What is a longitudinal design?
A longitudinal design examines observations across time to understand change, persistence or sequence. Some designs follow the same units repeatedly; others examine a defined population at successive points. Be explicit about which approach you mean.
Repeated observation can reveal patterns a single snapshot misses. It does not automatically establish that one event caused another. Timing, measurement and who remains in the study all matter.
Panel, cohort and trend designs
| Design | What is followed | Main caution |
|---|---|---|
| Panel | The same people or units repeatedly. | Attrition and changes in response behavior. |
| Cohort | A group sharing a starting event or characteristic. | Specify whether the same members are repeatedly observed. |
| Trend or repeated cross-section | A population through samples at different times. | Population trends are not individual trajectories. |
A workforce cohort might be everyone entering in one quarter. A panel would repeatedly observe the same participants. A population trend could use comparable samples of workers each year.
Plan the observation schedule
Choose intervals that match when change could occur and when the team can act. Immediate post-training learning and six-month application are different questions. More frequent measurement can add burden without improving the answer.
Record actual dates and exposure to the intervention. “Quarter two” may mean different elapsed time for people who entered at different points. See baseline data.
Keep measurements comparable
- Use consistent definitions, scales and scoring rules.
- Document changes in collection mode or question wording.
- Keep source dates and instrument versions.
- Use appropriate stable identifiers when following the same units.
- Check duplicates and ambiguous matches before analysis.
Do not promise anonymity if responses remain linked to an identifiable record. Use a design and access controls appropriate to the purpose and sensitivity of the data.
Worked example: attrition changes the picture
In an illustrative panel, 100 people answer at baseline and 70 at follow-up. Fifty of those 70 improve. You can report improvement among the observed matched group, but you do not know the result for the 30 missing people.
Compare relevant baseline characteristics of those retained and lost where possible. Explain likely sources of attrition. A statistical adjustment may help under particular assumptions, but it cannot make those assumptions disappear.
Analyze change at the right level
For paired data, distinguish within-person change from differences in group composition. For repeated samples, use comparable population estimates and report the sampling context. Account for repeated observations when selecting statistical methods.
Qualitative follow-up can explain transitions and barriers. Keep its sampling and interpretation limits clear. An account of why someone changed is valuable, but it is not necessarily a complete causal explanation.
Keep the history usable
Retain the original observations rather than overwriting the latest value. Store the time, source and relevant context with each update. This helps a team distinguish a genuine change from a corrected record.
WorldSkills’ published story describes building continuity across teams and event cycles. It illustrates a data-management need, not proof of a particular longitudinal research design. Read the story.
Distinguish a change in the person from a change in the record
Suppose a participant moves from site A to site B between observations. Keep both dated site relationships. If you replace every historical site value with site B, an earlier report may change even though no earlier observation changed. Decide whether each analysis groups people by starting site, current site or site at the time of observation, and state the choice.
A corrected score also needs a different treatment from a later score. Retain the observation date and the correction history. The revised value may change an earlier calculation; it should not appear as a new follow-up observation.
Different sites may collect useful local questions. Agree on the limited shared measures needed for comparison, including their definitions, eligible groups and observation windows. Keep local measures separate where comparability has not been established. A shared participant ID connects evidence; it does not make different questions measure the same thing.
For the operational plan behind that continuing record, work through the Case intelligence course. Keep the research question in charge of the collection schedule and the conclusions.
Frequently asked questions
How many time points are needed?
At least repeated observations are needed to examine change, but the number and spacing depend on the question and analysis.
Is a cohort always a panel?
No. Cohort describes a shared characteristic or starting event. Specify how members are observed over time.
Can I compare anonymous waves?
You can compare suitable group estimates, but cannot identify individual change without appropriate linkage.
Does longitudinal data prove causation?
No. It helps establish sequence but other explanations and design limitations remain.
