Residual Patterns For Detecting Trajectory Misspecification

Why Examine Residual Patterns?

A longitudinal model can produce plausible average trajectories while still getting their shape wrong. A fitted straight line, for example, may systematically underpredict early and late biomarker values and overpredict values in the middle. A single summary of residual size may not reveal this problem; plotting residuals against follow-up time or the fitted biomarker value can expose the pattern.

These plots are diagnostic evidence, not automatic tests of a particular alternative model. In joint models, repeated measurements from the same person are related, and event processes can affect which longitudinal measurements are observed. A pattern should therefore prompt a model check in context, not a causal conclusion or an unqualified claim that the trajectory is wrong.

The earlier material on conditional and marginal residuals distinguishes residuals formed from subject-specific fitted trajectories from those formed using the population-level mean. That choice matters here. Conditional residuals can help examine the trajectory after accounting for estimated subject effects; marginal residuals also reflect between-subject variation. Empirical estimates of subject effects are uncertain, so conditional residual patterns are not a perfect view of each person’s true trajectory.

What a Pattern Says About the Mean Trajectory

If the longitudinal mean structure is adequate, residuals should fluctuate around zero without a persistent systematic trend across time or fitted values. This does not mean every point must be close to zero, nor that the points are independent. Random scatter can be substantial, especially when measurements are noisy.

A smooth curve through residuals helps summarize the conditional mean pattern. A curve that stays near zero is compatible with an adequate mean trajectory, while a sustained departure from zero suggests that the fitted mean may leave some systematic structure unexplained. In a residual-versus-time plot, a U-shape or inverted U-shape can indicate omitted curvature. A roughly monotone trend can suggest an incorrect time trend, such as a missing time effect or an unsuitable time scale. A pattern against fitted values can reveal that the model’s errors change systematically across the biomarker range; curvature there may also indicate that the functional form for the mean is inadequate.

A useful justification follows from the meaning of a residual. Suppose the true conditional mean at time \(t\) is \(m(t)\), while the fitted mean is \(\widehat m(t)\). Ignoring estimation error for the moment, the conditional expected residual is

Thus, a systematic gap between the true and fitted mean appears as a nonzero conditional mean of the residuals. In an ordinary least-squares fit with an intercept, residuals sum to zero overall and are orthogonal to included predictors in the fitting sample. Those global constraints do not require the residual mean to be zero at every time or every fitted value: a curve can rise and fall while its total sum remains zero.

In joint models, this argument is a guide to interpretation, not a guarantee that a visual pattern has one cause. Residuals use estimated model parameters and, for conditional residuals, estimated subject effects. Also, measurements are repeated within subjects. A smooth plotted through all observations may look precise simply because some subjects contribute many measurements. Consider the number and timing of subjects’ observations, and where possible compare the pooled plot with subject-specific displays or summaries.

Distinguishing Time Trends and Fitted-Value Patterns

The two plot axes answer related but different questions. Time is the natural axis for asking whether the trajectory evolves incorrectly over follow-up. Fitted biomarker values ask whether residuals tend to be positive or negative in particular regions of the predicted biomarker range. In a monotone trajectory these axes may be strongly related, so seeing a similar curve in both plots does not provide two independent pieces of evidence.

A fitted-value plot is especially helpful when misspecification is not simply a function of time. For example, residuals that are mostly negative at low fitted values and positive at high fitted values indicate underprediction at one end and overprediction at the other. The same sign-pattern logic applies: residuals above zero mean the observed biomarker exceeds the fitted value, while residuals below zero mean the fitted value exceeds the observation.

Boundary Patterns and Sparse Support

The beginning and end of the observed time range deserve separate scrutiny. Flexible trajectory models, including spline-based models, are often less well supported near boundaries because there are fewer nearby observations on one or both sides. A residual pattern concentrated at the earliest or latest times may reflect an inadequate boundary fit, sparse data, influential observations, or an unsuitable time specification. It does not by itself establish which explanation is responsible.

Boundary evidence is particularly fragile when only a small number of subjects contribute measurements near the edge, or when observation times are shaped by clinical events. In a joint longitudinal and event-time analysis, later biomarker values may be unavailable after an event. A plot of observed residuals at late times therefore describes the available measurements, not necessarily the trajectory that would have been observed for all subjects had follow-up continued. Do not extrapolate a boundary pattern beyond the observed support on the strength of the plot alone.

A Practical Diagnostic Sequence

Interpretation is stronger when the plot is connected to the model’s prediction target, observed data support, and repeated-measures structure. A disciplined sequence helps keep visual clues from being treated as conclusions.

Residual pattern plots are mainly descriptive. Their visual smooths do not automatically come with valid uncertainty bands, especially if repeated measurements are treated as independent. If formal inference about a pattern is needed, its uncertainty must account for the dependence and model-fitting process. A diagnostic refit should also be compared on the same target and observed support; otherwise, apparent improvement may reflect a changed set of observations rather than a better trajectory.

Residual structure can have more than one source. A trend in the residual mean concerns the fitted trajectory, whereas serial dependence among residuals concerns their association across measurement times; the earlier material on residual autocorrelation addresses the latter. Likewise, a changing residual spread is not the same as a changing residual mean. Keep these questions separate when reading a plot, rather than assigning every visible feature to trajectory shape.

Common Pitfalls and Limitations

• Calling any curved scatter a misspecified trajectory. A few extreme observations can bend a smooth. Check the observations, subject contributions, and support before changing the mean structure.
• Treating all plotted measurements as independent. Many residuals from one subject do not provide the same evidence as the same number of residuals from different subjects. Use subject-level context when interpreting pooled patterns.
• Reading a boundary pattern as evidence beyond the data. Sparse observations near the end of follow-up cannot establish how trajectories behave outside the observed range.
• Confusing residual mean and residual spread. A systematic upward or downward smooth concerns average prediction error. A change in vertical scatter concerns variability and requires a different diagnostic interpretation.
• Assuming a better-looking fit is automatically better. More flexible trajectories can follow noise. A plausible time structure, adequate support, and careful reassessment are needed; visual fit alone is not a model-selection rule.

These plots are useful because they show where a fitted biomarker trajectory tends to miss, not merely how large residuals are overall. They have limits: the residuals are estimated, observations are clustered, and event-related observation patterns can constrain what the data reveal. Read a pattern as a reason to investigate, and describe its location, direction, and supporting data before proposing a change.