Why Inspect Residuals One Subject at a Time?
A population residual plot can appear unremarkable even when individual subjects have clear sequences of underprediction or overprediction. Those sequences matter: they may point to a time-dependent model mismatch, an unusual observation, or a subject-specific feature that disappears when observations are pooled. An individual residual profile makes the sequence visible by displaying each subject’s residuals against that subject’s observation times.
A profile is a diagnostic display, not a test that automatically declares a model adequate or inadequate. Its interpretation depends on what residual was calculated, how observations were ordered, and whether the plotted points are comparable. Earlier material on choosing residuals for standard diagnostics distinguishes common residual choices; here the emphasis is on what their subject-level sequences reveal.
Residuals are typically read relative to a horizontal zero reference. Positive and negative signs indicate departures in opposite directions under the residual convention being used; verify that convention before interpreting which direction corresponds to underprediction. If residuals are standardized or otherwise normalized, they may be dimensionless; raw residuals retain the units of the measured quantity. State the residual type and units in the plot or its caption.
Constructing and Reading a Profile
A useful profile puts time on the horizontal axis and one consistently defined residual quantity on the vertical axis. Each panel represents one subject, with observations ordered by actual time rather than by row order in a data file. If observations occur at the same time, use a clear plotting convention and retain their identities; arbitrary jitter can make a display easier to read but should not suggest that the recorded times differ.
A zero line helps identify sign changes. Connecting points can make runs easier to see, but a connecting line represents ordering, not a continuous trajectory between sampled observations. If times are widely spaced, a line should not be interpreted as evidence about what happened in the gap. For a large study, selected subjects may be shown for a focused inspection, while a report may also provide panels for all subjects. Selection criteria should be transparent: choosing only subjects with striking profiles can be useful for follow-up but cannot represent how common such patterns are.
Confirm the residual definition, sign convention, units, subject identifier, time variable, and any filtering applied to the diagnostic data.
Locate zero crossings, consecutive same-sign points, abrupt changes, and values that are extreme relative to the rest of the profile and the plotted scale.
Look for a recurring feature across subjects, while checking whether differences could reflect sampling times, residual scale, or a small number of observations.
State the subject or subjects, time region, direction, and evidence—for example, “three consecutive positive residuals”—before proposing an explanation.
A profile can reveal several kinds of behavior. A sign run suggests that consecutive observations lie on the same side of zero. An isolated extreme is a point far from the other points in its profile or from the broader display, and may merit data and model review. A repeated rise or fall over time may suggest a systematic temporal pattern. These descriptions are not interchangeable: a single very large point is not by itself a run, and a run can occur even when none of its values is especially large.
Worked Example: A Large Residual Inside a Positive Run
Suppose a hypothetical subject C has five normalized residuals in time order: \(-0.3,\ 0.2,\ 2.7,\ 0.1,\ -0.2\). The residuals are dimensionless, and the sign convention is known. Identify any sign run and describe the value \(2.7\).
The ordered signs are negative, positive, positive, positive, negative. The second, third, and fourth residuals are three consecutive positive values: \(0.2 > 0\), \(2.7 > 0\), and \(0.1 > 0\). Therefore, this profile contains a positive sign run of length three. The first and last observations are negative and do not extend that run.
The value \(2.7\) is also an isolated extreme by magnitude within the profile: it is much farther from zero than the adjacent values \(0.2\) and \(0.1\). It should be flagged for review as an extreme value, while the sequence should separately be described as a run of three consecutive positive residuals. The appropriate conclusion is not that \(2.7\) is a run by itself; the run is the three-point sequence.
This is descriptive evidence, not a formal significance result. Review the underlying observation, its timing and units, and the relevant prediction and residual calculations before treating the point or run as evidence of a particular mechanism.
Why Population Summaries Can Hide Subject Patterns
Pooling residuals across subjects can obscure opposing sequences. A population display may show the overall distribution or trend, but it does not necessarily preserve the identity and ordering of each subject’s observations. If one subject tends to have positive residuals in an interval and another tends to have negative residuals there, their pooled average can be close to zero even though neither subject’s profile is centered around zero in that interval.
Worked Example: Opposing Profiles Cancel in a Pooled View
Consider two hypothetical subjects observed at three common times. Subject A has residuals \(1.0,\ 1.2,\ 0.8\); subject B has residuals \(-1.0,\ -1.2,\ -0.8\). All values are normalized residuals. At each time, calculate the mean across the two subjects and compare it with the individual sequences.
At time 1, the pooled mean is \((1.0 + (-1.0))/2 = 0/2 = 0\). At time 2, it is \((1.2 + (-1.2))/2 = 0/2 = 0\). At time 3, it is \((0.8 + (-0.8))/2 = 0/2 = 0\). Thus a plot of these means would lie exactly on the zero line at all three times.
The subject profiles tell a different story. Subject A has three consecutive positive residuals, while subject B has three consecutive negative residuals. The pooled means conceal both sign runs because the departures cancel at every time. The mean is arithmetically correct, but it does not describe the within-subject sequences. Inspecting profiles alongside population-level displays is therefore valuable when the diagnostic question concerns subject-specific patterns.
This calculation does not show that either subject’s pattern has a particular cause, nor does it establish how often such profiles occur in a larger population. It shows precisely why a population-level summary can fail to reveal a pattern that is present in the individual data.
This cancellation is one reason to interpret individual profiles together with, rather than instead of, population goodness-of-fit panels and residual-versus-time displays. Population panels support questions about overall behavior; individual profiles answer questions about the identity, order, and persistence of departures within subjects. Neither view substitutes for the other.
