Like shoes and shirt, one size does not fit all: Evidence on time series cross-section estimators and specifications from Monte Carlo experiments

Political scientists often and increasingly analyze time-series cross-sectional (TSCS) data. These data come with signficant problems, such as accounting for unobserved variation across sample units and appropriately specifying dynamics. Furthermore, even though fixed-effects (or least squares dummy variable (LSDV) models) can address unit heterogeneity, least squares (LS) estimation of modelds with fixed-effects and lagged dependent variables are known to be biased. Alternative estimators, mostly from economics, and generally designed for short panels, have been proposed to address this bias, but it is generally not well known how these estimators perform in comparison to simple methods like LS and LSDV on TSCS data. The preliminary results we illustrate here suggest that LSDV is generally as good or better than instrumental variables (IV) approaches in terms of bias and efficiency. We examine estimator performance under conditions where the importance of the unit effects and the correlation of the unit effects with the independent variables are allowed to vary and find that LSDV performs well. Unfortunatley none of the estimators, particularly LS, perform well when the dynamics of the model are mis-specified. The lesson is that new estimators do not, in general, solve the problem of mis-specifying the model’s dynamics.