At a Glance
In a panel survey, you interview the same people in each round. In a repeated cross-section, you draw a new sample from the same population each round. Both can feed a baseline-to-endline comparison. They answer slightly different questions, fail in different ways and need different sample sizes. The baseline, midline and endline page introduces the timing; this page helps you choose between the two designs.
At a Glance
| Panel | Repeated Cross-Section | |
|---|---|---|
| Who is interviewed | The same people each round | A new random sample each round |
| Shows | Change for individuals, including who moved up and who fell back | Change in the population or group average |
| Precision of change | Often better, when answers are correlated across rounds | Weaker for the same sample size, because rounds are independent |
| Main risk | Attrition: the people lost are rarely random | Different samples may differ by chance, and the population itself may change in composition |
| Other risks | Panel conditioning: repeated interviews can change what people report or do | Needs a current sampling frame in every round |
| Cost pattern | Tracking costs between rounds; fewer new listings and selections | A new listing and selection each round; no tracking |
| Fits when | Individual change matters and people can be found again | The population is mobile, the gap between rounds is long, or tracking is not feasible |
Four Questions to Decide
1. Do you need to know who changed? If the evaluation asks how many people moved from unemployed to employed, or whether those with the lowest scores gained most, only a panel can show it. If it asks whether the employment rate in the area rose, a repeated cross-section can answer.
2. Can you find the same people again? Check phone coverage, migration, seasonal work, displacement and consent to be recontacted. If a large share is likely to be lost, a panel brings both cost and bias risk. See nonresponse bias for how attrition is assessed.
3. Will the population itself change? In a repeated cross-section, people who arrive or leave between rounds change who is sampled. In a panel, the original people remain the focus but may not represent later arrivals. Decide whether your target is the original group or the current population.
4. Does repeated questioning risk changing answers? Panel members who are asked about savings or diet three times may begin to think about them, or learn what answers the survey expects. If this is likely, a rotating design (below) can limit it.
Worked Case
Worked Case: Precision With and Without Attrition
A fictional program expects a rise of about 5 points in a 100-point household food score, and the team wants to estimate the change to within about 2.5 points either way. The sizes below are precision targets, not power calculations; a power calculation would use the same inputs but a different formula. The score has a standard deviation of 20 in both rounds. We assume the correlation between a household's two scores is 0.6, simple random sampling, and the same variability in each round. These are illustrative assumptions, not typical values; use evidence from comparable surveys where available.
For a repeated cross-section, the standard error of the change with n households per round is 20 x the square root of (2 / n). For a panel, it is 20 x the square root of (2 x (1 - 0.6) / n).
| Design | Households | Standard error of change | Approximate 95% half-width |
|---|---|---|---|
| Repeated cross-section | 200 per round | 2.0 | 3.9 points |
| Panel, no attrition | 200 households | 1.3 | 2.5 points |
| Panel, 25% lost by endline | 150 households with both rounds | 1.5 | 2.9 points |
To reach a half-width of about 2.5 points, the repeated cross-section needs about 492 households per round (984 interviews in total). The panel needs about 197 households interviewed twice (394 interviews). To end with 197 after a 25% loss, enroll about 263 at baseline, which means about 460 interviews in all (263 at baseline and 197 at endline).
The panel still comes out ahead in this case, but three cautions apply. The precision advantage shrinks as the correlation falls and is gone when it reaches zero. The calculation covers sampling variation only; households lost to attrition may differ from those retained, and that bias is not shown by the interval. And clustering would enlarge all the figures. The design effect for a change estimate can differ between the designs, because the same clusters are revisited in a panel.
For the arithmetic of intervals, see confidence interval. For clustering, see design effect explained.
Mixed Designs
You need not choose purely.
- Rotating panel. Keep part of the sample for several rounds and replace the rest. This gives some individual change data and limits fatigue.
- Refreshment sample. Follow the panel and add a new random sample at endline. Comparing the two helps you judge attrition and conditioning.
- Panel with replacement tracking. Define in advance what you do when a person cannot be found: another household in the same cluster is not the same household, and the report should say how it was handled.
With any panel, record contact details, consent for recontact and the rule for movers. Weights and attrition adjustment are covered under sampling weights.
Pitfalls
Common Mistakes
Calling a repeated cross-section a panel. If the endline draws a fresh sample, you cannot report that "the same households improved". You can report that the area average rose.
Treating a panel completer sample as the original sample. If one in four households is lost, compare completers with those lost at baseline before you generalize the findings.
Replacing lost panel members without saying so. Replacements are new units. Keep them apart in the analysis, or state how they enter the estimate.
Assuming a panel is always cheaper or more precise. Tracking is costly, and the precision gain depends on the correlation between rounds.
Ignoring the design in analysis. Both designs need the strata, clusters and weights in the variance calculation. Otherwise, intervals are often too narrow.
Letting an AI tool choose the design from a one-line prompt. Ask it to list the assumptions it used (correlation, attrition rate, design effect), then check each against your context. Keep the choice with the team.
What to Record in the Design Note
| Item | What to Write |
|---|---|
| Design chosen | Panel, repeated cross-section or mixed, and why |
| Target of inference | The original group or the current population |
| Change to detect | Size, outcome and unit |
| Assumptions | Correlation between rounds, expected attrition, design effect, with sources |
| Sample per round | Planned completes and the enrollment needed after expected losses |
| Tracking plan | Contact data, consent for recontact, rule for movers |
| Attrition checks | Comparison of completers and those lost at baseline |