Panel and longitudinal analysis

Repeated measurement is the most under-used asset in routine service data. A single cross-section can only compare people who differ in every respect; a second observation on the same person allows each of them to act as their own comparison, and removes at a stroke everything stable about them that would otherwise confound the estimate.

The choice between fixed and random effects is a question about what the estimate is meant to mean, and it is settled by argument rather than by a specification test alone. Fixed effects answer a question about change within individuals and discard the between-individual comparison entirely; random effects use both and buy precision with an assumption. Both are usually worth reporting, with the difference between them treated as informative rather than as something to resolve quietly.

Growth-curve and mixed models describe trajectories rather than endpoints — how quickly an outcome changes, whether the rate differs between people, and whether it differs systematically with anything observable at baseline. Serial correlation and unequally spaced measurements are modelled rather than averaged away.

Attrition is the standing threat in any panel, and people who stop being measured rarely stop at random. Drop-out is modelled alongside the outcome, and estimates are reported under more than one assumption about why the later observations are missing.