proportional-error illustration, an IWRES may be written as the observed-minus-individual-prediction difference divided by the residual standard deviation at that prediction. In a proportional-error model with proportional error coefficient \(\sigma\), that standard deviation is \(\sigma\) times the prediction. This simple expression is useful for hand calculation; it is not a universal substitute for the residual definitions used by a particular estimation program.
What a Time Trend Means
Suppose the model predicts concentration \(C_{\text{model}}(t)\) at post-dose time \(t\), while the expected observed concentration at that time is \(C_{\text{true}}(t)\). Let the residual standard deviation at that time be \(s(t)\). If the observation error has mean zero conditional on time, then the expected standardized residual is the standardized difference between the true mean profile and the model profile:
This follows by substituting \(C_{\text{obs}}(t)=C_{\text{true}}(t)+\varepsilon(t)\), where \(E[\varepsilon(t)\mid t]=0\), and taking the conditional expectation. If the model describes the mean profile and the error scale appropriately, the expected residual is zero at each time. If the model’s mean profile is systematically too low or too high at some times, the expected residual there is positive or negative, respectively. Standardization changes the scale of the discrepancy; it does not make a persistent pattern disappear.
In practice, residuals are noisy. A few positive values immediately after a dose do not establish early-phase bias, and repeated measurements from one person are not independent pieces of evidence. Look for a pattern across observations and subjects, and check whether it is consistent with the concentration-time data and study design. A smooth curve or grouping by post-dose phase can help the eye, but the underlying observations and their distribution should remain visible.
Reading Absorption-, Distribution-, and Elimination-Phase Patterns
The interpretation depends on the route, formulation, sampling times, and model structure. These phase labels describe parts of the concentration-time profile, not diagnoses that can be read mechanically from the residual sign.
| Profile region | Possible residual pattern | One interpretation to investigate |
|---|---|---|
| Early rising absorption period | Predominantly positive residuals | The model may predict concentrations that rise too slowly or are too low early. |
| Near the peak or on the early decline | Early positive values followed by negative values | The predicted peak may be delayed, too broad, or otherwise misshapen relative to observations. |
| Rapid early distribution after an intravenous dose | Positive values very early followed by negative values during the rapid fall | A model may not capture the rapid distribution component adequately. |
| Late terminal decline | Residuals become increasingly negative at later times | The predicted terminal decline may be too shallow, subject to adequate sampling and an appropriate error model. |
The first two patterns can arise from absorption-rate or lag-time misspecification, but may also reflect dose timing errors, formulation differences, or sampling problems. A distribution-like pattern is most interpretable when early post-dose samples are available and the administration route and model structure support that interpretation. A late trend is informative about the terminal profile only if the observations actually represent the post-distribution elimination phase. In all cases, residual patterns nominate explanations to investigate; they do not identify a unique parameter change.
Worked Example: Early Positive Residuals and a Possible Absorption Mismatch
Consider a hypothetical oral-dose profile. The model uses a proportional residual standard deviation of \(20\%\) of the individual prediction, so \(s(t)=0.20\times\text{IPRED}\). Calculate IWRES as \((\text{DV}-\text{IPRED})/s(t)\), where DV is the observed concentration.
| Time after dose (h) | DV (mg/L) | IPRED (mg/L) | IWRES calculation |
|---|---|---|---|
| 0.5 | 3.0 | 2.0 | \((3.0-2.0)/(0.20\times2.0)=2.50\) |
| 1 | 6.0 | 5.0 | \((6.0-5.0)/(0.20\times5.0)=1.00\) |
| 2 | 7.0 | 8.0 | \((7.0-8.0)/(0.20\times8.0)=-0.625\) |
| 4 | 4.5 | 5.0 | \((4.5-5.0)/(0.20\times5.0)=-0.500\) |
The first two residuals are positive and the later two are negative. For instance, at 0.5 h the residual standard deviation is \(0.20\times2.0=0.40\) mg/L, and the observed concentration exceeds prediction by \(1.0\) mg/L, giving \(1.0/0.40=2.50\). The early observations are higher than predicted, while the later observations are lower. This pattern is consistent with a predicted rise or peak that is too delayed or broad, so absorption timing is one reasonable feature to investigate. Four values do not establish that explanation; checking dose and sample timing, other subjects, and the full concentration-time profile is essential.
Worked Example: A Possible Distribution-Phase Miss
Suppose a hypothetical intravenous bolus dataset has CWRES reported by the model-fitting software. The values below are illustrative software outputs, not calculations from a simple proportional-error formula.
| Time after dose (h) | CWRES | Profile context |
|---|---|---|
| 0.05 | 2.1 | Very early concentration |
| 0.25 | 1.6 | Rapid early decline |
| 1 | -1.0 | Later in the early decline |
| 4 | 0.2 | Later profile |
The early positive values indicate that observed concentrations tend to exceed model predictions in the earliest part of this hypothetical profile. The negative value at 1 h indicates an observation below prediction. Together, this sign change may be consistent with a model that misses a rapid distribution component: it underpredicts very early concentrations and then predicts too much as the observed concentration falls. The near-zero value at 4 h does not undo the earlier discrepancy; it indicates only that the residual at that observation is close to zero.
This interpretation depends on the early samples being correctly timed and on the rapid decline actually representing distribution rather than another process. A one-compartment model is one possible source of this pattern, but the residual plot alone cannot establish that a second compartment is needed. The sampling schedule, model predictions, and alternative explanations should be reviewed before changing model structure.
