A longitudinal study collects data from the same people, or the same units, at two or more points in time and compares the measurements between those points. That shows how individuals change and which came first, which a one-time survey cannot. This guide covers the main types, four well-known studies with their follow-up data, and a worked example of attrition and change scores.
Key Takeaways
- Three conditions define longitudinal research: data from two or more distinct periods, the same (or comparable) cases each time, and an analysis that compares the periods.
- The main types are panel, cohort and retrospective designs. A prospective design starts before the outcomes happen; a retrospective one is built after some of them have.
- Longitudinal designs beat cross-sectional ones on two counts: they measure change within people and show temporal order. Neither design proves cause and effect on its own.
- Attrition is the price. In Add Health, 11,979 people completed Wave VI, a number equal to 57.7% of the 20,745 interviewed at home in Wave I.
- Report the number of participants at every wave, why people left, and how the analysis treated them.
What is a longitudinal study?
A longitudinal design measures the same individuals, households, schools or other units on two or more occasions and compares the results between occasions. Most are observational: researchers record exposures and outcomes without assigning anything. Measuring the same cases again is what separates it from a single survey, and the comparison across time is its purpose.
Menard’s widely used definition, quoted in Iowa State University’s research methods guide, has three parts: “data are collected for two or more distinct periods (implying the notion of repeated measurements); the subjects or cases analyzed are the same, or at least comparable, from one period to the next; and the analysis involves some comparison of data between or among periods.”
Health research often uses a narrower definition. The NIH’s NCATS Toolkit says a longitudinal study “follows participants prospectively over time with continuous or repeated monitoring of risk factors or health outcomes, or both.” Caruana and colleagues, writing in the Journal of Thoracic Disease, are broader: the design is “generally observational, however, may also be experimental”, with “quantitative and/or qualitative” data. A randomized trial that measures participants at baseline, 6 months and 12 months is longitudinal too. Our overview of types of research methodology places time beside the other labels a methods section uses.
Longitudinal vs cross-sectional: what is the difference?
A cross-sectional study measures different people once, at a single point in time. A longitudinal design measures the same people repeatedly. The cross-sectional version is faster and cheaper and suits estimating how common something is. The longitudinal version shows change within a person and the order in which things happened, at the cost of years of follow-up.
| Cross-sectional | Longitudinal | |
|---|---|---|
| Measurement occasions | One | Two or more |
| Who is measured | Different people in each group | The same people each time |
| What it shows | How common something is, and group differences at one moment | Change within people, and which variable changed first |
| Age comparisons | Mixes age with generation (cohort effects) | Follows one generation as it ages |
| Main threat | Non-response at the single survey | Attrition that grows with every wave |
| Resources | One round of data collection | Funding, staff and participant contact for every wave |
Caruana and colleagues state the main weakness of the one-time design: a cross-sectional analysis “provides no information with regards to the influence of time on the variables measured.” The OpenStax Psychology 2e textbook adds the age problem: compare 20-, 30- and 40-year-olds at one moment and the differences reflect the eras each group grew up in as well as their age. The cross-sectional study guide covers the one-time design and when it is the better choice.
What are the types of longitudinal studies?
The three main types are panel, cohort and retrospective designs. A panel follows a sample of a population, often a random one, and asks the same questions at each wave. A cohort follows people who share a defining event, such as birth in the same week; in epidemiology, a cohort study also splits them by exposure and counts who develops an outcome. A retrospective design is assembled after the events of interest have happened, from records or recall.
| Type | Who is followed | Direction | Example |
|---|---|---|---|
| Panel study | A sample of a population, often random, asked the same questions each wave | Prospective | Add Health, a national school sample interviewed in six waves since 1994 |
| Cohort study | People who share a starting event, such as birth in the same week or year | Usually prospective | The 1970 British Cohort Study, babies born in one week of 1970 |
| Retrospective study | People whose exposures are reconstructed after the outcomes occurred | Built from past records, still earlier to later | A hypothetical study linking old school records to adult hospital records |
| Repeated cross-section (trend study) | A new sample each time | Prospective, population level | An annual national survey with fresh respondents |
Prospective vs retrospective. A prospective design records exposures first and waits for outcomes, which avoids recall bias: people describe their habits before anyone knows who will get ill. Caruana and colleagues define a retrospective design as one “designed after at least some participants have already experienced events that are of relevance”. It is faster and cheaper, but it depends on records kept for other purposes, or on memory.
