Why Follow Residuals Across the Dosing Interval?
A goodness-of-fit display can look acceptable overall while revealing a consistent error at particular times after a dose. For example, a model may underpredict concentrations early in the interval and overpredict them later. Summarizing all observations together can hide that pattern: positive and negative discrepancies may cancel. Examining residuals against actual time since dose helps show whether the model’s errors have a structure that the current model does not capture.
The diagnostic task is not simply to decide whether one point is “bad.” It is to compare the pattern across observations and subjects, inspect both population and individual predictions, and check whether recorded data could explain the pattern. A coherent time trend that remains under both kinds of prediction can support investigating model revision. It does not, by itself, identify the missing model feature or prove that the data are correct.
In practice, software may report residuals with a model-specific variance adjustment rather than this simple standardization. The exact residual definition should therefore be checked before interpreting a diagnostic plot. The examples here deliberately use a simple additive error model: conditional on the prediction, the observation has standard deviation \(0.10\) mg/L at every displayed time. This makes the arithmetic transparent; it is not a claim that additive error is appropriate for every concentration dataset.
Under an adequate conditional mean model and a correctly specified residual distribution, standardized residuals should generally be centered near zero across relevant times, without a persistent time-ordered pattern. This is a diagnostic expectation, not a requirement that every residual be small or that every sequence alternate signs. Measurements from the same subject are repeated observations and may be dependent, so a visual pattern should not automatically be treated as a collection of independent statistical tests.
Worked Diagnosis Across One Dosing Interval
The following hypothetical dataset summarizes six subjects with observations distributed across a 12-hour dosing interval. The entries are group means at each time. Assume the same \(0.10\) mg/L predictive standard deviation for each observation. In this balanced illustration, the displayed standardized residuals are the differences between the group means divided by that common standard deviation.
Worked Example: Finding Drift in Population Residuals
For each time, calculate the population standardized residual using \(R^{P}=(DV-PRED)/0.10\), then assess its direction across the interval.
| Time after dose (h) | Mean observed (mg/L) | Mean PRED (mg/L) | Population residual |
|---|---|---|---|
| 0.5 | 8.20 | 7.80 | +4.00 |
| 1 | 7.00 | 6.80 | +2.00 |
| 2 | 5.20 | 5.10 | +1.00 |
| 4 | 3.00 | 3.10 | −1.00 |
| 8 | 1.40 | 1.50 | −1.00 |
| 12 | 0.80 | 0.90 | −1.00 |
At 0.5 h, the calculation is \((8.20-7.80)/0.10=+4.00\). At 4 h, it is \((3.00-3.10)/0.10=-1.00\). The other rows follow the same formula. Thus the early observations tend to lie above the population predictions, while the later observations lie below them. In concentration terms, the model underpredicts early and overpredicts later.
The first three residuals have mean \((4+2+1)/3=2.33\); the final three have mean \((-1-1-1)/3=-1.00\). Their difference is \(2.33-(-1.00)=3.33\) standardized residual units. This contrast describes the displayed pattern; it is not a formal significance test. The important diagnostic feature is the ordered change in sign across time, not a claim that a particular cutoff proves misspecification.
Compare Population and Individual Predictions
Population predictions use typical model parameters, whereas individual predictions incorporate subject-specific estimates. The two views answer different diagnostic questions. A population pattern can reveal a systematic mismatch in the typical time course. Individual predictions can show whether subject-specific parameter estimates account for some of that mismatch. Because individual estimates are informed by the same observations being evaluated, better individual fit is expected and is not, on its own, evidence that the population model is adequate.
Worked Example: Does Individual Fit Remove the Drift?
Suppose the same hypothetical group summaries have the following individual predictions. Use the same \(0.10\) mg/L standard deviation to calculate the standardized residual relative to the mean individual prediction.
| Time after dose (h) | Mean observed (mg/L) | Mean IPRED (mg/L) | Individual residual |
|---|---|---|---|
| 0.5 | 8.20 | 8.00 | +2.00 |
| 1 | 7.00 | 6.90 | +1.00 |
| 2 | 5.20 | 5.15 | +0.50 |
| 4 | 3.00 | 3.05 | −0.50 |
| 8 | 1.40 | 1.45 | −0.50 |
| 12 | 0.80 | 0.85 | −0.50 |
For example, at 0.5 h, \((8.20-8.00)/0.10=+2.00\). At 12 h, \((0.80-0.85)/0.10=-0.50\). The individual residual magnitudes are smaller than the population residual magnitudes, but their signs still shift from positive early to negative later.
This comparison supports a careful conclusion: subject-specific predictions reduce the discrepancy, yet the same time-ordered pattern remains in these summaries. That persistence makes it reasonable to investigate whether the population model’s time course is inadequate. It does not establish the cause. Nor should a good individual fit be taken as decisive if individual estimates are weakly informed or residual shrinkage is substantial; the earlier material on shrinkage explains why individual residual diagnostics require that qualification.
