This paper uses a novel dataset to test the capital–skill complementarity hypothesis in a cross-section of countries. It is shown that for the full sample there exists evidence in favor of the hypothesis. When we arbitrarily split the full sample into OECD and non-OECD countries, we find no evidence in favor of the hypothesis for the OECD subsample, but strong evidence for the non-OECD subsample. When we use Hansen’s [Econometrica 68 (2000) P. 576] endogenous threshold methodology we find that initial literacy rates and initial per capita output are threshold variables that can cluster countries into three distinct regimes that obey different statistical models. In particular, the regime with moderate initial per capita income but low initial education exhibits substantially higher capital–skill complementarity than the regime with low income and low education and the regime with high education. This cross-country nonlinearity in capital–skill complementarity is consistent with the time-series nonlinearity found by Goldin and Katz [Quarterly Journal of Economics 113 (1998) 693] using U.S. manufacturing data, and promotes the view that the phenomenon maybe a transitory one. Copyright Springer Science+Business Media, Inc. 2005
