Dynamic Panel Models in Politics

The underlying estimators are the same — Arellano-Bond difference GMM, Blundell-Bond system GMM, and fixed-effects-with-lag specifications all originate in econometrics. The political-science framing emphasizes time-series cross-section data with relatively few, often non-random units (countries, dyads, states) observed over moderate T, where the Beck-Katz tradition of panel-corrected standard errors and careful dynamic specification is as central as GMM. In short, the generic dynamic-panel-data-model entry covers the econometric machinery; this entry covers how that machinery, plus the TSCS-specific PCSE design, is deployed on the kind of data political scientists actually have.

Use GMM when N is large relative to T, when the persistence coefficient ρ is itself a quantity of interest, and when the Nickell bias would meaningfully distort it — and when you can defend the instrument and serial-correlation assumptions. Use Beck-Katz OLS with panel-corrected standard errors and a lagged dependent variable when T is long enough that the 1/T Nickell bias is small, when the dynamics are a nuisance to be controlled rather than estimated precisely, and when you value transparency and robustness over the efficiency gains of GMM.

The Nickell bias is the inconsistency that arises when a lagged dependent variable is combined with fixed effects (or within transformation) in a panel: the transformed lag is mechanically correlated with the transformed error, biasing the persistence coefficient downward by an amount of order 1/T. Because political-science panels often have short to moderate time dimensions, this bias can be large enough to qualitatively change conclusions about how persistent an outcome is, which is exactly why GMM and bias-corrected estimators were brought into the field.