From a Pattern to a Defensible Next Step
A residual trend is evidence to investigate, not a model-change instruction. First verify that the horizontal axis represents the actual elapsed time since the relevant dose. In repeated-dose data, confirm dose times, sample times, dose amounts, missed or delayed doses, and whether the model and records use the same time origin. Check units, transcription, assay flags, and any data-processing rules that affect concentrations or timestamps. A mismatch in time units or a systematically shifted sampling-time record could create an apparent trend without requiring a new structural model.
Next ask whether the pattern is coherent across subjects and whether it is visible in both population and individual prediction views. A few isolated points with large residuals call for careful record-level review. A similar early-to-late sign change across many subjects, with no evident recording or assay problem, is more consistent with a systematic model discrepancy than with a handful of incorrect observations. Still, shared procedural errors can affect many records, so consistency alone cannot certify data quality.
Check the residual definition, units, prediction type, and whether time is elapsed time since the actual dose rather than a nominal visit label.
Review dose and sample timestamps, units, duplicate or excluded records, assay flags, and data-processing steps. Look for specific evidence of an error rather than changing observations to make the plot look better.
Ask whether the time trend is present for population predictions, whether it remains for individual predictions, and whether improvement in individual fit could simply reflect estimating subject-specific parameters from the same data.
If a record-level problem is documented, correct it using the defined data procedures and reassess. If records are credible and the pattern is systematic, investigate model structure or the residual model. Keep the diagnosis provisional until the revised analysis is checked.
Worked Example: Model Revision or Data Correction?
Suppose a review of a hypothetical study finds the same early-positive, late-negative residual pattern in several subjects. Dose and sample timestamps agree with source records, concentration units are consistent, and the pattern appears under both population and individual predictions. No specific assay or transcription flag is linked to the affected observations.
There is no identified data error to correct. The pattern’s repetition across subjects, its time ordering, and its persistence under individual predictions support investigating model revision rather than deleting or altering observations. Possible questions for that investigation include whether the model describes the relevant concentration–time shape adequately and whether its error assumptions are suitable. These observations alone do not distinguish among possible model changes or prove that any particular mechanism is missing.
The conclusion would differ if, for example, the early samples were all recorded using a time field shown to be offset from source documentation. That is concrete evidence for a data correction. The correction should follow the data-management process, be documented, and be followed by a fresh diagnostic assessment. A correction is not justified merely because changing a timestamp or value would reduce residual drift.
Interpretation, Alternatives, and Failure Modes
The key distinction is between an observation that has an unusual residual and a pattern whose direction is associated with time after dose. An isolated large residual may reflect ordinary variability, a data issue, or an influential observation. A smooth or repeated sign change across the dosing interval suggests that the expected concentration profile may not be matching the observed profile. A residual variance problem can change how large standardized residuals appear, but it does not by itself explain a persistent signed trend in observed-minus-predicted values.
The main alternatives are therefore not simply “model” versus “data.” A time-axis or unit problem can masquerade as model misfit; a residual error specification can distort standardized residuals; and a structural or covariate feature may be inadequate even when the records are correct. Use the raw concentration and prediction displays alongside standardized residuals: standardization helps compare observations on the model’s error scale, while raw differences retain concentration units and may expose scale-related issues. Neither display alone identifies the cause.
- Do not infer a cause from the sign pattern alone. The trend motivates targeted checks; it does not uniquely indicate absorption, distribution, clearance, or any other particular process.
- Do not treat repeated observations as independent evidence by default. Observations within one subject can be related. A descriptive contrast is not automatically a valid inferential test.
- Do not prefer individual fit without qualification. Individual predictions use subject-specific estimates informed by the observations, so they can look better even when population-level structure remains questionable.
- Do not “clean” data to improve fit. Correct a record only when a documented data-quality finding supports that correction, not because the model predicts it differently.
The diagnostic conclusion is strongest when the residual definition is understood, time is aligned to actual dosing, the data have been checked, and the pattern is coherent across subjects and prediction types. Even then, the evidence supports a next investigation rather than a final causal explanation. The purpose of the diagnosis is to decide whether the model or the records warrant closer review, and to make that decision traceable.
Check Your Understanding
Answer each question before reading the solutions.
- An observation is \(5.6\) mg/L, its population prediction is \(5.8\) mg/L, and its predictive standard deviation is \(0.20\) mg/L. Calculate its standardized population residual and interpret its sign.
- In a 12-hour interval, early population residuals are mostly positive and late residuals mostly negative. What does this indicate, and what does it not establish?
- Individual predictions reduce residual magnitudes, but the early-positive and late-negative pattern remains. Give an appropriate diagnostic conclusion.
- Several observations show drift, and their timestamps are all one hour later than source records. What should be investigated before proposing model revision?
Solutions. (1) \(R^{P}=(5.6-5.8)/0.20=-1.00\). The model overpredicts this observation by one predictive standard deviation. (2) The ordered sign change indicates a time-dependent discrepancy worth investigating; it does not identify the cause or prove the records are correct. (3) Individual predictions improve the fit in magnitude, but persistence of the time pattern supports investigating model adequacy, with appropriate caution about individual-estimate information and shrinkage. (4) Compare the analysis timestamps with the source records and established time conventions. If the offset is confirmed, correct it through documented data procedures and reassess the diagnostics before attributing the pattern to the model.