Panel-Based Survey Research — Longitudinal Panel Survey Design

In a panel-based survey the same individuals are measured at each wave, so you can directly observe intra-individual change. In a repeated cross-sectional survey different random samples are drawn at each wave; you can estimate population-level change but cannot track whether specific individuals improved or declined. Panel designs are more powerful for causal questions but more costly and vulnerable to attrition.

A minimum of two waves is required to measure change at all. Three or more waves are needed to model growth trajectories with latent growth curve or multilevel models. The number of waves should be driven by the expected change process: fast-moving phenomena (mood, daily behaviour) warrant shorter intervals and more waves; slow-moving constructs (educational attainment, personality) may require fewer waves over longer periods.

There is no universal threshold, but attrition above 20–30% over the full panel period warrants serious scrutiny. More important than the rate is whether attrition is random or systematic. If those who drop out differ from those who stay on key variables (e.g., lower income, worse health), analyses restricted to completers will be biased. Always test for differential attrition and report it; apply inverse-probability weighting or multiple imputation when attrition is non-random.

Ordinary regression ignores the dependency of repeated observations within the same individual and treats the data as if each measurement came from a different person. This underestimates standard errors and inflates Type I error. Use methods designed for panel data: mixed models, fixed-effects regression, generalised estimating equations, or structural equation models with repeated factors, depending on the research question.

Panel conditioning occurs when repeated participation in the survey changes respondents' behaviours or attitudes — for example, a panel about health behaviours may make participants more health-conscious. It can be detected by comparing early-wave responses of long-term panelists with responses of fresh respondents recruited at the same time. If differences emerge that cannot be explained by actual population change, conditioning is a plausible explanation and should be discussed as a limitation.