Is a repeated cross-section longitudinal? Sources disagree. Caruana and colleagues list repeated cross-sectional studies, where participants are “largely or entirely different on each sampling occasion”, as a longitudinal form, and Menard’s “same, or at least comparable” cases let them in too. We take a stricter reading: a trend survey shows that a population changed, but not that any person changed, because nobody was measured twice. If your design draws a new sample each year, call it a trend design. A trend design also has none of the exposed and unexposed groups or relative risks of a cohort study, which has its own guide.
What can a longitudinal study show that a cross-sectional one cannot?
It shows change inside the same people, and it shows temporal order: whether the exposure was measured before the outcome appeared. Measuring exposures first also keeps later knowledge from colouring what people report. What it still does not do on its own is rule out a third factor that drives both the exposure and the outcome.
Caruana and colleagues list five benefits of a prospective cohort: relating events to the timing and duration of exposures, establishing the sequence of events, following change in particular individuals, excluding recall bias, and separating the effects of age, period and birth cohort. Temporal order is necessary for a causal claim, but not sufficient. If children with more self-control later earn more, family income could explain both, so researchers adjust for such confounding variables. The guide to correlation vs causation sets out the other explanations a causal claim has to rule out.
Worked example: attrition across waves and a change score
This example uses invented numbers to show two calculations worth repeating whenever you read a longitudinal paper: retention across waves, and the gap between a change score and a comparison of groups at one wave. Every input is in the tables, so you can check the arithmetic yourself.
Part 1: retention. A panel recruits 2,000 adults at Wave 1 and re-contacts them every two years. Assume nobody returns after missing a wave.
| Wave | Responded | Retention vs Wave 1 | Retention vs previous wave |
|---|---|---|---|
| 1 | 2,000 | 100% | |
| 2 | 1,700 | 1,700 / 2,000 = 85.0% | 85.0% |
| 3 | 1,480 | 1,480 / 2,000 = 74.0% | 1,480 / 1,700 = 87.1% |
| 4 | 1,310 | 1,310 / 2,000 = 65.5% | 1,310 / 1,480 = 88.5% |
Wave-to-wave retention improved from 85.0% to 88.5%, yet the panel lost 690 people, 34.5% of its starting sample. A paper that reports only the wave-to-wave rate makes attrition look smaller, so look for the cumulative figure.
Now suppose 600 of the 2,000 smoked at Wave 1, and 330 of them were still responding at Wave 4, against 980 of the 1,400 non-smokers. Smokers were retained at 55.0% (330 / 600), non-smokers at 70.0% (980 / 1,400). Smokers were 30.0% of the Wave 1 sample but 25.2% of the Wave 4 sample (330 / 1,310), so any Wave 4 average now describes a healthier group than the one recruited. That is differential attrition.
Part 2: change score vs between-group comparison. Among people measured at Waves 1 and 3, 100 took up running between the two waves and 400 did not. The outcome is resting heart rate in beats per minute (bpm).
| Group | Wave 1 mean | Wave 3 mean | Change (Wave 3 minus Wave 1) |
|---|---|---|---|
| Took up running (n = 100) | 69 bpm | 64 bpm | −5 bpm |
| Did not (n = 400) | 73 bpm | 72 bpm | −1 bpm |
| Difference | 4 bpm | 8 bpm | 4 bpm |
A cross-sectional look at Wave 3 finds runners’ resting heart rate 8 bpm lower. The Wave 1 data show the future runners were already 4 bpm lower before any of them started. The difference in change is 4 bpm (a 5 bpm fall against a 1 bpm fall), half the Wave 3 gap; the other half is a selection effect a single survey cannot see. The 4 bpm is still not a proven effect of running, because new runners may also have changed their diet or sleep.
How is attrition reported?
Attrition is reported as numbers, wave by wave: how many people were eligible, recruited, measured at each follow-up and analysed, with the reasons for each loss. Good papers state retention against a named denominator, compare the people who stayed with those who left, and say how the analysis handled missing waves.
The STROBE reporting guideline for observational research turns this into checklist items. Item 13(a) asks authors to “Report the numbers of individuals at each stage of the study”, 13(b) to “Give reasons for non-participation at each stage”, and 13(c) to “Consider use of a flow diagram.” For cohort studies, item 12(d) asks how loss to follow-up was addressed and 14(c) asks authors to “Summarise follow-up time”. The STROBE explanation paper cites a survey of studies in 10 general epidemiology, public health and medical journals in which only 49 of 154 cohort studies (32%) gave some information on participation.
Three things to check in any retention figure:
- The denominator. The Dunedin Study reports that 94% “of living Study members” took part at age 45. A rate over the living leaves out deaths; a rate over all 1,037 founding members would be lower.
