How do you analyze longitudinal survey data?
Connect repeated observations to the same people, align them to an appropriate time scale and examine both individual paths and group patterns. Retain missing waves, instrument versions and the length of follow-up. Choose descriptive summaries or statistical models that answer your question without treating repeated responses as independent people.
Use this reference when your question needs observations beyond a single program cycle. Extend a matched-record workflow, describe different paths and build a report that can be updated when another observation arrives. Return those decisions to your course plan.
Longitudinal analysis is useful when a training team follows graduates, a service team checks on continuing needs or an organization tracks experience over several years. It can show whether an observed gain remains visible at later checkpoints. It cannot establish continuous improvement between unobserved dates or explain a change simply by naming its shape.
Start with the question and the unit of analysis
“How has the average score changed?” and “Which people show a rise followed by a decline?” are both legitimate questions. The first concerns a group summary; the second concerns individual sequences. A third question might ask whether change differs by exposure or starting circumstances. Each requires a different level of analysis.
A panel follows the same people. Repeated cross-sectional surveys may sample different people at each wave. Both can describe trends, but a changing cross-section cannot directly show within-person movement unless records can be validly linked.
Count people and observations separately. If 100 people each answer four times, that is 100 people and 400 observations—not 400 independent participants. Also distinguish a participant's own account from a parent's, colleague's or case worker's account of that participant.
Write the analysis question, observation unit, relevant period and decision before building a dashboard. A useful statement is: “For people enrolled in this cohort, how does the same confidence measure change during the first 24 months, and how much follow-up is missing?”
Choose elapsed time, calendar time or both
For a rolling program, elapsed time since intake helps compare similar stages. Someone's third response might arrive at month 12 while another person's third response arrives at month 20. Calling both “wave three” conceals that difference.
Calendar time remains useful when the question concerns a policy change, reorganization, economic event or common service disruption. Age may be a relevant time scale in a developmental study. Keep actual dates so the analysis does not depend on one fixed labeling scheme.
Define the starting event clearly: enrollment, first service, completion or account activation can lead to different interpretations. If you group observations into windows, document the boundaries and the rule for multiple responses in one window. Retain the exact dates behind the grouped display.
Do not compare a person's first six months with another person's four-year history and label both “sustained.” State the horizon: “higher at the available six- and twelve-month checks” is more precise than “improved permanently.”
Keep one observation connected to its full context
A long-format analysis table commonly has one row per person, measure and observation date. Include the contact identifier, cohort, start date, elapsed time, score, instrument version and source. Add response status and eligibility information where available.
A wide-format table can still be useful for a fixed set of checkpoints. It does not inherently produce alignment errors, provided dates and rules remain visible. Choose the format required by the analysis, while preserving a reliable underlying record.
Keep originals when resolving duplicates or correcting a match. An unknown response should not be attached to the nearest-looking name just to complete a timeline. A joined record is useful only when the relationship is trustworthy.
If your instrument changed, use the instrument-version reference to assess the effect. A new question, translation or response scale can create a break in comparability. Preserve the change rather than asking AI to smooth it away.
Worked example: similar endpoints, different paths
The fictional records below use a 1–5 confidence item at months 0, 6, 12 and 24. They demonstrate observed sequences, not validated improvement thresholds or customer outcomes. Higher coded values indicate greater reported confidence.
| Person | Month 0 | Month 6 | Month 12 | Month 24 | Observed description |
|---|---|---|---|---|---|
| A | 2 | 4 | 2 | 2 | Higher at six months, then back to baseline |
| B | 2 | 2 | 2 | 2 | Same recorded score at each check |
| C | 2 | 3 | 4 | 4 | Higher at later checks |
| D | 2 | 2 | 2 | 4 | Higher only at the final check |
| E | 4 | 3 | 2 | 1 | Lower at successive checks |
| F | 2 | 4 | Missing | Missing | Later path unknown |
A and B both start and finish at two. Their intermediate records differ. That is a reason to retain the full sequence; it is not proof that the program reached A but failed B. The measure may miss changes that matter, and external circumstances may also affect responses.
Among A–E, who have all four observations, the coded means are 2.4, 2.8, 2.4 and 2.6. This group summary is useful alongside the individual paths. It shows a pattern in the observed scores without claiming everyone experienced it.
F has an early increase but no later data. Do not label F as having maintained or lost the gain. Including F at early waves changes the reporting base, so show that clearly if you plot all available observations.
Make the paths visible before grouping them
For a small teaching dataset, individual line charts help readers see how the same endpoints can conceal different paths. Keep points at their actual observation times. A connecting line is a visual aid, not evidence that the score moved smoothly between visits.
Fictional A and B: identical endpoints, different observed histories. B's dashed line is offset slightly for visibility; all B scores are 2.
For a larger cohort, show a manageable sample of paths, small charts by group or a distribution at each time point. Do not publish identifiable histories without appropriate permission. Always show the number of observations behind a group line and distinguish measured values from model estimates.
Use descriptive groups only when their rules help
Labels such as “early increase, later lower” can support a review meeting. They are optional summaries, not the definition of longitudinal analysis. There is no universal requirement to use four to six shapes or merge a rare pattern just because it represents less than five percent of the cohort.
If you use categories, specify the minimum follow-up, score rule, allowed gaps and tie-breaking rule. Separate “insufficient follow-up” from a stable or changing path. Do not force an uncertain record into a favorable category.
