A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Panel data combines observations across units (individuals, firms, countries) over time, enabling control for unobserved heterogeneity, identification of time-varying effects, and more precise estimation of relationships. The balanced/unbalanced distinction and time dimension affect estimator choice and interpretation.
When you studied panel data basics, you encountered data with both a cross-sectional dimension (many units) and a time dimension (repeated observations). Now it's worth understanding precisely *why* that structure is so powerful for causal inference. The key insight is that panel data gives you two distinct sources of variation — within-unit variation over time, and between-unit variation at a point in time — and you can choose which one to use depending on what confounds you're worried about.
The central advantage is control for unobserved heterogeneity. Suppose you want to estimate the effect of job training programs on wages. Workers who opt into training may differ from those who don't in ways you can't measure — motivation, work ethic, family support. With cross-sectional data, these differences corrupt your estimate. With panel data, you can compare each worker to *themselves* before and after training. Any time-invariant characteristic (motivation, innate ability) cancels out in this within-person comparison. This is the logic behind fixed-effects estimation: we absorb unit-level constants, leaving only the within-unit over-time variation to identify effects.
The notation encodes this structure explicitly. Observations are indexed by (i, t): i identifies the unit (person, firm, country), t identifies the time period. The full dataset is an N × T grid, though in practice it's rarely complete. A balanced panel has every unit observed in every period — N × T observations total. An unbalanced panel has gaps, often because units enter or exit the sample (attrition in survey data, firm births and deaths in company data). The balanced/unbalanced distinction matters because some estimators assume balanced panels and will give wrong answers applied to unbalanced ones.
The time dimension T relative to N also shapes which tools are appropriate. Short panels (large N, small T — like annual surveys of thousands of individuals over 5 years) are the classic setting for fixed-effects and random-effects estimators. Long panels (moderate N, large T — like monthly data on 20 countries over 30 years) start to behave more like time-series data, and issues like cointegration, cross-sectional dependence, and non-stationarity become relevant. Understanding where your data falls on this spectrum determines which estimator properties — consistency in N, consistency in T, or both — matter for your application.