Residual Diagnostics for Longitudinal QSP Predictions

Why Examine Residuals Across Time?

A model prediction is not assessed only by whether its values look close to observed data overall. In a longitudinal prediction, the sequence and structure of deviations matter: errors may be small on average yet consistently too high early and too low late, or they may drift together over successive visits. Such patterns can signal that the model is missing a relevant time-course feature, even when a single summary of overall error appears acceptable.

Residual diagnostics are descriptive checks on predictions and their deviations, not a stand-alone proof that a model is valid or invalid. The intended-use claim and the prediction conditions must be clear, and the validation data must be kept distinct from data used to tune the model. Here, the focus is how to read residual patterns across time and dose while allowing for ordinary measurement variability.

For a QSP biomarker measured in concentration units, a residual of \(2\ \mathrm{ng/mL}\) means the observed concentration was \(2\ \mathrm{ng/mL}\) above the prediction. This sign convention should be kept consistent throughout an analysis. A residual plot puts residual values on one axis and a relevant feature—often time, predicted value, dose, or subject—on the other.

What Patterns to Look For

Start with a plot of residuals against time, using a horizontal reference at zero. If predictions are well centered at each time in the target population, residuals should generally lie on both sides of zero without a persistent time-dependent shift. This is not a demand that every residual be small or that the observed sequence alternate randomly. Rather, it is a check for structure that is too systematic to be explained comfortably as ordinary variation.

• Bias over time: residuals predominantly positive or negative in a time region suggest underprediction or overprediction there. A change from positive early residuals to negative late residuals can indicate that the predicted time course is too slow or too persistent, though the precise mechanism is not identified by the plot alone.
• Temporal dependence: neighboring residuals that tend to share a sign or move together show serial patterning. A run of positive residuals may be more informative than the same number scattered across time. Repeated observations from one subject are not independent replicates.
• Dose dependence: residuals that shift or spread systematically across dose groups suggest that prediction performance differs with dose. Plotting residuals against dose, and separately by time where useful, helps avoid mistaking a time imbalance across dose groups for a dose effect.
• Changing variance: a residual cloud that widens or narrows with time, predicted value, or dose may indicate nonconstant variability. A trend in spread is distinct from a trend in the residual mean.

A plot should preserve the design information. Connecting residuals within a subject can make time ordering visible; different symbols or panels can identify dose groups. When several subjects contribute measurements, show the individual sequences as well as a pooled view where possible. A pooled mean near zero can conceal a positive bias for one dose and a negative bias for another.

Worked Examples

Why the Patterns Matter: A Short Justification

The diagnostic logic follows from the role of residuals. If predictions are appropriately centered for the prediction target and the observation process, then the expected residual at a given time and dose is zero. Write this conditional statement as \(E(r_{ij}\mid t_{ij},d_i)=0\), where \(t_{ij}\) is time and \(d_i\) is dose. If instead residuals tend to be positive in an early time window, their conditional mean there is positive, which is the definition of local underprediction under the observed-minus-predicted convention.

Likewise, a change in conditional spread, such as a larger \(Var(r_{ij}\mid t_{ij},d_i)\) at later times, indicates that residual variability is not constant across those conditions. This reasoning justifies plotting both residual location and spread. It does not imply that a finite sample will have exactly zero average residual in every subgroup; sampling variation is expected, especially with sparse measurements.

Measurement Noise, Prediction Uncertainty, and Diagnostics

Observed residuals combine several sources of discrepancy. Assay or measurement noise can make observed values depart from a model prediction even when the model captures the relevant systematic time course. Predictions may also be uncertain because of parameter uncertainty or biological variation across subjects. A residual plot alone cannot separate these sources. Its useful role is to show whether deviations look compatible with the stated observation and prediction variability, or whether they form a reproducible pattern that deserves further assessment.

If a defensible predictive standard deviation \(s_{ij,\mathrm{pred}}\) is available, a standardized residual may be formed as \(z_{ij}=r_{ij}/s_{ij,\mathrm{pred}}\). This puts deviations on a scale relative to the expected predictive spread. It is informative only if the denominator represents the relevant uncertainty for that prediction and is not near zero. If it accounts for measurement error alone while ignoring between-subject variation, for example, it can make ordinary subject differences look unexpectedly large. Standardization does not repair a misspecified uncertainty model.

Useful checks include verifying units, time alignment, dose assignment, below-quantification handling, and whether the prediction is for an individual or a population summary. Plot residuals against time and dose, inspect within-subject ordering, and compare the observed spread with the expected measurement and predictive variation. A smooth trend, repeated same-sign runs, or a widening fan is more compelling when it recurs across subjects or is larger than the stated uncertainty would plausibly generate.

Common Pitfalls and Alternatives

• Relying on a pooled mean: an average near zero can result from opposing time or dose patterns. Inspect residuals by relevant condition, not only in aggregate.
• Calling every nonzero residual a model failure: measurement noise and subject-to-subject variability produce nonzero deviations. Consider their expected magnitude and pattern.
• Calling a run of residuals independent evidence: repeated measures from a subject share context and may be correlated. Preserve subject identity and time ordering in plots and summaries.
• Pooling dose groups with different time schedules: an apparent dose trend may instead reflect which times were measured in each group. Compare like times or show the joint time-dose structure.
• Interpreting changing spread as bias: a wider cloud does not, by itself, mean predictions are systematically high or low. Assess the center and spread separately.
• Using an arbitrary standardized residual: dividing by an uncertainty that omits important variability can exaggerate apparent failures. State what the scale includes.

Residual plots complement, rather than replace, other approaches. A summary error measure can compare overall magnitude but may conceal ordering and direction. A time-course plot of observed values and predictions gives direct scale context but can make small systematic offsets hard to see. Residual plots foreground deviations; uncertainty intervals or predictive distributions add information about how surprising those deviations are under an explicit uncertainty model. No single display answers every question, and diagnostic patterns should guide targeted checks rather than substitute for them.

The main limitation is that a residual pattern is not uniquely diagnostic. A trend can arise from prediction error, data processing, timing mismatch, an inappropriate observation scale, or an unrepresented source of variability. Sparse sampling and small numbers of subjects also make apparent patterns unstable. Describe what the plot shows first, then state what it motivates checking; reserve causal or mechanistic explanations for evidence that can support them.