Written and maintained by CASRAI Editorial Board
Last updated
A panel survey asks the same people the same kind of questions more than once, so a change you see is a change within people, not a difference between two different samples. This page defines the design, shows how a panel is recruited and weighted using two public examples, and works through the retention arithmetic that decides whether a panel can still be trusted at its later waves.
What is a panel survey? (definition)
A panel survey collects data from the same sample of respondents at two or more points in time; each round of data collection is called a wave. Because the same units are followed, it measures individual-level change, which a cross-sectional survey of fresh samples cannot do. A panel is one kind of longitudinal design; a cohort study follows a group defined by a shared experience, whereas a panel survey is defined by repeated measurement of the sampled respondents.
Two real programmes show how different panels can be:
| Feature | Pew Research Center, American Trends Panel | U.S. Census Bureau, SIPP |
|---|---|---|
| Population | U.S. adults, roughly 10,000 panelists | Households; a nationally representative longitudinal survey |
| Rhythm | One or two surveys a month; each open about 6 to 12 days | A continuous series of national panels; each panel’s households are interviewed multiple times over a four-year period |
| Participation and output | Subsamples are drawn, so not every panelist answers every survey | Gives monthly data about changes in household and family composition |
Sources: the Pew Research Center and U.S. Census Bureau pages listed under Sources, checked 2026-10-29.
How a probability-based panel is recruited and weighted
Recruitment is the first thing to check, because a panel that people can volunteer for describes volunteers. Pew states that members of the American Trends Panel are randomly selected and cannot volunteer: it uses address-based sampling, and the adult with the next birthday in a sampled household is invited at the end of a survey. Contrast that with an opt-in list, where the selection step is self-selection and the risks described in sampling bias apply from day one.
Weighting then happens in two steps, as Pew describes them: a base weight reflecting each person’s chance of selection, followed by calibration that aligns characteristics of the sample to known population parameters. Pew also describes how it manages loss over time: some panelists stop responding, so new members are added each year; an annual profile survey keeps records current (skipping it ends invitations); and overrepresented members are occasionally retired.
Formula: retention rate across waves
Attrition, the loss of respondents between waves, is the central quality problem of any panel. The measure below is the working definition used on this page (this page’s own working definition for illustration, not a quotation from a standards body):
Rt = nt / n1 × 100 (cumulative retention) and rt = nt / nt−1 × 100 (wave-to-wave retention)
- t: the wave number (1 for the first wave).
- nt: the number of respondents who completed wave t.
- n1: completes at the first wave, the baseline.
- Rt: share of the baseline still responding at wave t.
- rt: share of the previous wave’s respondents who responded again at wave t.
Worked example with the arithmetic shown
The counts below are an illustration invented for this page. They are not data from any real survey.
Suppose a four-wave panel completes n1 = 1,000, n2 = 820, n3 = 700 and n4 = 610 interviews.
- Wave 2: R2 = 820 / 1,000 × 100 = 82.0%; r2 = 820 / 1,000 × 100 = 82.0%.
- Wave 3: R3 = 700 / 1,000 × 100 = 70.0%; r3 = 700 / 820 × 100 = 85.4%.
- Wave 4: R4 = 610 / 1,000 × 100 = 61.0%; r4 = 610 / 700 × 100 = 87.1%.
Interpretation: reading the result
| Quantity (from the example) | Value | What it tells you |
|---|---|---|
| R4 | 61.0% | Of every 100 baseline respondents, 61 are still in the data at wave 4; 39 were lost. |
| r2 | 82.0% | The largest single-wave loss (18.0 points) occurs between waves 1 and 2. |
| r3, r4 | 85.4%, 87.1% | Loss per wave is smaller later, so cumulative retention falls more slowly after wave 2. |
A retention rate says how many were lost, not whether the loss is harmless. The question that matters is whether those who left differ from those who stayed on the outcomes you measure. Pew’s published approach to this is to add new panelists each year and to calibrate weights to population benchmarks, so the raw count is only a starting point.
Assumptions and common mistakes
- Treating an opt-in panel as a probability sample. If members volunteered, design-based weights and margins of error built for random selection do not describe it.
- Reporting only the wave-1 sample size. Analyses of change use the respondents who answered the relevant waves; report the n for each wave you analyse.
- Ignoring who dropped out. Compare baseline answers of continuing and departing respondents before interpreting change.
- Assuming every panelist answers every survey. Large panels such as the American Trends Panel draw subsamples for each survey, so the wave n varies by study.
- Confusing a panel with repeated cross-sections. Repeated independent samples track population change; only a panel tracks the same individuals. Question wording across waves matters either way; see survey question types.
How to report a panel survey (APA example)
Report the design, the recruitment method, the number of waves, the completes at each wave you analyse, how weights were built, and how attrition was handled. Illustrative reporting sentence using the worked example’s invented numbers:
The four-wave panel completed 1,000, 820, 700 and 610 interviews, giving cumulative retention of 82.0%, 70.0% and 61.0% at waves 2 to 4 relative to wave 1.
Reference-list pattern for a web page from a group author, with no date, in APA style:
Pew Research Center. (n.d.). The American Trends Panel. https://www.pewresearch.org/our-methods/u-s-surveys/the-american-trends-panel/
How to calculate retention in Excel, R or Python
Excel. Put the wave completes in B2:B5. In C2 enter =B2/$B$2 and fill down for cumulative retention. In D3 enter =B3/B2 and fill down for wave-to-wave retention. Format both columns as percentages.
R.
n <- c(1000, 820, 700, 610) cumulative <- round(100 * n / n[1], 1) wave_to_wave <- round(100 * n[-1] / head(n, -1), 1) cumulative # 100.0 82.0 70.0 61.0 wave_to_wave # 82.0 85.4 87.1
Python.
n = [1000, 820, 700, 610] cumulative = [round(100 * x / n[0], 1) for x in n] wave_to_wave = [round(100 * n[i] / n[i - 1], 1) for i in range(1, len(n))] print(cumulative) # [100.0, 82.0, 70.0, 61.0] print(wave_to_wave) # [82.0, 85.4, 87.1]
FAQ
What is the difference between a panel survey and a cross-sectional survey?
A panel survey re-interviews the same respondents across waves; a cross-sectional survey measures one sample once, or draws a new sample each time. Only the panel shows change within individuals. See the cross-sectional versus longitudinal comparison.
How is a panel survey different from a survey panel?
The phrases overlap. A panel survey is the design (the same people over time); a survey panel is the pool of recruited respondents that surveys draw from. The American Trends Panel is a pool from which subsamples are drawn for each survey. For wider context see the advantages and disadvantages of surveys and the survey research methods entry.
Can anyone join a probability-based panel?
Not in the design Pew describes: only people randomly selected can join, and you cannot volunteer.
How long can a panel last?
It depends on the programme. The Census Bureau describes SIPP as a continuous series of national panels, with each panel’s households interviewed multiple times over a four-year period.
Sources
- Pew Research Center, The American Trends Panel, https://www.pewresearch.org/our-methods/u-s-surveys/the-american-trends-panel/ (checked 2026-10-29).
- U.S. Census Bureau, Survey of Income and Program Participation (SIPP): About, https://www.census.gov/programs-surveys/sipp/about.html (checked 2026-10-29).
The retention formula and the worked example are this page’s own illustration; the counts are invented, not survey results.
