TGM AL MAMUN
Empirical Research Guide · Panel Econometrics

Which panel model fits your data?

The right panel estimator depends on the properties of your panel, not on preference. Test a real dataset, upload your own, or enter your test results, and follow the path to the estimator your paper needs.

9 estimator guides6 real example panelsPython · statsmodels

Choose a dataset

Six real panels: three micro panels (people, counties, firms) and three macro panels (states, countries). Each one leads to a different outcome.

Estimator guides

Routes to a finished paper

Each guide takes one real panel through estimation, diagnostic tests and writing, with Python code, real output and sample text for the paper.

Design-based

DiD and event studies

Treatment timing and a comparison group

Define treatment and timing · parallel trends and anticipation · staggered adoption needs heterogeneity-robust estimators · cluster at the assignment level …

Open the Causal Inference Guide →
Limited outcomes

Binary and count panels

Outcome is 0/1 or a count

Conditional FE logit · RE logit / probit with distributional assumptions · FE Poisson for counts · report marginal effects · incidental-parameter problem …

Lecture: Panel data →
How the tool decides. The route depends on both the number of units N and the median number of periods per unit T. Panels with T below 15 are treated as short (micro) panels: the F test asks whether unit effects matter, the Breusch–Pagan LM test whether random effects improve on pooled OLS, and the cluster-robust Mundlak test (reported with the conventional Hausman test) whether the unit effects are correlated with the regressors; a dynamic model goes to GMM only if N is at least 30, and otherwise to bias-corrected fixed effects. Panels with T of 15 or more and at least 10 units are treated as macro panels: the Pesaran CD test checks for cross-sectional dependence, the Pesaran–Yamagata Δ test for slope heterogeneity, and each variable is tested for a unit root in levels and differences with CIPS when it is cross-sectionally dependent and with the Fisher-ADF (Choi Z) test otherwise. CIPS p-values come from simulated critical values. When every variable is I(1), cointegration is tested with the Fisher–Engle–Granger test if slopes are heterogeneous and with the Kao test otherwise. With T of 15 or more but fewer than 10 units, each unit is modelled as a time series. These are standard rules, but panel tests have low power when T is short and can be distorted by breaks, so treat the result as a starting point and justify the choice in your paper.