Must all subjects have equal follow-up times?

No; suitable methods can handle irregular or unequal follow-up under stated assumptions.

Is every repeated survey longitudinal?

Only when the relevant units are followed over time; surveying new people each wave gives a different design.

Does a larger row count guarantee greater precision?

No; dependence, measurement quality, and missingness affect the information gained.

AI Glossary / Statistics and probability / Intermediate

Longitudinal or serial data

Definition

Longitudinal or serial data contain repeated measurements of a response on the same subject or unit over time. The supplied glossary specifically describes measurements at more than one time after a defined time zero. Such records preserve an individual's trajectory rather than providing only a single snapshot. Their central statistical feature is that observations from the same unit may be related, so the number of recorded rows is not the number of independent subjects.

What it means

The subject can be a person, customer account, machine, branch, or another unit whose identity persists across observations. A baseline measurement may accompany several follow-ups. Measurements can occur at regular intervals or irregularly, and not every subject must have the same number of observations.

Time, subject identity, and measurement meaning all matter. A monthly survey of different customers is a repeated cross-section, not an individual customer trajectory. A single aggregate sales series has serial ordering but does not necessarily provide the multi-subject longitudinal structure intended by the supplied term. State which structure is present before choosing an analysis.

How it works

Organize records with a stable unit identifier, actual measurement time, response, and relevant covariates. Separate baseline information from variables updated during follow-up. Plot individual trajectories as well as group summaries to identify heterogeneity and possible changes in measurement.

Choose a model suited to the question and outcome type. Mixed-effects models can represent subject-level variation, while generalized estimating equations commonly target marginal response relationships with clustered observations. Their interpretations and assumptions differ; neither is a universal default.

Account for within-unit dependence and examine missing measurements and dropout. An absent follow-up may be related to the outcome itself. Methods that use incomplete records still need assumptions about missingness. For predictive evaluation, preserve the required separation between subjects or future times so repeated records do not leak information across the intended boundary.

Three practical examples

A learning platform records the same learners' proficiency scores across several follow-up assessments.
A manufacturer measures vibration repeatedly on identified machines as they age and receive maintenance.
A business follows monthly engagement for the same customer accounts, distinguishing changes within an account from differences between accounts.

Think of it this way

Think of a collection of short films rather than a pile of unrelated photographs. Each film follows one subject, and the ordering of its frames reveals change. Mixing all frames together might describe the collection's average appearance, but it loses who changed, when they changed, and which images came from the same source.

Worked example

Consider an original example with three accounts measured at follow-up months 1, 2, and 3. Account A has scores 10, 12, 14; B has 20, 22, 24; and C has 30, 32, 34. There are nine observations but only three accounts.

The monthly means are 20, 22, and 24. Each account increases by 4 between months 1 and 3, corresponding to 2 score units per month over that interval. The persistent differences between accounts are much larger than their observed monthly changes. An analysis that treats all nine rows as unrelated misses this structure.

Now suppose account C's month-3 score is missing. The observed month-3 mean becomes (14 + 24)/2 = 19. Comparing that with the month-1 mean of 20 suggests a decline of 1, even though every complete trajectory increased. For the two accounts observed at both endpoints, the month-1 mean is 15 and the month-3 mean is 19, a rise of 4.

This arithmetic does not prove that complete-case analysis is unbiased; dropout may systematically select accounts. It shows why changing sample composition can distort a naive time comparison and why the missingness mechanism needs attention.

Illustrative case study

Fictional case: Orchard Training tracks participants through three assessments after a course starts. The program team initially compares each assessment's available average. An analyst notices that participants with different starting skills have different follow-up completion rates.

The team preserves participant identifiers, documents reasons for missed assessments, and models trajectories with a method suited to the response and study design. It reports uncertainty and sensitivity to missing-data assumptions. Managers learn both how the observed participants progress and how strongly conclusions depend on who remains in follow-up, rather than interpreting a changing average as an automatic program effect.

Common misconceptions and mistakes

Repeated rows are not automatically independent evidence. More visits can improve information about a trajectory, but they cannot simply replace recruiting more independent subjects. Equally, not all within-subject correlation is captured by a random intercept alone.

A longitudinal association does not by itself establish causation. Time-varying confounding and changes in measurement can affect interpretation. Carrying the last observation forward is not a universally valid solution to missing data. Collapsing a rich trajectory into one binary outcome can also discard useful information, though a summary may be appropriate when it matches a carefully defined question.

Business use

Longitudinal data help businesses distinguish customer differences from changes within customers, monitor deterioration, and evaluate sustained outcomes. They support questions about timing and persistence that a single snapshot cannot answer.

Plan stable identifiers, timestamp quality, and follow-up procedures from the start. For AI applications, specify whether the model will predict future observations for known subjects or outcomes for entirely new subjects. Those are different evaluation tasks. Report the number of subjects, visits, and missing records separately so readers can understand the actual information available.

Check your understanding

1. How many independent accounts are in the worked example?

Answer: Three, not nine. 2.

Why does the available month-3 mean fall after C is missing?

Answer: The highest-scoring account no longer contributes. 3.

Is the paired increase of 4 automatically representative of missing accounts?

Answer: No; that requires additional assumptions.

Related terms

Generalized estimating equation
Dropout
Missing data
Repeated measures
panel data
within-subject correlation
mixed-effects model
generalized estimating equations
time-dependent covariate

Learn next

Practice plotting subject-level trajectories and distinguishing within-subject from between-subject comparisons. Then study marginal and subject-specific models, missingness assumptions, and evaluation designs that respect time and subject identity.

Sources

Frank Harrell, Glossary of Statistical Terms, “longitudinal or serial data.” https://hbiostat.org/glossary/ — original definition and discussion of longitudinal response information. Statsmodels, “Generalized Estimating Equations.” https://www.statsmodels.org/stable/gee.html — primary documentation for correlated panel, cluster, and repeated-measures data. The worked example and fictional case are original.

Previous: Longitudinal modified treatment policy Next: Longitudinal study