- Planned exclusions. Add Health did not sample participants who were in 12th grade at Wave I for Wave II. They were left out by design, not lost.
- Returners and new recruits. Add Health interviewed 15,170 Wave I respondents at Wave III and 15,701 at Wave IV, so some people came back after missing a wave. The 1970 British Cohort Study added 79 recruits at age 5, 294 at age 10 and 65 at age 16, born in the same week abroad.
In your own paper, recruitment and follow-up numbers belong in the methodology section, and the bias that dropouts may introduce belongs in the limitations of the study.
Famous longitudinal study examples
Four well-documented examples are the Framingham Heart Study (1948, heart disease), the 1970 British Cohort Study (everyone born in one week), the Dunedin Study (a one-year birth cohort in New Zealand) and Add Health (a US school sample from 1994-95). Each publishes its design and sample sizes on an official page, and each shows a different recruitment or retention problem.
| Study | Started | Who | Sample | Retention or follow-up detail |
|---|---|---|---|---|
| Framingham Heart Study | 1948 | Adults in Framingham, Massachusetts, target age 30 to 59 | 5,209 (4,469 from a random sample plus 740 volunteers) | Only 68.7% of the 6,507 people sampled came in; exams every two years |
| 1970 British Cohort Study | 1970 | Everyone born in England, Scotland and Wales in one week of 1970 | 16,568 babies at the birth sweep | 628 babies born in Northern Ireland were not followed after birth; sweeps have continued to age 51, and an age 56 sweep is listed for 2026 |
| Dunedin Study | Born 1972-73, enrolled at age 3 | Babies born at Queen Mary Hospital, Dunedin, New Zealand, from 1 April 1972 to 31 March 1973, whose mothers lived in the Dunedin district | 1,037 (91% of the 1,139 eligible) | 94% of living members took part at age 45 |
| Add Health | 1994-95 | US students in grades 7 to 12 | 20,745 in-home interviews at Wave I | 11,979 completed Wave VI (2022-2025), 57.7% of the Wave I number, at an average age of 44 |
Framingham shows that a sample can be biased before the first follow-up. Its own history says the planners expected 1 in 6 respondents to have cardiovascular disease and found 1 in 50, and that the final cohort “was not random and it was healthier than the general population.”
An annotated example: childhood self-control in the Dunedin cohort
Moffitt and colleagues (2011, PNAS, open access) used the Dunedin cohort to ask whether self-control in childhood predicts health, wealth and crime at age 32. It is a useful model because it reports retention, measures the predictor years before the outcome, adjusts for known confounders and states the limits of its design.
What they did. The team followed “a complete birth cohort of 1,037 children born in one city in a single year” to age 32 “with 96% retention”. Researcher-observers, teachers, parents and the children rated self-control at ages 3, 5, 7, 9 and 11, combined into one score. At 32 the team ran physical examinations, clinical interviews and a search of court conviction records.
The numbers. Children with more self-control came from better-off families (r = 0.25) and had higher IQs (r = 0.44), so the authors controlled for both. By 32, 24% of members had a criminal conviction. Children with low self-control were more likely to have started smoking by 15 (odds ratio 1.69, 95% CI 1.45 to 1.96) and to have left school with no qualifications (odds ratio 2.28, 95% CI 1.92 to 2.72). In a second cohort of 500 British sibling pairs, the sibling with lower self-control had poorer outcomes despite sharing a family.
Why retention matters here. The authors write that members “with low self-control and poor outcomes have not dropped out of the study.” A panel that had lost its most impulsive members would have missed the bottom of the gradient.
What it can and cannot conclude. The design shows temporal order (self-control measured by age 11, outcomes at 32) and an association that survives adjustment for IQ and social class. The authors call the design “observational and correlational”. Children who became more self-controlled by young adulthood had better outcomes, but “it is not clear that natural history change of the sort we observed in our longitudinal study is equivalent to intervention-induced change.” The study supports testing interventions; it does not show that one would work.
How long does a longitudinal study take, and what does it cost?
There is no minimum length: two waves a few weeks apart meet the definition. Most well-known projects run for years or decades, because the outcomes they track, such as heart disease or adult income, take that long to appear. The cost comes from keeping staff, funding and contact with participants going for the whole period, more than from any single wave.
The OpenStax textbook says these studies need “an incredible time investment by the researcher and research participants” and that attrition rates “are quite high and increase over the course of a project”, so researchers over-recruit. Plans also change: the Dunedin Study’s unit says the original idea was one assessment at age three. For a student project, a two- or three-wave panel over a semester is a realistic version of the design.