For this exercise, one descriptive label per person produces six records across six descriptions, including F's unknown later path. In real data, descriptions may overlap. Decide whether your table counts mutually exclusive people or nonexclusive features; totals mean different things in those two designs.
A large unclassified group may indicate an inadequate rule, but it may also reflect missing data or genuinely varied histories. Review the records before changing the categories. Keep both the original rule and any revised version so a change in method is not mistaken for a change in outcomes.
Handle missing waves and unequal follow-up explicitly
Late joiners have had less opportunity to contribute long-term observations. People who stop responding may differ from those who remain. Report the number eligible for each horizon and the amount of observed follow-up, not just the total enrolled.
If a score is higher at month six and lower at month eighteen, the decline was observed between those checkpoints. You cannot identify the exact month it began without additional evidence. Schedule follow-up based on a justified question and practical need, not a falsely precise turning point.
Do not turn a missing value into zero or assume the last score remained unchanged. Interpolation may be useful for a clearly labeled visualization or a justified model, but interpolated values must not become unmarked observations. The attrition lesson provides a fuller missingness review.
Compare a clearly defined complete-record summary with other defensible analyses when missingness could matter. State their assumptions and differences. A model that uses incomplete records does not automatically remove selection bias.
Know when a statistical model adds value
Descriptive paths answer what was observed. Models can help estimate change, compare groups or account for repeated measurements and other variables. They require assumptions about the outcome, timing, dependence and missingness.
UCLA's longitudinal analysis seminar compares several approaches and explains how multilevel models represent observations within people. Such models can accommodate unequal schedules in suitable specifications. That flexibility still requires careful construction and interpretation.
A growth model does not necessarily group people into trajectory classes. It may estimate an average path and variation around it. Rule-based labels, growth models and methods that identify latent groups answer different questions; do not present them as interchangeable versions of the same method.
Use qualified statistical support when the report requires uncertainty estimates, adjustment for time-varying factors or complex missing-data handling. Retain plots and checks that show whether a fitted model is a reasonable description. Do not select a model only because it creates a positive result.
Read comments and events beside the relevant period
For A, a later comment might mention a job change, a difficult placement or a different understanding of the confidence question. These are possible explanations to investigate, not facts supplied by the score. Compare the comment date with the event it describes.
Bring in attendance, service updates or documents when they are relevant and permitted. Keep contradictory accounts visible. A participant's experience may differ from a staff assessment; joining both to one record should preserve who said what.
Apply the method from connecting quantitative and qualitative evidence: place the numerical observation, supporting account, limitation and next question together. A theme found after a decline is not automatically its cause.
For named follow-up, distinguish an agreed support action from a label about the person. “Review a reported difficulty” is more defensible than treating an uncertain trajectory as a diagnosis or prediction.
Test a continuing record in your workflow
- Start with the six records. Confirm that A and B retain different intermediate histories while F's later path remains unknown.
- Add actual dates. Introduce a staggered start and verify elapsed-time alignment without losing calendar time.
- Add a later observation. Record its effect on the current summary while retaining the previous reporting snapshot.
- Change a rule deliberately. Version it, rerun the stated population and explain which results changed because of the rule.
- Trace the output. Follow a count or interpretation back to its observations, source versions and permitted evidence.
In a configured Sopact workflow, ask to see how new surveys, notes and documents remain associated with the contact and period, how analysis is reviewed and how access is controlled. Test the actual workflow with your rules rather than assuming a platform automatically produces a valid longitudinal model.
A pass means the records, dates, denominators and interpretations reconcile. It does not require every person to fit a shape, every group to improve or a dashboard to reveal an exact intervention date.
Watch: keeping evidence connected over time
Watch the video · 2 minutes 34 seconds. This companion explains the value of keeping qualitative evidence connected to continuing records. It is not a statistical modeling tutorial. Browse more videos in the video library.
Frequently asked questions
How many waves are required?
Two observations can describe a difference. More observations can reveal additional shape, but the number needed depends on the question, timing and method. No fixed wave count guarantees reliable classification.
Should I stop reporting the average trend?
No. A group trend can be useful when its reporting base and limits are clear. Pair it with distributions or individual paths when the question concerns variation between people.
Can I use people with incomplete histories?
Often, with a method suited to the question and explicit missing-data assumptions. Do not claim to know unobserved periods or silently compare groups with different follow-up lengths.
Is elapsed time always the right time scale?
No. Elapsed time helps compare stages since an event. Calendar time or age may better answer another question. Retain actual dates so these choices remain possible.
Are trajectory labels the same as growth models?
No. Labels summarize patterns under declared rules. Growth models estimate change and variation under statistical assumptions; they do not necessarily create groups.
Can written responses be analyzed over time?
Yes. Preserve source wording, respondent, date and question context. Compare themes or accounts with a documented method and review changes in meaning, not just keyword counts.
Does a later decline prove that an outcome was lost?
It shows a lower recorded value at that check. Whether an outcome was lost depends on the measure, definition and other evidence. The timing and cause may remain uncertain.
When the question is how long an outcome lasts
Carry the dates, observation rules and missingness record into measuring outcome duration and drop-off. The related lesson distinguishes evidence of persistence from assumptions used to describe what happens after the last observation.
Reviewed September 12, 2026. All participant records and paths in this lesson are fictional teaching examples.