Outlier Detection In Longitudinal Measurements

Why Screen Longitudinal Measurements?

A single biomarker value far from its fitted value can be a recording error, a genuine transient change, or a symptom of a model that does not capture the subject’s trajectory. Repeated measurements make the distinction more informative: one isolated extreme differs from a sequence of residuals that remain unusually large or have the same sign. Outlier detection in longitudinal data is therefore a screening and interpretation task, not a rule for automatically deleting observations.

The aim here is to identify unusual measurements relative to a fitted longitudinal submodel, then summarize residual evidence across visits without allowing one extreme point to dominate every subject-level assessment. The earlier material on standardized conditional residuals for Gaussian markers provides the basic residual idea; this tutorial applies it to outlier screening and emphasizes what subject-level patterns can and cannot establish.

Model and Residuals for Screening

Let \(Y_{ij}\) denote the biomarker measurement for subject \(i\) at visit time \(t_{ij}\), for \(j=1,\ldots,n_i\). Write the longitudinal model as

where \(m_i(t_{ij};b_i)\) is the fitted conditional mean given subject-specific random effects \(b_i\), and \(\varepsilon_{ij}\) is the measurement-level error. In a joint model, the longitudinal submodel is part of the larger model for the biomarker and event time; this screening concerns the longitudinal measurements, not event-component residuals.

A positive residual means the observed biomarker is above its fitted conditional mean; a negative residual means it is below. Standardization expresses the discrepancy in residual standard-deviation units, so a residual of \(3\) is three such units above the fitted value. Under a correctly specified Gaussian model with known parameters, independent errors conditional on \(b_i\), and the correct conditional mean and variance, the error divided by its standard deviation has a standard normal distribution. In practice, model parameters and random effects are estimated, so the fitted residuals need not be independent standard normal variables; the reference is approximate and depends on the fitting procedure.

Before interpreting a large residual, check that the model’s conditional mean and variance are suitable for the observations being screened. The simple reference above presumes the residual standard deviation is correctly specified and that the conditional error distribution is adequately represented. If residual variance changes over time or depends on covariates, a single \(\widehat\sigma\) can make ordinary observations appear extreme in high-variance regions and conceal unusual values in low-variance regions. If errors are serially dependent, a sequence of same-sign residuals may partly reflect that dependence rather than a trajectory defect. The earlier material on residual autocorrelation, heteroscedasticity, and trajectory misspecification addresses those patterns; they should be considered before treating a threshold crossing as evidence of an unusual measurement.

Screening Measurements and Subjects

A common first screen is to flag measurements with large absolute standardized residuals. For example, \(|r_{ij}|>3\) is sometimes used as a heuristic flag under an approximately standard normal reference. It is not a formal significance test, a guaranteed false-positive rate, or a universal cutoff. With many measurements, even a small per-measurement chance of crossing a threshold can yield several flags. Dependence among visits also affects how often flags occur together. A flag is a prompt to investigate the value, its data record, and the model fit.

To summarize a subject’s residuals, use more than the largest absolute residual. Let \(R_i=(r_{i1},\ldots,r_{in_i})\). Useful summaries include the maximum absolute residual, the count of threshold exceedances, and robust measures such as the median absolute residual and median signed residual:

Here \(A_i\) describes a typical residual magnitude, \(S_i\) its typical direction, and \(K_i(c)\) counts measurements beyond a chosen screening cutoff \(c\). The median absolute deviation around the subject’s median residual, \(D_i=\operatorname{median}_j|r_{ij}-S_i|\), is another robust summary of within-subject spread. These summaries have different purposes: a large \(A_i\) suggests that residual magnitudes are often elevated, while a large \(|S_i|\) suggests that residuals are persistently shifted in one direction. Neither summary replaces inspecting residuals against visit time.

A robust summary is relatively resistant to one extreme value, which is useful when trying to identify a persistent pattern rather than simply rediscover the largest single measurement. But it can also hide a genuine one-time anomaly: with many visits, one extreme may barely change the median. That is why subject-level summaries and measurement-level flags should be reviewed together.

Why the Pattern Across Visits Matters

An unusually large residual and a persistently unusual trajectory are not the same diagnostic finding. The distinction follows from looking at residual magnitude, sign, and time order together. An isolated large \(|r_{ij}|\) with neighboring residuals near zero is compatible with a single measurement anomaly, but also with a real short-lived biological change. Repeated residuals with similar signs suggest that the fitted conditional mean may be systematically high or low for that subject or period. A changing pattern—for example, small residuals early and increasingly positive residuals later—may point to a time-dependent mismatch rather than one aberrant observation.

A short justification explains the role of standardization. If the model has the form \(Y_{ij}=m_i(t_{ij};b_i)+\varepsilon_{ij}\), with conditional mean \(E(\varepsilon_{ij}\mid b_i)=0\) and conditional variance \(\operatorname{Var}(\varepsilon_{ij}\mid b_i)=s_{ij}^2\), then

The first equality follows by dividing the zero conditional mean by \(s_{ij}\); the second follows because the variance of a constant multiple is the squared constant times the variance: \(s_{ij}^{-2}s_{ij}^2=1\). Thus standardization puts residuals on a common scale when the conditional variance model is right. It does not make residuals independent, correct a wrong mean model, or make estimated residuals exactly standard normal.

Practical Review and Common Pitfalls

A defensible review combines the numerical flags with the observed data, visit times, fitted trajectory, and model assumptions. First verify the measurement, units, visit label, and any data-processing steps. Then compare the value with neighboring observations and the subject’s broader course, without assuming that abrupt biological change is impossible. Next consider whether the conditional mean and residual variance are plausible at that time. If the pattern is repeated or time-ordered, inspect the relevant longitudinal residual diagnostics before concluding that the subject or measurement is unusual.

• Calling every threshold crossing an outlier. A cutoff such as 3 is a screening convention, not a universal decision rule. Review multiplicity, dependence, model fit, and clinical plausibility.
• Using only the subject’s maximum residual. A maximum identifies a candidate measurement but does not distinguish a lone spike from repeated departures. Pair it with robust summaries and the time-ordered residual sequence.
• Calling a persistently shifted subject a set of bad measurements. Repeated same-sign residuals may reflect a trajectory feature the model does not represent. Investigate the pattern rather than treating each value as an independent error.
• Assuming standardization fixes model misspecification. Dividing by an estimated standard deviation is useful only if that standard deviation is appropriate. Wrong variance, dependence, or mean structure can produce misleading flags.
• Deleting a flagged observation without justification. Removing a genuine value can distort the fitted model and its conclusions. Correct a record only when the data source supports the correction; otherwise document and assess the result with the observation retained and, when appropriate, in a clearly described sensitivity analysis.

Compared with raw residuals, standardized residuals make discrepancies more comparable across visits with different modeled variances. Compared with a maximum-residual-only screen, robust subject-level summaries are less dominated by one extreme and can highlight persistent departures. Neither approach is a substitute for the other: the first helps identify individual candidate values, while the second helps characterize patterns across a subject’s visits. Residual plots and checks for autocorrelation, heteroscedasticity, or trajectory misspecification address broader model adequacy questions. These methods are limited by sparse visits, estimated random effects, and the possibility that a real transient change resembles a measurement error. Outlier screening itself does not determine which explanation is correct.