You can't always get what you want? : a Monte Carlo analysis of the bias and the efficiency of dynamic panel data estimators

We assess the bias and the efficiency of state-of-the-art dynamic panel data estimators by means of model-based Monte Carlo simulations. The underlying data-generating process consists of a standard theoretical growth model of income convergence based on capital accumulation. While we impose a true underlying speed of convergence of around 5% in our simulated data, the results obtained with the different panel data estimators range from 0.03% to 17%. This implies a range of the half life of a given income gap from 4 years up to several hundred years. In terms of the squared percent error, the pooled OLS, fixed effects, random effects, and difference GMM estimators perform worst, while the system GMM estimator with the full matrix of instruments and the corrected least squares dummy variable (LSDVC) estimator perform best relative to the other methods under consideration. The LSDVC estimator, initialized by the system GMM estimator with the full matrix of instruments, is the only one capturing the true speed of convergence within the 95% confidence interval for all scenarios. All other estimators yield point estimates that are substantially different from the true values and confidence intervals that do not include the true value in most scenarios.