Checking the Evidence and the Display
Before interpreting a pattern, verify that the data supplied to the plot correspond to the intended observations. Subject labels should be stable, time should use the intended scale and units, and residuals should be calculated consistently. A sorting error can create a false sequence; duplicate or mismatched records can create apparent extremes. Confirm that missing residuals or excluded records have not silently changed the sequence being viewed.
Comparisons also depend on scale. A shared vertical scale across panels makes residual magnitudes visually comparable, but a small profile may then be hard to inspect. Independent panel scales can expose within-subject variation but make equal-looking distances represent different values across panels. If scales differ, label or otherwise disclose that choice; do not infer that two points have equal magnitude from their visual height alone. Axis limits should not clip extremes without a clear indication.
Unequal observation schedules require particular care. A run means consecutive observations in the plotted sequence, not observations at equally spaced times. Three same-sign points collected minutes apart and three collected weeks apart both meet the descriptive definition, but their scientific implications may differ. Describe the time spacing and avoid treating a run as a continuous period of misfit.
Worked Example: Separating a Time Pattern from an Extreme Point
A hypothetical subject has normalized residuals at times 0, 2, 8, 20, and 22 hours of \(0.1,\ 0.4,\ 1.1,\ 0.6,\ -2.5\). Describe the sequence, taking both sign and spacing into account.
The first four residuals are positive: \(0.1 > 0\), \(0.4 > 0\), \(1.1 > 0\), and \(0.6 > 0\). They form a positive run of length four. The final residual, \(-2.5\), is negative and is separated from the preceding observation by 2 hours; it ends the run. Its magnitude is large relative to the other values, so it is an extreme point to review.
The positive points span from 0 to 20 hours, but their spacing is unequal: the gaps are 2, 6, and 12 hours. It would be accurate to report a four-observation positive run across that time span. It would be inaccurate to imply that residuals were observed continuously or that the positive pattern persisted at every intervening time. The final extreme may indicate a data issue or a local model mismatch, but the profile alone cannot distinguish those explanations.
Assumptions, Alternatives, and Limits
Individual residual profiles are most interpretable when observations have correctly assigned subjects and times, the residual definition is suitable for the diagnostic question, and plotted residuals use a consistent convention. A zero reference is meaningful when zero represents no signed discrepancy under that residual definition. If residual scales differ by observation or subject, comparisons of apparent magnitude require additional caution. Profiles are also easier to interpret when there are enough observations per subject to show a sequence; a one-point profile can show an extreme but cannot show a run.
No independence assumption is needed to describe a visible run. However, one should not treat a run as a formal probability statement without a model for the residual signs and their dependence. Consecutive residuals from the same subject may be related, and observations may have unequal times or different measurement conditions. This tutorial’s three-point screening convention is a practical descriptive flag, not a calibrated error rate or universal cutoff.
The main alternative is to inspect pooled residual panels, such as residuals versus time, which are better suited to broad population-level patterns. Individual fit panels show observations and predictions in subject context; residual profiles emphasize signed discrepancies and their temporal ordering. These views are complementary. A residual profile does not replace checking the underlying data, reviewing individual fits, or examining the population-level diagnostics.
A striking profile is a reason to investigate, not a diagnosis. An extreme value might result from a recording or data-processing error, an unusual observation, or model inadequacy. A run might reflect a time-structured mismatch, but it might also arise by chance or from design features. The plot alone does not establish causation, quantify population prevalence, or prove that a model is invalid. Record what is visible first, then use the relevant data and modeling context to determine what follow-up is justified.
Common Mistakes
- Calling an extreme point a run. A single large value may be extreme by magnitude, but a sign run requires consecutive same-sign observations. Describe both features separately when both occur.
- Calling all points near zero adequate. Repeated small departures on the same side of zero can form a run; magnitude and sign sequence are distinct aspects of the profile.
- Assuming pooled balance means each subject is balanced. Opposing subject profiles can cancel in a population summary, as the arithmetic example showed.
- Reading a line as continuous data. Connecting sampled observations does not show residual behavior between observation times.
- Comparing heights across panels with different scales. Check axis limits and units before comparing apparent magnitudes.
- Assigning a cause from the plot alone. A profile identifies a pattern to review; it does not determine whether the cause is data error, sampling variation, or model structure.
Check Your Understanding
Answer each question from the profile evidence and calculations shown in the tutorial.
- A subject has residuals \(-0.4,\ -0.2,\ 0.1,\ 0.3\) in time order. Is there a run of at least three consecutive same-sign residuals?
- Two subjects have residuals \(0.7,\ 0.9,\ 1.1\) and \(-0.7,\ -0.9,\ -1.1\) at three common times. What do the pooled means show, and what do the profiles show?
- A profile has values \(0.2,\ 0.1,\ 3.0,\ 0.4\). Describe its sign sequence and the value \(3.0\) without confusing a run with an extreme.
- Why should a sequence of three same-sign residuals collected at widely unequal intervals not be described as continuous misfit?
- What should be checked before interpreting a profile that appears to have a sharp jump between two observations?
Solutions. (1) No. The first two values are negative and the last two are positive, so neither same-sign sequence has three observations. (2) At every time the pooled mean is zero because the values cancel pairwise; the individual profiles show a three-point positive run for the first subject and a three-point negative run for the second. (3) All four residuals are positive, so there is a positive run of length four. The value \(3.0\) is also an extreme by magnitude relative to the other three; it is not the run by itself. (4) The run describes consecutive observed points, not unobserved values between them. Unequal spacing affects the time span and interpretation, and no continuous measurements are supplied. (5) Check that the observations are assigned to the correct subject and sorted by the intended time, verify residual definitions and units, and inspect the underlying records and time gap before proposing an explanation.