Reading Residual Patterns to Localize Model Failure

Why Residual Patterns Matter

A model can produce acceptable overall prediction error while systematically missing an important part of the concentration-time profile. For example, overprediction at early times and underprediction later can partly cancel when errors are summarized by one average. Residual plots preserve the ordering of observations, making such structured errors visible.

In pharmacokinetics, the location of a pattern is informative. A pattern across time after a dose may point toward a problem in the description of the concentration-time trajectory. A pattern across predicted concentration may instead suggest a scale-dependent bias or an unsuitable observation-error model. These are diagnostic clues, not unique signatures: more than one model feature can produce similar residuals.

The prediction used must be stated or understood. It may be a population prediction, which does not condition on an individual's estimated random effects, or an individual-level prediction that does. The two answer different diagnostic questions and can show different patterns. The residual sign convention here is always observed minus predicted. Standardization also depends on the observation model: it is meaningful only when \(V_i\) represents the relevant variance for the prediction being assessed.

How to Read the Axes

A time-ordered residual plot places residuals against time after dose, or against another prespecified time coordinate such as time within a dosing interval. When repeated doses or multiple occasions are present, the time coordinate should make the pharmacokinetic phase interpretable. A concentration-ordered plot places residuals against predicted concentration, not observed concentration, unless there is a specific reason to use another axis. Sorting by the model prediction helps show whether errors change with the scale the model expects.

Under a suitable model, residuals should fluctuate around zero without a persistent trend or phase-specific shape. Standardized residuals should also have a spread broadly consistent with their assumed scale. This expectation is not a claim that every point must be close to zero. Random variation, sparse sampling, assay limits, and dependence among observations can all produce irregular plots.

• Time-ordered structure: runs of residuals with the same sign, a smooth trend, or a repeatable wave suggest that the predicted trajectory may be wrong in some time region.
• Concentration-ordered structure: residual spread that grows or shrinks with predicted concentration, or a systematic change in sign, suggests scale-dependent error or bias.
• Isolated extremes: a single large residual may reflect chance, a data issue, an unusual observation, or an individual-level feature. It does not by itself establish a general structural defect.

Connecting Time Patterns to Pharmacokinetic Phases

A time plot is most useful when interpreted against the broad phases of a concentration-time profile. During absorption, concentrations rise toward a peak; around the peak, the model transitions from rising to falling; after distribution processes have shaped the profile, the terminal decline may be dominated by elimination. The exact boundaries depend on route, dosing history, and model structure.

If residuals are mostly positive during the rising phase, observations tend to exceed predictions there; if they are mostly negative, predictions tend to be too high. A mismatch concentrated around the rise or peak motivates checking the absorption-related part of the model and the time alignment of dose and samples. A mismatch in an early post-peak or intermediate region motivates checking whether the model captures the profile's changing shape. A persistent trend in the terminal region motivates checking whether the predicted late decline is appropriate. These are localization principles, not one-to-one diagnoses: a parameter change elsewhere can also alter a region's predictions.

The residual sequence should be examined both across the population and within individuals when the data permit. Opposite patterns in different individuals can disappear in a pooled average. Conversely, repeated signs over time in one person do not automatically indicate a population structural error. Residual plots identify where prediction errors occur; they do not by themselves establish why they occur.

Concentration-Ordered Patterns and the Observation Model

A concentration-ordered plot complements the time plot because the size of an observation error may depend on the concentration scale. Under a constant additive error model, the residual spread is expected to be roughly constant in concentration units. Under a proportional error model, the spread tends to increase with the predicted concentration. A combined additive and proportional model allows both components. These expectations concern variability; a systematic positive or negative trend in the residual mean is a separate sign of bias.

Raw residuals are useful for identifying errors in the original concentration units, but their spread can naturally increase with concentration under a proportional error model. Standardized residuals account for the model's specified variance and are more appropriate for checking whether residual scale changes beyond what the model predicts. Neither representation removes the need to inspect the other: standardization can hide the practical size of an error, while raw residuals can make expected heteroscedasticity look like a failure.

A Short Justification for the Diagnostic Expectation

The reason for looking for residuals centered around zero follows directly from the model's conditional mean. Suppose the specified prediction is \(E(C_{\mathrm{obs},i}\mid \mathcal{I}_i)=C_{\mathrm{pred},i}\), where \(\mathcal{I}_i\) denotes the information used for that prediction. Then:

If the model also correctly specifies the conditional variance \(V_i\), then \(E(z_i^2\mid\mathcal{I}_i)\) is approximately 1, since \(z_i=r_i/\sqrt{V_i}\) and \(E(r_i^2\mid\mathcal{I}_i)=V_i\) when the conditional mean residual is zero. This derivation explains the diagnostic target: no systematic conditional mean pattern and a scale compatible with the variance model. It does not imply that residuals are independent, normally distributed, or exactly centered in every small sample. Those properties require their own assumptions and checks.

From a Pattern to a Targeted Revision

A disciplined diagnostic sequence reduces the risk of changing several model components at once and then being unable to explain what helped. First verify the observations, dose history, units, and time coordinates. Then reproduce the pattern with clearly defined predictions and residuals. Identify the narrowest time or concentration region where it occurs, and formulate a model feature that could plausibly affect that region. Fit a focused revision and check whether it improves the localized pattern without creating new systematic errors elsewhere.

Alternatives, Failure Modes, and Limits

Residual plots and global performance summaries answer different questions. An overall error metric compresses discrepancies into a number and can compare predictive accuracy, but it usually does not reveal where errors occur. An observed-versus-predicted plot shows calibration across the range, but may conceal the ordering of errors over time. Prediction-interval coverage assesses whether observations fall within stated intervals, but adequate overall coverage does not guarantee that the model is unbiased in each phase. These approaches are complementary rather than substitutes.

Several failures can make a residual plot misleading. A wrong prediction scale or variance estimate can create apparent patterns; pooling different dose occasions can obscure phase-specific behavior; and very sparse sampling can make a smooth time pattern impossible to assess. Repeated observations from the same person are not generally independent, so ordinary runs or correlation tests that treat every residual as independent may give misleading uncertainty. Visual structure is a prompt for investigation, not a stand-alone significance test.

Most importantly, a pattern does not uniquely identify a mechanism. A positive residual run during the rising phase points to underprediction there; it does not prove a specific absorption process is wrong. A terminal trend indicates a late-profile discrepancy; it does not alone distinguish among possible structural explanations. Residual plots should be combined with design knowledge, parameter plausibility, fit diagnostics, and predictive checks. The detailed comparison of structural and random-effect explanations requires a separate analysis; a pattern alone cannot settle that distinction.