Nonlinearities and Robustness in Growth Regressions

Cross-country regressions are a well-established means of attempting to uncover the empirical determinants of economic growth. However, in an influential paper, Levine and Renelt (1992) demonstrate that the results of these studies are very sensitive to the choice of conditioning variables. Using a variant of Leamer's (1983) extreme bounds test, they show that almost no explanatory variables are robustly correlated with growth. In this paper, I show that this extremely pessimistic conclusion is partly due to the ad hoc assumption of linearity in the traditional growth specification. Specifically, under alternative (nonlinear) specifications, the number of robust variables increases substantially.