Evaluating Specification Search in Cross-country Growth Regressions: Extreme Bounds vs. General-to-Specific Methodologies

Abstract

This paper is a substantially revised version of our earlier working paper, Truth and Robustness in Cross-country Growth Regressions. The most important revisions concern the handling of missing observations in the cross-country data set. In the earlier paper, these had been handled through case-wise deletion-the method common to all previous studies. In this version, they are handled through multiple imputation-a method that retains substantially more of the information content of the data set. Two variants of Leamer's (1983) extreme-bounds analysis are evaluated for their ability to recover the true specification and compared to a cross-sectional version of the general-to-specific search methodology associated with the LSE approach to econometrics. Evaluations are based on a realistic Monte Carlo experiment in which the universe of potential determinants is drawn from those in Levine and Renelt's (1992) study. Levine and Renelt's method is shown to have low size and extremely low power: nothing is robust. Sala-i-Martin's (1997a, b) method is shown to have high size and high power: it is undiscriminating. The general-to-specific methodology is shown to have size near nominal size and high power. Sala-i-Martin's method and the general-to-specific method are then applied to the actual data from Sala-i-Martin's original study. The results are consistent with the Monte Carlo results and suggest that only a few of the 61 potential determinants of growth matter.