The stringent, the liberal, and the robust: Can specification search algorithms help us solve the cross-country growth puzzle?

Due to a lack of a unified theoretical framework for empirical research on the determinants of long-term growth, different specification search algorithms have been applied to this field in the past decade. These include the general-to-specific approach (Gets), Bayesian Model Averaging (BMA), and Weighted-Average-Least-Squares (WALS). These methods are critically assessed in the context of cross-country growth regressions. Their efficacy to find the correct specification is evaluated by means of a number of Monte Carlo experiments. Robustness with respect to nonlinearities and set of potential regressors is assessed as well. BMA is found to be stringent, but reliable, Gets is found to be most powerful, but liberal, and WALS is found to be most robust to size of the potential set of regressors. Evidence is found for a tight relationships between long term growth rates and the following variables: initial income, real exchange rate distortions, years open economy and initial fertility rates.