Worked Example: A Late Trend Consistent With a Shallow Predicted Decline
In another hypothetical profile, the observations are in the terminal decline. Assume a proportional residual standard deviation of \(25\%\) of IPRED, so \(s(t)=0.25\times\text{IPRED}\). Use the same simple IWRES calculation.
| Time after dose (h) | DV (mg/L) | IPRED (mg/L) | IWRES calculation |
|---|---|---|---|
| 4 | 4.00 | 4.00 | \((4.00-4.00)/(0.25\times4.00)=0.00\) |
| 8 | 1.70 | 2.00 | \((1.70-2.00)/(0.25\times2.00)=-0.600\) |
| 12 | 0.65 | 1.00 | \((0.65-1.00)/(0.25\times1.00)=-1.40\) |
| 24 | 0.12 | 0.25 | \((0.12-0.25)/(0.25\times0.25)=-2.08\) |
At 12 h, the residual standard deviation is \(0.25\) mg/L, and the observed-minus-predicted difference is \(-0.35\) mg/L, so IWRES is \(-0.35/0.25=-1.40\). At 24 h the difference is \(-0.13\) mg/L and the residual standard deviation is \(0.0625\) mg/L, giving \(-0.13/0.0625=-2.08\). The increasingly negative residuals suggest that the model is increasingly high relative to observations late in the profile. If these samples are genuinely in the terminal elimination phase, a predicted decline that is too shallow is one possibility. The conclusion would be less secure if absorption or distribution were still contributing, if late samples were near quantification limits, or if the error model were unsuitable.
Comparing IWRES and CWRES for This Diagnostic
Both residual types can be examined against post-dose time, but they are not interchangeable. IWRES use individual predictions, which incorporate individual parameter estimates. They can therefore be influenced by how well individual parameters are estimated, especially when an individual has few informative observations. CWRES are designed to account for the conditional residual structure under the population model. Their approximate standard-normal behavior depends on the model and conditions discussed in the earlier material on conditional weighted residuals; it is not a guarantee that every time-specific pattern must vanish.
A useful diagnostic comparison is whether a visible time pattern appears in both views or only in one. A pattern in both can strengthen concern about phase-specific mismatch. A pattern mainly in IWRES may be affected by individual fitting or shrinkage, while a pattern mainly in CWRES may reflect the conditional residual construction or the population-level fit. Neither outcome is a mechanical decision rule. Review the predicted profiles and residual calculation, and do not choose whichever plot looks more reassuring.
The alternatives answer different questions. A residual histogram summarizes the overall distribution but discards time ordering. A plot against time after dose preserves the sequence within the profile and can reveal phase-specific structure. A concentration-time plot displays observations and predictions in their original units and helps connect residual signs to profile shape. These views complement one another; none alone proves a mechanism or establishes that a model is adequate for every intended use.
Diagnostics, Failure Modes, and Practical Interpretation
Before assigning a phase interpretation, check the conditions needed for the plot to be meaningful. Dose and observation times must be recorded correctly; the post-dose times must correspond to the dose that generated the modeled profile; the residual definition and error model must be understood; and the observations must be compared on a consistent concentration scale. A trend can be created or obscured by sparse sampling, a few influential observations, assay limits, or a mismatch between the assumed and actual residual variability.
Repeated samples from the same individual are related through shared parameters and biology. A run of residuals from one subject should not automatically be treated as many independent confirmations of a population-wide problem. Conversely, averaging across subjects can hide opposing subject-level patterns. Interpret the display alongside the number and timing of observations, and ask whether the same phase-specific feature recurs across individuals.
A practical reading sequence is:
Confirm that positive values mean observations are above predictions and negative values mean they are below.
Identify whether the trend occurs on the rising limb, around the peak, during rapid early decline, or in the late profile.
Review timing, sampling density, individual contributions, quantification limits, and whether the residual scaling is suitable.
Describe where predictions appear high or low, then name plausible explanations as hypotheses to investigate rather than established causes.
Common mistake: “The late residuals are negative, so clearance must be too low.” This conclusion is too specific. Negative late residuals mean observed concentrations are below predictions on the residual scale. If the observations are in the terminal elimination phase, a predicted decline that is too shallow is one possibility, but the pattern does not uniquely identify clearance. Incorrect timing, residual variability, ongoing absorption or distribution, and other model features may also contribute. Diagnose the mismatch first; evaluate its cause using the model and data.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- With the usual sign convention, what does a positive IWRES say about the observed concentration relative to the individual prediction?
- A hypothetical oral profile has positive residuals on the early rising limb and negative residuals near and after the peak. Name one plausible model feature to investigate and one reason the pattern is not proof.
- In a proportional-error model, \(\sigma=0.20\), IPRED is \(5.0\) mg/L, and DV is \(4.0\) mg/L. Calculate the simple IWRES.
- Late residuals become increasingly negative. What does this directly indicate, and what additional condition is needed before interpreting it as a terminal elimination mismatch?
Solutions. (1) The observed concentration is above the individual prediction. (2) Absorption timing or peak shape is one feature to investigate; timing errors, sampling variability, or other model misspecification can produce a similar pattern, so the residuals do not uniquely identify a cause. (3) The residual standard deviation is \(0.20\times5.0=1.0\) mg/L, so IWRES is \((4.0-5.0)/1.0=-1.00\). The observed concentration is one residual standard deviation below prediction. (4) Observations are increasingly below predictions on the residual scale. To interpret this as a terminal elimination mismatch, the samples must represent the terminal phase rather than ongoing absorption or distribution, and other data or model problems must be considered.