Where Papero fits
Papero helps with the reading and citing around a longitudinal design. Its Find papers page searches more than 250 million papers through OpenAlex, so you can find published panels and cohorts in your field and save them to your library before you cite them.
With the PDFs in your library, chat answers questions using only your library and the open document, and cites the passages it used, with the page when it is certain. That helps when you compare how several studies reported their follow-up numbers. The find sources page shows the in-document search workflow, and the glossary entry on research design places this design among the others.
Frequently Asked Questions
What is a longitudinal study in simple terms?
It is research that measures the same people more than once over time. A researcher might ask a group of 20-year-olds about their diet, then ask the same people again at 30 and 40, and compare the answers. Because the same people are measured each time, the results show how individuals change, not only how groups differ.
Is a longitudinal study qualitative or quantitative?
Either, or both. Most large projects collect quantitative data such as test scores, blood pressure or income, but researchers also interview the same people repeatedly to see how their accounts change. What makes the research longitudinal is measuring the same people over time, whatever type of data is collected.
What is the difference between a longitudinal study and a cohort study?
A cohort study is one kind of longitudinal design. In epidemiology it starts with people who do not yet have an outcome, records who is exposed to a factor and who is not, and follows both groups to count who develops the outcome. Longitudinal research is the wider category, and also includes panels drawn from random samples and designs that follow the same schools or households.
Can longitudinal research prove cause and effect?
Not on its own. It shows that the presumed cause was measured before the outcome, which a one-time survey cannot, and it allows adjustment for factors that were measured. A factor nobody measured could still explain the link. Researchers combine temporal order, adjustment, consistent results across studies and, where possible, a randomized trial before making a causal claim.
Sources
- Caruana, E. J., Roman, M., Hernández-Sánchez, J., & Solli, P. (2015). Longitudinal studies. Journal of Thoracic Disease, 7(11), E537–E540. https://doi.org/10.3978/j.issn.2072-1439.2015.10.63
- Moffitt, T. E., Arseneault, L., Belsky, D., Dickson, N., Hancox, R. J., Harrington, H., Houts, R., Poulton, R., Roberts, B. W., Ross, S., Sears, M. R., Thomson, W. M., & Caspi, A. (2011). A gradient of childhood self-control predicts health, wealth, and public safety. Proceedings of the National Academy of Sciences, 108(7), 2693–2698. https://doi.org/10.1073/pnas.1010076108
- Vandenbroucke, J. P., von Elm, E., Altman, D. G., Gøtzsche, P. C., Mulrow, C. D., Pocock, S. J., Poole, C., Schlesselman, J. J., Egger, M., & STROBE Initiative. (2007). Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): Explanation and elaboration. PLoS Medicine, 4(10), Article e297. https://doi.org/10.1371/journal.pmed.0040297
- National Center for Advancing Translational Sciences. (n.d.). Longitudinal study. NCATS Toolkit glossary (toolkit.ncats.nih.gov, November 2026): https://toolkit.ncats.nih.gov/glossary/longitudinal-study/
- Iowa State University Library. (2026). Longitudinal. Research methodologies guide (instr.iastate.libguides.com, November 2026), quoting Menard, S. (1991), Longitudinal research: https://instr.iastate.libguides.com/researchmethods/longitudinal
- Spielman, R. M., Jenkins, W. J., & Lovett, M. D. (2020). Approaches to research. In Psychology 2e (Section 2.2). OpenStax: https://openstax.org/books/psychology-2e/pages/2-2-approaches-to-research
- Framingham Heart Study. (n.d.). Epidemiological background and design (framinghamheartstudy.org, November 2026): https://www.framinghamheartstudy.org/fhs-about/history/epidemiological-background/
- Centre for Longitudinal Studies, UCL. (n.d.). 1970 British Cohort Study (cls.ucl.ac.uk, November 2026): https://cls.ucl.ac.uk/cls-studies/1970-british-cohort-study/
- Dunedin Multidisciplinary Health and Development Research Unit. (n.d.). The Dunedin Study and The birth of the Dunedin Study (dunedinstudy.otago.ac.nz, November 2026): https://dunedinstudy.otago.ac.nz/studies/assessment-phases/the-birth-of-the-dunedin-study
- Dunedin Multidisciplinary Health and Development Research Unit. (n.d.). The Study Members (dunedinstudy.otago.ac.nz, November 2026): https://dunedinstudy.otago.ac.nz/studies/assessment-phases/the-study-members
- Add Health, Carolina Population Center. (n.d.). About Add Health (addhealth.cpc.unc.edu, November 2026): https://addhealth.cpc.unc.edu/about/
